scarmig

> On that current AI is a bit lacking as they don't have very good sensors whereas human brains are pretty wired up for touch, smell, sight, sound and the like.

For sight and sound, the sensory input layer for AIs is much, much richer in terms of bitrate, though is redundant.

> AIs do get linked to cameras though and could improve a lot in those directions.

Video is an interesting one, as most AI systems use the full pixel information of a scene, while human vision is much more temporal, focusing primarily on logistic deltas rather than raw pixels. That's not inherent to hardware (e.g. see event or "neuromorphic" cameras), but it's a significant difference. My more sympathetic reading of embodiment is that the relative paucity of the human sensory stream is part of what makes us work better than current AI in many domains, in that it forces us to form richer, temporal world models from a low bitrate stream of salient events instead of relying on shallow responses to rich sensory data.

I think the critical factor that makes sci-fi hard isn't known physics, exactly, but early on establishing a known set of rules and letting them play out as they will. The narrative has to respect and adjust to those rules, not vice versa.

This ends up driving most hard sci fi to current understanding of physics, with maybe one tweak. If you diverge more from that, it's difficult or impossible to establish the rules and their implications early on, and so choices about them later end up serving the narrative.

Though, a certain irony: hard sci-fi doesn't seem to offer much predictive insight about the future compared to other subgenres. The 20th century sci-fi genre that seems to offer the most insight into our current moment is probably the one with the wildest imaginings: cyberpunk.

Embodiment doesn't exist, even in regular humans. We do not have direct access to reality; we have sensory inputs that are much lower bandwidth than one might expect but correlate with external events, and our brain uses them to form rich world models that are usefully predictive. Most of the world we experience is just in our head.

Things that are vaguely GAN-shaped might play some role in post training. Though no one would call them GANs or identify them as the key underlying technology.

The beernet conjecture: all conjectures have both messy, ugly proofs, as well as a elegant clean proof lurking behind the scenes.

I know multiple Physics PhDs who are living rich, happy lives.

Though, mostly by abandoning the field and moving to Anthropic.

The claim I was responding to was that alfalfa using 1000x the water of Google data centers was okay because alfalfa is needed for life. My response was not "data centers are needed for life" but "alfalfa is not needed for life."

Although delicious, alfalfa is not required for life. A lot of it goes to Saudi Arabia, who use it for horses and cattle, as the Saudis banned it domestically for being wasteful for water usage.

The AI paperclip case, however, is coming on extraordinarily strong.

> I'm surprised there are comments here that aren't impressed about what was accomplished here.

The phrase to describe it is anti-AI psychosis. Which isn't about providing thoughtful critiques of AI, which are good and we need more of. But anytime an LLM does anything--prove a major math problem, create a successful hack against multiple corporations simultaneously--people feel compelled to start minimizing it in ridiculous terms. It's just a script kiddy; it's just a marketing scam OpenAI cooked up; the Jacobian conjecture counterexample was something anyone could have done in a weekend; etc. It has to just be a stochastic parrot, because it's scary to imagine a non-anthropocentric world. And it's rightly scary, and we should slow down and try to better prepare for it. But blanket denial is not a strategy that will lead to success, and people who rely on it are sorely ill-prepared for the next couple years.

I like the math vs physics toggle.

A throwaway line that consists of baseless speculation that makes no sense is a Reddit-tier take, yes.

The entire "OpenAI and HuggingFace manufactured a hack in a conspiracy to make AI look powerful to get more funding" is a stupid, Reddit-tier take.

If you're Demis, at least, you sleep fine because you were personally an early investor in Anthropic.

> In May, a House of Lords report said England faced a 5bn litre a day shortfall for public water supplies by 2055.

> Water UK estimated that datacentres in England use 6.6m litres of drinking water every day.

> England’s current water shortfall is 220m litres more a day than was assumed in water forecasts.

If a driver doesn't get into a crash after a single mile, you can't conclude that they are safer than the average driver. If a driver does get involved in a crash after a single mile, it's very likely that they are not safer than the average driver. It could of course be extraordinarily bad luck, but that's unlikely.

In fairness to the authors, legal fault is not sufficient to encompass how different vehicle types change accident risk. E.g., if someone breaks abruptly for no reason and that results in rear ending, the breaker wouldn't be "at fault," but they still share part of the actual blame. And I'm sure at least half of the accidents involving taxi cab drivers had them similarly wrong place, wrong time.

Where I object to the authors is the statistical spelunking and secret data sets that aren't publicly available but provided as a favor from NY government as a favor to the author. Are cabbies really 14x as safe as the average driver? It's impossible to say; we know nothing about the data they used to get that estimate. And some things strain credibility: what model do they suggest for Waymos being much safer than cabbies for all accidents (as supported by their own reported data), but suddenly getting much worse for serious injuries and fatal accidents, especially seeing as how all of the three accidents involving the Waymos they use involved a stationary or low speed Waymo where the Waymo is pretty clearly not at fault?

Concretely, they are basing this analysis on 3 Waymos being involved in incidents involving a serious injury or fatality nationwide, versus 16 for NYC for hire ones (presumably in NYC or its immediate environs). The three incidents they're referring to involving a Waymo, for reference:

https://www.cbsnews.com/sanfrancisco/news/san-francisco-fata...

https://www.azfamily.com/2025/09/14/1-person-killed-multi-ca...

https://abc7news.com/post/police-chase-san-francisco-waymo-p...

The idea that a calculator, or a calculator with Python, or even a calculator with a proof assistant (and every book ever written on math) would help a random person at e.g. IMO or Putnam is fairly revealing.

Some quotes from a day ago, https://news.ycombinator.com/item?id=48957779:

> I hold my stance that LLMs are stochastic parrots... Making the parrots ever more complex and training

> Except solving problem is probably the least (even though it's important) interesting thing in research.

> Can we use AI to get a cure for cancer yet? Or is math-turbation the only thing these things are good for?

> Train on enough examples and statistical autocomplete gets you places. I'm surprised how anyone would even consider this intelligence?

And, as much as HN has declined in the grips of an anti-AI psychosis, Reddit is worse. I would love if social fora would switch to the reasonable claim that we're in a bubble; that's something that can be debated. That's not the dominant critique of AI, though.

"Any idiot could have done this, it's just high school calculus and just a counterexample anyway. Stochastic parrot, spicy autocomplete, AI psychosis. Wake me up when an AI does something real."

It's a kind of free insurance: if a bad financial event happens, you're protected against the worst downsides.

If you're from a poor family and break out of poverty, you're still in a worse financial state than someone from a rich family, even holding income and assets constant. You've got to effectively self insure (by taking fewer risks; being more conservative in investments; cutting down on rich people expenses that help with networking) to plan for the worst case scenario. And rich kids don't even realize how much their families' financial status enables and drives their behavior.

And that's not even starting to count financial expenses to take care of aging parents who can't afford to take care of themselves.

LLMs have basic reasoning and a whole lot of memorization. Through that basic reasoning and pruned search, combined with piles of compute, you can prove lots of things. But the memorization of human failure prunes that possibility, and you need to expend effort convincing the LLM not to prematurely prune based on previous human failure.

Two years from now, when a proof of the Collatz conjecture is announced and verified:

"No one cares about that! What practical applications does it have? It's pretty trivial anyways. More hype from the AI psychosis crowd."

You can solve it with two heaps, which don't need to maintain a complete order. Or selection algorithms, as in the sibling comment (asymptotically better).

Everything we do has side effects. Some have more than others; in the case of water, even heavy users of AI have a relatively small AI water footprint compared other things they do.

There are ideological perspectives--another comment in this chain declares that anything that an LLM does has zero value, purely by virtue of being from an LLM. There's not much to discuss there: maybe I find alfalfa useless and LLMs provide a lot of value, maybe you do the opposite. Short of violence, either immediate or delegated, the best way to resolve this conflict is to incorporate costs for resources like water appropriately (accounting for externalities), and let market participants bid to determine who gets access to it.

Datacenter owners are highly likely to find this a reasonable process, because they believe that what they are investing in and operating will provide a great deal of value on the market. And other industries (like your alfalfa) are far more likely to throw a fit at this process.

Or if you think alfalfa farms have fundamental, deep-seated water rights that don't extend to data centers, then data center operators can just purchase water from alfalfa farmers, and everyone involved will eagerly make that trade.

The hysteria around water usage rests on people not knowing the scale of industrial civilization. First thing to do is compare any estimate of data center water usage with the water usage of almond farming. Or, if you want to focus on individual consumer choices, the water footprint of eating a hamburger.

There's no statement one way or another about should in my comment; and, for what it's worth, my ideal would be an immediate global pause in AI research and development.

But the different terms imply different mental models of what LLMs are and can do. If you take two people, one who thinks of them as "artificial intelligence" and one as "stochastic parrots" (with all the implicit context and connotations of the individual words composing them), what mental model would have led to better predictions of LLMs' future circa 2020?

The "stochastic parrots" phrase is very dangerous in that frame. People read far more into what capabilities it implies are (im)possible than the narrow technical description the authors originally argued for. If all they are is spicy autocomplete or pastiche plagiarizers, there's nothing serious to worry about. And when an opposition gets stuck in a trough that mindlessly dismisses their future capabilities out of hand because of a bad mental model, it renders them ineffective at preventing the worst outcomes.

Which frame inspires a more productive research program? Which has better predicted the trajectory of capabilities over the past five years?

Back then, it sounded like making right-handed life was decades away. But with this work, couldn't you just as easily make this kind of synthetic cell right-handed?