artrockalter

The internet became a single point of failure for all of civilization years ago and we're only realizing it now.

Part of the problem is that from that perspective there was no independent analysis done that could vindicate them. METR is absolutely part of the EA/LessWrong/rationalist ecosystem, so of course their investigation would validate that group’s arguments.

In my experience LLMs do find a lot of embarrassing bugs as Linus says but it can constantly turn into a game of whack-a-mole where most of the bugs it finds were written in previous LLM sessions. It's a huge struggle to get it to actually fix the root cause of the bug instead of patching the symptom.

I know the company I consult for (not cybersecurity) is not in these programs and if attacked would need to use open weight models.

The Hugging Face incident is a great example of why open source models with defensive cyber capabilities are needed. Hugging Face did not have access to cyber-capable frontier models and kept hitting safeguards. Only by using the open source GLM-5.2 were they able to survive an attack. A world where open source models are banned is one where cybersecurity is impossible if you're not on OpenAI or Anthropic's allowlist.

The null is surely different in this case because we know that supplemented creatine goes to the brain and the brain uses it, which is definitely not a claim all supplements can make.

The fact that there's a biologically plausible mechanism plus mixed statistical evidence makes "maybe" a perfectly reasonable conclusion.

Is this methodologically similar to Anthropic's recent J-Space paper?

The task could be verifiable in the environment so limiting its CPU and RAM could be to discourage brute forcing the answer.

Both are true! The Confederacy did secede largely to preserve slavery, but the war was started to bring the Confederacy back into the Union, initially without the goal of also immediately abolishing slavery.

LLMs are good at writing complex regex, from my experience

When I use ChatGPT for work it frequently reads my Slack DMs even if they’re not directly relevant, so I’d question a lot of the premises of the article.

“they’ve been folded into Frankenstein positions that demand constant multitasking, social performance, sensory endurance, and emotional labor on top of technical skill”

My sense is that this is due to automation, not “neoliberal capitalism” as the author says. It’s much easier to automate a job if it’s a single task that’s done in a deterministic way.