Building AI Bookkeepers, DeepSeek, Gemini 2.0 // BRXND Dispatch vol. 064
See you at CES?
You’re getting this email as a subscriber to the BRXND Dispatch, a newsletter at the intersection of marketing and AI.
Luke here. Happy New Year. I hope everyone is feeling rested and ready for 2025. The first big event on our calendar is CES. Team BRXND will be in Las Vegas next week, 1/7-9, discussing all things AI. If you’ll be attending, let us know (just reply to this email or hit us up on the contact form)—we’d love to try to find time to connect. We have a suite at the Bellagio, if you would like to meet or get away from the craziness of the showfloor.
After CES comes BRXND LA 2025 on 2/6. If you’re interested in attending our West Coast Marketing x AI summit, there are a limited number of tickets available here. We announced our first wave of speakers and sponsors over the holidays (here). So far, the response has been fantastic. If you’d like to get involved as a speaker or sponsor, please follow those links.
What I played with over the break (Noah):
The last week in December was met with the annoying news that Bench, our bookkeeper, shut down. That sent me into a bit of a rabbit hole trying to build a tool that could do cash flow forecasting. If you’ve ever run a business, you know the pain that exists between your budgets/forecasts and your bookkeeping. This is made much more complicated when taking into account payment schedules (net 30/60/etc.). Anyway, needing a fun Christmas break project, I started hacking on building something that would:
Grab all my transactions (this is pretty easy because we use Mercury as our business bank, and they have APIs).
Allow the creation of “forecasted transactions” that are created for the future and account for spending
Ability to reconcile real transactions from the bank against those forecasted transactions (thanks, AI)
Ability to take an SOW and extract all the invoices to create forecasted transactions (I covered this back in November)
As exciting as all this must be for you (bookkeeping software!), what really struck me is how accessible it’s become to build your own apps to solve your own specific problems in your own specific ways. I think this is what Satya Nadella was referring to when he suggested SaaS would collapse. It increasingly feels like the architecture of the future is going to be a huge data warehouse with all of a company’s data, some code and AI in the middle, and interfaces either custom-built with something like v0 or tools like Airtable or Bubble. What was striking about my Christmas accounting experiment was how quickly I was able to get it all up and running. This is obviously partly due to my ability to write code, but it’s also because I’ve become incredibly adept at using Cursor (the AI coding assistant) as a way to skip the repetitive steps of building. Anyway, there’s still more to do, but once it’s done, I think I might open-source it for other Mercury users to be able to run this hybrid forecasting/accounting system.
What caught my eye this week (Luke):
DeepSeek’s new open-source DeepSeek V3 is beating many closed AI models on key performance benchmarks, in particular math and language tasks, at a fraction the price. It reportedly took the Chinese lab only two months and $5.57M to train V3. The model isn’t perfect; for starters, it seems to think it’s ChatGPT, suggesting it was trained using GPT-4 datasets. But the $5.57M price tag, if accurate, is wild – not just because it’s dramatically less than what we’ve heard about other models, but because it suggests we’re entering an era where building competitive AI models might not require tech giant resources.
A new Microsoft paper estimates the size of GPT-4o (~200B parameters), ChatGPT (~175B), Claude Sonnet (~175B), and other models. Hard to say whether these figures are real or not, but they seem directionally accurate.
I enjoyed Timothy B. Lee’s primer on Gemini 2.0 Flash Thinking. After throwing a bunch of NP-hard problems and geometric reasoning tasks at it, he concludes Google’s new thinking model isn’t quite on par with o1, but it might be able to undercut OpenAI on price.
If you haven’t seen this walk-thru of Gemini 2.0’s live code tutor yet, you should give it a watch. Real-time coding assistance is going to be a game-changer for work.
Wired is tracking every copyright battle in the AI industry in a single handy chart.
From tracking cancer to helping win a Nobel Prize, the NYT reports on how AI hallucinations are actually speeding up scientific breakthroughs: “Now, A.I. hallucinations are reinvigorating the creative side of science. They speed the process by which scientists and inventors dream up new ideas and test them to see if reality concurs. It’s the scientific method — only supercharged. What once took years can now be done in days, hours and minutes.”
Perplexity CEO: The future of advertising is AI agents vetting brands on behalf of consumers.
As always, if you have questions or want to chat about any of this, please be in touch.
Thanks for reading,
Noah & Luke