Clive Chan leaves OpenAI's custom silicon program to join Anthropic's in-house chip initiative

@tenobrus @bubbleboi said Intel, which seems more likely given Ant's natsec focus

@tenobrus Mass vertical integration is inefficient. But it does happen for defence critical industries.

@tenobrus > these *may* only be for inference, but i would find this unlikely

Reach for the stars land on mars

> if it's not already happening then they'll train for it

Call me cynical but I don't think they have the talent to do it, at least not better than Nvidia

@tenobrus this has long been my "short nvda" thesis, even though i do own some. sufficiently powerful agi nukes their moat

@tenobrus If we end up in a world where NVDA gets nuked to the tune of trillions by the prospect of a chip-designing superintelligence while Anthropic and OAI are still outside SPY I might literally burst into ecstatic sobs 🥹

@tenobrus Surprised you're not also mentioning ASML here, who from what I know are the true bottleneck on the whole stack

@tenobrus I feel like Nvidia’s moat is supply chain at this point. Getting hundreds of suppliers together to make Blackwell racks and make Ultra updates a year after is no joke.

@flawedaxioms @tenobrus Part of asml is based in USA and necessary

It's only a matter of time, maybe 10 years, before people start to form lifelong or near lifelong allegiances to a small number of "Corpos" or TransNats where the social hierarchy is built around progressively increased access to dense SOTA models.

It will come down to allegiance to the group being the determining factor of what quality model you're allowed to use.

@tenobrus Tsmc won't get any more value out of it than a Nvidia chip tho ? It's just changing customers, the only winners are labs, who will stay private until ipo at 5T, nobody quitting underclass

@tenobrus the anxiety of leaving nvda for one of them....the ups and downs of controversy

@tenobrus As cool as AI is as an achievement, EUV chip fabrication is the most mindfuck thing I’ve ever read up on and it’s truly insane that humans are able to manufacture at scale with such a process that powers ordinary devices in my home/pocket. Long TSMC+INTC.

@tenobrus You skipped CUDA. Nvidia is a software company

@teortaxesTex software has this extremely unusual culture of giving away all its knowledge for free

to my understanding, hardware and the likes lock their learnings behind NDAs, tacit wissenschaft, and other obstructions.

Unfortunately this will prove to strongly resist first-principles AI

@tenobrus Samsung will eventually get their shit together too

@flawedaxioms @tenobrus Well the thing about bottlenecks is that it’s not a bottleneck until it matters.