The illusions of thinking produced by a "thinking trace" is just as ignorant of reality as the first and last tokens. All they know of reality is the tokens in their context.
The problem is, LLMs have such little understanding of the world around them. "Find exploits in specific software on this device" may as well be "find exploits".
The point isn't what happens after something is solved. Where is the motivation for humans to study/find a potential solution if your kids are going to starve while you do it? These things can brute force problems that have a clearly and feasibly searchable answer, but they can't make creative leaps. It will be a severe dark ages for mathematics if humans stop contributing.
Getting the kids to insist they be paid what their contributions are worth would raise the boat for all developers. Of course that would take some sort of big bad union, which the oligarch vc asshats really don't like.
#2 (prove those chats changed model behavior) is pretty straightforward if the anonymized data from chats can be actively searched by a model. In fact, it could be very clear if the provenance of context is traced. If anonymized data from chats leak into the context of an actively running model it would clearly influence the answer.
If the model has access to the "anonymized" data from chats, and the model is capable of building its own context from data that it can search through, including this data. Then it looks pretty damning. An independent review of the data traces from CoT and tool use involved in producing the result should make it clear one way or the other. Seems like discovery in a civil lawsuit could be very productive.
But you get more funding when you call it Recursive Self Improvement. Even better if you call it RSI so it doesn't evoke pesky skynet scenarios outside of AI safety circles.
This only applies if LLMs aren't making mistakes 20% of the time and that's the problem. When you're only saving time on the easy part, it doesn't matter if you're working twice as fast because review of the tricky parts is still going to take 80% of what it would have taken to do the whole thing. Total effort ends up being more rather than less if you want the same quality.
You're right that Nazi doesn't mean Nazi anymore. It means neonazi / white supremacist / white nationalist, which is a much broader group of people that, for some baffling reason, are under the impression that people don't care about their fascism and racism anymore.
The nasa guides are a great resource. Designed with the knowledge that a bad soldering job can cost lives. Maybe overkill for your average diy, but great for a thorough high-quality deep dive. Here is the reference standard context of that guide (also pdf): https://nepp.nasa.gov/docuploads/06AA01BA-FC7E-4094-AE829CE3...
Crypto mining would be one application. But also, combined with a sandbox escape would make it particularly devastating. Usually full control of a device takes at least two exploits given the layers of security present in OS and browser environments.
Not entirely. Lack of regulations mean recycled water facilities can end up being consumption facilities with pollution risk. Also Meta plans to restore as much water as they use by 2030. Which, seems like a while to break even. That's primarily because ai data centers require a lot more cooling water and power than traditional ones. So, you can scream fake news all you want, but it is a concern and it is a good way to oppose ai in general. And power consumption, noise pollution, and pollution are also serious issues. "Whatabout this worse thing" is a pretty terrible argument. You can whatabout anything, but it doesn't make that thing any less bad. aka two wrongs don't make a right.
Exactly right. It's a lot easier to prevent an asshat venture than it is to dismantle it. Now that the public is aware, I wouldn't be surprised if new golf courses start getting the same treatment.
Opposition to AI data centers is completely rational, which is why both liberal and conservative folks are against it. These data centers are being forced on communities with soon-to-be-former leaders agreeing to nondisclosure agreements. They raise electricity prices. They are being slapped together with poor water recycling, and noisy diesel generators that pollute. The data centers for AI use a lot more resources than traditional ones. And the tech oligarchs think they can ram rush these things through, but they can't. People are sick of the lack of transparency. They're doing it all wrong to save a buck and a minute.
> its self-improvement loop ends where everyone’s does: “Humans review those suggestions as PRs on the factory definition and merge improvements."
That's because it's still painfully clear that genAI has no taste. It's a median content generator. And the median kinda sucks. Of course you still need people to review the output.
The problem with the down detector main reporting page is that all of the graphs are scaled to the same size. The OpenAI spike was nearly 40,000 and the Google spike was just over 100 (just over 400 for Gemini). They look the same in the reporting page.