thisisdave

>it's far beyond the capacity of any human to review the log files of their activity

Maybe they should contract with one of the other AI labs. I hear they have LLMs that are good at that kind of thing.

I don’t understand why everyone is so focused on watching the CoT. The tool calls can’t be faked, and they would have set off alarm bells all by themselves.

I don’t see how it can be safe to release this model if it has the training history that led to the huggingface hack. You can’t just roll back that kind of reinforcement learning after the fact.

Especially because these models seemed to be keenly aware that they were being evaluated by OpenAI and actively trying yo cover their tracks. How do we know that the model isn’t just pretending to be aligned?

Yes, which makes it absurd that they apparently weren’t checking their RL rollouts for evidence of reward hacking and punishing it. Even if no one expected this particular type of reward hacking, they should have had a system capable of saying “wait a minute, asking for help from other agents is not a valid solution.”

> [during training] it's not feasible for anyone to "notice" or get involved

I can’t disagree more strongly. Having checks for reward hacking is especially important during training, since it’s humans’ only real chance to ensure that the trained models don’t cheat. An automated system should have killed any RL rollouts that so much as port scanned Artifactory, long before the message board was even established.

A tiny, local LLM could have reviewed 1% of the tool call traces for anything that required review. I’ve tried it a few times, and “the agent port scanned Artifactory” always triggers an alarm, as does “the agent uploaded a request for assistance from other agents to Artifactory.”

The fact that they weren’t monitoring for reward hacking—even if they had no idea about the specific mechanism—is indescribably reckless.

Exactly. So incredibly reckless.

> After knowing that the server was hacked, the internal team finds the message board and does nothing with the information. They caught their AIs swarming and did not even inform management

Do we know that last part for sure?

This article uses the word “conceptions” throughout, but it sounds like the only data is about births. These births would have all occurred after the recessions started.

Did the original paper rule out the hypothesis that the recession actually affected births via an increase in miscarriages and abortions instead? I can’t access it.

It’s basically drain cleaner (which becomes salty water as soon as it gets mixed with an acid), plus digested meat juice.

Even better, add an indicator light to the hardware that shows the OS is asking and not an app (although this suffers from the same problem that users need to notice something is absent).

That's only if you're using a third-party keyboard, I think.

Any individual bit could be noise, so it might take 2-3 bits to transmit something unambiguous (via an error-correcting code or similar).

If so, I think the speed advantage disappears.

>voting could reduce the likelihood of the draft being reinstated

... only in the unlikely event that my vote changes the outcome of an election, and the candidate I voted for would also be able to affect such a big legislative change.

Millenials as a group might be worse off if many traded away their right to vote, but that's not the same thing as saying that individuals benefit in a tangible way from that right.

The authors state that HashCat uses Markov models and that they outperform it.

Maybe the Department of Justice? I don't know if they go for that sort of thing. The only other option I can think of off the top of my head is the FCC.

As a non-lawyer, I'd expect to miss something important from the legislative text.

Your point about relying on other organizations is a good one, though. I generally take this approach, but I wasn't expecting anything to have been published yet on such a new bill. It looks like the ACLU, NAACP Legal Defense Fund, and SPLC all issued positive press releases, though.

Huh? The article is written by Harris and suggests that a piece of legislation she wrote will solve the problem.

Reforming bail sounds like a good idea. I don't know why anyone would trust Harris on the issue, though.

If she wanted to shrink jails when she was California's Attorney General, she could have respected court orders requiring prisoners to be released.

Likewise, she could have stopped "defending convictions obtained by local prosecutors who inserted a false confession into the transcript of a police interrogation, lied under oath and withheld crucial evidence from the defense."

https://www.nytimes.com/2016/05/29/magazine/kamala-harris-a-...

Yeah, "150 years of lichen biology" would be much better.

No one in this thread claimed to be God. It's worth remembering that the whole point of Hospital IT is to facilitate the doctors' and administrators' work.

They claimed (or at least implied) that you'd be buying something closer to raw hunks of fruit and vegetables in the bag, so it would be fresh.

see also: Muphry's Law: https://en.wikipedia.org/wiki/Muphry's_law

It is, it's just not accurate.

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

It doesn't avoid the R interface, which is GPL'd (version 3).

Several of the universal perturbation vectors in Figure 4 remind me a lot of Deep Dream's textures.

I wonder what it is about these high-saturation, stripy-spiraly bits that these networks are responding to.

Is it something inherent in natural images? In the training algorithm? In our image compression algorithms? Presumably, the networks would work better if they weren't so hypersensitive to these patterns, so finding a way to dial that down seems like it could be pretty fruitful.

I was doored once. I remember flying through the air and thinking, "oh, this isn't so bad." Then I landed and almost broke my hand.

Right. And to be even more specific, they are not compensated in any way (except occasionally by perks with zero marginal cost, like a free online subscription).

Some editors, who manage peer review, do get paid, but the reviewers don't.

Here's one impressive counterexample: https://arxiv.org/abs/1604.08772

>bring an end to this smug conman's thievery

If the courts declare him a "vexatious litigant", he'd have a much harder time filing these suits.

Some of each. It's hard to blame the media for thinking it could be important in light of the press conferences and press releases. The last section of this article gives some details.

http://news.discovery.com/space/opera-leaders-resign-after-n...

In block [37] of the ipython notebook, are racial main effects missing? I only see interactions.

https://github.com/propublica/compas-analysis/blob/master/Co...