Confirmed — he posted this Sep 17, second AI security thread in 10 days.$G $UNI
*Buterin vs AI doom thesis:*
1. *Sep 7:* Rejected Liron Shapira's claim AI will crash BTC 50% in 2 years. Said BTC doesn't need social consensus for network-layer fixes — client/mining pool patches enough, chance of breaking
2. *Sep 17:* New long-form post on verification:
- *Core idea:* "AI gives you ability to write large volumes of code at cost of accuracy, and formal verification gives you back accuracy" — sloppy AI code is transient problem, because verified proof is permanent, bug is one-time
- *Cypherpunk defense:* Defenders have advantage. Bleak future where AI makes open-source indefensible breaks internet freedom ethos
- *What he's building:* Arklib = formally verified STARKs, evm-asm = EVM made verifiable, plus live examples — AI agents already found *CVE-2026-34219* remote crash in libp2p Gossipsub, project http://better.codes integrates AI + machine-checkable proofs
- *90% net worth in crypto* — says holding crypto = bet defense keeps pace with offense
*Limitations he admits:*
- Spec risk: proof worthless if spec wrong, definitions can be 1000+ lines
- Hardware can't be formally verified like software, plus compiler, side-channel, governance, oracle failures
*Offense side real though:* Anthropic Sep threat report — observed AI automating vuln research, exploit dev, multi-target campaigns; 2,300 AI-found vulns across 392 OSS projects, 832 accounts using AI across kill chain, some pipelines producing zero-days validated in test envs
So his call: AI doesn't make attacks unstoppable — it makes *verification cheap enough to shrink trusted core*. Not panic, but "keep secure core small, even shrink it."#HKMAPlansWholesaleCBDCByYearEnd #BOJHikesRatesTo31YearHigh #ChainlinkPowersBottomlinePaymentPlatform
*Buterin vs AI doom thesis:*
1. *Sep 7:* Rejected Liron Shapira's claim AI will crash BTC 50% in 2 years. Said BTC doesn't need social consensus for network-layer fixes — client/mining pool patches enough, chance of breaking
2. *Sep 17:* New long-form post on verification:
- *Core idea:* "AI gives you ability to write large volumes of code at cost of accuracy, and formal verification gives you back accuracy" — sloppy AI code is transient problem, because verified proof is permanent, bug is one-time
- *Cypherpunk defense:* Defenders have advantage. Bleak future where AI makes open-source indefensible breaks internet freedom ethos
- *What he's building:* Arklib = formally verified STARKs, evm-asm = EVM made verifiable, plus live examples — AI agents already found *CVE-2026-34219* remote crash in libp2p Gossipsub, project http://better.codes integrates AI + machine-checkable proofs
- *90% net worth in crypto* — says holding crypto = bet defense keeps pace with offense
*Limitations he admits:*
- Spec risk: proof worthless if spec wrong, definitions can be 1000+ lines
- Hardware can't be formally verified like software, plus compiler, side-channel, governance, oracle failures
*Offense side real though:* Anthropic Sep threat report — observed AI automating vuln research, exploit dev, multi-target campaigns; 2,300 AI-found vulns across 392 OSS projects, 832 accounts using AI across kill chain, some pipelines producing zero-days validated in test envs
So his call: AI doesn't make attacks unstoppable — it makes *verification cheap enough to shrink trusted core*. Not panic, but "keep secure core small, even shrink it."#HKMAPlansWholesaleCBDCByYearEnd #BOJHikesRatesTo31YearHigh #ChainlinkPowersBottomlinePaymentPlatform
