Pinned Tweet
We've been building Kyoto, an AI for your taxes.
It scores 100% on TaxCalcBench v2: all 50 test cases correct across federal and state returns, on every scored line. The best frontier model scores 64%.
We're building towards this world, where advice is cheap and verifiable.
Advice is about to get very, very cheap
It should be obvious to everyone that information is now fully commoditized.
Advice is next, and it’s coming fast. The generic stuff — “Do I need an NDA before I show someone my idea?”
“C-corp or LLC?”
“How much equity does employee number three get?” — those kinds of answers are going to become instantly available to a founder in Lagos who doesn’t know a single other entrepreneur.
And I think that’s fantastic.
If you're a founder in BLR who recently got their co incorporated in the US please help me with a few questions, BLR's best coffee on me.
Nakshatra Saxena retweeted
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.
Nakshatra Saxena retweeted
I'd say ~75% of my issues with AI coding disappeared once I gave my agent the ability to verify its own work
Zodiac is my favorite David Fincher movie of all time. On its cover, it features karl the fog and has the most interesting shot of the Golden Gate bridge.
Nakshatra Saxena retweeted
This #book is as niche as it can get.
It’s about the history of Lean, the math formalisation system that’s enabling AIs to solve Millennium Prize problems.
Lean started as an attempt to do formal verification of software in order to find bugs. But it was quickly adopted by the math community that wanted to bring rigour to math proofs which were becoming longer and unwieldy over time.
Lean uses type theory to enable machine verification of math proofs (search for Howard-Curry correspondence for a beautiful rabbit hole!).
This book is inspiring because it shows how a humble v0.1 attempt by a single individual can compound to attract a large community and eventually change the course of history within a decade or so.
Nakshatra Saxena retweeted
Introducing Workers IO
Agents are making it possible to turn ideas into software faster than ever. To keep up with that, we need better ways to understand how our software will behave in real-world.
Workers IO runs your entire system through millions of simulated futures, exploring what happens when machines fail, networks slow down, and events arrive in the wrong order. It searches for the combinations that break your system before your customers encounter them.
Every failure it encounters is perfectly reproducible, helping teams understand what happened, work on a fix, and verify against the same scenario.
This ability changes what teams are willing to take on. Difficult engineering decisions become things that you can explore and verify.
That’s the future we’re working toward: small teams building increasingly ambitious systems, with the evidence to stand behind.
I can’t wait to see what you build with that confidence.
Nakshatra Saxena retweeted
Many people in the AI industry believe that harnesses don’t matter, models have become powerful enough to solve problems with minimal scaffolding.
Well, that’s true only if problems are fully specified upfront like they often see in benchmarks.
Most real world work is ill-specified and humans discover what they want as they go along. There models do require harnesses that optimally help the human explore and discover what she wants AI to do.
So harness design must be now thought from human cognitive system point of view, not from the model point of view.













