Ivo Wever

Do not assume opinions or character from a few tweets. I will point out good/bad arguments for positions I disagree/agree with.

[Ivo Wever @Confusionist]

@paulg There are more than 3 students for whom the numbers are consistent with not cheating if you take the normal difference between midterm and finals scores into account. Not many, but more than 3 need to be given the benefit of the doubt here

[Ivo Wever @Confusionist]

@SpencrGreenberg Assuming the reason they are unethical is that you would also act on those beliefs (there is no thought crime), then: things that are contingently true, which would improve the world when turned false.

[Ivo Wever @Confusionist]

File under 'Note that even experts make mistakes.' If you can't afford mistakes, you need something more than just one expert guarding quality.

[Matteo Collina @matteocollina]

Hey everyone! I triaged a @nodejs vulnerability, wrote the fix, and pushed it through the security process.

I now think it shouldn't have been a CVE. And the fix broke a big part of the ecosystem.

🧵

[Ivo Wever @Confusionist]

@robinhanson @TheZvi Yes, and additionally it implies it is an unexpected concurrence, a connotation that 'and' lacks.

[Ivo Wever @Confusionist]

@theo @MoonWalker2000D Would be interesting to see how the 'advisor mode' from CC performs, so Sonnet 5 that can consult Opus 4.8

[Ivo Wever @Confusionist]

@sporadica 1. Dario spoke for all of OpenAI then.
2. They worried about *risks*. Risks are not certainties. No one thought it would 'fuck us up'.
3. They didn't know what GPT2 was capable of. Imagine it was as capable as 5.6 is now, without anyone prepared by the past few years.

[Ivo Wever @Confusionist]

@paulg Step 3: be incredulous and indignant when the next administration does that in ways you disagree with

[Ivo Wever @Confusionist]

@Meaningness In addition to some things already said, I've always taken the hourly chances to not be independent. So you're literally missing information to calculate daily chance from hourly chances.

[Ivo Wever @Confusionist]

@tszzl @robbensinger More unpleasant with a higher p(doom), which means people are likely to deceive themselves to depress their p(doom).

[Ivo Wever @Confusionist]

@Austen They were already in serious decline before AI hit

[Ivo Wever @Confusionist]

@charlesarthur @Pinboard Is it 1998 again?

[Ivo Wever @Confusionist]

@fchollet You do not understand a complex system unless you also understand the edge cases and failure modes, because they are part of the contour of the system. But they absolutely do not *define* it. You could perhaps say they *complete* it.

[Ivo Wever @Confusionist]

@JohnWittle @allTheYud The concern is whether the AI is being honest, not whether it's correct.

[Andrew Curran @AndrewCurran_]

A new, more capable version of Mythos has emerged from training. I don't know whether it will be called Mythos 5.1 or Mythos 6, or if Anthropic will keep it internal to accelerate further development - but it has arrived.

Stopping models like Fable 5 or Mythos 5 from being served to the public does nothing to slow down development. In fact, it probably speeds it up slightly by freeing up resources. There are also no rules preventing the labs from continuing to advance capabilities while any current model is under embargo - or from keeping progress quiet until they choose to release it. None of them can afford to pause or slow down. We need only look at how capable GLM-5.2 is as proof of this. To protect their business models, the frontier labs must continually train increasingly capable systems to stay ahead of open source, and each other. The current continues to rage beneath the ice, and we continue to race toward our destination.

[Ivo Wever @Confusionist]

@ATabarrok @AmandaAskell Overtreatment consistently happens, also when only professionals interpret the data. What Stokes attempts to portray as sarcasm is unfortunately just reality. .

[Ivo Wever @Confusionist]

@SterlingCooley Oh look, a peddler of misinformation feels called out and totally Streisands himself, being eviscerated in the ensuing discussion.

[Ivo Wever @Confusionist]

@AmandaAskell @Bob866088873493 Do you believe, discounting hindsight, that, given the symptoms, they were right not to think an MRI would turn up anything? The desirable false positive rate for not prescribing an MRI is not 0 after all. Though perhaps not for so long..

[Ivo Wever @Confusionist]

@tszzl You don't think it was the kind of thing that was 'in the water supply'? It would be interesting to hear someone from Anthropic on how far off that idea they think they were.

[Ivo Wever @Confusionist]

@TheZvi Fixing ‘this particular jailbreak’ puts Anthropic in the position of having to forever promptly keep fixing whatever they consider a ‘jailbreak’, which would quickly become arbitrary, capricious and impossible.

Dario skipped ahead to the inevitable endgame.

[Ivo Wever @Confusionist]

@BasilEsq_ @TheZvi You are very confused or a liar.

Coinbase has existed since 2012 and the vast majority of its code has been produced by humans, before LLM's were even a thing.

LLM's don't even help that much in setting up KYC capabilities. Code is not the limiting factor.