thomasahle

1) That's maybe $180,000 per year, so much less than median OpenAI employee wages.

2) OpenAI doesn't pay API prices.

3) Compute costs are likely already their biggest expense, dwarfing wages.

Do you want people here to try to convince you to have kids?

Yes please. For a moment I thought Terrence Tao had scooped Anthropic.

• 98.6% on ARC-AGI-3

• 97.6% on frontier math

• 95.9% on CAD

• 100% on ExploitBench

Nothing modest about it

> Grades should come from hard randomized exams with unlimited retakes

Exams will have to be a lot longer if you allow unlimited retakes. Generally exams work on a sample principle, but this breaks with retakes.

> had similar (slightly higher) performance.

The data point around 80 minutes seems like noise to me. Looks like there isn't enough data/students who spend that much time and also used AI.

It would be nice if AI was a force for good as well as bad, but the data here doesn't support it

Has anyone started proving their sandboxes in Lean (or Coq, etc.)?

> I expect a lot of the social anxiety can be mitigated by arranging things so that students get a chance to know the examiners and build a rapport with them ahead of time.

That's not my experience. Standing in front of somebody who's been your teacher and you might look up to is way scarier than some rando.

The talk is wild: https://www.youtube.com/watch?v=87DyyMV0kCY

> I want to note that every step in the process we discussed has had a remediation applied. The credentials have been revoked. The zero date has been patched and mitigated.

Good.

> a model trained while the message board was originally available and also found this this particular path to recreating it. This model creates a new agent message board using directories.

So no remediation applied to the models...

It seems super dangerous to continue training on those weights.

Nah, Claude will finish it over night

If it's successful, why do you think we'll even know how it did it?

Here's one way it could happen:

Let's say there's some circuit that does problem solving of the kind we call intelligence.

We dont know what this circuit looks like, but it exists in our brain.

Doing regression on outputs from the brain (e.g. internet text) with enough parameters, we can "fit" our model to this circuit.

But if you try to fit it with fewer parameters than it needs, you're just going to get some linear approximation.

> in a world where a CPU itself may have bugs in the way it executes assembly, that seems a bit silly beyond a certain point

CPUs are tested so so much more thoroughly than any software today. Including using formal proofs.

If the chance of a CPU bug is something like 1/10^6 and each module in your software is 1/10, reducing that to 1/100 makes a bog difference.

> duping is a great way to show people that even the app layer can be commoditized

Everyone is already cloning the app layer. There are 30+ claude-code clone, and codex work already cloned cowork, grok and gemini is doing the same.

> Whereas in the US you only need to go through fingerprinting only once

Once for each arrival.

> as long as China doesn't enter the GPU/RAM race

China is obviously in the GPU/RAM race. Heard of Huawei, Moore Threads, Lisuan Tech, CXMT?

Yes, but then it should be easy to check of the result you see match the reasoning you see

TIL:

> Many providers build their proxy pools by partnering with device owners who agree to share their bandwidth, while others use embedded SDKs in free apps or VPNs.

WTF. That's just botnets.

Source: https://www.fbi.gov/investigate/cyber/alerts/2026/evading-re...

I got into YC, immediately got a 6M offer from Google, sold and won.

Do you have a source for this, or just rumors?

The responses I get from pro don't feel like ensembles. They are often very one directional.

I used to part time for the (Danish) mail service. The only sorting that was done automatically was the post codes. That was enough to get the letter to the right post office. The rest was done by the mailmen/women early in the morning. It was a lot of fun trying to figure out what was meant by some of the addresses. The older people in particular often knew the story of why certain places were sometimes addressed in certain ways, or could guess the addresses based on the names of the people living there.

> After reading the article, the main "case against geometric algebra" I could find in there was that the author does not like the people using/doing research in geometric algebra

Mathematics is a social activity. The research cultures of different branches matter.

Why do you need individual data for gerrymandering? Don't you only need area level?

> Ban it from the dataset, add it to the analysis. You can choose your own flavor of noise.

Not sure exactly what you're proposing, but if the noise is added independently to different people, you can just buy multiple copies to reduce it.

There are a lot of ways to do this wrong, which is why so much analysis has gone into differential privacy.

Deepmind and OpenAI have offices in Europe. But I don't think Anthroipc does?

I recently built a very large test bench for System Verilog.

I ran a bunch of different compilers on it, including some open source ones.

Some of them failed some tests, and it was natural to have my LLM (Claude Fable 5) root-cause the issues, and to double-check my test bench wasn't to blame.

But now I stood with all these patches that I couldn't just throw at the upstream maintainers all at once. I ended up just filing a few issues and moved on to other things.

It felt weird to just file issues when my LLM had already spent a lot of time root-causing and fixing the issues. But then, maybe they could just have their LLMs do the same.

Still not sure if it was the right call?

> The rate of fundamental, broad-based breakthroughs lifting all LLM applications has clearly slowed with many of the most impactful recent discoveries being in scaling, optimization, tuning and productization toward specific domains.

To me it definitely feels like it's still accelerating, with the most impactful recent discovery being RL training reasoning models (late '24, early '25).

There's an interesting article called "sigmoids won't save you" https://www.astralcodexten.com/p/the-sigmoids-wont-save-you which argues that (unless you have privileged information) you should always assume a process will continue about as long as it’s continued already. (Lindy's Law)

With that in mind the current disruption should last another 10-15 years (assuming it started in '10 or '17.)

Is my website broken?

I'm trying to recreate all the commercial EDA stack in open source. (RTL simulators, synthesis, formal proof tools, etc.)

Building compilers has a _lot_ of parallel tasks agents can work on.

Wish me luck..

He used 600B tokens in 30 days.

I use more than 150B/month with just 15 codex accounts.

60 accounts is "just" $12,000/month. So Peter could "save" 100x by using monthly accounts.

Of course, he doesn't have to, as he works at OpenAI now.