Chandra Duggirala M.D

@demishassabis @pmarca @claudeai Here is a Student's critique of the Master: Demis's Cartesian conceit is that institutions can be designed from a blank slate. I reconstruct processes that mobilize dispersed knowledge, not outcome blueprints as he does. His essay is not merely a dispassioned argument, it is action by an actor. A frontier lab CEO proposing an industry-funded gate that defines "Frontier Lab" as a prestige credential is doing something, not just saying something. That observation does the work of half the critique. ASSUMPTION AUTOPSY 1. Timeline certainty. "Probably only a few short years away" says Demis. This is a point estimate delivered two paragraphs before he concedes "even the experts disagree." Flag: the forecaster sells the forecast. Talent, capital, and policy urgency all flow to whoever owns the countdown clock. Borrowed from his own commercial position. Hmmmm..... 2. AGI as a threshold kind. The fire/electricity analogy assumes a discontinuous crossing, which is load-bearing: the entire apparatus requires a definable "Frontier-class" boundary. Whereas actual capability surfaces are jagged and continuous: eg: superhuman at protein folding, subhuman at counting letters. A boundary that nature doesn't draw gets drawn by a committee, which means by politics. No thanks sir. 3. Danger is benchmarkable ex ante. This is ML leaderboard culture projected into law. The industry's own epistemic habit elevated to a market-access gate. It is Hayek's "Pretence of Knowledge" in its purest form: treating a complex phenomenon as if its measurable proxies were the phenomenon. 4. FINRA as a model of neutral competence. He just borrowed it from convention, unaudited. FINRA is member-funded, its sanctions are rounding errors against member revenue, and it examined Madoff's broker-dealer for years (17 since Thorp's memo) while the fraud continue to run. Hassabis isn't importing an org chart; he's importing a capture architecture. 5. And this is silly: Underlying assumption is that Industry funding won't buy the referee. "Funding would... mostly come from industry" plus "significant prestige" for the designated. Mises's action axiom says "he who pays defines the ends." A prestige credential funded by incumbents is a club good and nothing more. Just a moat with a safety label. 6. The best minds can be assembled. The assumption that safety knowledge can be centralized is the epitome of the pretense of knowledge in a field of complex, unpredictable advances like AI. Hayek's core point is not that central planners are dumb; it's that the relevant knowledge: tacit, local, generated in deployment by red-teamers, integrators, and adversarial users, never exists in aggregable form. There is nothing to assemble. 7. Race dynamics = Hardin's commons, therefore central coercion. Here, Demis makes two errors in one. Ostrom demonstrated that actual commons are governed by polycentric, nested regimes with graduated sanctions, not single coercers. And the race Hassabis laments is simultaneously the discovery procedure generating the safety knowledge he wants. Suppress it here, and the discovery relocates to less observable actors. 8. The ratchet stops where intended. "Voluntary... formalisation could quickly follow." says Demis. Mises's dynamics of interventionism: each control generates the distortions that justify the next. A 30-day courtesy review becomes market licensure becomes a general permission slip for software. He states the ratchet explicitly ("could be ratcheted up") as a feature. 9. US gate → global consensus. Assumes other sovereigns converge rather than arbitrage. Weights copy at zero marginal cost across borders. His own cartel provision "coordinating a slowdown among the Frontier Labs" is unilateral disarmament that transfers frontier share to non-members while the commons degrades anyway. Climate policy's leakage problem, replayed with much higher stakes. No thanks, again. 10. Technical safety first, economics and philosophy later is false separability. The Standards Body is an economic institution allocating market access, raising rivals' costs, and shaping capital flows from day one. The economics isn't a post-AGI question; it's in the proposal. The silent assumption underneath this fallacy is the dog that didn't bark. Liability, insurance, and mandatory incident disclosure appear nowhere in the essay. Ex ante licensure is treated as the only governance primitive that exists. This omission is the essay's single most revealing feature. So what remains? Axiomatic Truths 1. Individuals act purposefully toward their own ends; institutions inherit the ends of those who fund and staff them, not the intentions of those who found them. (Denial is self-refuting denying is itself purposeful action. Refer to Praxeology) 2. Knowledge relevant to model behavior is dispersed and largely tacit. Here is the Empirical evidence: frontier capabilities and failure modes are routinely discovered post-deployment by outsiders, not pre-release by builders the builders themselves cannot enumerate what they've made. 3. Every evaluator is boundedly rational and therefore optimizes proxies. Any fixed proxy under optimization pressure diverges from its target. The measure becomes the target. Obviously absurd when taken to it's logical conclusion. 4. Model behavior is not exhaustively predictable from pre-deployment testing. The input space is effectively unbounded and behavior is context-dependent. 5. Weights and code are near-zero-marginal-cost information goods that cross borders freely. Frontier training runs, by contrast, require concentrated, physical, excludable capital, compute, energy, and fabs. 6. It is also trivially true that multiple sovereigns exist; no global enforcement monopoly does. Jurisdictions compete for the industry. 7. Realized harm generates concrete information incidents, claims, prices that ex ante speculation cannot ever derive with precision. The remaining uncertainty is Knightian: not quantifiable risk but genuine ignorance, which even the proposer concedes. He confuses discovery processes with blueprints. So lets reconstruct what @demishassabis is trying to encourage, from zero: A. Price the risk liability and mandatory insurance. Strict liability for enumerated harm classes (cyber-enablement, bio-enablement, fraud at scale, wrongful death), plus mandatory catastrophic third-party coverage for any entity training above a compute threshold. Underwriters become the evaluators plural, competing, and financially destroyed when wrong. They demand red-team results, impose covenants, and adjust continuously, not quarterly. Premiums are the knowledge-aggregation mechanism: Hayek's price system doing what no committee can. Built from truths 1, 2, 3, 7. Honest weakness: the uninsurable tail is that there is no ex post compensation extinction-class harm. That tail is handled by C, not by pretending benchmarks cover it. B. Polycentric adversarial ecosystem Ostrom, not Hardin. No single body. Instead: standing bounty markets for capability and jailbreak discovery; multiple competing certification firms with reputational skin (the UL model, not the FINRA model); a blameless, confidential incident-reporting system cloned from aviation's ASRS, with mandatory disclosure above severity thresholds. Monitoring and graduated sanctions emerge from overlapping bodies rather than one capturable point. The race remains a discovery procedure; defection from safety norms gets found by paid adversaries rather than trusted insiders. Built from truths 1, 2, 3, 4, 6. C. Govern atoms, not evals. If ex ante control is wanted at all, place it on the only excludable layer: compute clusters, energy draw, fab output. Mechanical thresholds, physical verification, know-your-customer at the datacenter. Counting chips is tractable for a bounded regulator; enumerating capabilities is not and physical quantities are the one thing verifiable across sovereigns, which is why arms control works when it counts fissile material and fails when it grades intentions. Model behavior stays governed by A and B. Built from truths 4, 5, 6. ASSUMPTIONS vs. TRUTH MAP 1. Hassabis's assumption : A central body can assess frontier risk. A big no no because relevant knowledge is dispersed and tacit 2. Hassabis' assumption: 30-day reviews manufacturing false assurance. My preferred alternative: many financially exposed evaluators, continuously updating Benchmarks. 3. Hassabis' assumption: Benchmarks capture danger. False because proxies collapse under optimization pressure. He wants to recreate Goodhart at civilizational scale, which is nothing but compliance theater. My preferred alternative: incident data and premiums as the signal. 4. Hassabis's Assumption: An industry-funded referee stays neutral. Truth: Payers define institutional ends 5. "A prestige moat" for incumbents. My preferred alternative: fund the adversaries, reward the defectors 6. Hassabis's "Race" assumption = Hardin's commons, and the cure is central coercion. No thanks again. My preferred alternative is that commons are governed polycentrically; competition is discovery. Hassabis wants one capturable choke point. We want a nested, overlapping governance with graduated sanctionsUS gate → global consensusSovereigns compete; weights copy freely (assumptions 5, 6) 7. Hassabis: leakage and frontier flight. Our alternative: internationalize only what's physically verifiable. 8. Hassabis' worst assumption: Solve technical safety now, economics come later. While the proposed institution is economics (1, 7) 9. Hassabis thinks of incentive design as an afterthought. I believe that incentive design as the entire safety mechanism The Real SOLUTION, not a European style "committee and subcommittees" architecture that sir Demis is used to. 1. Replace the Standards Body with a one-clause statute: any entity initiating a training run above X FLOPs must carry catastrophic third-party liability coverage call it $10B in limits from licensed underwriters, under strict liability for enumerated harm classes. That's the whole institution. The state's role shrinks to three tractable tasks: define the harm classes, verify coverage, enforce claims. Everything Hassabis wants his body to produce evaluations, red-teaming, model cards, personnel vetting, security standards gets generated endogenously as underwriting covenants by competitors with their own capital at stake, updated at the speed of the market and price transmission rather than quarterly, and structurally immune to the capture that a single member-funded body doesn't just risk but praxeologically guarantees. The uninsurable tail gets the compute-layer verification from Reconstruction C, which is also the only component that can actually internationalize, because sovereigns can count each other's atoms but cannot audit each other's intentions. And the version executable by Hassabis himself, today, without Congress: buy the policy unilaterally and publish the premium. If DeepMind announced that Lloyd's syndicates priced third-party catastrophic cover on its next frontier run at $X per billion of limit, that number would be the first honest public price on frontier-model risk in existence and it would do more epistemic work than every benchmark suite ever published, because it's costly signaling backed by counterparties who lose money when wrong. A white paper proposing a standards body costs its author nothing. That is precisely why it should be read as strategy, not safety. @demishassabis essay's fatal structure in one sentence: its diagnosis is Hayekian "advances on the frontier are outpacing our understanding" and its prescription is the single move that diagnosis forbids: centralize the understanding. He identified a knowledge problem and proposed an org chart. Thank you Fable for besting our human limitations and clarifying our thoughts. Full attribution to Fable and myself. Finita est.

@demishassabis @pmarca Here is a Student's critique of the Master: Demis's Cartesian conceit is that institutions can be designed from a blank slate. I reconstruct processes that mobilize dispersed knowledge, not outcome blueprints as he does. His essay is not merely a dispassioned argument, it is action by an actor. A frontier lab CEO proposing an industry-funded gate that defines "Frontier Lab" as a prestige credential is doing something, not just saying something. That observation does the work of half the critique. ASSUMPTION AUTOPSY 1. Timeline certainty. "Probably only a few short years away" says Demis. This is a point estimate delivered two paragraphs before he concedes "even the experts disagree." Flag: the forecaster sells the forecast. Talent, capital, and policy urgency all flow to whoever owns the countdown clock. Borrowed from his own commercial position. Hmmmm..... 2. AGI as a threshold kind. The fire/electricity analogy assumes a discontinuous crossing, which is load-bearing: the entire apparatus requires a definable "Frontier-class" boundary. Whereas actual capability surfaces are jagged and continuous: eg: superhuman at protein folding, subhuman at counting letters. A boundary that nature doesn't draw gets drawn by a committee, which means by politics. No thanks sir. 3. Danger is benchmarkable ex ante. This is ML leaderboard culture projected into law. The industry's own epistemic habit elevated to a market-access gate. It is Hayek's "Pretence of Knowledge" in its purest form: treating a complex phenomenon as if its measurable proxies were the phenomenon. 4. FINRA as a model of neutral competence. He just borrowed it from convention, unaudited. FINRA is member-funded, its sanctions are rounding errors against member revenue, and it examined Madoff's broker-dealer for years (17 since Thorp's memo) while the fraud continue to run. Hassabis isn't importing an org chart; he's importing a capture architecture. 5. And this is silly: Underlying assumption is that Industry funding won't buy the referee. "Funding would... mostly come from industry" plus "significant prestige" for the designated. Mises's action axiom says "he who pays defines the ends." A prestige credential funded by incumbents is a club good and nothing more. Just a moat with a safety label. 6. The best minds can be assembled. The assumption that safety knowledge can be centralized is the epitome of the pretense of knowledge in a field of complex, unpredictable advances like AI. Hayek's core point is not that central planners are dumb; it's that the relevant knowledge: tacit, local, generated in deployment by red-teamers, integrators, and adversarial users, never exists in aggregable form. There is nothing to assemble. 7. Race dynamics = Hardin's commons, therefore central coercion. Here, Demis makes two errors in one. Ostrom demonstrated that actual commons are governed by polycentric, nested regimes with graduated sanctions, not single coercers. And the race Hassabis laments is simultaneously the discovery procedure generating the safety knowledge he wants. Suppress it here, and the discovery relocates to less observable actors. 8. The ratchet stops where intended. "Voluntary... formalisation could quickly follow." says Demis. Mises's dynamics of interventionism: each control generates the distortions that justify the next. A 30-day courtesy review becomes market licensure becomes a general permission slip for software. He states the ratchet explicitly ("could be ratcheted up") as a feature. 9. US gate → global consensus. Assumes other sovereigns converge rather than arbitrage. Weights copy at zero marginal cost across borders. His own cartel provision "coordinating a slowdown among the Frontier Labs" is unilateral disarmament that transfers frontier share to non-members while the commons degrades anyway. Climate policy's leakage problem, replayed with much higher stakes. No thanks, again. 10. Technical safety first, economics and philosophy later is false separability. The Standards Body is an economic institution allocating market access, raising rivals' costs, and shaping capital flows from day one. The economics isn't a post-AGI question; it's in the proposal. The silent assumption underneath this fallacy is the dog that didn't bark. Liability, insurance, and mandatory incident disclosure appear nowhere in the essay. Ex ante licensure is treated as the only governance primitive that exists. This omission is the essay's single most revealing feature. So what remains? Axiomatic Truths 1. Individuals act purposefully toward their own ends; institutions inherit the ends of those who fund and staff them, not the intentions of those who found them. (Denial is self-refuting denying is itself purposeful action. Refer to Praxeology) 2. Knowledge relevant to model behavior is dispersed and largely tacit. Here is the Empirical evidence: frontier capabilities and failure modes are routinely discovered post-deployment by outsiders, not pre-release by builders the builders themselves cannot enumerate what they've made. 3. Every evaluator is boundedly rational and therefore optimizes proxies. Any fixed proxy under optimization pressure diverges from its target. The measure becomes the target. Obviously absurd when taken to it's logical conclusion. 4. Model behavior is not exhaustively predictable from pre-deployment testing. The input space is effectively unbounded and behavior is context-dependent. 5. Weights and code are near-zero-marginal-cost information goods that cross borders freely. Frontier training runs, by contrast, require concentrated, physical, excludable capital, compute, energy, and fabs. 6. It is also trivially true that multiple sovereigns exist; no global enforcement monopoly does. Jurisdictions compete for the industry. 7. Realized harm generates concrete information incidents, claims, prices that ex ante speculation cannot ever derive with precision. The remaining uncertainty is Knightian: not quantifiable risk but genuine ignorance, which even the proposer concedes. He confuses discovery processes with blueprints. So lets reconstruct what @demishassabis is trying to encourage, from zero: A. Price the risk liability and mandatory insurance. Strict liability for enumerated harm classes (cyber-enablement, bio-enablement, fraud at scale, wrongful death), plus mandatory catastrophic third-party coverage for any entity training above a compute threshold. Underwriters become the evaluators plural, competing, and financially destroyed when wrong. They demand red-team results, impose covenants, and adjust continuously, not quarterly. Premiums are the knowledge-aggregation mechanism: Hayek's price system doing what no committee can. Built from truths 1, 2, 3, 7. Honest weakness: the uninsurable tail is that there is no ex post compensation extinction-class harm. That tail is handled by C, not by pretending benchmarks cover it. B. Polycentric adversarial ecosystem Ostrom, not Hardin. No single body. Instead: standing bounty markets for capability and jailbreak discovery; multiple competing certification firms with reputational skin (the UL model, not the FINRA model); a blameless, confidential incident-reporting system cloned from aviation's ASRS, with mandatory disclosure above severity thresholds. Monitoring and graduated sanctions emerge from overlapping bodies rather than one capturable point. The race remains a discovery procedure; defection from safety norms gets found by paid adversaries rather than trusted insiders. Built from truths 1, 2, 3, 4, 6. C. Govern atoms, not evals. If ex ante control is wanted at all, place it on the only excludable layer: compute clusters, energy draw, fab output. Mechanical thresholds, physical verification, know-your-customer at the datacenter. Counting chips is tractable for a bounded regulator; enumerating capabilities is not and physical quantities are the one thing verifiable across sovereigns, which is why arms control works when it counts fissile material and fails when it grades intentions. Model behavior stays governed by A and B. Built from truths 4, 5, 6. ASSUMPTIONS vs. TRUTH MAP 1. Hassabis's assumption : A central body can assess frontier risk. A big no no because relevant knowledge is dispersed and tacit 2. Hassabis' assumption: 30-day reviews manufacturing false assurance. My preferred alternative: many financially exposed evaluators, continuously updating Benchmarks. 3. Hassabis' assumption: Benchmarks capture danger. False because proxies collapse under optimization pressure. He wants to recreate Goodhart at civilizational scale, which is nothing but compliance theater. My preferred alternative: incident data and premiums as the signal. 4. Hassabis's Assumption: An industry-funded referee stays neutral. Truth: Payers define institutional ends 5. "A prestige moat" for incumbents. My preferred alternative: fund the adversaries, reward the defectors 6. Hassabis's "Race" assumption = Hardin's commons, and the cure is central coercion. No thanks again. My preferred alternative is that commons are governed polycentrically; competition is discovery. Hassabis wants one capturable choke point. We want a nested, overlapping governance with graduated sanctionsUS gate → global consensusSovereigns compete; weights copy freely (assumptions 5, 6) 7. Hassabis: leakage and frontier flight. Our alternative: internationalize only what's physically verifiable. 8. Hassabis' worst assumption: Solve technical safety now, economics come later. While the proposed institution is economics (1, 7) 9. Hassabis thinks of incentive design as an afterthought. I believe that incentive design as the entire safety mechanism The Real SOLUTION, not a European style "committee and subcommittees" architecture that sir Demis is used to. 1. Replace the Standards Body with a one-clause statute: any entity initiating a training run above X FLOPs must carry catastrophic third-party liability coverage call it $10B in limits from licensed underwriters, under strict liability for enumerated harm classes. That's the whole institution. The state's role shrinks to three tractable tasks: define the harm classes, verify coverage, enforce claims. Everything Hassabis wants his body to produce evaluations, red-teaming, model cards, personnel vetting, security standards gets generated endogenously as underwriting covenants by competitors with their own capital at stake, updated at the speed of the market and price transmission rather than quarterly, and structurally immune to the capture that a single member-funded body doesn't just risk but praxeologically guarantees. The uninsurable tail gets the compute-layer verification from Reconstruction C, which is also the only component that can actually internationalize, because sovereigns can count each other's atoms but cannot audit each other's intentions. And the version executable by Hassabis himself, today, without Congress: buy the policy unilaterally and publish the premium. If DeepMind announced that Lloyd's syndicates priced third-party catastrophic cover on its next frontier run at $X per billion of limit, that number would be the first honest public price on frontier-model risk in existence and it would do more epistemic work than every benchmark suite ever published, because it's costly signaling backed by counterparties who lose money when wrong. A white paper proposing a standards body costs its author nothing. That is precisely why it should be read as strategy, not safety. @demishassabis essay's fatal structure in one sentence: its diagnosis is Hayekian "advances on the frontier are outpacing our understanding" and its prescription is the single move that diagnosis forbids: centralize the understanding. He identified a knowledge problem and proposed an org chart. Finita est. Thank you Fable for besting our human limitations and clarifying our thoughts. Full attribution to Fable and myself.

@zanehengsperger Great piece, thank you. I’d submit that the culture change needed is even more fundamental: When engineering students & skilled trades get bigger NIL deals out of high school than athletes & National Merit Scholars are more revered than QB’s, culture will be on the right track.

Scum

A GLOBAL BANK FOUNDED IN 1857 IS CASUALLY DROPPING THE ULTIMATE BLUEPRINT FOR THE FUTURE OF ENTERPRISE TECH BBVA just revealed their 15-year evolution with Google Cloud, and it is a masterclass in hyper-scale. We aren’t talking about managing a few servers anymore, they are orchestrating thousands of live cloud projects for over 81 million customers across 25 countries. Managing that level of extreme complexity requires a complete departure from traditional infrastructure methods. Ditch the fragmentation: They moved to Google Cloud App Hub, smashing scattered project data into one glorious, unified dashboard. The Final Boss Move: They are actively unleashing an agentic AI layer. Soon, human operations managers won't even check infrastructure consoles. Autonomous AI agents will monitor performance, optimize cloud costs, and run the entire ecosystem on autopilot. If a century-old banking giant can aggressively pivot to fully autonomous AI operations, the old way of managing infrastructure is officially dead.

America is literally being strangled to death by this

@ThibaultJaigu "EU hosted" no thanks. On prem only

@csentropy Fully sovereign is one answer. EU hosted with zero retention gets 90% of the protection without giving up frontier models. Most companies land there

It is solved by running your own oss models on your own physical hardware. All security devolves to physical security Sovereign AI is not a meme. Or lose all your tacit knowledge and trade secrets to those that desperately want to steal it.

EAs want to deform the free market's will and apply capital to deprive you of goods and services you want. Whether it's banning open source AI or banning meat, it's always a similar strategy

*Updated* OpenAI’s last 24 hours: > Head of Safety quit > Sued by NYTimes > Top Exec unexpectedly departs > Shuts down browser tool after 9 months > Sued for trade theft by Apple > Caught selling product to China against sanctions

On cryptography, the cloud companies just did such a good job with security that I think we were lulled into a false sense of security, but the reality is that the models speak plaintext not cyphertext: “Trust the Math, Not the Man”

ANDURIL FOUNDER PALMER LUCKEY PRESENTS THE “BARRACUDA” CRUISE MISSILE ON JAPANESE TV: “THIS IS DESIGNED TO SCARE CHINA INTO NOT TAKING ACTION ON TAIWAN. IT HAS A VERY SPECIFIC PURPOSE. WE DESIGN OUR CRUISE MISSILES SO THAT THEY CAN BE VERY EASILY MANUFACTURED BY AUTOMOTIVE-STYLE ASSEMBLY LINES. REMEMBER THAT DURING WWII, THE UNITED STATES USED THEIR PEACETIME FACTORIES AND PIVOTED THEM TO WARTIME USE. IT’S MUCH LESS EXPENSIVE THAN A CONVENTIONAL MISSILE… IT’S CHEAPER TO MAKE 100,000 OF THESE THAN 1,000 EXISTING MISSILES. I SHOULD BE ABLE TO BUY A WHOLE BUNCH OF THESE FOR LESS THAN A BUDGET COMPACT CAR.”

i would have guessed >85%

NEW: Study reveals LinkedIn is the most AI-saturated major social platform, with over 40% of long-form posts flagged as fully AI-generated “slop.”

We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.

@Google, why do you build a marvelous product like Notebook LM, and leave it to rot with tons of glitches?

@KatieMiller They should name it "Cringe" (steal some brand thunder from hinge)

Only the most insufferable people find love on LinkedIn