What caught our eye this week // BRXND Dispatch vol. 063
A new weekly roundup of links and ephemera from the intersection of brands and AI.
You’re getting this email as a subscriber to the BRXND Dispatch, a newsletter at the intersection of brands and AI.
I hope you’re all having a relaxing holiday. In an effort to create some more regularity with this newsletter in 2025, my friend Luke Hopping is going to be helping out with all things BRXND in 2025. Luke and I have worked a bit together in the past on Ride Review, and he’s also leading all things for Ride AI (more about that in the future). In addition to my regular missives about building and playing at the intersection of marketing and AI, we will be publishing weekly roundups that are mostly made up of the links and conversations that fill our Slack daily. If this format resonates, we’ll make it a regular feature in 2025 alongside the typical explorations, analysis, interviews, and other randomness that has made up the bulk of this newsletter for the last few years. Thanks for joining us on the journey! - Noah
Before we dive in: BRXND LA 2025 is coming up on 2/6. If you’re interested in attending, tickets are available here. We just announced the first batch of speakers — a stellar mix of marketers and technologists from renowned companies like Amazon, PepsiCo, Zillow, Getty Images, Airtable, and more. We still have a few slots open for interesting demos and case studies, so if you’d like to suggest a new speaker, please reach out. (Oh, and if you’re interested in sponsoring, we’ve got a form for that too.)
What caught my eye this week (Noah):
While much of the month’s focus was on OpenAI’s o3, the newest version of their “O-series” reasoning models that performed absurdly well on the ARC advanced reasoning test, I want to rewind back a few weeks to Meta’s Llama 3.3 release. While this model is nowhere near o1 (and definitely not o3), it’s pretty close to GPT 4o, and it’s a fraction of the size. Where things get really interesting, though, is that because it’s open source, you can access it through services like Groq and Cerebras, which offer ridiculously fast inference. While everyone is going into 2025 focused on what it will mean when models can be significantly smarter than GPT 4o/Claude 3.5, I’ve been thinking more about what it will mean when you can get instantaneous answers. Just last week, I implemented Groq for a client project and brought the time to process from 20 seconds to just three. We’ve already got essentially free intelligence with these models, but what happens when it’s also instant?
Finally, if you didn’t read Anthropic’s analysis of building AI agents, I highly recommend it. It cuts through a lot of the complexity (and nonsense) of the space and offers a fairly specific definition of what an agent actually is, as well as some really helpful set of workflows that perform well when building with AI. I’m partial to the orchestrator:
I don’t agree with everything in the post, but I’m really glad they’re trying to put some boundaries around the language. This is something we’ll be covering at BRXND LA in some depth with a talk from Langchain’s Lance Martin, who will be teasing apart the fact and fiction of agents.
A few other things worth checking out (Luke):
Going back to o3 for a minute, OpenAI’s new model broke barriers with its high score on the ARC advanced reasoning test. But while François Chollet designed ARC to test AI systems that could learn efficiently like humans, o3 achieves its results by throwing massive compute resources at the problem. As Melanie Mitchell points out, this raises important questions about whether we’re actually advancing toward general AI or just getting better at pattern matching with brute force.
Gemini 2.0 Flash Thinking is taking an interesting approach by showing its reasoning work explicitly. This isn’t just interesting to observe – it’s genuinely useful when you’re trying to understand whether the model is on the right track.
tldraw is what happens when you reimagine the whiteboard for the AI era.
Anthropic unveiled a new tool called Clio to understand how people use their AI assistant, Claude, while keeping user information private. Instead of using actual user conversations, they’re using aggregated data to identify areas where Claude is being overly cautious, then generating entirely new training data for those topics – a clever solution to the privacy-improvement paradox.
This chart of AI-native revenue from Sapphire offers an interesting snapshot of where we are. I’m particularly curious to see how these numbers evolve in 2025.
I’m looking forward to spending more time getting to know Alibaba’s Qwen in the upcoming year. It’s the most serious open-source model maker outside Meta. They’ve got some of the best small models.
Finally, check out ‘M.A.R.S.’ – an entirely AI-generated music video (lyrics, music, visuals, etc) that seems lab-grown to get stuck in listeners’ heads. While AI won’t replace human creativity, it’s fascinating to watch it create its own distinct cultural artifacts.
Let me know if you find this format useful. As always, if you have questions or want to chat about any of this, please be in touch.
Thanks for reading,
Noah & Luke