Setting aside the disagreement, I was very interested to see the net pricing of the discovery:
> All told, the week-long effort consumed 300 billion output tokens — $22.5 million worth of compute, if charged at current Astra rates.
> The Navier-Stokes existence and smoothness problem is one of the seven Millennium Prize problems — a set of major unsolved math problems, each carrying a $1 million bounty
I know openAI isn't solving these problems in order to make profit, but it's interesting to guess how close we are to these things becoming profitable. Eg, if you think their public pricing for Astra is ~2x as expensive as their internal price, then they lost ~10M on net for this proof. That's not profitable, but it is much better than I would have expected, which is exciting for the other Millennium prize problems! Of course the fundamental approach (which they may have plagiarized from Buckmaster and Alpoge) might have added cost to that as well. Nonetheless, I wouldn't be too surprised if they're all solved within the next 3 years!
Wow I wonder how much of that differential is them operating at an extreme loss, and how much is the relative cost of their new models vs old. If a significant portion of that is net losses, I'm very concerned for their business model :|
I don't think I'm discounting the human breakthroughs, the cost of employees, or the previous attempts. All of that is significant work, and has lead us to a scenario where the marginal cost to solve one millenium prize problem appears to be ~$10M. I feel like you're taking a very "to bake an apple pie you must first create the universe" perspective on this; I make no claim about how much money was spent in the past, on whether it was well invested, etc. Regardless of how we feel about it, it's already been done, right?
Ya, setting aside any judgement about what's "right", this strategy doesn't seem profitable. There's no real moat between one model/provider to another, switching is relatively easy. I don't understand how this is supposed to work for them.
It strikes me as similar to UPS / USPS / Fedex -- everyone uses the mail, and they mostly use whichever is cheapest for their requirements. I don't think there's much loyalty to specific services, and people are happy to switch between the options
just dont fall in love with goofy memory style features and its likely gonna be fairly easy to just plug and play whatever model for a ton of use cases.
I dont see a lot of love for weird memory like features on HN, but on provider subreddits its constantly talked about.
Every project I have tends toward coding agents over time partly for this reason - if nothing else, I want have both control, visibility and portability over memory and a log of the reasoning that went into the current state of things.
> Degree is over-valued as a life goal in a time where we no longer really value tertiary education the way we should
What does “the way that we should” mean here? Are you disagreeing with the free market valuation of a good? If not the equilibrium created by supply and demand, what do you propose instead?
My family is Persian and I grew up in Turkey. You definitely develop an appreciation for carpets, and it's sooooooo heartbreaking to see people use art as an entryway rug, simply because they don't know :(
> I am confused at how Waymo engineering can be so robust as to yield an astonishingly good safety record, and yet so slapdash as to happily drive into deep water.
I feel this is actually somewhat straightforward. I assume deep water on roadways is not commonly in the training set, because frankly it isn't common in real life, and when it is common people do not drive and do not gather that training data. As a result the proper response has not adequately been beaten into the models. There are probably also challenges of world-sensing, since water can act as a mirror, and maybe other complications. So waymos are bad at handling deep water on roadways. However, deep water on roadways is also not common in the areas where waymos are deployed. As a result, waymo's have a great safety record, and at the same time they make mistakes that are obvious to a human.
A common criticism of AI discourse is that people act as if LLM's "think". I don't want to be a vocabulary purist, but I suspect that's related to the astonishment here -- the Waymo doesn't know what flooding is, it doesn't fear drowning, it doesn't think. So unless it's been repeatedly trained, or a special case has been hard coded by manual effort, it doesn't know that flooded roadways are dangerous.
I have made a lot of assumptions here, and I don't truthfully know what the training data looks like. Feel free to push back if you think my assumptions are wrong. I'd especially be interested if somebody can show that water on roadways _is_ in the training data
People drive into deep water all the time - some states specifically have laws making them financially liable for the cost of rescue because it’s such a stupid thing to do. But still, they do it.
I think the fact that there are laws about it is not good evidence that it should be in the training data. Laws often cover weird edge cases, and if an edge case happens 50 times in 100 years, there's likely a law covering it. At the same time, that's probably not enough occurrences for it to naturally end up in a dataset -- the edge case would probably need to be intentionally sought out. I'm not saying that driving-in-deep water only happens 50 times in 100 years, it's certainly more common than that, I'm just saying that despite laws on the topic, it may still be too rare to be well represented in training data. For example, in real life I've only seen a car drive into deep water once. Even if we include recordings that I've seen, that would maybe bring it up to 20?
Without a good way of getting it into the dataset via simulation, or a more speculative approach (world models?), edge cases and unusual circumstances could cause failures.
I think my broader point is people don't do so well in unusual circumstances either: blizzards, heavy rain, dust storms, etc. They'll hydroplane, drive into stopped traffic, etc. We need to decide if we'll hold self-driving cars to some unreasonable standard of perfection or accept them once they are X safer than a human benchmark, even if they still have Y rate of failure per million miles.
My thinking here is that Waymo has logged hundreds of millions of miles at this point (and even more in sim), and there are a lot of nines in their safety record. So even the rare edge cases should have come up. You make a good point that the deep water scenario may be not only rare, but also under-represented in the training data. On the other hand, you'd think they would have thought to explicitly train for it; it's not that obscure of a possibility.
Ya, this does strike me as pretty high on the list of the thing I'd intentionally seek out if I was doing this job, along with object-in-the-road, icy road, tornado/hurricane, wildfire, and perhaps hail.
Upon further thought, I realized this should be counted as a correct prediction for Zitron! Zitron did indeed “keep writing this stuff until…proven wrong”. Zitron’s prediction says nothing about what happens after being proven wrong.
On the other hand, this seems like a super interesting path for an (unglamorous) PhD -- replicate the N most important findings in a field, subject to some restrictions (maybe focus on the ones that are "easy", or haven't been replicated yet, or the ones that could be patched together to make an acceptable PhD thesis). This sort of stuff really appeals to me, I think I might end up doing something like that in 5-10 years.
I share your enthusiasm for this. Creating a map of the most influential research across all the fields to identify the highest priority work, any work replicating that work or validating that work by building from it would be a good start. Ideally funded by governments, but institutions, companies, and individuals could put up bounties for replicating particular research.
I think you're right that this will be the most common objection, but wrong that it will be exceedingly difficult to find an advisor that would support this. I assume that if you _can_ overturn some foundational result, you won't face criticism over novelty (potentially you can argue that overturning important accepted results is inherently novel). Then the rest is just math:
Assume that n% of foundational studies are demonstrably incorrect if replication were attempted. Then if you replicate X foundational studies, there's a 1 - (1-n/100)^X chance that you'll show one foundational study to be incorrect. Then I think you just need to find an advisor who believes the incorrectness rate n, and the replication attempts X is high enough for you to revert a foundational study over the course of your PhD.
> > The basic shaping is estimated to have taken around 200 hours, and in total the recreation likely took more than 370 hours.
> Nearly a year of "full time" work at 40hrs/week went into this
Sorry, I think you may be mistaken? From the section you quote, it looks like it's about 370 hours, or 9 weeks of full time effort. That's still a lot, but certainly not nearly a year.
I can also imagine that someone with a lot of practice might be significantly faster than some unskilled researchers.
I also don't think it rules out it being something for a child. It's not so rare to do big projects for you kids even today, over many evenings and weekends. And when you don't have any books to read or screens to watch...
> All told, the week-long effort consumed 300 billion output tokens — $22.5 million worth of compute, if charged at current Astra rates.
> The Navier-Stokes existence and smoothness problem is one of the seven Millennium Prize problems — a set of major unsolved math problems, each carrying a $1 million bounty
I know openAI isn't solving these problems in order to make profit, but it's interesting to guess how close we are to these things becoming profitable. Eg, if you think their public pricing for Astra is ~2x as expensive as their internal price, then they lost ~10M on net for this proof. That's not profitable, but it is much better than I would have expected, which is exciting for the other Millennium prize problems! Of course the fundamental approach (which they may have plagiarized from Buckmaster and Alpoge) might have added cost to that as well. Nonetheless, I wouldn't be too surprised if they're all solved within the next 3 years!
reply