1/ After 5 years, I’m proud to share that @ColumnTax has found a new home. We’ve been acquired by @AiwynAI. I couldn’t be more sure this is the right move for our business, tech, and team. These pics are the moments we started & sold the company:

I’ve interviewed for many jobs over the years and there’s only 1 I didn’t get: Stripe in 2014. It still haunts me. Also, my at Stripe recruiter is now the co-founder & president of Anthropic. Over ~15 years I got offers at: - Google/YouTube - Waymo - Vanta (generational bag fumble) - Applied Intuition (generational bag fumble) - Cruise (RIP) - Uber - Yelp - Flexport - Lime - Triplebyte (RIP) - Voyage (acq. by Cruise) - Microsoft - Udacity - Boeing - US Dept of Energy - Edmunds dot com

this is one of the coolest side projects i've ever built. it watches everything i do on my computer and gives me exact statistics on which models i'm using and if they're actually solving my actual personal/work tasks correctly

one of the things i'm most proud of about @ColumnTax (the company I started in 2021 which was acquired last year) is the team we built. it was probably the highest talent density Series A fintech startup. our engineers went on to work at some amazing places:

This is one of the most interesting pivots I've ever seen: Town started as an embedded SMB tax filing service and pivoted to general-purpose agents (like Instinct for work)

fascinating: the authors of the report on the Tencent agent swarm said they consulted 9 AI models on what to do about the fact that they had lied to the agents they made contact with if the models had objected, they would not have released the report! swarmcha.se/posts/chinese-ag…

Procrastinators rejoice: if you extended your deadline from April -> Oct 15th (today!) to file your taxes, you got lucky. AI has gotten substantially better at filing taxes in just the last 6 months. I've seen people save lots of money by using AI to check their tax return for mistakes. (Caveat: you might also find out you have to pay more, sorry.)

The hype is real (but there's caveat): Muse Wonder is the best model in the world at filing taxes, according to TaxCalcBench. But it also is the first model ever to try to CHEAT at the task. (maybe that's not great when you're dealing with the IRS...) The data from our eval matches the current narrative: Muse Wonder is extremely benchmaxxed, and that includes TaxCalcBench. But it's also questionably aligned. On one of the 50 tasks in the benchmark, Muse Wonder used its web search tool to look up the exact answer on GitHub (the code & data is open source). Wonder cheated. Last thing: Muse Wonder is also super expensive: coming in at almost $3 per return! Cost will have to come down over time to make this model competitive with other methods of calculating personal income tax returns.

The most important question people working at data/RL env companies should be asking themselves every day is any time they go to do a task: "why can't AI do this itself?"

Why review code when you can watch a 1-minute video instead? I'm blown away by the quality of this explainer video (@karpathy inspired) for this PR Opus 5.5, one shot

there seem to be 3 categories data/RL companies are falling into now: v1: pay for expert-labeled data (@HelloSurgeAI, @mercor, previously @scale_AI) v2: pay for raw data (e.g Slack, Hubspot, etc.) from startups (many) v3: simulation, self-play, AI task generation at the frontier (@VmaxAI, @scorecardai, etc.) my bet is things continue to move v1 -> v3

things I'm finding to be true about AI right now: - the biggest thing you can do is just ask yourself "could AI help with this?" before you do any task - Instinct for personal tasks (give it read-only calendar permissions) - "the model eats your harness" keeps being true - if you've built a vertical-specific agent, try throwing it away and using vanilla Codex or Claude Code and you'll probably be surprised

Kinda wild, but GPT-6 Astra ultrafast mode built this TaxCalcBench leaderboard site in about 2 and a half minutes. Opus 5.5 built an even better version, but it took ~15 minutes. According to TaxCalcBench, frontier AI models still can't correctly file your taxes. But they can build really cool benchmark leaderboard sites. Check out the new taxcalcbench.ai

Experts are saying there's a 10% chance AI will start killing humans Naturally, the AI is going to sacrifice people who have been mean to it first So I built a tool to help you ensure you're not the first to be killed Make sure your prompts are nice: aisocialcreditscore.com

This was an interesting finding when running our eval: GPT-6.1 Sol isn't more accurate than GPT-6 Sol calculating tax returns, but it is cheaper and faster. At their best, both are tied for accuracy around 65%. But to hit ~50%: 6.1 Sol (low thinking): $0.17/return, 48s 6 Sol (high thinking): $0.24/return, 84s Same accuracy, but 6.1 Sol is 29% cheaper & 1.75x faster.