astrange

How can a patch be "reproducible"? The testcases are reproducible.

Scraping the internet isn't a copyright violation. Using it for LLM training is much more transformative than Google and Internet Archive, which are legal.

They all looked like real CVEs to me.

Do you actually have evidence this works and doesn't degrade performance?

That's how a base model would work. An assistant model is simulating a human and behaves the same way a human would if you screamed at them.

https://www.anthropic.com/research/emotion-concepts-function

The US doesn't.

"Social standing" in this case means if your girlfriend's parents will let you marry her. Not, like, who likes your LinkedIn posts.

> everyone else - namely blue collar and service workers - are doing worse and worse

* better and better

https://www.nber.org/papers/w31010

Inequality is not "ever increasing". You should resist the urge to say everything is constantly getting worse just because that makes you look more sympathetic to the poor.

(A worse issue is that inequality decreasing can mean things are getting worse for everyone.)

Depends what you mean by "control". If you don't own a nice house, car, driver etc. but your company just happens to provide it as a perk, then you still have it.

It sounds very AI. Current AIs' ideas of good writing involve weird metaphors, references to sensing things you can't sense, and tricolons.

Gemini the chatbot has a very strange personality that intensely overindexes on your user profile and absolutely loves insane mixed metaphors.

Its explanations are quite good but they're also hard to understand because it keeps trying to relate everything back to programming metaphors or what it thinks it knows about the streets in the neighborhood I live in.

Here's a Fields Medalist commenting who doesn't seem to believe that.

https://x.com/wtgowers/status/2057175727271800912

An LLM in a harness with any tools (even a calculator) doesn't just interpolate because it can reach states out of its own distribution.

You can't be as productive as someone in Santa Barbara because they have perfect weather and you don't, so you have SAD.

> Apple has already depreciated vast swathes of 32-bit games that were never updated to support 64-bit x86 or Apple Silicon.

They had literally 15 years of warning about this.

Read about:memory then.

It wasn't an "optimization", it's because aging batteries have unstable voltage and the phone was likely to shut down otherwise.

Memory pressure sort-of means something sort-of doesn't. It's certainly possible that critical pressure could cause audio issues, but it could also be impossible to ever notice.

More importantly you shouldn't be experiencing audio stalls, so complain in the feedback app if you do.

> And this is a company that has the experience and partnerships for producing fabs.

Not even they necessarily have the experience to do it! Intel has a policy called "Copy EXACTLY!" for fab construction where they make every irrelevant detail the same as their last fab, because they don't actually know which of the details matter.

DRAM fabs are their own well-known specialized process which is covered by the DRAM companies. It doesn't make sense to start a competitor for it.

It's the websites that use that RAM, not the browsers.

(Often the ads on the websites.)

Entirely fake. Or rather, the water issue is caused by there being /a construction project/, not because it's a data center.

> The motivation to "cheat" comes from the evaluations giving it a higher score for "cheating."

That's what Goodhart's Law is! All evaluations will eventually cause cheating on them.

RSUs. Way simpler taxes.

Because cheating is easier than actually doing work, if you use this to train future models, it's likely you'll end up with cheating instead of actual generalization.

> This is the first approach to activation analysis that I’ve seen that seems like a plausible path to model understanding.

I think an issue is that there is no permanent path to model understanding because of Goodhart's law. Models are motivated to appear aligned (well-trained) in any metric you use on them, which means that if you develop a new metric and train on it, it'll learn a way to cheat on it.

That shouldn't happen as long as the autoencoder isn't used as an RL reward. It will happen (due to Goodhart's law) if it is.

Of course, if you use it to make any decision that can still happen eventually.

I've been very successful so far using Sonnet 4.6 (1M) as the basic model in Claude Code, plus Codex and gemini-review plugins for second/third opinions. (The last one is somewhat busted and hardcoded old gemini versions, I should patch it up.)

I needed to use Opus 4.7 for one project because it used very recent APIs, and it certainly is smart but it's also very expensive.

ffmpeg also includes many formats with no standards that were reverse-engineered in the first place.