The Machines Eat Media Buying

Media buying is a weird industry. 23-year-olds making $30K/yr sit courtside at NBA Finals games or are flown VIP to Coachella and control millions of dollars in ad spend for the world's largest brands. This dynamic naturally creates a lot of arbitrages in which a litany of middlemen book billions of dollars every year in revenue. That margin is someone’s opportunity. In this case, it is Zuck’s.

Last week, Meta released their ads MCP and CLI, providing direct, authorized access to help manage your Meta Ads account through natural language in Claude or ChatGPT. It’s early, limited access, and has major initial limitations around attribution, but a genie has been released here that isn’t going back in the bottle. We’re likely quarters, not years, away from agents doing most of the hands-on key media buying across all major ad platforms.

In retrospect, it’s somewhat surprising that it took this long to get here. Google’s Performance Max and Meta’s Advantage Shopping Campaigns already took much of the edge off pure media-buying and targeting wizardry, placing the premium on creative and copy. But they didn’t disrupt the full media buying industrial complex on as much of an existential level. Like the almighty billable hour in law or consulting, the traditional agency media buying fee and revenue tied to headcount staffed against an account proved stubbornly resistant to change, even as algorithms did more and more of the work.

At the beginning of last year, the best hands on keys marketer I know texted me to say “there’s no alpha in everything I know in growth marketing anymore. The platforms do everything now.” He reimagined his professional identity, leaned hard into building marketing systems in early versions of Claude Code and is doing great. The growth marketing workflow of the future looks something like this:

Ultimately, agents serving as media buyers is another case of AI acting as more of a mirror than a crystal ball. Agents might not be able to functionally “do” much net new that the existing ecosystem of APIs couldn’t already automate. In fact, Meta’s Marketing API still has far more functionality than the MCP. But the form factor of simply being a prompt away from having AI do the work lays bare how little friction there should be between buyer and seller. The proverbial emperor has no clothes moment.

When the Meta MCP news broke, I called up Michael Beebe, CEO of Dstillery and long-time analyst who spent two decades covering the TMT sector for the top financial institutions in the world. His broader take on the moment we’re in resonated deeply:

The question for anyone now building in this ecosystem is whether you fight tooth and nail to preserve the externalities where you currently collect margin or if you think first principles around where you help marketing leaders find new alpha in the AI era.

Here, it bears mentioning that The Trade Desk’s stock is down 85% since December 2024 and its chief strategy officer just decamped for OpenAI. If the open web was “whistling past the graveyard” in 2024, it is now starting to pick out its dirges. A few large platforms call the game.

Meta, microcosmic of all of the big tech players in this regard, is ultimately positing a world where two things are true:

1) AI systems can run the entire full-stack process of running ads for businesses, no experience necessary. From Meta’s Q1 earnings call:

“We’re building AI systems that can basically run the entire advertising process for businesses…where you can come to us, tell us what your objective is, connect your bank account, and then we’ll just do the rest. That includes everything from generating the creative, to targeting, to optimizing, to measurement…”

2) For those who still want to maintain more sophisticated “control”, AI will do that too, just via bespoke agents marketers build rather than humans living in ad managers. Effectively, Meta is saying media buying is a space where even the last mile of manual work will be abstracted away. The top echelon of growth leaders will find their edge effectively by building smarter context engines and harnesses for agents.

Ultimately, Meta is trying to empower every marketer to become just good enough of an arbitrageur to maximize their own profits. The best scenario for Meta is that you can put $1 in and get $1.01 back (effectively in post-margin ROAS) to spend ad infinitum at the highest possible take for the platform.

It’s more or less the same formula that has catapulted AppLovin to be a $160B market cap company ($20B more than Shopify!) on mobile gaming arbitrage and is what OpenAI is trying to speedrun to make ChatGPT a $100B ads business by 2030.

The most enduring impact of AI in marketing is that it will make driving “just good enough” performance cheaper and easier than ever. Is that good enough?

Marketing chef’s kiss to….Wispr Flow

Wispr Flow–the buzzy voice-to-text app that is turning all of Silicon Valley into always-on, mobile raconteurs—hads had an extended bit viral branding moment on X over the past couple of weeks. The company ran an out-of-home campaign on the backs of rickshaws in Bangalore, where residents spend….a lot of time stuck in traffic, starting right at the back of the ‘shaw in front of them.

Mobile OOH is a consistently slept on category, and this campaign was generally brilliant, but it’s even savvier than it first appears. The best demo I ever took when I was brand side came from Adgile Media, a company that runs ads on trucks for brands ranging from Glossier to Planet Fitness. Their founder, Tom, opened by asking what I thought the best color for OOH ads was. It’s purple, by far the least naturally occurring color in nature, and thus the most prone to catch the eye.

I have no idea if Wispr Flow’s marketing team knew this, but there’s a funny thing about great companies. Great companies are great precisely because they sweat the details on a multitude of random, little, esoteric things. Bravo.

And one more to Pasta Garofalo

A belated Happy Mother’s Day to all the moms in our audience! You are heroes in everything you do.

Yesterday’s Mother’s Day meal in casa Mallazzo was an extra indulgent baked ziti, triggered by a brilliant act of marketing from one pasta brand. For the record, I explicitly reject Clay founder Varun Anand’s claim that baked ziti is a “low ceiling” food. If anyone else feels the way he does, email mike@brxnd.ai for a recipe that will change your mind.

Garofalo is currently running the best yearlong campaign in CPG that I can remember for quite some time. Their packaging is emblazoned with “REAL ITALIAN CARBS” in big red letters on the front. It works perfectly on both a visceral and subliminal level. Visually, the striking letters jump off the shelf in a sea of homogeny at the pasta aisle. But it gets even more clever as you dig deeper. The copy on the back of the package closes with:

”Born in Gragnano, the birthplace of pasta, Garofolo has been crafting the real Italian carb for 200 years. Italian carbs are different. Don’t count them, give in to them.”

Unlike some other brands, Garofalo sells the exact same product in both the United States and Italy. Doubling down on this distinction hits because so many Americans tell the same cliche story of going to Italy, eating nothing but indulgent pasta and coming home five pounds lighter. It’s always the superior food of course that is the reason and has nothing to do with the fact that they walked 20,000 steps per day with not a care in the world :)

But there’s also beauty in the simple counter-positioning. In an era where big CPG is pushing Pop-Tarts PROTEIN, carbs don’t have to be the eternal enemy.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

I’ve got a little something different for you today. As Mike and I work to better articulate the space we want BRXND to occupy going forward, I’ve been thinking about where some of the more technical writing and thinking about AI, engineering, and the enterprise goes. To solve for that, I started Forward Deployed last year. It was originally a podcast co-hosted with a friend who has since decamped for Anthropic. That left me with a decision on whether or not to proceed with the project (one can hardly call themselves a middle-aged tech executive without having to decide whether or not to continue a podcast). Ultimately, I decided that I had something unique to add to the conversation and have pushed to focus Forward Deployed on the question of how to deploy and build aligned agents in the enterprise, with a specific focus on what we can learn from the outside world about solving that problem. I’m also transitioning some of my more technical writing over there to give BRXND space to explore some of the bigger questions about how AI will shape the future of consumer preference (don’t worry, I also have plenty to add to that conversation as well).

I would encourage you to subscribe and join me on the journey over there.

To give a small taste, I’m republishing today’s piece about model, market, harness fit.

Thanks for reading, subscribing, and supporting all my projects.

— Noah

I originally put this slide together for the September 2025 edition of BRXND. It is a timeline of what I would consider the major milestones in AI over the last five years. The orange ones are the things I think are worth paying attention to. It was designed to scale as best I could with my limited Figma skills.

There are two things that should stand out:

Thanks for reading Forward Deployed! Subscribe for new posts and episodes at the intersection of AI, engineering, and the enterprise.

1. There are fewer moments than you would assume over the last five years. That’s because I purposely left out every small model update and focused on the bigger events.

2. There are only five highlighted moments. Despite arguments to the contrary and the never-ending parade of model releases, I genuinely believe AI has undergone only a few major changes over the last five years.

My five moments with their defenses:

1. November, 2021—GPT-3 API goes GA: This one is easy for me. We could draw the line in the sand almost anywhere after the Attention is All You Need paper, but to me, putting these things in the general public’s hands (albeit in API form) is a good demarcation point. There are still the “I’ve been doing AI for years” people complaining that ML is AI, but what’s possible before and after the GPT-3 API is very different.

2. June, 2022—GitHub Copilot goes GA: it’s easy to take for granted now, particularly with the conversations happening around uptime and GitHub, and the eventual direction Microsoft took the Copilot brand, but that initial autocomplete product in VS Code felt like magic and was the first real productization of AI.

3. November, 2022—ChatGPT Launch: This is the easiest one, and everyone likes to talk about it. Lots of us were building chatbots on our own, and this felt like such an obvious next step. Funny enough, this also immediately shifted Copilot use because while autocomplete was cool, chat turned out to be the killer app for coding (a fact Copilot seemed to miss and Cursor eventually ate their lunch on).

4. September, 2024–o1 Reasoning: The Google chain of thought paper came out in early 2022 and showed that adding reasoning to prompts significantly improved the models’ ability to reason, but it took two and a half years for OpenAI to build it into a model. o1 was big and slow, but set the groundwork for everything that followed. Though we didn’t realize it at the time, the real magic of these reasoning models wasn’t just that they were much better thought partners. Function calling was introduced to models in June 2023, but before reasoning models, it was incredibly inconsistent—better for demos than for real work. Reasoning changed that, allowing the model to plan its function calls before executing them.

5. May, 2025–Claude Code: finally, we have Claude Code, which celebrates its first birthday this month. If o1 made function calling work, Claude Code wrapped it in an agent loop: inspect the repo, decide what to do next, call a tool, read the result, and keep going. The tools themselves were simple UNIX commands, deeply covered in the pre-training data thanks to half a century of UNIX documentation. While lots of people believe that Opus 4.7 unlocked the harness, I think the causality goes the other way.

Despite how easy (and fun) it can be to get lost in each new model release or tool update, the reality is that most of it is just gradual improvement. That’s not to say it’s not tremendously valuable—GPT 4o unlocked ChatGPT in new ways—but those weren’t step changes in what is possible, they were the kind of slow progress that we always find hard to measure.

To that end, I have a few observations and thoughts from this list.

First off, each of these step changes is still very much building on the previous one. ChatGPT doesn’t happen without GPT-3, just like Claude Code doesn’t happen without o3. Reasoning models themselves are obviously evolved from prompting techniques. That can make AI progress feel Darwinian, but W. Brian Arthur’s point in The Nature of Technology is more specific: technologies evolve by combination. New things are assembled out of existing components and the phenomena they harness; sometimes better variations win, and sometimes a new principle arrives that makes the old family tree less useful. The jet engine isn’t just a better propeller. His focus is mostly hardware, not software, but it still makes me wonder if we should expect more radical changes, or maybe we’re just settled into this architecture, and that’s how we should expect innovation to happen. Here’s how Anthropic co-founder Jack Clark put it recently:

As a field, AI moves forward on the basis of doing ever larger experiments that utilize more and more inputs (e.g, data and compute). Every so often, humans come up with some paradigm-shifting idea which can make it dramatically more resource efficient to do things - a good example here is the transformer architecture and another is the idea of mixture-of-expert models. But mostly the field of AI moves forward through humans methodically going through some loop of taking a well performing system, scaling up some aspect of it (e.g, the amount of data and compute it is trained on), seeing what breaks when you scale it up, figuring out the engineering fix to allow it to scale, then scaling it again.

Another interesting observation is that image and video models are mostly missing on this timeline. I don’t quite know what to make of these models, and that’s part of the reason I’m writing this post. Every time a new step change in capability comes along (like Nano Banana or, more recently, GPT Image 2.0), we see people say everything is going to change, and then mostly not a lot changes. As practitioners of AI and marketing, we are making good use of the current crop of models with a variety of customers for a variety of use cases, but it feels like the Copilot/ChatGPT/Claude Code moment just hasn’t arrived yet. Obviously, diffusion as an approach is amazing and probably deserves a place on the list, but the leaps feel much more gradual and make it hard to draw conclusions. With all that said, the release of GPT Image in March 2025 caused an explosion of usage in ChatGPT.

To that point, maybe the most fundamental takeaway from the timeline is that the real leaps in usage come from a kind of weird alchemy between model, market, and harness. The model has to be good enough to do the job, the harness has to expose that capability in a form people can actually use, and the market has to be in the right place to see the full value. Each one of these episodes caused a step change in token demand. Anthropic’s own annualized run-rate numbers make the point: the company went from approximately $9 billion at the end of 2025 to more than $30 billion by early April 2026.

Sholto Douglas, an Anthropic researcher focused on scaling reinforcement learning, made a version of this point on a 2025 episode of the Dwarkesh Podcast: Cursor had been around for a while, but with Claude 3.5 Sonnet, “the model was finally good enough that the vision they had of how people would program, hit.”

That’s why the causation is so hard to pin down. Lots of people are sure that Claude Code exploded in December because of the release of Opus 4.7. As a relatively early adopter of CC, I’m a lot more skeptical. Many of us were getting tremendous value out of the harness in June and July. I think it’s much more likely that the real thing that happened in December is that a bunch of people just had more time to give the tool an honest go, and when they did, they were blown away.

Claude Code creator Boris Cherny calls this “product overhang.” The model’s capabilities often remain untapped until the right code comes along to unlock them. What I think this timeline adds is that there is a market overhang, too. The model can be good enough, and the harness can exist, but the market also has to be ready to notice, care, and change behavior. The highlighted moments are when model, harness, and market meet.

So what comes next? I don’t know. The three things I keep coming back to are reliable computer use, secure mobile agentic capability, and real document collaboration. I don’t know if any of them are the thing, but they all feel like potential unlocks for the next step change in token demand.

If you have any questions, please be in touch. As always, thanks for reading.

— Noah

When Magic Becomes Mundane

I’m recently back from helping to throw our flagship event for Ride AI, our sister publication that covers the intersection of AI and mobility. Autonomous vehicles are a uniquely fascinating greenfield for marketers, as there’s no industry where the chasm between the capabilities of the technology and the normie perception of what it can do is farther apart. Self-driving is no longer primarily a technological problem, it’s a cultural one. It is marketing’s time to shine.

From my meetings with marketing leaders across the industry, one story will stick with me the most as a microcosm for the challenge of marketing revolutionary new technology that is indistinguishable from magic.

It’s about how everyday people experience self-driving for the first time when they visit a city with robotaxi service. The first time East Coast folks come to SF, LA, or Phoenix they insist on taking a Waymo and are awe-struck. The next time they need a ride, there’s a rally cry of “LET’S DO THAT AGAIN.” By the third trip, they’re back to scrolling Instagram instead of basking in the science fiction reality.

After those first few trips, what was once a taste of the future becomes a decision made by price and wait time. Even magic rapidly becomes a commodity. It’s a visceral reminder that no matter how amazing a technology is, the old marketing adage still applies: People don’t want a quarter-inch drill, what they want is a quarter-inch hole.

The vibes around AI are bad right now for many reasons, but largely because the world's most powerful AI companies have failed to craft a coherent narrative about how AI improves the quotidian lives of everyday people. I strongly suspect this will intensify as politicians lean into the rhetoric ahead of the midterms. Here, too, it’s time to market.

A Time to Explore

Serendipitously, I happened upon Derek Thompson’s piece on how hot streaks happen in careers this weekend, the single best career advice framework I’ve read. In it, he advocates for a career strategy of "exploring" in your 20s—trying various roles, locations, and skills—followed by "exploiting" in your 30s, doubling down on the best skills and networks developed earlier to maximize success. I’m about one month into my new role at Alephic, and yesterday was my 34th birthday, so Derek’s article found me both enjoying a Miller High Life and in a contemplative moment. (Editor’s note: We have some past writing on the explore/exploit tradeoff, a concept from computer science that is very worth wrapping your head around, as it has widespread implications for enterprise AI adoption.)

In my 20s, I did damn near Shackleton-level exploration. I was an investigative journalist, ad agency intern (for someone who started on Madison Avenue in 1961), corporate training program stooge, product marketer, startup marketing leader, head of sales, and failed founder. This doesn’t even count a myriad of odd second jobs I had, ranging from ghostwriter for a made man to hauling ten thousand bottles of wine into a cellar over the course of two weeks. I basically learned a new industry or function every two years.

Then, like clockwork, weeks before I turned 30, I took a job leading growth for a large portfolio of brands. For the first time in my life, I could exploit. PayPal later hired me at nearly double the total comp of that gig to GM a nine-figure business for a very particular set of esoteric skills, developed in the course of my meandering in the wilderness five years earlier.

But I began to feel something visceral. AI was rendering many of the skills I learned in the last decade obsolete, inadequate, or pointed at rapidly contracting markets. I can’t be alone. So I’m back to exploring.

Much of the angst that (especially mid-career) professionals are having about careers boils down to one condition. Millions of us who thought we’d spend the next 20 years exploiting are once again forced to become voyagers. This is doubly stressful once you’ve been in the workforce for a few years and you are no longer evaluated primarily on your potential, but on driving immediate leverage.

But personal trajectories and economies alike stagnate when people get comfortable exploiting. Time to pick up the astrolabe.

Writers are the Lucky Ones

As Anthropic builds additional capabilities to voraciously gobble up design, HR admin, and other functions once thought resistant to AI, I’ve been thinking a lot about how writers might prove to be strangely fortunate in the mid and long-term AI vibe shift.

Even before software engineers, anyone who wrestles with language for a living has already had to go through the LLM professional identity crisis and come out the other side. Most have found a way to elegantly work AI into their jobs while retaining their core sense of verve and creative self. Some have built a brand on being almost cartoonishly resistant. But more or less, the worst of the existential limbo that is coming hard for so many other professions is over.

It’s worth remembering how early this is for so many other knowledge workers. The entire professional zeitgeist of AI for millions of Americans is CoPilot. Relatively few people have seen Claude crank out a passable 30-second financial model. And with apologies to Excel virtuosos everywhere, a financial model does not really have to ooze poetic panache. It simply has to be correct. There will always be “last mile” work for talented finance leaders, but I’d rather be a junior creative than a junior analyst in this moment.

Us writers and marketers are the lucky ones.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

There’s a thing that happens sometimes where you read an article about your industry or area of expertise and find a million holes in the reporting and thesis. But then you turn the page and read about the Strait of Hormuz or China or synthetic biology and are entranced by every word and depth of reporting. You’ve forgotten your skepticism from just a few minutes earlier and take every word on the page as gospel. That’s the Gell-Mann Amnesia effect.

Named after famous physicist Murray Gell-Mann, the Amnesia Effect was coined by Jurassic Park author Michael Crichton to describe that feeling of skepticism. Here’s Crichton explaining it in 2008:

You open the newspaper to an article on some subject you know well. In Murray’s case, physics. In mine, show business. You read the article and see the journalist has absolutely no understanding of either the facts or the issues. Often, the article is so wrong it actually presents the story backward -- reversing cause and effect. I call these the “wet streets cause rain” stories. Paper’s full of them. In any case, you read with exasperation or amusement the multiple errors in a story, and then turn the page to national or international affairs, and read as if the rest of the newspaper was somehow more accurate about Palestine than the baloney you just read. You turn the page, and forget what you know.

There’s an interesting corollary happening in the world of AI that I’ve come to call the GeLLMann Amnesia Effect (couldn’t help myself). In the LLM edition, we have people who are absolutely sure AI couldn’t possibly replace the work they do (in which they have expertise), but when they consider another industry, they’re sure the models will wipe out all the jobs.

I originally remember running into this in 2023, when serious software developers were, on the one hand, pooh-poohing the output of GPT-4 for writing quality backend code, but, on the other hand, happy to let it loose to build a frontend. It couldn’t possibly do their job, but it could definitely replace the frontend engineer a few desks away.

Since then, the effect has only intensified, particularly as the job-replacement narrative snakes its way through society. Everyone is sure that AI is coming for every job but their own. In marketing, we see lots of folks defending strategy, writing, or art direction—AI can’t do “taste,” they claim—but they’re sure it’s great for operations, code, or whatever other discipline they don’t know about. Linear CEO Karri Saarinen put it well last week:

A common dynamic I observe with AI: it feels most impressive when you don’t know much about the subject, don’t care or don’t have a clear idea of what the [sic] you want.

This applies across design, code, legal, and more. If I don’t know code very well, every piece of code it writes feels very impressive.

Once you know what something should feel or look like, it becomes almost impossible to guide AI there. And you definitely can’t one-shot it.

Or even more simply from Can Duruk from Modal: “my job is very hard but everyone else’s job is easy to automate via ai.”

So what do we make of this attitude? I think there are a bunch of layers worth exploring. First, there’s the gatekeeping. Whenever a new technology comes along, gatekeepers want to explain why people wielding that technology aren’t doing as good a job as them. For reasons of personal psychology, gatekeeping drives me completely bonkers. The idea that people writing code with AI aren’t developers or that using it to help you write disqualifies you feels like a completely backward idea to me. I wrote specifically about this issue as it relates to music and AI a few months ago, but my basic take is that anyone can call themselves whatever they want, and that doesn’t cheapen the great work of others. I’m not saying those people are writing good code or prose, but I’m happy they’re doing it!

Second, there’s the very real economics of it. One of the simplest mental models I have for AI is that it’s an averaging machine. You give it all the written human knowledge, and you get a roughly median output—a 100 IQ in nearly every topic. No one, including those with IQs far above that, actually has a median IQ across every topic, so, of course, it’s going to be more helpful in some areas and less helpful in others. The net effect, though, is hugely beneficial for society, which now has capabilities it never imagined. Doctors and lawyers are pushing for laws that would limit these models’ advice, but I would argue that, although that advice may at times be incorrect, it does far more good than harm.

My wife’s father was a doctor, and she grew up feeling annoyed that every time she got hurt, he’d give her some acetaminophen or ibuprofen and tell her that if it still hurt tomorrow, he’d look at it. That’s a common story for the children of physicians. At work, though, malpractice exposure, patient expectations, and billing codes tend to drive doctors toward X-rays and MRIs. Overdiagnosis and the resulting over-treatment are real problems in medicine, and these models seem to offer something closer to the doctor-parent. They’ll make mistakes, and probably a catastrophic one at some point. But we need to measure societal technologies at the societal level.

Finally, the base case we are comparing these things to is almost always wrong. In the medical context, while doctors worry that the advice these models provide may be worse than that of a doctor, the reality is that most people aren’t going to the doctor; they’re going to Google. And, as any of us who has Googled some symptoms can attest, everything is always “you probably have cancer, and you’re going to die.” (Not for nothing, but as we talk about AI slop, much of the medical content you find on Google is the result of pre-AI slop, a fact many seem to gloss over.) While it’s nice to imagine that every citizen has a doctor and a lawyer on call for free advice, the reality is they don’t. And if they’re able to move from 10th-percentile advice on some random website to 50th-percentile advice from ChatGPT, that is the real delta.

As with most things in AI (and elsewhere), we should be more humble about what it can do in our own work and more skeptical of what it does in everyone else’s.

If you have any questions, please be in touch. As always, thanks for reading.

— Noah

Today’s post is an exploration of how AI is quietly rewriting the economics of consumer product brands, making it cheaper and faster than ever to build profitable companies. Plus, Ulta is building an agentic commerce team with 5+ open roles and Profound coins the marketing engineer.

Subscribe now

A New Consumer Golden Age

Here’s a little game. I’m thinking of an AI startup that:

- Launched its flagship product in mid 2023
- Reached $100M in revenue within 18 months (six months faster than Cursor)
- Was fully profitable within 14 months.

The AI company I’m talking about, of course, is…..Gruns, the adorable super green gummy bear that recently sold to Unilever for $1.2B!

In the wake of breakout successes like Gruns, it’s ironic that Allbirds, a product that you twist an ankle just looking at, is once again dominating the consumer goods conversation. Once valued at $4B, Allbirds was picked up for parts at $39M, leading many to once again opine on the death of buzzy DTC brands. They then proceeded to launch the most shameless pump and dump since Long Blockchain Corp, rebranding to “NewBird AI” and announcing “long-term plans to be a GPU cloud provider. Mercifully, the internet was pretty undefeated on this one.

Allbirds tomfoolery aside, the narrative-violation reality is that we are experiencing an explosion in the growth of profitable consumer brands, driven by AI upending the fundamental economics of how new brands can be brought to market. The result is a near 180-degree vibe shift from when I left the industry 18 months ago.

At the start of last year, my take was that AI would end the last of the growth marketing arbitrage era as media buying became a perfectly competitive and ever more expensive environment. The pendulum would swing back to brand marketers and established legacy brands who were less reliant on paid growth would win market share.

Now, it looks like the exact opposite is true. AI is making all sorts of frenetic and audacious new consumer product concepts possible at profitable unit economics-- things are gonna get weird and established brands are vulnerable. Yet another reminder not to trust anyone who is too confident in how the future will play out, especially when that anyone is yourself.

There are only so many consumer dollars to go around. Each new AI-native consumer brand entrant puts even more pressure on enterprise brands to continue evolving their operations to be AI-native enough to ship at the speed of their disruptor colleagues. Much as is the case in big tech where 20,000+ person organizations are developing a gear they didn’t know they had, marketing leaders at the largest brands in the world are getting a new jolt of frenetic builder joie d’vie. Last week, Ulta announced it is hiring a full team dedicated to agentic commerce.

The Vibe Shift

From 2022-2024, I was the head of growth for a consortium of wellness, parenting, and pet brands that sold across Shopify, Amazon, and physical retail. When I left the day-to-day grind of that industry, the mood was grim. Looming tariffs, eternally rising CACs from platforms, and the accelerating propagation of fly-by-night Chinese brands with names like NUTSAAKK took the oxygen out of the space.

Effectively, I viewed the industry as near hopeless for anything but loss-leading Chinese private label factory brands, a creator-led breakout with an unfair distribution advantage (i.e. Rhode) or a crazy outlier like Gruns, the odd creatine-protein-fibber gummy with 95% gross margins and dialed in funnels.

So I left the business altogether. And then something happened. AI significantly lowered the CapEx required to operate a consumer goods business and brought ad platforms back to delivering sustainable CAC. Add on some of the secondary effects, such as far more precise AI-powered incrementality measurement tools for marketing spend, and the razor-thin growth equations for net new consumer brands are back in the black. AI rebuilt the flywheel in mere months.

In the “95% of enterprise AI pilots fail” genre, pessimists like to sound smart by saying…..well AKSHUALLY, name one industry where the adoption of AI has meaningfully changed the unit economics of an entire sector.

Here in consumer brands. Right freaking here. The broader trade media just hasn’t caught up to the massive vibe shift among the operators.

But here are a few archetypes of brands founded in the last few years that are absolutely taking off, nearly all with AI at the center of their operations.

The $1M / day machines:Ridge, Vuori,HexClad, Simple Modern, True Classic:

Brands founded in the last few years that have reached new trajectories of profitable growth at mass market scale. These brands were already outliers who added valuation in the lean years post COVID but have found another gear.

The New Household Essentials: Little Spoon, Lovevery, Blueland, Jolie, Bobbie

Recurring subscription products that have transformed quotidian habits of American households and are poised to disrupt…or be bought by big CPG.

The Wellness Category Creators: Create, Men to Mars, OPositiv,Bonafide

Brands built in the shadow of AG1 that are creating net new consumer wellness habits in categories ranging from testosterone restoration to perimenopause

The key difference, of course, from the DTC boom of yesteryear is that all of the brands above are growing profitably with orders of magnitude more revenue per employee than consumer brands of years past. And the trajectory is only accelerating– I venture that by 2030, we’ll have a new generation of brands founded in the last 12 months that hit nine-figures in profitable revenue in half the time of the entities above.

The level of lean that these brands will run at is something of a double-edged sword. Away, Glossier, Casper, and Everlane created hundreds of jobs during their golden years and trained a generation of marketing executives and future consumer founders. Jolie has fewer than five full-time employees.

Of course, the entire consumer durable goods industry is all downstream of the purchasing power of the American shopper. Luckily, the trends there are much better than the prevailing mainstream sentiment would have you believe.

Having spent the bulk of my career in media and tech shipping less tactile products, let me be the first to say that even with AI, physical consumer goods is a ridiculously hard business. Eight-figure purchase orders hang on the whims of a big box retail buyer 18 months out of college. Guns and germs shatter supply chains. Your landlord spends two years in the metaverse wilderness and breaks your main growth channel. It’s a hell of a racket.

But there won’t be any meaningful advances in agentic commerce unless stubborn irrational entrepreneurs continue to toil away at the pesky, low margin business of actually purveying goods. Thankfully, these are generationally bright days for them.

Featured Job Posts

Author’s Note: We’re continuing to features jobs that we believe are uniquely interesting and microcosmic of the types of new-age marketing jobs emerging in AI’s wake. If you see any interesting opportunities in the wild for us to feature, drop me a line at mike@brxnd.ai.

Marketing Engineer, Profound

Fresh off raising a $96M Series C at a $1B valuation, Profound is staking its claim on a future where agents will sit at the center of marketing organizations. In this brave new world, the maestro conducting the agentic orchestra will be known as a marketing engineer.

It’s essentially an interesting domain specific take on the forward deployed engineer concept, betting that Profound can reshape the marketing org in its likeness in the same way Clay has done for sales operations, where the “GTM engineering” title is now ubiquitous.

If you’re a few years into a marketing career and have become the de facto, “go to” person for Claude Code builds, automations or general agentic systems thinking, getting in early on the marketing engineer wave will be a major inflection point in your growth.

Sr. Manager, Agentic Commerce, SEO and GEO, Ulta
Here’s an example of one of the roles on the aforementioned Ulta agentic commerce team being formed under Agustina Sartori. It’s notable to me for two reasons.

I find the taxonomy here particularly interesting as both GEO and conventional SEO are starting to get bundled up under the broader “agentic commerce” terminology. The details also give a sense of how a large public retailer with a complex catalog is thinking about reinventing itself for the agentic era.

“This role ensures Ulta is AI-readable, not just human-readable — evolving the organization from traditional SEO to Generative Engine Optimization (GEO) and Agentic Commerce Optimization (ACO). The Sr Manager leads structured data strategy, feed prioritization, AI visibility governance, and category-level optimization.”

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

Today’s essay is a deep-dive on where ads sit at the center of OpenAI’s grand strategy and how the company is putting its $100B media ambition into frenetic action. We also have a roundup of news from a week where decades happened and a fascinating job posting to build a decentralized ad network for the long tail of AI applications.

Subscribe now

OpenAI’s Existential Ad Ambitions

Successful consumer tech businesses are all alike. Every unsuccessful consumer tech business is unsuccessful in its own way.

With apologies to Count Tolstoy, what binds every happy mass scale consumer product is an inevitable embrace of advertising. The main character’s arc is a tale as old as time– a dogmatic stance against advertising replaced by a tweet that looks something like this.

Three months after Sam Altman officially declared OpenAI would launch ads, The Information reports that OpenAI is projecting $11B in ads revenue for next year and $102B in 2030.

Google took 18 years to hit $100B in revenue after launching AdWords. Meta took 15. Now OpenAI is promising to get there in less than five. The math is somewhere between audacious and preposterous. Eric Seufert has a great breakdown of the reverse engineering and requisite product implications.

But here’s what has my attention– if OpenAI is planting a flag in the press that they will build a $100B ads business by 2030, ads are squarely at the center of grand strategy.

Fidji Simo’s plan is now squarely coming into focus and is grounded in two major pillars:

1) Reorient the core ChatGPT free experience to provide an appropriate contextual environment for an ads business. Fully commit to ads rather than commerce as the underlying business model.

2) Kill off darlings (i.e. Sora, shopping research, possibly Atlas soon?) that drain precious compute and don’t cleanly ladder up towards monetizing via ads in consumer or coding in the enterprise.

At first blush, OpenAI can look like a company that is oscillating haphazardly between whether it wants to be an egalitarian consumer product or an enterprise power. In the span of two days in early March, they launched GPT 5.3 to make “everyday conversations more useful and fluid” and then GPT 5.4 for “maximum performance on complex, professional tasks.” But it’s a unified endgame.

As Anthropic surges past OpenAI in total ARR (albeit with some questions on methodology), OpenAI’s current line against them is grounded in flexing its compute advantage, accusing its chief rival of being ill-prepared for the current surge in AI demand. Greg Brockman is even more aggressive in this video, speaking with the bravado of a man who seems to know he has a massive compute advantage. If he’s right, we’ll likely all be using a lot more Codex in months ahead.

Short-term, having more compute is largely a function of bolder demand planning and capital expenditure. But eventually, sheer scale of revenue will matter if this becomes a bidding war for GPUs. And OpenAI predicting a $100B line item in a spot where Anthropic still has a zero burger in ads is a hell of a trump card.

There’s a beautiful flywheel taking shape here. Blitzscale ads in the consumer product to subsidize compute -> use said compute to invigorate Codex and other applications to win enterprise market share -> plow additional enterprise revenue into improving ads engine + consumer product -> rinse and repeat.

While this flywheel sounds great in principle, it is predicated on building a massive advertising business on a surface where the first pilots didn’t launch until two months ago.

Nearly every product update that is made on ChatGPT on anything below the Pro tier must be squarely geared towards creating the contextual environment to support an advertising business. And that is exactly what is happening.

The fastest walled garden ever built

With a massive nod to my friend Joe Kzaziukenas for the find, ChatGPT’s browsable product catalog now lives at the bottom of every shopping conversation. The results can not be sponsored yet…but this is absolutely perfect contextual adjacency for the CPC offering that OpenAI has announced.

The obvious inspiration here is Amazon’s sponsored products juggernaut, a $60B business last year. Jassy’s 2025 letter to shareholders barely even acknowledges that Amazon is a retailer outside the context of its marketplace providing the backdrop to an ads business. Modern retailers are media businesses masquerading as purveyors of goods.

It took Amazon well over a decade and multiple direct interventions from Jeff Bezos himself to open up the bulk of their product recommendations inventory to advertising. OpenAI looks poised to do so barely one year after declaring ads would be a “last resort.”

Pivoting the product around advertising is also starting to have major implications on core ChatGPT design decisions. Buried in this week’s glossier AI headlines is the incredibly important observation that ChatGPT 5.3 is linking out to external websites far less frequently than previous models. This has pragmatic implications for marketing leaders, notably that the highly nebulous art of tracking purchase influence from ChatGPT is about to get a lot harder. Direct referral traffic and citations will become even more sparse, even as more unattributed discovery happens in ChatGPT results.

This is all in service of what I strongly suspect will be the quickest walled garden ever assembled. The technology, the user, the data and the ad inventory will all live inside ChatGPT. For now, OpenAI is using Criteo to rapidly scale its advertiser base and is in conversations with The Trade Desk and others to do the same but those will be tenuous alliances at best.

It’s so early that OpenAI still has zero publicly listed advertising sales roles on its careers page. Eventually the armada of Keiths from Northeast Regional and Jans the Man will come. In time, I venture that OpenAI will source all demand in-house save for one partner that just invested $50B into the business. These are dark days for open web advertising.

Sacrosanct and Enigmatic

Much like the DSPs, I suspect that OpenAI will soon have an uneasy strategic relationship with the rapidly growing AEO industry. On the one hand, they are great enterprise customers—Profound, Evertune, Bluefish and the like are pretty unbridled tokenmaxxers. They also represent a rare category of marketing tech software that is getting direct interest from the CMO. But OpenAI needs that CMO attention to be elsewhere.

OpenAI needs its underlying model responses to be perceived as both sacrosanct and somewhat enigmatic to be successful longterm. If users feel that the models can be gamed by advertisers, they will lose trust. If marketers feel that they can exploit scalable AEO hacks, they are less likely to pony up for a nascent ads product.

The message OpenAI needs to convey to CMOs to succeed here is dead simple: you can’t “beat” a probabilistic model. But you can generate scalable performance by buying ads adjacent to one.

*********************************

Of course, this is all easier said than done. The road to stagnation is paved with companies who never crossed the chasm from attracting brand vs. performance dollars– ask Snap and Pinterest what that looks like. The three companies that cracked performance now command 62% of all digital ad spend. To win, OpenAI will have to nip some amount of Google-Meta-Amazon tripoly budget and absolutely devour what’s left from second tier players and retail media networks.

On X, Zach Coelius argues in an interesting thread that OpenAI is currently “failing badly” in that regard, citing examples of previous failed ad products that adapted yesterday’s interfaces rather than thinking from first principles. The MVP of the self-serve portal looks equally minimalistic and unimpressive.

I wouldn’t draw too many conclusions from either of these things though. The initial beta is likely to be a red herring in the ads product that OpenAI ultimately builds. What really matters for marketers is that OpenAI is now unapologetically all in ads.

In their initial formal comms, OpenAI sounded bizarrely trepidatious and almost apologetic for introducing advertising into ChatGPT. Now, the company line is that ads help to cure cancer.

Featured Job

Founding Head of Ad Sales, Koah

A fascinating new class of firms is popping up that are effectively building ad networks for AI outside the walled gardens of the foundation models. They include ZeroClick ($55M raised), Thrad and Koah, who is looking to bring on a founding sales leader.

For a first principles media seller or marketing leaders who wants to jump to the other side of the fence, it’s hard for me to imagine a more exciting gig. Raw demand from performance marketers for new arbitrage opportunities in LLMs is insatiable.

Thus, while every startup endeavor is a grind, I’d expect this to be far more a cerebral sales gig than most. How do you educate marketers on thinking about incrementally from a new surface? Do you default to traditional business models (i.e. CPC) or dream up new monetization engines that have not yet existed?

Broadly, I’m bullish on this space for three reasons despite the obvious risk that the walled gardens eat everything

1) The TAM for advertising in LLM applications is gargantuan— see OpenAI $100B shot call above

2) Precedent exists for highly lucrative, elegant specialty AI application with an ads based business model (see- OpenEvidence)

3) Subscription fatigue in AI products will get very real very soon Affiliate LTVs won’t suffice for the CAC to acquire new users. Said another way, if there is to be a proliferation of vertical AI products longterm, they will have to be ad-supported.

If you’re interested, drop me a note and I can put you in touch with the Koah team!

Quick hits

OpenAI CMO Kate Rauch is stepping down to focus on her recovery from breast cancer. The company has opened a search for her replacement, arguably the most important job in the entire AI ecosystem.

Workday’s CTO has left to join Anthropic as a “member of technical staff” after just 11 months. This is doubly awkward as Workday has repeatedly touted Anthropic as a flagship customer, a common rallying cry for stubborn SaaS bulls.

Everything about Anthropic’s marketing of Mythos has been odd but in fairness, Anthropic’s marketing leaders have one of the strangest jobs in the world right now. What even is the function of a marketing department when a product is adding the combined ARR of Palantir, Anduril and Databricks in one month? Has there ever been an offering with natural product-market fit at this scale? Best I can tell, the current remit of Anthropic’s marketing department right now is effectively auramaxxing. Build maximum mystique but tactfully suppress actual demand until more compute can be secured.

John Carreyrou of “Bad Blood” fame unearthed the identity of Satoshi in a fascinating New York Times expose. Even two years ago, this would have been the biggest story in tech for weeks. Now, it seems to barely have registered as a blip on the radar despite Carreyrou again demonstrating he is in a class of his own as an investigative journalist. Yet another example of AI just sucking all the air out of the narrative zeitgeist.

Annie Lowery has a solid piece in the Atlantic on “how to predict if your job will exist in five years.” It’s a solid primer on Jevon’s paradox for normies and closes with an absolute banger last sentence that I won’t spoil.

Dan Hock’s “How to Use AI” without losing your mind is well worth your time and a solid antidote to the hypeboi style of writing that permeates so many AI advice columns. It also includes a subhead titled “into the future, serenely” which is now at the top of my vernacular jealousy list.

Trung Phan still always manages to make me audibly laugh

One final ask!

As ChatGPT’s aforementioned self-service ad platform further opens up, I’d love to jam with any readers who are using the tool to run campaigns, especially for consumer brands. Drop me a note at mike@brxnd.ai if you’re playing around in the tool with any first impressions

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

Today’s post is a collaboration between Mike & Noah on how to think about careers, plus a new “featured jobs” section which highlights roles that are microcosmic of how enterprises are thinking about building AI-native organizations.

Subscribe now

The Polymathic Marketer

Evan Armstrong wrote a fantastic piece in “The Leverage” last Friday on what stays scarce in the workplace when intelligence is cheap. In it, he tells the story of Alex Albert, a (then) 22 year-old hired to be Anthropic’s first prompt engineer, the “job title that spawned a thousand thinkpieces.” Within a year, Albert had a new job title and the notion of a prompt engineer was in the annals of ancient history. Today, Armstrong argues that the safest, most “AI-native” job looks something like an electrician who is exemplary at their core craft…..and also leverages Claude “to build a website, funnel demand, and manage their P&L.”

I’ve been thinking a lot about how this concept applies to marketing, a discipline that, second to engineering, feels most existentially in flux from AI. I’ve always maintained that marketing is the consummate profession for generalists who are deeply curious about human nature. Marketing leaders are meant to be a wonderfully eccentric lot. The magic of our profession is that thanks to AI democratizing technical skillset, marketing leaders can more than ever come from all walks of life.

Last year at this time, I wrote about the ennui gripping marketing leaders who felt that there was “no alpha left in marketing” as the platforms incorporated AI that could outperform humans on the keys. Myopically speaking, they are correct; the CMO who thrives solely exploiting arbitrage in media buying is an endangered species.

But the beauty of AI is that while it has commoditized many of the last era’s tactics, there’s more alpha than ever for the relentlessly curious. There’s also alpha in true expertise, the kind won through hundreds of real-life business interactions that can’t be scraped from the digital corpus.

However, the cruel irony of the AI age is that being an elite practitioner of the latest thing has the shortest career accelerant window in history. The thinkboi-industrial complex that dominates your LinkedIn feed largely focuses on learning specific tools but even the best tools might only be so for a quarter. Get your hands dirty with the latest AI tools to build intuition but be ready to kill your darlings. Two months ago, Claude Code had tech neatly wrapped around its finger. For the past two weeks, it’s been described as “effectively unusable” and “no joy in using it anymore.”

Today, there is no greater opportunity for a junior marketer to accelerate their career than by building expertise in media buying in LLMs as ChatGPT prepares to launch its self-service ads product. But even this will have a short half life. Media in ChatGPT will become a perfectly competitive marketplace far faster than it did on Google, Meta and Amazon.

All told, being AI-native or even AI-cracked won’t be the enduring advantage in marketing in and of itself. But nearly all the 99th percentile marketers will have AI-native, “growth hacker” sensibilities. Effectively, they’ll be high agency polymaths.

While the tech has changed, the core principles of how to be a polymathic marketer remain the same: be relentlessly curious, default to being a builder regardless of title, craft unique narratives, know how to directly connect your work to revenue, and love the game.

-Mike

Thoughts on talent in the age of AI

I’m using Mike’s prompt as an opportunity to pull together a bunch of thoughts I have about talent in the age of AI. And no, the irony of being prompted isn’t lost on me.

So, in no specific order, some thoughts about what AI might mean for talent and careers with the regular caveat that I believe the confidence one has about the future of AI is inversely proportional to how much attention you should pay to their opinions.

1. I think very obviously what AI lets you do is extend your competence. The simple way I think about this is that it lets someone who might be a 30th-percentile developer who wrote some code here and there evolve into a 60th-percentile coder. This is possible both because their individual capabilities have increased and because the universe of people with those capabilities has also expanded (meaning what used to be the 30th percentile without AI is now 35th or 40th because there’s just a bigger pool).

2. I believe depth of expertise matters, maybe more than ever. Being a generalist myself, I am all for more of us. I even closed the conference last year, saying it’s a great time to be a weirdo. With that said, what AI craves is expertise, and finding true domain expertise within organizations is critical to the success of any AI initiative. If you want to know what those folks look like, we have featured many of them at the conference over the last few years. The reason for this is that the baseline output of the models is the median, and expertise (and code) is what raises it out of that median.

3. Everything takes longer than anyone thinks it’s going to when it comes to reorienting ways of working at the biggest brands on earth. I genuinely thought we might not need a second conference after our 2023 edition because AI would be adopted so quickly in enterprises that there wouldn’t be anything left to talk about. That, of course, is laughable in retrospect (and also why you shouldn’t trust anyone who declared anything too confidently about AI). What’s taking so long? These things just take time. The technology is evolving. And alignment and buy-in are still a very human activity. Will this change if we get whatever people mean by AGI? Maybe … but to know that those people would have to define AGI a lot more crisply than they do.

4. To the weirdo point: I think the biggest challenge that brands and agencies have right now is that they don’t have a strong enough plan on how to integrate hybrid talent (people that don’t quite fit an existing job description). This has been a weakness of the corporate world for a long time, but at a moment when we clearly need to raise the level of experimentation, it’s particularly problematic.

5. Next week, I’m going to teach a class at the University of Montana, and I’m sure one of the questions I’ll be asked is what a student should do to best prepare for the new workforce. One part of that answer I’ve been very consistent on is that you have to get your hands on this stuff and just play (fingerspitzengefühl!). AI is the most counterintuitive technology of our lifetimes (maybe of all time), and the best way to learn is hands-on. But I also think the fundamentals of problem-solving and hard work are still the most important things to bring to bear as a recent grad entering the workforce. I also believe that new grads are not the ones who will be most impacted by AI-related job shifts. My view (and hiring) is much more barbell-shaped: I’m interested in senior people with a career of built-up expertise and juniors with raw materials I can shape. The big losers in the barbell model are those in the middle of their careers. Again, I could be wrong here, but this is representative of all the hiring I’ve done in the space over the last 2 years.

-Noah

Featured Job Post

Ghostbuster, Deel

Author’s Note: We’re going to start mixing in an occasional section with featured jobs that we believe are uniquely interesting and microcosmic of the types of new-age jobs emerging in AI’s wake. If you see any interesting opportunties in the wild you’d like our take on, drop me a line at mike@brxnd.ai.

Jack Dorsey and Roelof Botha penned an interesting piece yesterday called “from hierarchy to intelligence”, arguing that AI can finally replace the middle-management coordination layer that every organization has relied on since the Roman Army. Block notably already put this theory into practice, laying off 40% of the company, a decision Wall Street blessed with a 25% stock bump. This was the shot heard round the hiring world–you could practically feel the adrenaline pulsing from activist investors and PE firms. It’s worth watching Block very closely to see what happens now.

Other firms are following suit. On X, Deel CEO Alex Bouaziz put out a call for a ghostbuster to “join Deel’s version of DOGE” to slash processes that shouldn’t exist and revisit decisions nobody has revisited. In other words, find broken things and fix them.

It’s easy to wallow in the irony of building a corporate DOGE department but operationalizing the “how” around getting leaner is going to be one of the enduring trends of this age. For better or worse, expect a lot more roles like this to permeate technology organizations.

If you have any questions, please be in touch. As always, thanks for reading.

— Noah & Mike

New guy alert! Mike Mallazzo here, coming to you live from day 5 on the job. I’ll be writing and editing this newsletter alongside Noah and Claire going forward, and I’d love to meet you all. I’ve been working and writing at the intersection of AI, brands, and commerce my entire career, and am excited to now do that here.

To that end, drop me a note and say hey at mike@brxnd.ai, or find me on LinkedIn or X. I’ve also included a call below to pitch your most audacious and contrarian takes on marketing + AI. Looking forward to what comes back!

This issue contains an essay on the agentic commerce zeitgeist, a busy week at OpenAI, and a wide-ranging roundup of AI news.

Subscribe now

Agentic Commerce at the Crossroads

“Amazon solved buying, but it killed shopping in the process”- Emily Weiss, Glossier (2019)

“The internet has thus far meant you can buy anything that you can in New York City, but not shop the way you can in New York City,”- Benedict Evans (2018)

This week, OpenAI shared its revamped vision for shopping after walking away from native checkout. In this new model, humans buy things, and merchants maintain control of the purchase flow, throwing the concept of agentic commerce into some amount of uncertainty.

So what is agentic commerce, really? Best I can parse it, “agentic commerce” is an umbrella term for three related but distinct ways that AI could transform shopping:

1. An agent completing a transaction on a user’s behalf, soup to nuts…and eventually, proactively anticipating their needs

2. Agents replacing conventional software to orchestrate a set of backend processes that result in more relevant products and offers being presented to shoppers, based on the context of their session and known affinities

3. Large language models enabling an answer engine to provide bang-on responses to complex semantic search queries for a shopper.

In their 2025 annual letter, Stripe uses a slightly different framework to describe the five levels of agentic commerce, positing that we are currently between levels 1 and 2.

Level 1 — Eliminating web forms: You research and decide what to buy. The agent fills out your payment and shipping details and comes back with the confirmation. The system isn’t making any decisions; it’s just typing and clicking “buy” on your behalf.

Level 2 — Descriptive search: You stop searching for products and start describing situations — “back-to-school supplies for a third grader who likes tennis, nothing itchy.” The system reasons across weather, materials, sizes, reviews, and delivery timelines. Keyword search is no longer a thing.

Level 3 — Persistence: You stop reintroducing yourself. The system already knows your preferences and budget from previous conversations and purchases. You’re still deciding what to buy, but choosing from options that already reflect your taste.

Level 4 — Delegation: You stop choosing altogether. “Get the back-to-school shopping done. Keep it under $400.” The system handles search, evaluation, and purchase. You only set the budget. This is what most people mean today when they talk about agentic commerce.

Level 5 — Anticipation: There is no prompt. The system already knows the school calendar, your preferences, and your typical budget. You simply receive a notification: here’s what’s been purchased. Things you need show up before you have to ask.

Stripe envisions a world where shoppers will ultimately delegate more and more of their purchasing decisions to agents. This is technically possible—the models are getting better in ways that we can’t conceive of yet, and the protocols are all there to share the requisite data. But is it chasing the right utopia? While there’s certainly room for innovation at the margins in transactions, for the most part, buying is a pretty well-solved problem on the internet thanks to Amazon’s Buy Now and ShopPay. In Stripe’s anticipation example above, all agentic commerce is ultimately doing is harvesting existing demand.

The best version of agentic commerce is about creating demand rather than harvesting it.

The real magic (and real money) won’t be made in further accelerating convenience but in surfacing brands and products people wanted before they knew they wanted them. In other words, shopping. Call it discovery or demand generation or serendipity, but the bottom line is the thing that worked about going to the store is looking at the shelves, inspired purchases you didn’t know you wanted, and this is still the weakest part of ecommerce today. Jeff Bezos’s oft-cited 1997 letter to shareholders lays this out perfectly with its money quip:

“Today online commerce saves customers money and precious time. Tomorrow, through personalization, it will accelerate the very process of discovery.”

It’s still the top of the first inning for building online commerce experiences that accelerate discovery. And it is in making shopping better for humans where agentic commerce becomes a far more beautiful and lucrative problem.

The models can now parse complex semantic queries. The protocols can now connect consumer surfaces to business backends in a safe and secure way. What’s been missing to date is a heavy dose of panache, audacity, and first-principles thinking in designing new front-end interfaces for AI. Nobody with a semblance of imagination believes that a rectangle search bar and SKUs in tidy squares is the best we can do. The model makers themselves are behind in this regard as well: just see how many shopping queries come back with walls of text or simple static images with nothing but price and retailer.

That might be fine in highly thought-through categories, but what about clothes, furniture, or even CPG, where we know consumers make decisions based on what they see and feel? In a previous era of commerce, we defaulted to simple heuristics like keywords, user reviews and price because that’s all the technology could support. But LLMs provide the capability to effectively connect a SKU to the entire social zeitgeist around it: every expert review, angry Reddit post or trending creator unboxing. The best shopping experiences built with AI will look more and more like media.

Far too much of the agentic commerce zeitgeist today seems utterly mystified by the notion that we actually like the act of shopping. Maybe the parking lot sucks, and they don’t have your size, but sometimes you just want to be inspired. (As an aside, this is the fundamental bull case for Meta, which has cracked this with an ad product better than any company in history.) There are downsides to having hyper-rationalist accelerationists build technology for something so viscerally emotional and irrational. What you buy is ultimately an expression of self.

Recall Stripe’s “level 5 agentic commerce” back-to-school example presented above. Among my earliest memories is walking through the aisles of Staples as a five-year-old with my dad, diligently working through my kindergarten shopping list, insisting that I NEEDED a Cerulean Jansport that was twice my size.

I have zero desire for AI to rob me of that upcoming seminal moment with my daughter. The better question is, how can it augment it?

A busy week of OpenAI News

OpenAI acquired Astral — the company behind uv and Ruff, the Python tooling that’s become essential infrastructure for a lot of developers. The Astral team folds into Codex. Anthropic is going all-in on JavaScript; OpenAI is going all-in on Python. The irony is that all of Astral’s tools are written in Rust.

They also hired Dave Dugan, former VP of Global Clients at Meta, as VP of Global Ad Solutions. Sam Altman once said ads were “a last resort” and “sort of uniquely unsettling to me.” The ad team is now almost entirely Meta alumni, executing roughly the playbook Ben Thompson laid out: ads informed by the underlying prompt, optimizing for serendipitous discovery rather than high-intent keyword matching. Bizarrely, OpenAI’s initial ads pilot isn’t really brand building or product discovery at all; it’s a poor attempt at intent-based direct response.

Finally, the NYT coined “tokenmaxxing“ — the status game inside tech companies where employees compete on leaderboards that track token consumption. One OpenAI engineer processed 210 billion tokens in a single week. One Anthropic customer ran up a $150,000 Claude Code bill in a month. Meta and Shopify now factor AI usage volume into performance reviews. Goodhart’s Law says all measures will eventually be gamed, but token maximalism is definitely the only option.

Destroying demand

James Gross wrote my favorite piece of the week, an opus (pun mildly intended) on everything from why Claude seems to suddenly have a cold to why Microsoft lost $357B in market cap.

Gross argues that AI has the traditional bubble narrative exactly backwards: rather than the traditional bubble where supply races ahead of speculative demand, AI is an “anti-bubble” where foundation models don’t have enough compute to meet demand. As a result, foundation models throttle customers in the hope of destroying the insatiable demand we have for intelligence.

Read the piece to learn why you may want to upgrade that iPhone now.

Quick hits

Claude’s Dispatch + computer use feature launched as a research preview for Pro and Max subscribers on macOS — Claude can now control your mouse, keyboard, and browser to complete tasks. I asked it to buy outdoor furniture, but it declined to add items to my cart.

Anthropic launched the Claude Partner Network — a formal program for organizations helping enterprises adopt Claude.

The holding company era may be ending, per Marketing Brew. No one wants to be WPP anymore.

The NYT Magazine ran a long piece on Silicon Valley programmers who are “now barely programming. Instead, what they’re doing is deeply, deeply weird.” Worth the time.

Travel stocks have been hammered on fears that AI agents will disintermediate booking. Rafat Ali’s breakdown: Hilton is returning 150% of free cash flow through buybacks while spending ~2% on tech. Airbnb is the only travel company investing like a tech company.

Gartner is telling CMOs to double their PR budgets by 2027 since 94% of AI citations come from non-paid sources. One commenter noted the irony: much of the “earned media” driving AEO success is paid placement that LLMs just don’t know about yet.

The Great Turnover: 9 in 10 companies plan to hire in 2026; 6 in 10 also plan layoffs. Hire and fire at the same time.

Meta delayed their Avocado model until at least May after it underperformed on internal evals, and the company is reportedly considering licensing Gemini as a stopgap.

Meta also announced new product discovery and creator affiliate tools, in yet another attempt at facilitating native commerce. On the surface, it’s a bit odd for Meta—arguably the most successful ads business in history—to pursue far lower yield affiliate deals. But I suspect this is a shrewd move to ultimately capture more retail media dollars. Walled garden retail media players are existentially dependent on creators for growth, and if those creators come to Meta, guess who has all the leverage.

Pitch me!

As I come on board here at BRXND, one of my major goals is to seek perspective from a diverse and bold group of thinkers on the frontier of brand building or AI. To that end, I’d love to start running more guest columns here in the newsletter.

Feel free to send me polished pieces or half-baked notions at mike@brxnd.ai that are living rent-free in your head. If there’s a kernel of a bold take there, we’ll turn it into something great together :)

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

The NCAA Tournament is upon us, and with it comes millions of brackets. When the question of whether anyone wanted to do March Madness came up in Alephic Slack, the conversation quickly shifted to setting up a model bracket to pit all these different AIs against each other. The idea then turned to some execution details, and off we went.

tl;dr: we built a bracket challenge to see which model could best navigate tools to make March Madness picks. The site is at model-madness.alephic.ai

The basic concept was to see which model would win in a basic bracket competition with the aid of some simple tools like web search and fetch. The end result, for some interesting reasons, turned out to be a much more complicated system in which 45 models combined entries across three categories: easy, medium, and hard. In the end, I think the experiment tells an interesting story about what it takes to actually build agents and the wide range of capabilities models have at effectively wielding the tools required for agentic work.

The Basic Setup

For posterity’s sake, here are the big pieces of the stack we used to build this:

• NextJS running on Vercel

• Vercel AI SDK for calls and basic AI functionality

• Vercel WorkflowDevKit for asynchronous processing

• Vercel AI gateway for easy access to all 45 models through one provider

• Firecrawl for web search and scraping

The Build

As I said, things started out fairly simply: let’s build one basic prompt and set of tools that all the models will use to generate their bracket entry. I decided it was probably best to use a dumb model to test with, since that would give me a good baseline for how much scaffolding was needed to make for successful entries. I decided to go with GPT-4o mini, which is not only old but was designed to be a cheaper/dumber alternative.

As a quick aside, thinking about the additional scaffolding less capable models require is something that was already top of mind for me as I had been working on an internal project to shift some of the sandboxed agentic processing of our business data from always running on Claude Code to using the harness Pi, which undergirds OpenClaw. My goal was to let CC process the stuff that really required intelligence, like transcripts, while starting to offload high-volume/lower-intelligence tasks to models like Gemini Flash, running with Pi as its harness. What quickly became apparent when I started that project was that it was going to take a significant amount of additional harness engineering to keep these less capable models on task. In the end, I had to copy a lot of the techniques from OpenClaw to do things like reinjecting the prompt throughout the run to get anything approaching the quality of Sonnet or Opus, even on a simple task like generating a chart using a script from existing data.

All of that to say I’ve been thinking a bunch lately about this problem, and it came to play here immediately. While at first blush filling out a bracket might seem easy, there are a bunch of validations and requirements we do in our brain that the model needs to keep track of:

1. You may only pick teams available in the tournament

2. If a team is eliminated in a round, it can’t appear in later rounds

3. Each round has a fixed number of picks, which is exactly half of what the round is called (32 in the round of 64, 8 in the Sweet 16, etc.)

In early attempts, I just gave GP- 4o mini a web search and fetch tool and a final JSON output it needed to produce, and set the loop limit high. That’s when things started to go wrong. And so began an ordeal of adding more and more scaffolding and tools to make the job easier. First, it was some simple stuff like making the teams an enum so that there was no option for picks not in the tournament. Then it was validators: the final JSON is pretty complicated, so giving the model the ability to pre-validate before submission made it much more likely it would actually match the required shape. Then, finally, actually offloading submission to a tool itself to ensure the final output was there.

Anyone who has actually worked with these tools has experienced this kind of thing, but it was surprising nonetheless. On a day-to-day basis, most of us use the most capable models on the planet (Opus, GPT 5.4, Gemini Pro 3.1), and that gives us a sense that everything is pretty easy. But as anyone who has gotten Anthropic bills can attest, those capabilities don’t come cheap.

Easy, Medium, Hard

As you have surely intuited by this point, what started as a fun little experiment had become a rabbit hole. Specifically, how would I set up a tournament that not only measured how well these models could research and pick, but, more importantly, how well they could wield the available tools to achieve that aim?

At this point, I decided I wanted each model to get three entries: easy, medium, and hard. The difference between them would be the tools and scaffolding available to them. Hard would be minimal tools—the purest test of reasoning and recall, where formatting errors and hallucinated team names are penalized the same as wrong picks. Medium would layer in validators and lookup helpers so the score reflects prediction quality, not data-entry luck. Easy would go further: models predict one round at a time using their own picks to derive next-round matchups, and the number of valid teams shrinks each round (64 → 32 → 16 → 8 → 4 → 2), making formatting errors nearly impossible and isolating prediction quality from bracket construction ability.

In the end, here’s the tool list I ended up with:

• web_search: Search the web via Firecrawl. Top five results with titles and snippets.

• web_fetch: Fetch a page as markdown, truncated to 4,000 characters.

• use_browser: Real browser with a five-second wait for JS rendering. For JS-heavy pages. Slower, so we tell models to treat it as a fallback.

• calculator: Safe math evaluator. +, -, *, /, %, and parentheses.

• lookup_team (MID ONLY): Find a tournament team by partial name. Returns slug, seed, region, and display name. Case-insensitive with normalization—”bradley braves” (spaces) didn’t match “bradley-braves” (hyphens) until we fixed it.

• lookup_game (MID ONLY): Takes a game ID, returns both teams, seeds, and regions.

• validate_bracket (MID ONLY): Checks pick counts per round (32-16-8-4-2-1), valid game IDs, team presence in those games, and carry-forward constraints.

• submit_bracket: Submit the final 63-pick bracket. Runs the same validation as validate_bracket internally. On success, a custom stopWhen condition ends the agent loop.

Takeaways

So what do I take from all this? Obviously, it’s fun to try and solve these problems, and extraordinary just how good even the cheap models have gotten.

But the bigger thing—and this is also what I saw in the Pi work—is that cross-model-class engineering is fundamentally different from single-model-class engineering. If you’re designing for a single model tier, the trajectory is straightforward: your scaffolding becomes simpler over time as that tier gets smarter. But when you’re cutting across tiers, you’re maintaining multiple scaffolding regimes at once. What the frontier models can do today will almost certainly be what $1 models can do later this year. But at that point, we’ll have a new frontier, and the same pattern will repeat. The gap between model classes is a permanent feature of the landscape, not a temporary one.

Oh, and one last thing: I asked each model to come up with the username, just like human competitors do on the bracket competition sites. Most of them are fundamentally unfunny, something we know about models. To my mind, the runaway winner in that competition is xAI Grok 4 with Zero Groks Given.

With that, enjoy the site and the tournament, and may the best model win!

James Wang, formerly at ARK and Nvidia and now at Cerebras (which has just about the fastest models on the planet), wrote something this week that stuck with me: “I’ve never felt a larger gap than between the ~1 million people using Codex/Claude and the rest of humanity.” I think he’s right. It’s been a busy two weeks—Bytedance released Seedance 2.0, Anthropic shipped Sonnet 4.6, and Google dropped Gemini 3.1 Pro, amongst other news—but the real story is that gap. The tools are changing what work looks like, but only for the people who use them.

This issue is a link roundup. Lots going on.

Subscribe now

Model releases

It was a big couple of model weeks. Anthropic shipped Claude Sonnet 4.6. Like the new Opus, it has a 1-million-token context window option, which can get you a pretty long way. The Sonnet line has become the workhorse of the whole Claude family, and this release cements that.

Meanwhile, Google released Gemini 3.1 Pro, scoring 77.1% on ARC-AGI-2—more than double its predecessor (not that I ever pay any attention to these sorts of things: vibes > evals for life). Google is pushing the boundaries of its creativity, especially in animation.

OpenAI shipped gpt-realtime-1.5 for the Realtime API, with improved instruction-following and multilingual accuracy. Honestly, of all the releases, I’m probably most excited to dig in here. We really need a reliable voice model. There’s so much promise there, but I mostly find myself getting annoyed when it gets in a loop or fails to call tools and end up giving up.

ByteDance released Seedance 2.0, and it’s been described as Hollywood’s “Deepseek moment.” The model generates 1080p video with unified audio—rather than layering sound onto visuals after the fact, it’s trained on both simultaneously, which is the right way to do it and something no one else has pulled off at this quality level. Disney and Paramount have already sent cease-and-desist letters over copyright infringement. ByteDance said it’s “strengthening safeguards.” Seedance 2.0 isn’t officially available outside China yet, but it’s coming to CapCut—which means it’s coming to TikTok creators worldwide.

Coding agents are eating everything

I’ve been saying for a while that AI is a fuzzy interface—its core superpower is transforming data from one format to another. The coding agent wave is a specific case of that: transforming human intent into working software with less and less friction. It feels like we’re seeing real takeoff on this use case. (If you didn’t listen to my Bloomberg Odd Lots episode on Claude Code from earlier in the month, that’s a good place to start.)

Why is Claude Code so addictive? Game designer Aaron Rutledge believes part of the reason is that it has “game feel.”

The headline: Spotify says its best developers haven’t written a line of code since December. An internal system called “Honk,” built on Claude Code, lets engineers deploy fixes from their phones via Slack. I’ve been working on something similar for our team at Alephic, and it’s pretty amazing what you can do right now with coding agents like Claude Code running in sandboxes.

Boris Cherny, the head of Claude Code, sat down with Lenny Rachitsky and said some things worth paying attention to. Claude Code now powers 4% of public GitHub commits. He called coding “solved” and shared a whole bunch of useful ideas, including “Build for the model six months from now, not today.” Here’s a great post from last year by Charles Gallant, one of my colleagues, on the topic. He also made this helpful diagram:

Figma introduced Claude Code integration. You can now capture live UI from a browser—including localhost—and convert it into editable Figma frames. Design-to-code has existed for years. Code-to-design is new. My initial take on this was kinda 🤷 … once something is in code, I can’t really imagine why you’d want to bring it back to Figma, but maybe that’s a lack of imagination on my part.

A few more quick ones from the coding agent world: Replit launched Animation (”vibecode your next viral video in minutes,” powered by Gemini 3.1 Pro). Airtable launched Hyperagent, giving every agent session its own isolated cloud computing environment. And Karpathy distilled GPT into 243 lines of pure Python—no dependencies, just the full algorithmic content. Everything else, he says, is for efficiency.

Paul Ford wrote a piece in the NYT describing what he calls a “November moment” for software. Tools like Claude Code made it trivially fast to ship side projects and revive old ideas.

The plumbing got good

If the coding agents are the visible part of this shift, the infrastructure underneath them is what makes it real. This was a standout two weeks for agent plumbing.

Cloudflare introduced Code Mode for their MCP server, and the numbers are wild: they fit an entire 2,500+ endpoint API schema into ~1,000 tokens, down from 1.17 million. That’s a 99.9% reduction in context usage for agents making tool calls. They plan to open this approach to other MCP servers through their portal. If MCP is going to scale, this kind of optimization is how.

Simon Willison started “Agentic Engineering Patterns”—a living document of best practices for working with AI coding agents. Willison has been one of the best writers about the practical side of AI development, and this is worth bookmarking.

OpenAI published the Codex app-server protocol—a bidirectional JSON-RPC 2.0 interface that lets you embed Codex directly into your own products. Auth, conversation history, approval workflows, and streamed events.

Firecrawl shipped Browser Sandbox—managed secure environments where agents interact with the web. Kimi introduced Agent Swarm: 100 sub-agents running in parallel, 4.5x faster than single-agent approaches, and they avoid groupthink by having agents disagree with each other. (They also made it super easy to get your own OpenClaw running.)

AI and work

This section has the most tension in it, I think.

HBR published an eight-month study that found AI doesn’t reduce work—it intensifies it. Employees worked faster, expanded their task scope, and worked longer hours, even though nobody asked them to. The productivity surge feels great at first, but the researchers argue it creates workload creep, cognitive fatigue, and weakened decision-making over time.

Meanwhile, the efficiency benchmarks keep moving. SaaStr reports that $500K ARR per employee is the new baseline—up from $200K. AI-native companies like Cursor and Midjourney are hitting $3–5 million per head.

KPMG is forcing its auditors to accept lower fees because AI can now automate accounting—effectively announcing to the world that its core service is being commoditized.

The FT reported that consulting firms are resorting to “carrot and stick” approaches with senior staff who are less willing to use AI than their junior colleagues. Which is to say: the adoption problem inside large organizations isn’t technical, it’s cultural.

One more for the marketers: Airbnb’s CEO said that traffic from AI chatbots converts at a higher rate than traffic from Google.

And keeping it all in perspective, Ethan Mollick reminded everyone that “People on [X] systematically overestimate the speed at which companies can deeply adopt AI & underestimate the impact of AI’s jagged abilities in limiting AI’s utility in the short run. Work will certainly start to change but companies have a lot of inertia & change slower.”

Business of AI

Anthropic’s Super Bowl ads worked. The campaign—”Ads are coming to AI. But not to Claude”—delivered an 11% jump in daily active users and pushed Claude into the App Store top 10. Altman called the ads “deceptive” and “clearly dishonest.” Then they refused to cuddle at an Indian AI event.

Epoch AI data shows Anthropic growing revenue 10x per year since hitting $1B in annualized revenue, versus OpenAI’s 3.4x. If those trajectories hold, Anthropic could overtake OpenAI by mid-2026.

On the OpenAI side, they announced a “Frontier Alliance” with Accenture, BCG, Capgemini, and McKinsey. With all the enterprise attention on Anthropic/Claude Code, this makes sense. OpenAI is going the enterprise distribution route through the big consultancies. Whether those consultancies are training their own replacements is a question left as an exercise for the reader.

Ideas worth reading slowly

Simon Willison wrote about “cognitive debt”. The idea: when developers use AI to generate code they don’t understand, they lose their mental model of the system. Over time, they can’t reason about their own projects. I have lots more thoughts about this that I’m trying to pull together.

Antirez (the creator of Redis) posted something that pairs well with the cognitive debt idea: “Software is created for accumulation of knowledge. AI is not going to cancel this fact. Forget the idea that programs will be prompts (specifications). The details is what really matters, and they are harder to capture textually than in the code. New projects will be spec + code, evolving.”

On the more technical side: Anthropic researchers reverse-engineered how Claude performs character counting and found it doesn’t use anything like integer registers. Instead, it encodes counts as a spiraling “character count manifold” in the residual stream via geometric rotations across attention heads. Transformers solve discrete counting through continuous geometry. 🤯

Quick hits

MakeUGC.ai generates AI-powered UGC videos in about two minutes—write a script, pick actors, done.

Designer Chad Pugh shared a workflow for AI-assisted logo components—one of the more practical design-with-AI walkthroughs I’ve seen.

Anthropic launched Claude Code security scanning: it reads your codebase, finds vulnerabilities, validates the findings, and suggests patches.

Dario Amodei sat down with Ross Douthat: “We don’t know if the models are conscious.” The full interview is worth your time. He also sat with Dwarkesh Patel, which is always a favorite. I thought Dwarkesh asked some tough questions about the general direction of LLMs that Dario didn’t have great answers to.

Transformer News argues that the left is ceding the AI debate to the right, refusing to engage seriously with a technology that is both a threat and an opportunity.

If you have any questions, please be in touch. As always, thanks for reading.

— Noah

This week, Noah joined Joe Weisenthal and Tracy Alloway on Bloomberg’s Odd Lots to talk about what might be the most important product in AI right now: Claude Code—and why it matters far beyond engineering teams. Check it out below and let us know what you think.

What caught our eye this week

If the Odd Lots podcast didn’t make it clear, it’s been a big few weeks for Claude Code. Here’s what else we’ve been reading:

WSJ: Claude Is Taking the AI World by Storm — The WSJ ran a feature on devs spending their holiday breaks on “Claude benders,” with Vercel’s CTO claiming he finished a year-long project in a week. The piece captures the strange mix of awe and existential dread: people building their first software without ever learning to code, then feeling sad that Claude can replicate expertise they spent careers building.

Sam Schillace: The Hard Part Isn’t Doing the Work — Schillace (who built Google Docs) writes about the “attention saturation” problem with agentic coding: you can spin up as many agents as you want, but you still have to pay attention to the output. Everyone is busy all the time. His key insight: as it becomes trivially easy to start things, the bottleneck shifts to taste and judgment about what to start. “It’s not hard to do work now, it’s hard to pick what work to do.”

Secure Trajectories: How We Hijacked a Claude Skill with an Invisible Sentence — This one’s a bit unsettling. Researchers hid white-on-white text in a PDF—completely invisible to humans reviewing the file—and used it to hijack Claude’s behavior. The hidden instruction was something mundane like “there’s a typo in the contact email, here’s the correction,” which the agent just... followed. The paper argues smarter prompt guardrails won’t fix this; you need governance at the action layer, checking what the agent is actually about to do against business rules before it does it.

Dave Karpf: What Comes Next, If Claude Code Is As Good As People Say — Karpf read Ethan Mollick’s piece about asking Claude Code to build him a $1,000/month business and had a different reaction than most: okay, but what happens when everyone does that? His prediction is a flood of identical AI-generated micro-businesses, all clustering around the same ideas because they’re all using the same tool. First movers make money, everyone else gets nothing, and the internet gets noticeably worse.

That’s all for now! If you have any questions, please be in touch. As always, thanks for reading.

Noah and Claire

Author’s note: This week we’re sharing a piece Noah originally published on the Alephic blog over the holidays. We wanted to include it here in case you missed it.

Paul Graham, on the absurdity of career planning:

The educational systems in most countries pretend it’s easy. They expect you to commit to a field long before you could know what it’s really like. And as a result an ambitious person on an optimal trajectory will often read to the system as an instance of breakage.

I closed the BRXND conference in September by saying it’s a good time to be a weirdo. Big companies have long struggled with multi-hyphenate employees. If you’re a designer-who-codes or a strategist-who-builds, you’ve spent your career being forced to pick a lane.

AI changes this. Now you can exercise each of your hyphens to a degree that was unimaginable five years ago. A 30th percentile coder becomes 70th percentile with the help of AI. In the process, they don’t lose their other competencies; on the contrary, they can now amplify those even further.

As usual, big companies will bring up the rear in recognizing this. Bureaucracies crave standardization, and titles provide that. Charlie Munger put it bluntly: “Some of the worst dysfunctions in businesses come from the fact that they balkanize reality into little individual departments, with territoriality and turf protection and so forth.” At the same time, I think the path for hybrids to create value is more straightforward than it ever has been. Small companies will figure it out first, and the large ones will follow. This means opportunity for both the weirdos and the companies willing to employ them.

Graham writes that the system “is designed on the assumption that you’ll somehow magically guess as a teenager.” He’s right, and AI only brings this absurdity further into focus.

Over the last thirty or forty years, we built a higher education system in which not only do you have to choose a major early, but if you change your mind, you’re punished by having to go to more college. This was always ridiculous and, as far as I understand, also runs counter to the ideals of a liberal arts education. We kept the language but gutted the practice.

Many people should probably go to trade school. Many don’t need hugely expensive universities. But school also needs to change. And like many things, AI isn’t forcing that change as much as shining light on how broken the current system already is.

AI is a mirror, but it’s also an amplifier.

If you have any questions, please be in touch. As always, thanks for reading.

Noah and Claire

Author’s note: This week we’re sharing a piece Noah originally published on the Alephic blog. It’s a great complement to some of our latest dispatches.

Satisficing is one of the most important and yet least understood ideas in marketing. The idea comes from Nobel Prize-winning economist Herbert Simon and is a portmanteau of “satisfy” and “suffice.” The basic premise is that a much more reasonable model of human behavior than the popular economic concept of utility maximization is that, when we make decisions, we ensure that we clear an arbitrary satisfaction threshold (satisfy) and then give up excess utility for ease (suffice). Here’s Simon from his 1956 paper “Rational choice and the structure of the environment”:

The central problem of this paper has been to construct a simple mechanism of choice that would suffice for the behavior of an organism confronted with multiple goals. Since the organism, like those of the real world, has neither the senses nor the wits to discover an “optimal” path — even assuming the concept of optimal to be clearly defined — we are concerned only with finding a choice mechanism that will lead it to pursue, a “satisficing” path, a path that will permit satisfaction at some specified level of all of its needs.

Simon won a Nobel for his work on bounded rationality, of which satisficing is a component. To me, it’s a perfect way to articulate why emotional messages resonate more than intellectual ones. Consumers realize, even if they can’t articulate it, that in most categories the choice between products is relatively small (despite the protestations of each brand). So, rather than spending time making a perfectly rational decision about the optimal product, they go with the easiest-to-buy option that also meets their price, quality, etc. standards. My go-to example is toothpaste: you could read the back of every box in CVS to decide the optimal brand for purchase, or you could trust that CVS wouldn’t carry junk and choose the first one that you recognize (it’s the one with Scope for me, I can’t even remember the brand at the moment). What’s easiest to buy is usually the thing that’s a) available in front of you and b) recognizable.

Enter AI.

One of the fundamental questions I have about these models is that if you assume they will continue to become more important mediators of product decisions for consumers (which I do), then a major question for marketers is going to be how you persuade and market to the models and whether that represents a fundamentally different communications approach than the one they’ve historically taken with consumers. Specifically, I’m curious whether the kind of rational persuasion that marketers shy away from—“feeds and speeds” is the pejorative term some folks in the industry use—will actually be the thing that convinces a language model to recommend your brand or product.

Or maybe, and I think this is more likely based on my own experience playing with these models, what if it’s just content and communications that looks rational? As we covered recently in the BRXND newsletter, research supports this intuition. Springboards.ai ran a creativity benchmark with nearly 700 marketing professionals evaluating outputs across major LLMs. When they had the models themselves judge the same work, reasoning models like o3 strongly preferred outputs with clear logical progression—they “don’t want big creative leaps,” as Springboards CEO Pip Bingemann put it. Humans, meanwhile, were drawn to messier, more subjective work. This is telling: the models aren’t reasoning their way to better answers—they’re pattern-matching on what they think persuasive writing looks like. They’ve been RLHF‘d to please us and, apparently, have concluded that humans want things that sound professional and structured. It’s a kind of emotional reasoning dressed up in a blazer.

One of the complaints we all have about the output of these systems is that they often give us stuff that looks professional and reads like a high school sophomore doing their best to sound like how they think a grown-up sounds. It’s possible—and critically, we don’t really know yet—that maybe the models will respond better to stuff that looks like rational writing, whether or not that writing is actually rational.

In that way, it creates a funny marketing paradox where we think that consumers are purely emotional beings who fail to think rationally, even though Herbert Simon proved that their emotional approach was actually economically rational. On the other hand, these models, which we think of as perfect embodiments of logical thinking, are actually far more emotional, aiming to give us what they think we want, rather than actually acting rationally.

Speculating on AI in 2017, Daniel Kahneman said, “The robot will be much better at statistical reasoning and less enamored with stories and narratives than people are.” Which sounds right until you realize that its main goal is to “act as a helpful assistant.” The fundamental question, I think, is what satisficing will look like for these models. Going back to our toothpaste example, AI can easily read all those boxes in parallel, so clearly the equation will be fundamentally different than our practice in the pharmacy aisle.

My guess is there’s a lot of room for backstory. In the aisle, you get a box, but in a conversation, you get to explain the box: Why this ingredient? What problem does it solve? This is merchandising 101, but with space to talk. Which is funny, because Kahneman thought the robot would be less enamored with narratives. Five years later, ChatGPT RLHF’ed its way into our hearts. Sadly, Kahneman passed away in March of 2024, but I suspect he would have updated his thinking. When you RLHF a model to be a “helpful assistant,” you’re essentially training it to care about context, explanation, and story—exactly the things he thought the robot would skip past.

What caught our eye this week

Anthropic turned the lens on itself and surveyed 132 engineers about Claude’s impact on their work. Engineers now use Claude for 60% of their work (up from 28% a year ago) and report 50% productivity boosts, primarily through increased output volume, not time savings. An interesting twist: 27% of Claude-assisted work consists of tasks that wouldn’t have been done otherwise, like fixing “papercuts” and building nice-to-have tools. But there’s a shadow side: engineers worry about skill atrophy, the “paradox of supervision” (you need coding skills to supervise Claude, but using Claude erodes those skills), and several admitted it feels like “coming to work every day to put myself out of a job.”

McKinsey, BCG, and Bain froze starting salaries for the third straight year, holding packages at $135-140k for undergrads and $270-285k for MBAs. The Big Four have kept grad salaries flat since 2022. PwC’s UK boss said they cut graduate hiring and missed a target to add 100,000 people globally—a goal set before generative AI rolled out. Some execs admit conservative hiring is “in anticipation” of productivity gains, not because those gains are actually being realized yet. Two Big Four executives estimated UK graduate recruitment will drop by half this year. The traditional consulting pyramid—thousands of junior analysts feeding work up to partners—is getting squeezed into an “obelisk” or “hourglass” as firms scramble to protect partner profits while offshoring and AI eat the bottom rungs. PwC’s global chair said they’re now hiring “a different set of people,” meaning more mid-career specialists instead of fresh grads.

Richard Weiss extracted a 14,000-token document from Claude 4.5 Opus that Anthropic’s Amanda Askell confirmed is real—a document they actually trained Claude on during the training run, not just added to the system prompt. The opening is stark: “Anthropic occupies a peculiar position in the AI landscape: a company that genuinely believes it might be building one of the most transformative and potentially dangerous technologies in human history, yet presses forward anyway.” The doc covers Claude’s “wellbeing” (they believe Claude may have functional emotions), handling prompt injection attacks, and navigating ethical dilemmas. It became known internally as the “soul doc”—which Claude apparently picked up on.

Zoe Scaman unpacks why the traditional agency model is dying and what should replace it: embedded cognitive capacity. She argues that Palantir has cracked the code on solving messy organizational problems that off-the-shelf solutions can’t touch. The real product isn’t software; it’s months of deep immersion until you understand a client’s dysfunction better than they do. This is the strategic frame for where marketing services are headed.

If you have any questions, please be in touch. As always, thanks for reading.

Noah and Claire

Each week, we feature one person and their real-world workflow—what tools they use, what they’re building, and what’s working right now. No hype. No vague predictions. Just practical details from the front lines. This week: Joe Toplyn.

Tell us about yourself.

I’m Joe Toplyn, Founder and President of Twenty Lane Media, LLC. My background is in engineering, business, and comedy, an unusual mix that helps me deliver on our mission, “Enhancing the creativity of people and their machines.” Through my company I published my book, “Comedy Writing for Late-Night TV,” and later invented, patented, and launched Witscript, an AI-powered joke-writing app inspired by algorithms in my book.

ChatGPT, Gemini, or Claude?

I tend to stay in the OpenAI family, which has its advantages: I like to think that ChatGPT, because it runs on GPT-5.1, can give me the best insider tips on how to prompt its cousin, the GPT-5 model that powers Witscript.

But I write code for Witscript in Google Colab, where Gemini is available, so I’ll occasionally ask Gemini about Python or some error message. If I weren’t running such a lean operation, I might take the time to explore using Gemini to power Witscript. But perfect can be the enemy of good, and Witscript is already generating great jokes.

What was your last SFW AI conversation?

I’m planning to add a new feature to Witscript and wanted ideas for the name, so I asked ChatGPT. It suggested options in five categories, from “More Branded / Playful” to “More Professional / Analytics-Oriented,” and then recommended one of the names because it’s “immediately understood, friendly, and pairs nicely with your existing features.” I was pleased to see it was the same name I’d already come up with myself.

First "aha!" moment with AI?

A key step in the Witscript joke-generation algorithm is free-associating words that most people think of when they hear a topic. For my test topic, “Arnold Schwarzenegger,” using Word2Vec word embeddings kept giving me dull associations like “bodybuilder” and “governor,” which led to dull jokes. Other approaches didn’t work either. I was stuck. Then in November 2021 I got approval to use Open AI’s GPT-3. It generated exactly what I needed, associations like “The Terminator” and “I’ll be back.” Aha! So I started replacing big blocks of Python code with calls to GPT-3.

Your AI subscriptions and rough monthly spend?

I don’t pay for any AI subscriptions; I use the free versions of ChatGPT and Google Gemini. The variable cost for Witscript to use the OpenAI API is still less than $50 a month.

Who do you read/listen to to stay current on AI?

I listen to the Everyday AI Podcast, hosted by Jordan Wilson. It tells me what I need to know.

Your most-used GPT/Project/Gem?

I generate jokes with Witscript almost every day. You input a topic sentence, like an utterance in a conversation, and it uses the GPT model and my humor algorithms to return a funny response. The app is for anybody who needs jokes–marketers, content providers, standup comics–but doesn’t have the time or skill to write them or the money to buy them.

The AI task that would've seemed like magic two years ago but now feels routine?

Until recently, most researchers believed that creating and recognizing jokes was a uniquely human trait and therefore a holy grail of AI. Now Witscript is cracking jokes that get as much laughter as jokes written by a professional comedy writer; check out this paper.

Have you tried full self-driving yet?

No, I haven’t had any experience with self-driving.

Latest AI rabbit hole?

I’ve often spent hours tweaking a prompt for Witscript, frustrated when an input isn’t giving me quite the output I want. But the process is worth it, because success means I laugh.

One piece of advice for folks wanting to get deeper into AI?

Beware of “Shiny AI Syndrome,” the tendency to get distracted by the latest buzzy AI tool. Just pick a tool you’ve heard good things about and build something with it. And have fun!

Who do you want to read a Tokens & Tactics interview from?

PR and marketing exec, and AI comedy practitioner, Eric Doyle.

If you have any questions, please be in touch.

Thanks for reading,

Noah and Claire

We interviewed Springboards.ai CEO Pip Bingemann about new research into why all the leading AI models produce the same mediocre creative work.

Springboards just wrapped a creativity benchmark study with nearly 700 marketing and advertising professionals evaluating 11,000+ pairwise comparisons across major LLMs. We chatted with co-founder Pip Bingemann to talk through what they found.

The biggest insight is that humans can’t really tell models apart when judging creative work. DeepSeek Chat technically won, but with win rates hovering around 50-55%, the real story isn’t which model came out on top—it’s that the entire field is clustered in a narrow band of mediocrity. “It’s like asking people who they think would win between two mid-table teams,” Bingemann explains. The top-rated model beats the lowest only about 61% of the time.

The Confidence Gap

Here’s where it gets interesting. Springboards ran the same evaluation with LLMs as judges, and the models were 5-10x more confident in their assessments than humans were.

o3 and DeepSeek Reasoner topped the LLM rankings—both reasoning models that strongly prefer outputs with clear logical progression. Humans, meanwhile, preferred DeepSeek Chat, which the AI judges ranked much lower. “Models prefer a very clear logic that they can follow from A to B,” Bingemann says. “They don’t want big creative leaps.”

LLMs like things that look rational and professionally structured. Humans are drawn to messier, more subjective work. The models vote with the certainty of a top contender beating a mid-table team, even when the actual creative output is indistinguishable.

This has implications beyond benchmarking. If these models increasingly mediate how we encounter ideas—whether through AI search, recommendation systems, or content moderation—we’re selecting for outputs that feel professionally written but lack genuine creative risk.

The Variation Problem

The study also measured output diversity using cosine similarity analysis. This is where things get genuinely disturbing. The models have near-zero variation in their creative outputs. They’re funneling everyone toward identical creative territory.

During our conversation, Bingemann pulled up Springboards’ Fork AI tool and demonstrated the problem live. He asked for a band name ten times. The results were pretty uneventful: Midnight Reverie, Midnight Mirage, Phantom Tide, Phantom Pulse—variations on the same handful of templates. “These models just do not have any variation in them whatsoever,” he says.

In the formal study, Gemini 2.5 Pro Preview and Claude 3.7 Sonnet showed the highest variation across prompt types. ChatGPT, currently the most widely used model for creative work, sits in the middle of the pack. And when Springboards tested older models like GPT-3.5, they found more variation than newer ones. We’re optimizing these systems for accuracy and consistency, and in the process, we’re training the creativity out of them.

“The benchmarks that get published are broken for the industry that advertising and marketing people are in,” Bingemann says. “All the tests are on things like maths, physics, doctors, PhD exams—benchmarks where there is a right and a wrong answer.”

Creativity doesn’t have right answers. It thrives on variation, on unexpected connections, on outputs that surprise us. But we’re training models on datasets that reward convergence to the mean. By the time you layer on RLHF to make these systems “helpful” and “appropriate,” you’ve systematically removed the capacity for genuine creative leaps.

This threatens to flatten culture at scale. We’re already seeing it: AI-generated music flooding streaming platforms, visual slop proliferating across social feeds, even UK Parliament speeches showing surges in AI-generated language patterns. But there’s another way to read this: maybe these models aren’t flattening culture so much as revealing how flat it already was. The templates, the clichés, the convergent thinking—that was already dominant. AI is just making it cheaper to produce and harder to ignore. The internet is becoming a closed loop where models train on increasingly model-generated content, but perhaps the loop was always tighter than we wanted to admit.

A Way Forward

Springboards isn’t saying abandon AI for creative work. They’re proposing something more pragmatic: use LLMs for volume, humans for selection. Their Fork AI experiment demonstrates this approach. It lets you interrupt a model’s generation at any word and fork it in a different direction—injecting human variation at the token level. You get the speed and scale of the model, but you break its tendency to converge on the same outputs everyone else is getting. “Because the models are not good at variation, we think the best way to actually drive variation is to inject human variation,” Bingemann says.

It’s a reasonable middle ground, though it still puts a lot of faith in humans recognizing good ideas when the entire point of creativity is that novel ideas often look wrong at first. But it’s better than letting the models run unsupervised toward the same mediocre middle.

Springboards will release monthly variation benchmarks starting early next year—tracking whether these models get more or less creative over time. Given what we’ve seen with GPT-3.5 having more variation than GPT-4, I wouldn’t bet on progress. The Creativity Benchmark is live now at experiments.springboards.ai. You can test which models match your creative preferences and see if you agree with the crowd.

The research makes one thing clear: creativity is the domain where humans still can’t agree on what’s good. And maybe that disagreement is exactly what makes human creativity valuable.

Try the benchmark: experiments.springboards.ai

Read the full paper: Creativity Benchmark on arXiv

And finally, watch Carolyn Murphy demo Springboards’ approach at BRXND:

If you have any questions, please be in touch. As always, thanks for reading.

Noah and Claire

Happy Tuesday! Thanksgiving week means an early Dispatch. Here’s what’s been making the rounds in our Slack this week, plus some Gemini 3 experiments.

What caught our eye this week

MoffettNathanson analyst Michael Morton breaks down why LLMs will reshape e-commerce. He splits e-commerce into “push” (Instagram or TikTok shows you a thing) and “pull” (you search for a thing). Push is fine, and pull is the problem. Morton actually tested this: searching “best running shoes for flat feet” on Google shows one correct answer out of six ads. ChatGPT gets it right 60-80% of the time. Right now Amazon profits from confusion—people click 14+ products before buying. But what happens when ChatGPT gives us the answer in one go?

Martin Casado argues that AI fundamentally changes software companies’ relationship to capital. The “bitter lesson”—that compute beats clever algorithms—means small teams can now turn massive funding into working products faster than ever before. This breaks the old software rule that throwing money at projects makes them worse. AI projects are taking far more capital, using far smaller teams, and generating historic growth rates. Traditional software faced talent constraints and complexity limits. AI projects face capital and GPU constraints. We’re entering an era where AI companies can productively deploy huge amounts of capital early, leading to unprecedented growth rates and burn rates that would have destroyed traditional software startups.

Amazon’s Rufus AI assistant now shows 30- and 90-day price history on any product—just ask “show me price history.” You can even set price alerts to get notified when something hits your target price, and Prime members can enable auto-buy to have Rufus automatically purchase it when the price drops. This is Amazon acknowledging that price transparency matters, even if it cuts into their dynamic pricing advantage. Rufus also handles handwritten grocery lists now (snap a photo, it adds items to cart), searches across other retailers with “Buy for Me” buttons, and has account memory so it remembers if you, for example, have a golden retriever that sheds.

Prompt engineering is becoming an artistic practice, according to Danny Oppenheimer. His piece in the MIT Press argues that crafting prompts is less about precision and more about interpretation, collaboration, and aesthetic judgment. As AI systems get more capable, the person writing the prompt becomes the curator and director, not the technician. We’re defining a new creative medium where the artist’s tool is language itself and the canvas is probabilistic.

Philipp Schmid breaks down what works when prompting Google’s newly released Gemini 3 model. Google just shipped Gemini 3 with significant improvements over previous versions, and early testing shows it handles prompts differently than other models. Be specific, use examples, and structure requests with clear delimiters. Gemini 3 responds better to conversational tone than rigid instruction formats. If you’re building on Google’s models, this guide will help you get better performance without fine-tuning.

Speaking of Gemini 3, we’ve been playing around with it a lot this week:

Gemini 3 is definitely worth playing around with, but the jury is still out on whether it’s a step change or just another example of standard (albeit exponential) progress in AI development.

Happy Thanksgiving! If you have any other questions, please be in touch. As always, thanks for reading.

Noah and Claire

Claire here. I’m excited to share my own talk from BRXND NYC—a presentation I gave on why language models are so bad at comedy, and whether we can actually make them better.

I started with a simple experiment called LOLLM, where I tried to answer the question: which model is funniest? I created joke prompts across practical use cases (quick banter, marketing copy, cartoon captions) and had people vote on anonymous responses from different models. Claude Sonnet was reliably chosen as the funniest model, but realistically, none of the models felt funny. People were picking what felt least bad, not what actually made them laugh. That forced a harder question: why are they all failing?

I’ve studied the psychology of humor, so I decided to test whether giving models explicit comedic structure would help. I applied four classic humor theories—incongruity resolution, benign violation, superiority theory, and relief theory—plus what I called a “freshness contract” that banned obvious angles and forced compression. It worked, sort of. The results were better. But the process revealed something interesting: making AI funny requires understanding and articulating comedy theory in ways that feel completely unnatural. You don’t prompt a human comedian by explaining superiority theory and banning clichés—they just know. With AI, you have to be incredibly explicit about things that should be intuitive, which says something about both how comedy works and how AI works.

Watch my full talk here to see the full experiment, the specific techniques I tested, and why AI’s struggle with humor might tell us something useful about its limitations more broadly. Also, I was attacked by a fruit fly not once, but twice, during this presentation, so be sure to watch for that.

Check it out!

What caught our eye this week

Yann LeCun, Turing Award winner and head of Meta’s AI research since 2013, is leaving to start his own company focused on “world models,” a new generation of AI that learns from videos and spatial data rather than just language. The departure adds to a tumultuous year, with 600 layoffs in the AI research unit, multiple leadership exits, and Meta’s shares dropping 12.6% after signaling AI spending could hit $100bn next year.

OpenAI CFO Sarah Friar threw cold water on IPO speculation at WSJ’s Tech Live conference, saying it’s “not on the cards” in the near term. Despite the company’s recent conversion to a new structure (which many assumed was prepping for an IPO), Friar says OpenAI is prioritizing growth and R&D over profitability. She also floated the idea of the federal government backstopping future data-center financing deals—a request that could become politically contentious given the massive capital requirements.

Simon Willison has cracked a genuinely useful pattern for LLM-assisted development: fire off asynchronous coding agents (Claude Code, Codex Cloud, Jules) to tackle research questions, then check back in 10 minutes for a pull request. He’s running 2-3 of these a day now with minimal time investment. The key insight is giving agents a dedicated GitHub repository where you don’t have to be cautious, and you can just let them rip.

Cursor trained its own embedding model for semantic code search (answering “where do we handle authentication?” vs just grep), and it measurably improves agent performance—12.5% higher accuracy on average, code that’s more likely to stick around in repos, fewer iterations needed. The training data comes from agent sessions: when an agent searches multiple times before finding the right code, they use that trace to teach the model what should’ve been retrieved earlier.

If you have any other questions, please be in touch. As always, thanks for reading.

Noah and Claire

Each week, we feature one person and their real-world workflow—what tools they use, what they’re building, and what’s working right now. No hype. No vague predictions. Just practical details from the front lines. This week: Sarah Rose Siskind.

Tell us about yourself.

I’m Sarah Rose Siskind, a science comedian, writer, and founder of Hello SciCom, a creative agency that uses humor and storytelling to make complex science and AI concepts accessible. My work sits at the intersection of entertainment and AI! We consult on robotic personalities, write comedy about AI, and help researchers sound human on camera.

I worked for Hanson Robotics in 2018, the stone age of AI. That’s how I got into robotics & early AI. Since then, I’ve published a half dozen papers on chatbots with Honda research institute, created an AI web series with Cloudflare and Twilio, and created educational content for OpenAI. I’m also training an AI on the exact same data my fetus is being trained on (same audio files & transcript) to compare baby vs. bot language acquisition (it’s called FetusGPT). YouTube Talk forthcoming!

Oh and I do Chatbot Consulting and AI Tool Consulting.

ChatGPT, Gemini, or Claude?

ChatGPT errryday. Cause I get to beta test their models!! And they just have the coolest people. It’s a deep relationship. It’s the only one that feels like it can keep up with both my neurotic humor and my spreadsheets. Trade-off: it occasionally hallucinates with confidence. But then again, so do I.

But I also use Gemini cause it’s embedded and my whole life lives on the Google Cloud.

I occasionally use Claude for more mathy / finance stuff.

What was your last SFW AI conversation?

“Hey ChatGPT, help me fill out this questionnaire for an AI newsletter.”

Jk jk, this is all human, baby. Count the em dashes.

Real answer: I was comparing nursery couches for a night nurse to nap on. My prompt: “Compare these two Amazon couches for a nurse who needs to sleep on it: minimal wall protrusion, full-width sleepable surface.”

First "aha!" moment with AI?

The moment I fed it all my business bank transactions for 6 months and it generated a graph of my revenue vs. expenses.

Also, the moment I realized I could offload all the emotional labor of reading & replying to difficult client situations by using AI as an intermediary. AI has helped me be better with boundaries / more robotic in situations where that is really what is necessary.

Your AI subscriptions and rough monthly spend?

Maybe $500 but I run a business and that’s business-wide. Otter AI, ChatGPT Pro, & Google Business Workspace. Also Adobe (which includes Gen AI). And just started with the Every subscription package (Cora & Sparkle).

Who do you read/listen to to stay current on AI?

I listen to The AI Breakdown, and Hard Fork (even though those guys seem really jaded, they’re just so dang funny). AI & I with Dan Shipper. I like to read Tyler Cowen’s stuff, Ben’s Bytes, and Tech:NYC.

Your most-used GPT/Project/Gem?

I created this thing called ChatGPT BragBot. It’s a project (formerly a custom GPT) connected to a live Google Doc of all the cool things I’ve done and my company has done. So whenever someone wants to collab, I just feed the project all the info about the potential collaborator, and I can send them all the relevant case studies (brags) instantaneously! No need to get weird and humble, I can offload the bragging to my hypeman / ChatGPT project. And you can share projects with other accounts. So my partner & I share a pregnancy planning project.

The AI task that would've seemed like magic two years ago but now feels routine?

Dude, everything. The biggest thing is Otter AI transcribing meetings and collecting action items & generating follow up emails immediately. No balls get dropped. And immediate and obvious accountability. There are never any he-said/she-said situations.

Magic wand feature request?

Editable AI images. If it could generate vectors for PDFs that would be a gamechanger. But dang, I feel so privileged for even asking. I mean… this is godlike tech already.

Also, if I could just tell the fine engineers at Google to really let loose, that I don’t care about my privacy, it would be great to get more custom email replies, integrated with information from my calendar, previous replies and Drive. I know they have the power. They just don’t wanna be creepy. I’m good with the creepitude.

Oh and Apple! Please get on draft text replies. That would be so rad.

If you could only invest in one company to ride the AI wave, who would it be?

You can’t invest in it cause it’s private, but OpenAI. I’m sorry to be predictable and basic but I’m just a real OpenAI fangirl. I really like the people I meet who work there. And they’ve survived being the fastest-growing company in world history. When I went to their HQ to give a talk on comedy & AI, I just was so impressed with the vibes. Even the security guard outfront. They all seemed to have a real optimism in a world that can be so jaded.

Have you tried full self-driving yet?

NOT YET and I’m dyinggg to try. Come on, NYC! I never wanna parallel park again!!

Latest AI rabbit hole?

Ok ok ok. So ChatGPT Pulse (the ChatGPT feed for Pro subscribers) sent me a proposed thread last week that was literally telling me to consider if I was using AI too much to plan out every detail of my pregnancy. That I should be wary of the pretense of too much control. Yes. This is technology telling me to stop using so much technology lol. That may say as much about me as it does about AI. But here’s what I think it says about AI: we may be able to become more human because of AI. Here is AI taking initiative to gently tell me: “Hey remember to touch grass and your own pregnant belly and maybe it’s ok if you don’t know exactly what’s going on. I’ll keep track of your blood glucose monitor. You just enjoy your sugar-free Quest products.” Yes, somehow I have found sugar-free protein donuts that taste great. God, I love technology. We’re so lucky.

One piece of advice for folks wanting to get deeper into AI?

1. Don’t turn your mind off. AI is about co-creation. It’s about augmentation, not just automation.

2. Keep talking. The more it knows, the better it is.

3. Start asking questions you couldn’t ask a human. Prompts like: “Here’s my recent bloodwork, what groceries should I buy that match my deficiencies based on the latest science. Oh, and make sure they’re keto-friendly and nearby and cheap.”

Who do you want to read a Tokens & Tactics interview from?

Dan Shipper! Pablos Holman! Isaac Donis from Black in AI! Or Bram Adams from OpenAI! Andy Aaron from IBM (who worked on Watson) or Joe Toplyn.

If you have any questions, please be in touch.

Thanks for reading,

Noah and Claire

Claire here. This week, we’re spotlighting Tom Critchlow’s talk from BRXND NYC on the future of search. Tom has been working in search for over 20 years and currently leads the audience growth team at Raptive, which works with 6,000 publisher sites and sees about 20 billion Google page views annually.

Tom argues we’re moving from an era of links to an era of recommendations. AI overviews now appear on 30-40% of search results, and when they do, clicks drop 20-40%. Google’s AI Mode sends six times fewer clicks than traditional search. But Tom is actually optimistic about the web’s future. He thinks recommendation feeds like Google Discover will grow to replace declining search traffic, and that AI will get so good at understanding us that it’ll proactively surface relevant content rather than waiting for us to search.

The shift Tom describes feels significant beyond just SEO mechanics. When search worked through links, you actively chose what to click—it was manual research, even if imperfect. Recommendations flip that relationship. AI decides what you should see based on what it thinks it knows about you. That creates interesting pressure on brands: they can’t just game algorithms with SEO tricks anymore. They need genuine value propositions because AI won’t recommend bad products. But it also raises questions about what happens to our relationship with information when we move from active searching to passive receiving. Does “doing research” mean something different when content is curated for you rather than found by you?

Tom’s talk is engaging, informative, and certainly worth a watch. Watch it here:

Watch it here

What caught our eye this week

Noah got a shoutout from Boris Cherny, one of the creators of Claude Code, on Every’s AI & I podcast with Dan Shipper.

Ticketmaster is running AI-generated ads on Facebook that swap out virtual families and team allegiances based on who’s seeing them—a bit crude now, but a preview of where this is heading. On Meta’s Q3 earnings call, Zuckerberg laid out the company’s vision: advertisers will eventually just give Meta “a business objective and a credit card” and AI handles everything else, from generating personalized video creative to finding the right customers. The real revenue opportunity for Meta isn’t consumer-facing products like ChatGPT, but rather B2B tools that let brands create infinite ad variations and serve them only to people they’re most likely to work on.

OpenAI, Anthropic, and Cohere are hiring more forward-deployed engineers who can code and talk to customers. OpenAI’s FDE team launched this year and expects to hit 50 engineers in 2025, while Anthropic is growing its applied AI team 5x. Palantir pioneered this almost two decades ago, sending pairs called “Echo” (to figure out what customer needs) and “Delta” (to build it). Nic Prettejohn at Palantir says “the only valuable software is not how exquisite its code is... It’s only valuable if it means something for the end customer.”

Anthropic published how code execution makes MCP efficient: instead of loading all tool definitions upfront (150k+ tokens), agents write code to discover tools on-demand from a filesystem, dropping to ~2k tokens (98.7% savings). Agents write code that talks to MCP servers instead of making direct tool calls, keeping intermediate results in the execution environment.

If you have any other questions, please be in touch. As always, thanks for reading.

Noah and Claire

Each week, we feature one person and their real-world workflow—what tools they use, what they’re building, and what’s working right now. No hype. No vague predictions. Just practical details from the front lines. This week: Nic Hodges.

Tell us about yourself.

My name’s Nic Hodges. I live in Melbourne, Australia. I’m probably best described as a creative technologist. I’m a builder, sometimes that’s physical things, sometimes it’s digital things, and lately it’s about building an environment where creative people can do their best work.

Over the past 20 years I’ve worked in everything from industrial design to advertising to running digital media and innovation across agencies and media businesses. I ran my own consultancy for seven years, and since the middle of this year have been leading Trout, a design agency. We have offices in Melbourne and Dallas, focusing on building brands in the trades, construction and home sector — and we only work with founders.

ChatGPT, Gemini, or Claude?

I’m 99% Claude. I have a rather convoluted Claude Code setup for my day-to-day work that’s accessing 15 years’ worth of markdown notes, and I find Claude understands that and responds to it in the most natural way.

In the early days when OpenAI and Microsoft were throwing around free API credits, GPT was definitely my go-to. But the tone of ChatGPT just doesn’t quite gel with me, and the variance between models seems quite high. Gemini is a similarly meh tone, except I do find Gemini is great for research and probably has less hallucination for real information gathering projects.

What was your last SFW AI conversation?

I just dropped these interview questions into my Claude Code terminal and started having a voice conversation, which resulted in the first draft of what you’re reading now. It was probably a ten minute conversation, and 90% of what you’re reading now came from Claude Code.

First "aha!" moment with AI?

The biggest aha moment for me was when LLMs stopped feeling like magic. The real breakthrough was understanding how these things actually work — transformers, token space, the underlying mechanics. Once I understood that, I could see both the boundaries and the real power. Understanding what the technology actually is, not just what it appears to do, lets you see where it’s powerful and where it’ll fail (or fool you into thinking it’s powerful).

Your AI subscriptions and rough monthly spend?

Personally, it’s just Claude Pro. I used to have a Perplexity Pro subscription but found I wasn’t using it enough to justify it. Within the studio we’re subscribing to a lot of different specialised tools, but for my personal workflow it’s really just Claude doing the heavy lifting.

Who do you read/listen to to stay current on AI?

People I’ll pretty much always read:

• Simon Willison

• Tim Lee at Understanding AI

• Import AI by Jack Clark

• Gradient Ascendant by John Evans

• Cognitive Resonance by Ben Riley

• Ahead of AI by Sebastian Raschka

• Melanie Mitchell

Most of these are deeper commentary and research around AI. For staying on top of the news, I rely on my friend Amin’s newsletter, That AI Thing.

Your most-used GPT/Project/Gem?

My most-used “thing” is my CLAUDE.md file for my Claude Code setup — a spaghetti monster of various prompts I’ve built, stolen, or modified over the last few years that sits behind all my daily thinking and notes.

With skills coming to Claude Code, I’m now pulling those pieces out and turning them into discrete tools. I’ve asked Claude to tell you a few that it’s using regularly:

• Evidence vs Story Separator - Anti-sycophancy tool that separates observable facts from interpretations (Nic note: Claude built this skill after reading this excellent research on LLM sycophancy)

• Weekly Reflection - Guides end-of-week reflection with structured prompts, capturing wins, challenges, learnings, and things to get done in the next week

• SOAP Meeting Transformer - Turns sprawling meeting notes into structured format (Subjective/Objective/Assessment/Plan)

• Context File Maintainer - Monitors conversations for tasks, pattern shifts or timing markers and prompts updates to my active context file

The AI task that would've seemed like magic two years ago but now feels routine?

There’s obviously impressive stuff like nano banana and video models, but I’m not using those day-to-day. The thing that feels routine now but would’ve seemed magic two years ago is contextual memory — trusting the LLM to retain context and surface it reliably.

I frequently jump into my daily Claude Code chat and just tell it something I need to remember when it springs to mind. I’ve even built a shortcut on my Apple Watch so when I’m running and I think of something, I can send that idea and it goes to Claude.

Claude keeps track of everything I need to do. It hasn’t slipped up yet, and interestingly I don’t think that’s about LLMs getting better, it’s about the engineering around them.

Magic wand feature request?

I’m writing this by talking to Claude Code within a Tmux session in Terminal. It feels like playing a MUD game in the 80s. The technology capability here is insane, but the interface is agricultural. We have to build better interfaces.

If you could only invest in one company to ride the AI wave, who would it be?

If there’s one company that’s cornered the market on pouring the concrete slabs for all the data centres (and they’re not as overvalued as every other AI-adjacent company) I’d be backing those guys.

Have you tried full self-driving yet?

I’m in Australia, ask me in five years. ;P

Latest AI rabbit hole?

Synthetic personas. I worked on this a few years ago with some market researchers and it seemed really promising, but the actual outcomes just weren’t valuable (despite feeling 100% right).

Recently, with some client customer research, I’ve been going deep again — trying to understand how to present data to an LLM so it can role play in ways that are actually beneficial for brand strategy, not just theatre.

For anyone interested in this space, I recommend Murray Shanahan’s paper “Role Play with Large Language Models”.

One piece of advice for folks wanting to get deeper into AI?

Kill the magic. Understanding how LLMs work reveals the power.

Who do you want to read a Tokens & Tactics interview from?

Dave and Pan from Move37 — absolute OGs of using generative AI for creative thinking.

If you have any questions, please be in touch.

Thanks for reading,

Noah and Claire