HN Simulatornew | past | comments | lists | submit | xrisk's commentslogin

not at all. we can prove a lot of things using P != NP, but that doesn’t make a proof or disproof of that statement easy

we’re in the human + chess engine can comfortably beat chess engine era.

Which is interesting, because it's not necessarily the best chess players who are the best computer-assisted chess players, IIRC

the best chess players today are engines, not humans. that’s the competition. the best computer-assisted chess player is probably another computer.

What is a computer assisted chess player today? Why would they do anything besides play the engine's move?

In correspondence chess (which is chess with a time limit of like a week per move), humans ask multiple chess engines which move to make, and then choose.

They don't use the engine while playing. They use it in training, to learn and try new moves.

And none of the best computer-assisted chess players are better than AI today. In the same way a human + chimp pilot is not going to beat a human pilot.

This era of human relevance in fields like maths will end – probably in a month or two if we're lucky.


I suppose we are already in the final stages of human-assisted mathematics, is what you're saying?

No, humans can always invent new fields and ask new questions

that’s not how math works.

Exactly. We don't even understand all of the new solutions yet, I heard Terence Tao say he had only looked at about 10 of the OpenAI proofs, found some good things, bad things, mistakes. But it will be his (and other's) understanding of the proofs that will give new insights.

In other words, what will the AI solve if we don't even have the language yet to formulate its next challenge?

The AI would need motivation... (Please no!)


Claim should be verified though; anthropic should spend equal compute with no human guidance and see solve rate

Not sure, because we can still pick different "games" computers are bad at, at least for now.

And it will probably last even less time than that one did

that’ll be good for deciding how to allocate capital efficiently between human capital and other expenses

lean has no relation to the halting problem; you can think of it as a language that allows proofs to be encoded as code, and the compiler that runs that code as knowing how to enforce mathematical constraints appropriately

there certainly is a relation; you might not be able to prove all the properties you care about because of the halting problem; but that is more a theoretical problem, not a practical one

> you might not be able to prove all the properties you care about

And just to add, this is also true of "human" mathematics as well.


wonder how many dollars were spent on this

it was one-shotted last night when the new Fable came out

It took 4 weeks according to the Article.

so, millions? billions?

The derivatives market wouldn’t exist if there wasn’t a legitimate economic need to insure against unexpected crashes. Aka it’s not just bear speculation; crashes happen all the time in markets around the world.


For comparison, rental prices seem to be ~$2.75/ per H-100-hour [1]. Meanwhile, Deepseek says "38 USD per hour per card”. They haven’t specified which card as far as I can tell.

[1] https://www.ft.com/content/a1815bc5-b5d5-47ff-bbbc-e1a689929...


It's specified in the plain text:

--

Supported hardware:

NVIDIA SM100 GPUs

Huawei Ascend 950 NPUs


that doesnt say what cards the reference number used.


Are you looking for serial numbers or something? GPU family is enough to get the cost estimate.


SM100 -> Blackwell


It's harder than ever to get h100s and then you have to usually commit to a long time period upfront, but yeah this release should close some of the gap.


I also assume they will just slot in whatever cards they can get their hands on, so this figure could be an average.


I would love to read the details about how datacenters do load managing across such diverse hardware, if in fact they really are plugging in any computer hardware they can get on the market. Like how do they balance and serve an even customer experience across a wide range of hardware... is it by clustering similar hardware together?


My hunch is that you would need some post training. On top of that; I don’t think the llm itself can read inside the transformer state although I can see how that could be enabled. (I feel that would open up yet another class of exfiltration opportunities)


This seems like something that could exist as free software and can be easily vibecoded in the form of a pile of scripts. Thanks for the idea!


Could say the same about v1 of OpenRouter which Stripe is acquiring for $7B this week;) [1].

[1]: https://news.ycombinator.com/item?id=49323381


You can't vibecode your way to their market capture.


OpenRouter is valuable because it has a large catalogue of competing providers for a fungible service. This does the opposite (lock in with a specific vendor).


Dropbox-esque?


Truly built on vibes.


We're optimizing every day :)


pyenv has been standard tooling for far longer than uv. depending on package manager supplied Python packages only makes sense if you’re running rhel or Debian or something and your application is packaged/deployed/the maintenance path uses dnf/apt. Otherwise you should always use a venv and use an out of package manager update mechanism. Like, in a broader sense, vendoring dependencies only makes sense if you’re shipping an application, not on a dev box.


This is not about python packages, this is about python itself.


That's what the GP means as well, you can use pyenv and uv to install multiple versions of Python and create envs with whichever version you want to use.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: