It seems that 19th-century tile makers sometimes used clay tiles as memo pads. Cawarden Reclaim (who found that dismissal notice) stumbled upon another tile that contains an order list:
This is an old handmade pan tile, made on 21st July 1802. [...] Reading the information, it seems that this is an order list for the maker. With no paper and pen to hand, the words were jotted down on a freshly formed clay tile, which was waiting to be dried out and fired. This tile was fired, used on a roof, and remained in place for approximately 220 years.
Does the math check out, though? For reference, Nvidia added $17B of non-marketable securities in 2025, and its total revenue was $215B. So, assuming all these equity investments went to GPU purchases, they only amount to 8% of its annual revenue.
Since these firms (neoclouds and AI startups) will get consolidated if Nvidia puts too much money into them, I assume the majority of their funds are coming from elsewhere.
> Corporate development-led AI equity investments alone exceeded $40 billion in just the first four months of 2026
The accusation in the headline is that this activity contributed to the general shortage of GPUs which would, if true, increase their pricing power and profit margins even on chips sold to other customers. So I don't think it's a direct comparison to their equity deals.
Is this the only reason Nvidia made money? No, obviously. But is it adding fuel to the fire? If there is a bubble will this make the fallout worse? Seems likely.
More importantly this suggests a second plausible reason the board might have wanted to wrest control of the company away from him. (The other hypothetical one being that he was doing illegal stuff.)
According to Gemini, Nvidia added $43 billion in direct equity stakes,$36 billion in AI compute partnerships, and $100+ billion in loan lease guarantees, which works out to $63 billion this year, $20 billion for the next 3 years, and $10 billion for the last 6 years based on it's assumptions.
So about a third, a sixth, and a twelth of their annual revenue respectively.
I think the larger concern is that both the revenue and assets/agreements depend on Nvidia maintaining sales, market share, and their share of capacity at TSMC/etc... If they run into trouble on any of those fronts, their revenue and assets could take hits from multiple directions.
I actually got a very different impression. He could publish this interesting result, exactly because he had been working on this project i.e., had an interesting dataset to feed into Claude and had the expertise to interpret its output.
Compare him to those who are burning tokens to generate cringeworthy AI slop on social media. The value of a tool depends on the skill of the user.
I’m very curious what the major bottleneck is here. I can only think of three possibilities: 1) model capabilities, 2) human researchers, or 3) available compute resources.
Given that frontier models are leapfrogging and massive infrastructure buildouts are ongoing, I don’t see why any of these three couldn’t be overcome.
Side note: I’m also curious why people (myself included) don’t seem to be getting much more intelligent as AI becomes widely available. As if humans are bound to the earth, while AI is leaping for Mars.
Apparently, when Japanese farmland is abandoned, the landscape generally transitions into a dense and dark forest. So species that prefer open fields or ponds tend to disappear, because humans stop cutting trees and irrigating rice fields.
Overall, I got a faint impression from the article that it’s like saying “abandoning a zoo reduces biodiversity”. Strictly yes, but these species were there only because humans constantly intervened in nature.
Yes. Reading the reasoning traces in the original post, I got the impression that LLM weights are generally not interpretable even with frontier models. It's impressive that Astra can perform a small surgery on Qwen to remove its bottleneck, but it's not the same thing as being capable of fully understanding its weights.
Maybe it's bit like neurosurgery: We humans can treat certain pathological cases in our brains, but aren't exactly sure how our brains really work.
Why do we die of old age in the first place? My money's on it being a strongly positive evolutionary force for a species. Almost all organisms age. It would make sense to me that the ones that tried not aging are gone now, likely because they were much much worse at adapting to changing conditions than ones that could reroll the dice every generation. I would further guess that the healthy oldsters filled up their ecological niche and blocked the young individuals from gaining a foothold.
But that argument doesn't prevent limited mechanisms, and conditional mechanisms in particular would still be rewarded. Conditional mechanisms provide much more short term adaptability to changing conditions, without damaging the long term species-wide adaptation.
There are intriguing parallels with centrally planned vs market economies.
Natural selection doesn't work like that. Having two offspring is good. Having three offspring is even better. Having ten offspring is really great. More than than ten is greater still...
There's a common theory that for mice, they're unlikely to survive for many years because of predation, so there's not much benefit to preventing aging in later years. Small benefit means any counterbalancing disadvantages (lower vigour?) could outweigh the anti-aging benefit, or that the force of selection is insufficient to overcome random mutations that degrade the anti-aging mechanism, but "that's good enough, no need for more" is not how evolution works.
Do you think AI is as good at writing bug-free code as finding bugs? If that's the case, this should be a one-time event during human coders -> AI transition. It's like catching a lot of bugs during a handover review.
What categories of bugs would you say we eliminated going from macro assemblers to compilers?
(Operating systems, by the way, are the exception to the rule for assembly: the Linux kernel today has an x86/x64 assembly source component around the same size as MSDOS 5.0: ~250kloc of it. Windows is smaller, based on the NT4 kernel leak at around 70kloc. As a percent of total they're much less, of course, but they also do much more. And have correspondingly more bugs, too.)
> What categories of bugs would you say we eliminated going from macro assemblers to compilers?
Entire classes of 'safety' bugs that just don't exist in modern compilers.
E.g.: forgetting to "POP" something you "PUSH"-ed, especially when it was a conditional allocation for a temporary.
Similarly, type errors were much more common, especially signed/unsigned, or accidentally overwriting things in scenarios that just don't occur even with ordinary C code. Function args were easy to push "correctly enough" that it "worked", but not really.
The code it writes in the first attempt is at least as buggy as a human's. So the question is more whether it is as good at spotting bugs in its own code as it is at spotting bugs in human code
I'm tempted to say "no, it isn't". But I don't have numbers to back that up
A. The capability is just not there. OpenAI and Anthropic can produce a brute-force Lean proof, but their models aren’t capable enough to digest the proof as human mathematicians do.
B. They haven’t seriously tried it yet. They prioritized grabbing the land first, and have not spent much effort to polish things up.
C. It’s fundamentally impossible. These theorems are provable only in a very ugly manner. This is the reason why human mathematicians haven’t been able to crack them before AI in the first place.
As a programmer, what I’m seeing is (C). Many recent cyber hacks are products of AI trying & combining every possible tricks, and I don’t think there is any elegant short path that human engineers can replicate.
The four-color theorem appears to be case C. That was the first really long computer-assisted proof. No one has since found a concise way to do that proof, and it's been half a century now.
I'm not sure how Terence Tao's "digestion" concept holds up if the upcoming AI math results are all like that. It would be almost like humans are chewing bones under the table, while LLMs are feasting.
D. They seek disruption as way to market dominance and worldly power. It's not unfortunate, in their minds, but the means to their ends.
Disruption has been a SV buzzword for a long time now. It's not a byproduct but a strategy for your new technology product. SV has long perfected disruption and the people behind the AI products are masters at it.
How can so many who have seen it (and used it) for so long be so taken in by this well-worn tactic. Over and over, they choose what is most disruptive. They could have dealt with the mathematical output in many ways, including working with mathematicians to polish and publish; you see what OpenAI chose.
Similarly was the recent events where AI software 'broke containment' and hacked others. Obviously they could have controlled it if they valued it; they got the result they wanted. Notice how Anthropic (or were they first and OpenAI followed?) urgently publicized that their product, too, 'broke containment'. 'We're disruptive too!'
For A, a common misconception is that models write out lean proofs from the start. They write/develop the Natural Language proofs, and then translate (sometimes not even the same model) to lean afterwards.
This is correct. Out of the $80.7B loss, $72.4B was due to changes in the fair value of their convertible notes (and similar financial arrangements).
Going a bit meta, I'm curious why we don't observe a faster convergence of opinions here. I assume that HN users are well-versed in AI tech, and the current generation of LLMs will happily break down the relevant information in Form S-1. Nevertheless, we see a chaotic storm of opinions, with wildly varying S/N ratios.
Is it a limit of current LLMs, or is the bottleneck perhaps human nature?
Yes but those convertible notes only exist to grant them access to compute they would otherwise need to pay fair market value for.
If I drive a taxi making $200/day and need $150/day of gas.. but convince Shell to take a convertible note instead of $150.. Do I make a) $200/day profit, b) $50/day profit, or c) $200/day - (change_in_value[convertible_note]) in profit?
Some are arguing that C is somehow fake, but its obviously more real than A!
Another big component of the $80B loss was $3B in stock-based compensation. This line item is very expense-like (“instead of paying your drivers in cash, give them shares of your taxi company”), and Anthropic also adjusts it away when reporting their adjusted net income. Without this adjustment, their non-GAAP operating income was negative in 2026 1H.
Overall, I feel like a man in the parable of The Elephant in the Dark while writing this. Not only is my sight limited, but this elephant (Anthropic) is dancing wildly!
Human nature. Obviously we know some things are just true. It's common sense! Why would I invite a non-deterministic hallucination machine that can't think to challenge my world view with something inconvenient like facts?
People are emotionally invested in a deflation and slowdown in AI because they feel threatened by AI. They'll cling to reeds of hope and wilfully misread financial stats to make themselves feel better.
I imagine that it was like writing a dismissal note on the white board after a heated office argument.
reply