Carlos E. Perez

A proposed theory of physics reveals that the universe is a living process!

Reviewing this, I created it 11 months ago (time flies). Still very relevant.

A real challenge for software developers (and actually anyone) is how to reduce the compounding atrophy induced by cognitive surrender.

Without regulation of social media and AI, the US population will likely become dumber and dumber. The invisible hand of the market is not aligned with a smarter public.

A bold proposal for reality and the universe as a living process.

AI will remember Patrick Stewart as Jean-Luc Picard.

My favorite actor has passed away.😢😢

🧵 1/ Crypto got Proof of Work. AI just got Proof of Model. For the first time, an AI provider can prove which model actually answered you, with math instead of a promise. 👇 2/ The problem: every time you pay for a frontier model, you're taking it on faith. You can't verify whether that answer came from the model you paid for or from a cheaper stand-in. 3/ Enter FlashProve, from Anthropic, EPFL and MIT. Every response can come with a cryptographic receipt that says "this exact model produced this exact output." You can check it on a laptop CPU in about 17 seconds. It's been demonstrated on a 400B-parameter model. 4/ The twist: the receipt reveals nothing about the model. The weights stay secret, and so do the internals and which experts ran. Closed models can now be audited without being opened. That used to sound like a contradiction. 5/ Proof of Model doesn't depend on trusted hardware or chip-vendor enclaves. It rests on hash functions alone, and it's plausibly post-quantum secure. Even a provider whose own proving code is buggy can't forge a proof. 6/ Who should be nervous? • Routers quietly swapping in cheaper models • Resellers passing off quantized knockoffs as premium • Anyone whose margins depend on you never being able to check 7/ Who benefits? • Regulators: they can check that only approved models are served • Evaluators: they can confirm that the model they tested is the one deployed • Security teams: tampered servers fail verification • Nations: they can verify AI agreements without having to trust each other's chips 8/ The underrated use is proving a negative. A model can prove it spent a set token budget hunting for bugs and found none. "Trust us, it's secure" becomes something you can check. 9/ Now the honest part, because hype ruins good science: • Proving costs about 72–94× normal serving compute • Proofs cover only 1,024 tokens • It proves a quantized model with greedy decoding (always the top token), not the exact production output • It catches cheating after the fact; it doesn't block it 10/ But look at the trajectory. Earlier work reportedly took 14.3 hours to prove 1,000 tokens on this same model. FlashProve runs at about 110 tokens per second. At roughly 100× overhead, random spot checks are enough to make cheating a losing bet. 11/ One more detail: the C++ prover and verifier were generated by AI agents from a formal Lean specification. So AI is writing the proofs we'll use to verify AI. 12/ Proof of Work secured money. Proof of Model could secure intelligence.