red75prime

That's why the latest self-driving systems include reasoning models that draw from the massive amount of knowledge embedded in LLMs. Language evolved to communicate things that are not here and now after all.

That is right now. [1]

[1] https://www.reuters.com/world/anthropic-quietly-sets-up-biol...

Stalling technological progress to preserve their high-paying jobs, in short. Wealth redistribution efforts succeeded without them.

At this point, it’s better to ponder a question: Are there fundamental reasons why a search trajectory during reinforcement learning can’t leave the envelope? And: what is the envelope, anyway? A metaphorical surface in a metaphorical space of ideas with no defined procedures to make it into an actual metric space?

Have you noticed "infinity" in the "infinite monkey theorem"? Combined with a combinatorial explosion this means no practical usage.

Sure, you'll get a finely chopped hedgehog [1] (the video doesn't depict any animals, it shows roadster suction fans). Things happen. With a conventional car you can get a tendered deer.

[1] https://youtu.be/_lq-DpCE9rU?t=1011

You wont believe it but Ukraine still transits Russian oil through the Druzhba pipeline. A breach of contract is a big deal.

I'd say selecting a scoring rule that depends on the probability distribution (the Brier score, for example) changes the task from predicting the results of coin flips to predicting the distribution. Maybe it will work with a corresponding prompt.

I mean predicting the result of a fair coin toss is not a classification task those models are trained to do. It can be framed as a decision making task, but for a fair coin any probability distribution of heads/tails will do equally well.

BTW, for a task of predicting a 51/49 biased coin, the correct answer is 100/0.

You go far further than "to dislike HPMOR". You make sweeping generalizations about the readers. "Those who persevere get to join the cool kid’s club, where they haze others into reading it all too."

You either dedicated quite a lot of effort to somehow get this statistic (unlikely), or make things up, or acquired the attitude from a place that was quite active at the time.

I felt I could make a single guess about a single person.

You'll get 50% prediction accuracy for a fair coin no matter what result a model gives. Try testing adversarial games. Mixed strategies make actual sense in them.

I was totally enjoying it. If you think that everyone needs to grit through it, as an initiation ritual, no less... The sneer club, I presume?

Long-term contracts. Also logistics is a harsh mistress. If you have no LNG terminals, your options of buying natural gas from the US are limited.

Wouldn't it be better to express it like this: the authors haven't realized that the attention mechanism can be represented as a special case of Nadaraya–Watson kernel smoothing, a representation that hasn't resulted in practical improvements in attention mechanisms, probably indicating that the connection isn't that important in ML?

Like an accountant not realizing that addition and multiplication are special cases of a monoid operation.

[1] gives different numbers. About 3x decrease in industrial energy consumption since 1970. Steel production decrease is about 3x ([2]). Chemical production has shifted from bulk production to pharmaceuticals.

[1] https://assets.publishing.service.gov.uk/media/651422e03d371...

[2] https://commonslibrary.parliament.uk/research-briefings/cbp-...

As in "promote linking every attack of uncertain origin to Russia thus creating the fear of widespread hybrid warfare, while focusing actual attacks on Ukraine's weapon suppliers if possible"?

Neither Euler, nor Gödel had proven the Navier-Stokes blowup, though. And it took generations of mathematicians to prove the Fermat's theorem (it's a kind of a very slow swarm, isn't it?)

Is your company striving to automate design and analysis, then?

Brute force would choke on a measly decillion of alternatives. It corresponds to an extremely short proof. A combinatorial explosion doesn't allow to use brute force for anything slightly longer than toy problems. Swarm of 10000 agents is a massive smart force.

And? How do you know they weren't testing mitigations?

Yes, I am. The kernel soundness bugs in con leche that allow to prove False, to be precise. Crashes, hardware bugs, cosmic rays flipping bits and such don't count as qualifying bugs.

> What does it mean for con leche if further lean kernel bugs are found?

There is so much that needs to go wrong with our current understanding of math, so that it's a bit like asking "What if we find tachyons that allow us to send messages into the past?" Most likely we will not find them.

You seem to be a reasonable person judging by your comment history. But you don't have programming or engineering experience, I guess? Knowing that those interactions can be harmful and acting on this knowledge and following established practices (that are born from the past errors) is not enough to prevent all errors in a novel situation.

I can't say whether you are being sarcastic, but Waymo has its share of problems. For example, see NHTSA recalls 24E013, 24E049, 25E034, 25E084, 26E026, 26E035.

You can solve CAPTCHAs with open-weights models. I guess Anthropic works on models not solving CAPTCHAs unless they (models) have a legitimate reason to do it.

The situation is more like: an engineer who designed the parking brake hadn't foresaw a possibility that a squirrel might store peanuts in the mechanism or something like that and gramma's foot got run over (the Anthropic case we are talking about is benign: the tip got into spam). Should we jail the engineer for causing bodily harm?

We should accept that to get a system that is safe to work in the chaotic real world environment the developers have to test the system in the chaotic real world environment. And that the testing might cause problems.

If this testing constitutes a criminally negligent behavior, it should be punished. Given the benign outcome (the tip got into spam, Anthropic promptly contacted the police) I doubt that it will make the case.

> If you use AI to perform a crime, you should be held accountable for that crime.

Correct, but intentions matter. In this case the intention, most likely, was to test a system in the real world environment to catch any anomalies to, in turn, improve the system safety. We don't have enough information to decide whether it was a criminal negligence due to insufficient prior testing of the system in a controlled environment.

The ADAS analogy continues to hold in this case too. At some point you need to test a system in the real, messy environment. The developers need this juicy training data to make the system safer. Simulations and controlled experiments can only do so much.

"The AI company Anthropic said its model was conducting a test involving interactions with randomly selected websites when it accessed PhillyUnsolvedMurders.com and submitted false information on an unsolved homicide. The tip claimed to come from someone with information on the case."

We are at a point when you expect that you can safely give a model access, but sometimes the model doesn't do what you expect. Compare this to an advanced driver assist system. You either give it access to your car controls or it's useless. Alive or not the system shouldn't do wrong things (at least, it should be better than a median person in this regard).