
some favorite books: docs.google.com/document/d/1… my substack: creon.substack.com regenerative agriculture: docs.google.com/document/d/1… health & life extension: docs.google.com/document/d/1…

NASA for 32 years Dir. of R&D @Planet Labs since 2015 -- Science is poetry written about nature in the language of mathematics. -- Tweets are PERSONAL views

some favorite books: docs.google.com/document/d/1… my substack: creon.substack.com regenerative agriculture: docs.google.com/document/d/1… health & life extension: docs.google.com/document/d/1…

The op against Jensen suggests we've got a classic "Baptist/bootlegger alliance": The Baptists are the decels: "Prohibit AI -especially open weights." The bootleggers are... Anthropic: they want to be the only ones selling AI. For now, they're aligned en.wikipedia.org/wiki/Bootle…



All this fuss recently about editing an LLM's weights directly to add or remove facts: Chris Hay basically had it seven months ago, using a very nice interface. (link in thread below)

Just another reminder of the vastness of high-dimensional space: in D dimensions there are D! permutations, each of which is one shadow on the wall of Plato's cave, representing the same underlying structure, up to relabeling.

Between progress in quantization, distillation, sub-quadratic attention, looping, the transformer itself, and future advances it will likely turn out - as usual - that algorithmic improvements eclipse the improvements due to hardware. The reverse bitter lesson, learned many times before.

If "neurons are clean" (one neuron per concept), and the platonic representation hypothesis (PRH) is true then the at least the final layer of different models trained on different modalities with different seeds and different downstream tasks should be identical up to a permutation ("neurons are clean"). A permutation is always an orthogonal transformation (the PRH).

If you're an AI enthusiast, an anime enthusiast, or just want an entertaining & thoughtful fictional epic focused on new technology, watch "Pantheon" on Netflix. You can thank me later. Similar ending as "Dancers at the End of Time" (obscure reference)

US federal budget (2025): $10.2 trillion US federal + state govt budgets (2025): ~$12 trillion SpaceX 2025 revenue: ~$19 billion So at present it is the government which is "hoarding" 99.8% of those resources.
All I'm saying is that if you had a tribe of say 100 people and 5 of those people hoarded 95% of all the resources while half the tribe starved the rest of the tribe would eventually nationalize SpaceX

The 5 people here are NOT "hoarding the resources". They're deploying the resources efficiently to make things people actually want: EVs, high-speed internet, products to your door same day, space launch... and creating jobs. Government seizure & redistribution destroys this.
All I'm saying is that if you had a tribe of say 100 people and 5 of those people hoarded 95% of all the resources while half the tribe starved the rest of the tribe would eventually nationalize SpaceX

The processes were on the same physical machine! No "air gap". It's an exploit from a time before AI agents: a cross-VM cache side-channel similar to Spectre (2018). This has a doom-hype (e.g. "Ban open weight models!") smell to it. And yes, we need better cyber-security.
OpenAI CEO Sam Altman: "I don't even know a dramatic enough word to use… models can find bugs that are outside of the normal threat model of what we think a hacker can do… hardware-level, insane stuff… I no longer think we're gonna be able to just patch all the bugs." It's over.




DM me here to get on the invite schedule: one online seminar per month: ~40 minutes + Q&A. No spam. Also DM me here if you 're seriously interested in presenting to the group. Here is a partial list of previous seminars with recordings & other info: foresight.org/seminar/space-… Here's the youtube channel (all the recordings) youtube.com/playlist?list=PL…