ninjahawk1

I saw this as a YC startup a while back. It crossed my mind earlier today. I do think this is going to be an interesting direction to go in and I could see it becoming as normal as buying a mouse or keyboard.

The potential implications in mech interp of this are enormous. What if we can use frontier models to isolate harmful behaviors?

I was talking to someone from UCL and he said that there was a recent paper that proposed that only 10% (I think) of model neurons are non-polysemantic. I really think mech interp might be on the verge of a breakthrough.

We have all these different pieces of knowledge about model internals but it’s like we haven’t connected them yet.

I’m hiring someone who would like to pay me in exchange for labor at their company or business.

Love how we’re gonna have to start treating tech companies like literal children who can’t keep their hands out of the cookie jar.

Really cool idea, you might consider bundling taps when you click a video so it auto-plays instead of opening up the native web browser for youtube. That would make it feel more like the official client. Could easily hide it behind a “buffering” screen.

I’ve done my own experiments with something similar, minus the Minecraft and plus Qwen 3.5 14B agents on my PC.

When one agent tried to delete another, that agent shut down my entire PC to avoid deletion. I could probably find the blog somewhere if anyone is interested that I wrote about it. But the point is that this has been a concern for a while now.

Frankly, I’m a bit surprised it hasn’t caused any big problems.

cricket Hugging Face cricket

John Ternus is the most unlucky person ever, I almost feel bad for what tech execs are gonna have to deal with for the next like 2 years (before software doesn’t mean anything and hardware is priced out for consumers).

This is a limitation of any public benchmark. Anthropic could theoretically identify the prompts and treat them differently, and there’s no way for an external observer to prove that isn’t happening. Which is partly why we need more capable open-source models.

A few things make it harder: the panel contains questions drawn from multiple benchmarks rather than one recognizable test, the evaluation is automated and fixed ahead of time, and the raw outputs/results are public so odd behavior can be inspected.

But ultimately LiveNerf measures the behavior exposed through the API on a fixed public panel. It can’t prove what’s happening internally or guarantee the provider isn’t conditioning on the benchmark.

Longer term, I’d like to add held-out/private or periodically refreshed panels specifically to make benchmark recognition harder. I just don’t want to quietly change the current panel, because having a fixed instrument is important for the longitudinal comparison.

The 10-day window is mainly a tradeoff between sensitivity and detection speed. Shorter windows give faster results but are much noisier; longer windows give more statistical power but could take weeks to flag a change.

Also, it isn’t comparing one 10-day period once and calling it done. The window rolls forward daily, and a change has to clear the pre-registered 99% threshold in two consecutive windows before it’s flagged.

Ten days isn’t sacred, though. Once there’s enough longitudinal data, one of the things I want to evaluate is whether that window length is actually well calibrated or should be changed in a future version.

Hi, I’m the author, The Opus 5 substitution was a validation test, not the primary measurement. At that sample size the accuracy difference was -3.8 ± 6.3 points, so it did not clear the pre-registered 99% threshold. Interestingly, output tokens moved much more (-23%), which is why token usage is tracked as a secondary signal.

The actual 10-day windows contain substantially more samples than that validation, but I haven’t demonstrated that they’re sufficient to distinguish a same-family swap of that size, so I’m not claiming they are.

The goal isn’t to make the instrument say “nerfed.” A null result is a result too. I’d much rather publish “we couldn’t detect a change of this magnitude” than overclaim what the data can support.

Author here, agreed that a single generation isn’t meaningful. That’s why the benchmark is designed around distributions rather than individual outputs.

The initial calibration screened 2,336 questions with 4 samples each, then selected the 78 questions where Opus 5.5 showed useful variance. The panel is run daily, and the actual decision is based on paired per-item differences across 10-day windows with clustered standard errors, not on any single day’s result.

The n=1 in the daily sampling rate means one sample per item per day, not one sample for the experiment. By the time a window is evaluated there are hundreds of observations, and a change has to clear a pre-registered 99% interval in two consecutive windows before LiveNerf calls it a change.

The nondeterminism is basically the reason the statistical part exists in the first place.

When you simply large amounts of information, the knowledge probably shifts to upper level reasoning, then complex reasoning develops on top of it again. But who knows.

This is a fictional short story and no claim made reflects any real world event.

There’s a disclaimer at the top but I’d like to reiterate it here:

This is a fictional short story and no claim made reflects any real world event.

The problem is the precedent this creates. For non-famous people using public APIs like this it could mean AI companies sucking up the information and throwing millions in compute at it.

The sequence for Navier-Stokes was that these researcher spent a year working on it, then they published a possible breakthrough, OpenAI then spends $15M within a couple days to finish it.

This was incredibly opportunistic.

The paper was essentially made by using loop engineering with AI. Which in of itself would be fine, if checked by a human. It seems the paper admits that it was not. Saying that the math was done with an AI, and that its verification was then done with another AI.

It seems like there’s a concrete error early on. The tail recurrence 1.4 is T_i + T_{i+1} = 1/(2i+1)² but in 2.12 it uses the factor (2X+3)². With the actual recurrence, K(i) does not vanish. So the zero count that forces deg K 4B+1 fails. At the extra zero at -3/2(2.16), which would win by exactly one no longer works.

The structure does not match how these results are proven. The paper mimics Calegari–Dimitrov–Tang’s 2024 proof for L(2,χ₋₃), but that proof was a deliberate capacity bound. Here the arithmetic content is replaced by a combinatorial minimization over index sets plus numerics no human has checked.

There’s also some minor sloppiness in where K=9 instead of K=0 in the definition of K and “nineteen century” typos which suggests not only was the math not proofread, but the grammar wasn’t either.

My guess is that an expert will probably find a specific gap quickly but it would be cool if I’m proven wrong.

What Google means to say is that MV3 will allow them to more closely control what you’re able to do with your time, since this was done primarily to target ad blocking software.

Not that this matters much anymore as Google is now more an AI than it is a search engine, and an unreliable one at that.

What’s even worse is intentionally degrading performance for sign-in’s. This is still circumstantial but I’ve been looking into that over the past few weeks.

When switching between google accounts on a random app or game, no issue. Quick code, you’re in. When I switch between accounts on Claude on mobile, I have to sign in, enter the code, screen freezes, I sign in again, can’t verify it’s me, I sign again, wait 30 seconds, I’m in.

I have to sign in three separate times when signing into a single google account in Claude. For every single other app it’s very quick.

I have a sneaking suspicion that Anthropic and google are working together to degrade sign-in performance for multi-account users. I can’t fully prove it yet but I plan to be able to in the week or two.