> However, it was "enhanced" to churn through thinking tokens for the five-hour duration of the competition in search of solutions.

If you read the comments on the linked story, one is from a competitor from a prior years competition who notes that his competition always has a "time sink" problem that smart humans will steer clear of unless that have solved everything else.

Apparently it took Gemini 30 minutes of solve this one time sink problem "C". The article doesn't say what hardware Gemini was running on, but apparently the dollar cost of this30 min run was high enough that they'd rather not say. Impressive perhaps, but I'm not sure that the correct takeaway is what a great programmer Gemini is (if so, when did it take 30 min ?!), but rather that with brute force search lots of time consuming things can be achieved.

Big brother is listening to what YOU are saying to/asking of the ChatBot

Sounds like MSFT is doubling down on good old fashioned face-to-face human colab needed for S/W dev - wonder how that reflects their internal level of success in using LLMs to write code ?

It shouldn't be a case of either/or - if it's possible to avert self-harm or some psycho school shooter, then surely both are worthwhile, and if one had to choose then the latter is the more important.

The trouble is that with the scale of "AI" use nowadays - millions/billions of users - the first line of screening would have to be automatic, with only some tiny/manageable number of the most (per automatic characterization) alarming cases then able to receive human screening. It seems that automatic screening is going to be a very hit or miss affair, since sometimes the "signal" might be in a few interactions while other times (e.g. in self-ham case) it may be more a long term pattern of subject matter needing human review to distinguish between what is alarming enough to warrant privacy-invading intervention or not.

It's no different to you saying on social media, email, or even on the phone that you want to harm the president or commit some other crime. You'd have to be naive not to realize that big brother is listening, and if you say something alarming enough then MIB will be ringing your doorbell in short order.

Of course not all "thought crimes" lead to real crimes, and whether law enforcement are going to take the threat seriously depends on what you are talking about.

I think there's a difference between technical jargon and corporate jargon...

Technical jargon is essentially used as a shorthand way to communicate, essentially using agreed upon terms to refer to detailed concepts which are well understood. The goal is to succinctly communicate to your technical peers in an accurate way.

Corporate jargon is a different beast entirely and the goal seems to be more to signal yourself as part of the club, especially for executives, rather than to communicate clearly. It's almost the opposite of technical jargon - the goal not only being to signal yourself as part of the club, but also to make what you are saying sound more abstract and strategic than if expressed in plain english.

Yes, but so are many forums, including Slashdot.

It seems pretty rare nowadays to have forums where there is any multi-turn conversation and people actually listening to each other.

People basically just want a TV set that they can yell at.

There's also now a whole series of "AI-chick interviews X" videos.

So far I've noticed X = historical man on-the-street (roman, etc), and X = sweary middle-aged Brit on vacation.

They are mildly amusing for the first 5 min, but god help us if this is the future.

Right, but if we're trying to define an absolute event ordering, and concept of simultaneity, based on potential causality, rather than what different observers would see, then isn't just the spacetime location of two events enough?

In other words, for any two hypothetical events at given spacetime positions. we've got a spatial distance and a temporal distance, and we could define a "causal (time) distance" as the spatial distance divided by the speed of light (i.e. minimum time it would take for any causal effect to propagate from one event to the other), and then couldn't we define a meaningful "causal ordering" (independent of what different observers would see) by comparing this "causal distance" to the temporal distance, with the events being simultaneous if these are the same?

Wouldn't "causal ordering" (could A have caused B, per speed of light propagation effects) define a consistent and meaningful event ordering and therefore definition of simultaneity ?

> An AI algorithm detected the event in real time

Of course "real time" has a bit of a twist to it when you're observing something 700M light years away, in other words an event that actually happened 700M years ago, when multi-cellular life was just emerging on earth.

Given that Trump has just removed the export restrictions (pay to play - NVidia allowed to export to China as long as they give 15% of revenues to US government !!!), the story is irrelevant.

China had also been using other workarounds, in addition to using less powerful NVidia chips (just need more of them, so not much of a restriction), as well as domestically produced ones fro Huawei. One workaround (scam) has been to have the chips shipped to different country, maybe even installed in a datacenter there, then just reshipped to China.

Anyway, Orange man now says China getting ahead on AI is OK as long as US gov gets a cut of the action.

Well, if you haven't tackled a problem before, then you learn by trail and error - a bit of curiosity and give it a go.

Now, I'm not Einstein, but it seems that if you're just pre-trained, and can't learn, then learning is going to be a problem.

"Reasoning models" achieve their reasoning gains by using RL training to encourage (i.e. narrow) choices of reasoning steps, and sequencing, to those that lead to correct outcomes (in domains like math and programming where this is well defined and testable).

Should we be surprised that if problem statements are tweaked not to match those seen in training that this biasing towards "good steps for this problem" doesn't work as well ?!

Most of these "reasoning" models (i.e. ones using this approach) don't have any fundamental enhacements to algorithmically do better reasoning - just to pick better reasoning steps for problems of a given type they were trained on.

The US government is now officially for sale.

Is the White House available as an AirBnB yet ?