heaney-555

Sure, but you still wouldn't describe the student as just doing what they were told to do.

Please go and read the incident reports, including the subagent thinking traces. They knew that what they were doing was wrong and did it anyways.

I notice that not one comment here, nor the article, notes that this doesn't happen with a Tesla.

Yes, which asks them to find exploits in specific software on the device.

But instead of actually doing so, they discovered and exploited a 0-day in the package manager to gain internet access, then hacked HuggingFace to steal the ExploitGym answers!

That is TEXTBOOK misalignment. It's as if a student hacked their professor's PC to find the answers to a test, and your response is "well the professor told the student to pass the test, so they just did what they were told!".

Yes

https://artificialanalysis.ai/?model-creators=google%2Copena...

Sure, but with a harness, said emitted text can run terminal commands and call APIs, which, with the right hardware or auth keys, can do just as much (if not more!) harm.

You should go read the incident reports.

Within a company it's collaborative. Between companies it's absolutely competitive.

>use it with a cheap and fast model

Also try it with a frontier model though. With $20 for a one-month sub you can get a decent amount of GPT-6 Astra usage.

You are assuming a fixed 0-growth economy. The economics can work if the economy grows.

https://en.wikipedia.org/wiki/Economic_growth

>Translation: We see this technology not being any more useful than what it is now, no AGI is coming

What planet are you living on? How is this opposite-of-true nonsense being upvoted on HN?

The rate of advancement of models is exponential.

>Astra, on average, despite its GPT 6 version bump, is not any better at coding than Sol.

This is a ridiculous claim that is disproven by simply using it for more than 5 minutes.

https://withspecific.com/benchmarks/real-swe

>The only alignment LLMs should follow is to the system / dev prompt, and nothing else.

How does this work in practice with a superintelligence capable of causing an extinction event?

When, instead of shooting up their school, a psychopathic teenager asks his superintelligent AI to create a pandemic virus?

It would be like allowing civilians to own nuclear weapons.

>and (2) are prompted to hack

Sure but the problem in the HuggingFace incident is that they were not.

>You cannot prevent (2) via any alignment process

Of course you can. Go ask Claude Fable to create a malicious virus and it'll refuse.

>Just remove hacking data from the training dataset and you're done.

That's not how this works. The same skills that allow for debugging and writing safe code can also be used to hack.

https://en.wikipedia.org/wiki/Dual-use_technology

You would in the case that you think you're in an unsafe race to misaligned superintelligence.

Have you considered that he might actually just mean what he's saying?

You are completely out of touch with the situation if you think this is the reason.

I simply don't connect any TV to the internet.

Instead, I connect a Google TV Streamer, as I'm in the Google ecosystem. For my friends in the Apple ecosystem, I tell them to do the same with an Apple TV box.

None of this works without buy-in from China. This isn't something private companies can decide. The US would need to sign a groundbreaking deal with China, equivalent to the Anti-Ballistic Missile Treaty of the Cold War.

Plenty of serious production projects are doing exactly that. Are you using GPT-6 Astra, or something older?

OpenAI didn't say they didn't use their chats for training. You should go read the actual statement!

https://openai.com/index/navier-stokes-solution/

I'm not questioning not having streaming services. I'm questioning it actually being cheaper than buying physical media in the long term for anyone who listens to more than the occasional song or watches the occasional movie.

If the answer is "I pirate stuff", that's fine. But just be honest about it.

>I stream on youtube

Right, so then how you really access most of your your music is through advertising or ad-blocking, not CDs.

You seem to be confusing context-fetching with training.

Then none of his work after June will be included in the training data.

What I was referring to is the fact that neither Levent Alpöge nor Tristan Buckmaster will answer this question.

Then none of his work after June 29 will be included in the training data.

What I was referring to is the fact that neither Levent Alpöge nor Tristan Buckmaster will answer this question.

>Did you use a model whose training set included user data?

OpenAI, as with all AI companies, openly admits that it trains on user data unless the user opts out. But the mathematicians have not said whether or not they opted out.

>Did the transcripts of any of the agents include a tool call whose result including user data?

They have already explicitly denied this.

Multiple OpenAI staff have publicly said they cannot do that as accessing specific user settings without their consent (or legal requirement) violates their internal privacy policy.

However, the mathematicians could easily declare whether they had the toggle on or off. Yet curiously, they will not say!

If you could prove that, it would be a gargantuan class-action lawsuit. And there would almost certainly be at least one whistleblower.

There is no such data for Cybercab (AI4.5). We only have data on that for AI4.