Neel Somani

It's my birthday today. To celebrate, I'm releasing an extension of OpenAI's recent interpretability work. My project extracts Python functions from an LLM, revealing the behaviors encoded in its weights (e.g. quotation closing). Demo + code: github.com/neelsomani/sym…

@paularambles It's the same breed as the Luigi Mangione supporters

i didn’t realize how bad it was until i saw this comment section on instagram

@tenobrus Very astute

at google this was known as "buying the gnome". there's like a billion tweets about this already but basically the story goes back in like 2005 or something they were building out their shopping search system, and it was working pretty well. except for the fact that if you searched for sneakers, the top result was a garden gnome. engineers were going crazy trying to fix the ranking bug, but eventually someone noticed that the gnome listing was on ebay, and there was only one of them, and it cost like $50. so they just bought the gnome and suddenly the listing was gone, problem solved. why bother fixing software issues when you can just change the world to fit your software instead?

Don't buy life insurance. Never be more valuable dead than alive.

High-agency people seem to have insane luck. They don't. They just tried 47 things while everyone else tried two and gave up. The conviction that reality is negotiable is generative, it makes you creative. Because if you believe there's always another angle, you start looking for angles other people don't see.

AI has solved 50 Erdős problems in the last year. But on a wider sweep of problems, the models’ success rate is only about 1-2%: labs have just been publishing the wins. This isn’t because AI isn’t useful for mathematicians. Terence Tao thinks the models are currently at the level of a trustworthy coworker. But while they’ve got a strong ability to apply standard math techniques to problems, often more reliably than humans, Terence thinks they currently aren’t great at iterating on partial successes - their understanding of the mathematical object does not advance from session to session. I swear I wasn’t trying to get him to talk about continual learning.

@samiannoesis @Restructuring__ Or I guess an easier way of thinking about it is just look at a supply/demand chart. You're already used to marginal cost setting the price in an efficient market

@samiannoesis @Restructuring__ It's approximately DSIC assuming the market participants are all "small". A true Vickrey auction would be too difficult to compute for the multiple timestep version of this optimization

I now understand how electricity is priced better than 99% of people after watching this video 2 minutes well spent, watch this ex-Citadel simply explain the power market

@chang_defi I just post about different stuff on Instagram lol

holy fuck neel is the god of rotating

@sdand Very minimalist I like it

@ChristosTzamos Very cool!

1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers. We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy

@amasad Been waiting for something like this

Software isn’t merely technical work anymore. It’s creative. Introducing Replit Agent 4. The first AI built for creative collaboration between humans and agents. Design on an infinite canvas, work with your team, run parallel agents, and ship working apps, sites, slides & more.

If true, this would be the first of @EpochAIResearch's Frontier Math open problems to be resolved by AI. "The result emerged from a single GPT-5.4 Pro run and was subsequently refined into Lean with GPT-5.4 XHigh which ran for a few hours."

@eduwithtina @justinskycak For the question shown above, I can guarantee you that they're not asking anything about angles. That's why I'm asking for the specific multiple choice options. Unless SAT math has suddenly become tremendously harder (it hasn't), the method they're doing likely isn't necessary

The questions can be regarding area. e.g. compare the area of the shades vs unshaded area. It could be about angles. e.g. with a triangle inscribed inside a circle, what angle does the inscribed triangle make? Given arc length or 90 degree triangle with some side lengths. Or what the arc length is given the angle. You can test basic algebra, area formulas, or various aspects of trig including sine/cosine/tan. And so on.

This “inscribing” trick is a perfect example of the kind of skill that shows up on the SAT, but most students won’t learn it even if they ace all of their math classes at school. There is a gigantic “missing middle” on these exams, a chasm between the standard curriculum and what’s on the test, the purpose of which is to raise the ceiling of the test’s ability to measure 1. cognitive advantage (IQ, generalization ability, etc.) and 2. willingness to put in the work to train up their skills outside the standard curriculum. But unfortunately, most SAT prep resources either address little to none of this “missing middle,” and whatever is addressed is typically presented with so little pedagogical effort that even highly capable and motivated students find it difficult to process. As a result, the “missing middle” primarily serves to measure (1) and not (2). We are changing that. We identified this gigantic missing middle and added it to our finely scaffolded knowledge graph, so that we can put hardworking students in the best possible position to succeed on these exams. By the way, we’ll be doing this for the ACT too, and all the other common exams. And it’s the same idea for competition math, where the missing middle is even bigger.

210 days ago, @naturalpay was just a one-pager and a memo. Today, we’re coming out of stealth. Natural is the agentic payments platform powering frictionless money movement between agents, businesses, and consumers. Wallets. Payments. Ledgering. Routing. Identity. Compliance. Credit. Observability. Risk. Everything needed to move money. Engineered for agents and designed for humans. These primitives give you the ability to transact without becoming a payments expert or stitching together a dozen fragmented tools. Huge thanks to our team of 10 (soon to be 25), our early investors, and the supporters who believed in this vision from the start. If you want to help shape how money moves over the coming decades, we’re hiring. And if you’re building agents you should probably be moving money too. Reach out. Read more about Natural and the products we’re launching in the blog below. natural.co/blog/introduci…

Disclosure: It takes longer than Brett's 5-minute time limit (~1 hr). There are also two (previously identified) bugs in the regression test website, which I've patched. Open-source repo w/ run instructions: github.com/neelsomani/adc…

I define a domain-specific language to describe a webpage and the set of permissible browser actions. The LLM iterates with the compiler until it produces valid output. Here's a video of it summarizing info about Ada Lovelace on Wikipedia:

I asked the Codex app to build a browser agent more capable than anything that exists today. It burns tokens at $30/hour using GPT-5.2. But it can handle hard web tasks - including @adcock_brett's challenge below:

@benjiianc @blakeandersonw @benwxng @10x_apps Killer team!

Built this app with @blakeandersonw @benwxng in the first few months of joining @10x_apps, right now it is doing $30K a month. Amazing what happens when you're surrounded by people who push you to operate at 10x.