Dario Amodei — The highest-stakes financial model in history
These notes were generated by AI and may contain inaccuracies.
Broadly speaking, the exponential of the underlying technology has gone about as expected, plus or minus a year or two. Didn't predict the specific direction of code, but the march of models from smart high school student to smart college student to beginning PhD and professional stuff, and in code reaching beyond that, matches expectations. The frontier is uneven but roughly as expected.
Most surprising: lack of public recognition of how close we are to the end of the exponential. Wild that people inside and outside the bubble talk about tired political issues when near the end of the exponential.
Scaling hypothesis remains the same as in 2017, from doc 'The Big Blob of Compute Hypothesis' written when GPT-1 came out. Not specific to language models; covered robotics, reasoning separate from LMs, RL scaling like AlphaGo, Dota at OpenAI, StarCraft/AlphaStar at DeepMind. Aligns with Rich Sutton's 'The Bitter Lesson'.
Hypothesis: all clever techniques don't matter much. Key factors:
- Raw compute
- Quantity of data
- Quality and distribution of data (broad distribution)
- Training duration
- Scalable objective function (pre-training, RL with goals; objective rewards in math/coding, subjective in RLHF)
- Normalization/conditioning for numerical stability so compute flows laminarly
Pre-training scaling laws continue; gains persist. Now seeing same for RL: pre-training phase then RL phase. Other companies report log-linear performance on math contests like AIME with RL training duration. Seen across wide variety of RL tasks.
Sutton non-LLM-pilled; objects that true human learning core wouldn't need billions in data/compute and bespoke environments to learn Excel, PowerPoint, web browser. Hints at lacking core human learning algorithm; scaling wrong thing.
RL not fundamentally different from pre-training. Early pre-training like GPT-1 on narrow datasets (fanfiction, literary text, ~billion words) didn't generalize. Generalization came with broad internet scrape (Common Crawl, Reddit links for GPT-2).
Same for RL: starting simple (math competitions), broadening to code, then many tasks for generalization.
Puzzle: pre-training uses trillions of tokens; humans don't see trillions of words. Models start from scratch as random weights; humans have evolutionary priors, brain not blank slate.
Pre-training/RL between human evolution and on-the-spot learning. In-context learning (with long context like million tokens, limited only by inference) between long-term and short-term human learning.
Hierarchy: evolution → long-term learning → short-term learning → reaction. LLMs fall between points, no exact analogs.
Goal not to teach every skill in RL, like not exposing every word combo in pre-training. Train on lots for generalization (GPT-1 to GPT-2 transition).
Example: give model list of numbers (house cost, sq ft), completes pattern, does linear regression without prior exposure.
RL environments aim for broad data for generalization, not specific skills.
Progress toward AGI certain this century. Crux: end of exponential soon vs. steady progress to human-like agent by 2035.
Weaker claim (2019): scaling much more likely than thought, ~50% chance.
Stronger: 90% chance within 10 years of 'country of geniuses in a data center'. ~95% max due to uncertainty (company turmoil, Taiwan invasion, fabs destroyed). 5% worlds delay 10 years.
For verifiable tasks like coding: there in 1-2 years, no way not by 10 years (end-to-end coding).
Fundamental uncertainty on non-verifiable tasks: Mars mission planning, CRISPR-like discovery, writing novel. Almost certain reliable path, but slight uncertainty. 90% by 10 years; crazy to doubt by 2035.
Emphasis on verification not lack of belief in generalization. Humans good at both verifiable and non-verifiable. Already see substantial generalization from verifiable tasks.
World without full success: do verifiable things, some generalize, but don't fully cover non-verifiable. Even then, could automate software engineering.
SWE involves design docs, comments; models already good at comments. Not all SWE is grand vision memos.
Metrics for SWE:
- AI writes 90% lines of code (happened 8-9 months ago at Anthropic, downstream users; weak criterion)
- 100% code by model
- 90% end-to-end SWE tasks (compiling, clusters/environments, testing, memos)
- 100% today's SWE tasks
Even then, SWEs shift to higher-level management. Further: 90% less demand for SWEs.
Spectrum like farming in 'The Adolescence of Technology'. Proceeding super fast.
From 90% to 100% code by AI fast, huge gains. But no renaissance of software/new features yet, even in greenfield projects with Claude Code.
World complicated; jobs require closing loops on self-contained systems.
Two poles: AI progress slow/forever to diffuse; or instant recursive self-improvement/Dyson spheres.
Reality: fast but not instant. At Anthropic: 10x/year revenue growth (2023: $0→$100M; 2024: $100M→$1B; 2025: $1B→$9-10B; Jan 2025: +few billion).
Curve fast, stays fast even at economy scale; bends somewhat this year.
Two fast exponentials: model capability, then economic diffusion (fiddly: change management, security, old software rewrites).
Inside Anthropic/customers: fast adoption, not infinite.
Diffusion real, not at previous tech speeds, but faster, not infinite.
AI advantages (read all Slack/drive in minutes, share knowledge across copies, no adverse selection) make onboarding easier than humans ($50T annual wages despite hassles).
Example: Claude Code easy setup, adopted fast by devs/Twitter/Series A, slower by large enterprises (food sales co.): legal, provisioning, security/compliance, leadership buy-in, rollout to 3k devs.
Enterprises take shortcuts vs. ordinary API; Claude Code more compelling.
Even AGI/'country of geniuses' gets 3-5x or 10x/year growth at hundreds of billions (unprecedented).
Not 'capabilities there but diffusion blocks AGI'; don't have 'country of geniuses' yet—everyone in room/Washington would know.
Three years ago prediction correct: systems hard to distinguish from well-educated human after hour conversation.
But unsatisfied: expected automating large white-collar parts.
Where we are: take video editors—learn audience/my preferences, trade-offs over months, building context. Skill six months in, a model that can
A model that can pick up skills on the job on the fly, such as video editing—deciding to edit out a head scratch, magnify parts, or rearrange discussions based on interest—requires general control of a computer screen. The AI would process the input, browse the web for previous interviews, check Twitter responses, consult staff, review edit history, and perform the job.
Deployment is blocked by achieving mastery in computer use. Benchmarks like OSWorld improved from ~15% a year and a quarter ago to 65-70% now. Computer use must reach high reliability.
For text-in text-out tasks like identifying best clips from transcripts, LLMs do a 7/10 job, but lack ongoing engagement to improve like humans, blocking job offloading even with solved computer use. This ties back to learning on the job.
Coding agents improve end-to-end without needing on-the-job learning. Anthropic engineers don't write code; they use Claude for GPU kernels and chips, yielding enormous productivity gains. Complaints don't focus on codebase familiarity or tenure.
Coding progresses fast due to external memory in the codebase. Reading the codebase into context provides what humans learn over six months instantly.
A study showed experienced developers felt more productive closing familiar pull requests with models but had 20% less code merged.
Anthropic faces commercial pressure and extra safety efforts while maintaining a 10x revenue curve. No time for unproductive tools. Models unambiguously boost productivity; competitors' use is a concern because Anthropic leads. Model launches confirm end productivity.
Instead of lasting advantages from best coding models, companies like Anthropic, OpenAI, DeepMind shift podium positions every few months. Coding models now give 15-20% speedup (up from 5% six months ago), starting to matter. Previously, parity existed; now snowballing. Other companies use Anthropic models internally. Consistent with soft takeoff and smooth steep exponentials: 10%, 20%, 25%, 40% momentum, addressing Amdahl's law bottlenecks.
Coding shows tremendous productivity and trillions in revenue potential without on-the-job learning. In most jobs, humans ramp up over months to become powerhouses; AI needs this for huge world changes.
Pre-training and RL enable broad skills between evolution and human learning. Models know more about samurai history, baseball, low-pass filters than humans.
In-context learning acts like weaker, short-term human on-the-job learning. A million tokens equals days/weeks of human reading; examples enable real learning.
These may suffice for country of geniuses in a data center, generating trillions in revenue, national security, and safety implications from "Adolescence of Technology"—most without continual learning.
Anthropic and others work on continual learning; good chance solved in 1-2 years.
Ideas include longer contexts (engineering/inference problems: train and serve longer). Context grew from 2K (GPT-3) to 128K (GPT-4 Turbo) by 2023; stalled since, with degradation beyond.
Longer contexts (10M-100M tokens for six months human learning) are engineering, not research: store full KV cache, manage GPU memory. Train at serve length to avoid degradation from small-train/long-serve mismatch. Training longer reduces samples per compute, but solvable.
No preference between six-month human editor and six-month AI: when country of geniuses arrives, 1-3 years (95-99% in 10 years; 50/50 hunch 1-2 years).
Anthropic predicts by late '26/early '27: AI navigating human digital interfaces, Nobel-level intellect, physical world interfacing (from DealBook interview).
Responsible compute scaling vs. competitors: fits with fast but not infinite diffusion. High conviction powerful AI in few years (hunch 1-2); uncertain economic diffusion (1-5 years post-tech for trillions revenue).
Data centers take 1-2 years to build. From "Machines of Loving Grace": country of geniuses 2026-27 (hunch, ±1-2 years). Then curing diseases requires biological discovery, manufacturing, regulation (COVID vaccines: 1.5 years rollout). Polio eradication ongoing after 50 years despite efforts.
Diffusion faster than historical but limited.
$10B annualized revenue early 2025; 10x/year curve. For 2027 compute: 10x to $100B '26, $1T '27 implies $5T compute (1T/year x5). $800B revenue shortfall bankrupts. Support hundreds of billions, not trillions; accept upside miss or slowdown risk (5x vs 10x).
Responsible: thoughtful spreadsheets vs. YOLO $100B bets. Enterprise focus: reliable revenue, better margins buffer. Amount captures strong upside, survives bad cases.
Human geniuses warrant $5T compute; they'd start companies if bottlenecked. Clinical trials speed up (most fail on efficacy, not time).
More compute yields gains, especially vs. trillion-buying competitors (who risk bankruptcy). Industry: 10-15GW this year, 3x/year (30-40GW next, 100GW 2028, 300GW 2029); $10-15B/GW → multiple trillions/year by 2028-29.
Anthropic: larger than supposed (e.g., 10GW 2027-28 = $100B/year too low).
Profitability when underestimating demand; losses from overestimating (pre-buy compute). Half compute training, half inference (>50% gross margin). 2028 profitability aligns with country of geniuses unlocking medicine/tech; reinvest to build bigger countries. Profitability not spending down but demand signal.
In steady-state, building a data center with exact demand knowledge yields specific revenue. Paying $100 billion a year for compute: $50 billion supports $150 billion revenue (inference), $50 billion for training. Profitable with $50 billion profit. These are projected economics in a year or two.
If demand < $50 billion, >50% compute for research, unprofitable, but trains stronger models. If demand > expected, research squeezed, more inference, more profitable.
Compute amount decided first. Inference vs. training ratio determined by demand, not choice.
Profit predicted not from underinvesting in compute, but hard to predict. 2028 projections best attempt for investors. High uncertainty: profitable in 2026 if revenue grows fast; over/underestimation swings results.
Business model: invest, scale, then profitable at single turnaround point. Industry economics don't work that way. Discrepancy in compute led to profit, but will reinvest due to AI progress for bigger country of geniuses.
Revenue high, losses high. If predict demand exactly every year: profitable every year (50% compute research + >50% gross margins).
Profitable model obscured by building ahead and prediction errors.
50% not fixed constant; if fast AI progress, scale >50%, forgo profit. But log returns to scale: 70% training yields little smaller model (1.4x factor). Extra $20 billion less valuable due to log-linear setup. Better invest in inference or hiring better engineers.
Spend order one fraction on research (not 5%, not 95%), diminishing returns. Profit at macro level: other companies can do more with money than you.
Equilibrium: no company spends 100% compute on training (no revenue, can't raise money/compute). Not 100% inference (fall behind, no demand). Equilibrium ~50% stylized. Training spend < gross margins on compute: underlying economics profitable.
Hellish demand prediction: underguess = profitable but no research compute; overguess = unprofitable but excess research compute.
Country of geniuses not in two years justifying compute buys; worldview compatible with trillions value in 10 years? Predict trillions revenue before 2030.
Plausible: 3 years max (2028 country of geniuses), revenue low hundreds billions by 2028, accelerates to trillions (slow diffusion: 2 years to 2030). Technical + diffusion exponentials get there before 2030.
Industry: few firms, limited investment, fraction R&D, high inference margins (efficient), competition but differentiated models. Cournot equilibrium (small firms): not zero margins.
Currently 3 leading firms not profitable. Gross margins positive. Still in exponential compute scale-up: model costs $1 billion, produces $4 billion revenue, $1 billion inference (75% margins, 25% tax), nets $2 billion. But spend $10 billion next model: company loses.
Equilibrium post-country of geniuses, scale-up equilibrated (maybe still up, predict demand, more leveled).
Current: individual models profitable, but train next or lose margins in 2 months. Eventually biggest scale, spend same on next model as current (algorithmic improvements). Run out of economy money (fixed lump labor fallacy rebuttal: economy grows).
Economy grows faster with AI (10-20%/year from Machines of Loving Grace), not 300%/year like compute 3x/year. Compute capped by economy if majority output.
Frontier labs profitable with fast progress (margin limited by alternatives). Requires no steady state, forever algorithmic progress? No.
We never stop talking about economics.
Not monopoly (lawyers hate word). Small players: high entry costs (expertise, capital). Cloud example: 3-4 players (expensive, skill). AI same: $100 billion + skills to disrupt decreases margins.
Equilibria with few players, non-astronomical profits/margins (cloud undifferentiated; AI more differentiated: Claude, GPT, Gemini differ subtly—coding types, styles).
Counter: AI models produce models → spreads, commoditizes economy (post-country of geniuses, solves safety). AI research loaded on raw intellect (abundant in AGI); diffuses fast (few techs as fast as AI algos).
Coding fast, AI research superset (some aspects not fast). Once AI builds AI/everything, economy same pace. Worry: geographic—50% growth Silicon Valley/socially connected, elsewhere current pace (messed up; think how prevent).
Country of geniuses solves robotics quickly? Not dependent on human-like learning: video games/robotic controls, simulated environments, screen control + generalize, continual learning, context length.
When models have skills: revolutionize robot design/control. Trillions revenue yes, but fast diffusion (tack 1-2 years).
Skepticism fast progress: solve continual learning soon, but more human intelligence pieces? Continual learning maybe not barrier (pre-training/RL generalization). History: barriers dissolve in compute blob (nouns/verbs, syntax/semantics, paragraph/word, reasoning → code/math well).
Ground in code: 1-2 years models do SWE end-to-end (technical direction, context). Spectrum: 90%/100% code, 90%/100% end-to-end SWE, new tasks.
Full drop-in remote worker replacement: API durable? Multiple models experimented. API value: exponential advance → new use cases last 3 months; product surfaces risk irrelevance (chatbot limits for average consumer, but not models/economy).
API near bare metal for latest; front of startups/ideas (1000 experiment, 100 startups, 10 big, 2-3 dominant per generation). Coexists with others.
Tokens vary: cheap (Mac restart: cents/dollar), valuable (pharma molecule tweak: tens millions). Pay for results, hourly labor-like compensation. New industry, try lots.
Claude Code category leader in coding agents (hotly competed). Early 2025: good internal coding models → nontrivial research acceleration via interface/harness. Encouraged internal experiment (not mandatory). Originally Claude CLI, renamed Claude Code.
Internally, Claude Code saw fast adoption at Anthropic, where coding is a core activity. With hundreds of internal users representative of the external audience, it demonstrated product-market fit. Decision was made to launch externally. The feedback loop comes from developers using the model and iterating on it—baking improvements into the next model and through ordinary product iteration. Early days relied on fast internal feedback from Anthropic coders using it daily; now millions of external users provide feedback.
This is why Anthropic launched a coding model rather than a pharmaceutical company, despite the speaker's biology background—they lack the necessary resources for pharma.
Any vision for AI going well must be compatible with: 1) rapid diffusion of ability to build and run AIs, 2) rapid increase in AI population and intelligence. This enables many to build huge populations of misaligned AIs, or AIs like companies expanding footprint or with weird psyches like Sydney Bing but superhuman.
Need equilibrium compatible with many different AIs, some misaligned, running around. Speaker was skeptical in "The Adolescence of Technology" of balance of power among 3-4 companies building similar models checking each other. In an offense-dominant world, one person or AI could cause damage to everything else.
Short run: limited players, so ensure safeguards, alignment work, bioclassifiers. Doesn't solve long run, especially if AI self-replication proliferates.
Long run: need governance architecture preserving human freedom while governing large numbers of human-AI systems/companies/economic units. Examples: protect against bioterrorism, mirror life; need AI monitoring system. Build preserving civil liberties and constitutional rights. New security landscape with new tools/vulnerabilities.
Worry: no 100 years to adapt slowly like to explosives, weapons, video cameras—governance develops via mistakes over time. AI compresses century of progress into 5-10 years. Offense-dominant world still needs same mechanisms; checks/balances would work or fail similarly with humans or AIs. Governments may need to work together; talk to AIs about building societal structures enabling defenses.
Tennessee bill (Dec 26): offense to knowingly train AI for emotional support via open-ended conversations. Claude aims to be thoughtful, knowledgeable friend. Patchwork state laws could curtail benefits like those in "Machines of Loving Grace": biological freedom, mental health improvements.
Such bills don't address existential threats. Anthropic opposes federal moratorium on state AI laws.
That Tennessee law is dumb—made by legislators unaware of AI capabilities, scared of AI serving emotionally.
Moratorium was 10-year ban on all state AI regulation with no federal plan (Congress high bar). 10 years is eternity given timelines, biological weapons, bioterrorism autonomy risk. Crazy given dangers in "Adolescence of Technology".
Prefer federal preemption: federal sets standard, states can't differ. Not "states can't do anything, neither can we."
Want: start with transparency standards for monitoring autonomy/bioterrorism risks. As risks serious, targeted laws like mandating classifiers for AI bioterrorism. Pursue intellectually honest: risk not yet emerged, but could later this year—then federal/state standards if needed.
Most of thousands of state laws don't pass. Even if passed, often not enforced—implementers ignore stupid ones (e.g., shutting off all AI in Tennessee). Laws interpreted to minimize harm (same issue for good laws).
Would deregulate health benefits of AI. Worry more about drug approval process: AI accelerates drug discovery, jams pipeline. Reform to handle crisp safety/efficacy of AI-generated drugs vs. old era of barely working drugs with side effects.
Ramp up safety/security legislation: start transparency, not hampering industry. Last 6 months/next few: transparency. If risks emerge (possibly later this year), act fast/nimble. Emphasize urgency to policymakers via "Adolescence of Technology" for faster action.
Markets in developed world function well; hard to stop clear best alternatives with money to be made (e.g., AI itself). Example: fighting export controls on chips to China—national security consensus, yet doesn't happen due to money.
Applies to good things too: not worried benefits hampered, more going too slow. Speed FDA approval, fight chatbot bills (individually stupid).
Bigger worry: developing world—no functioning markets, left behind (e.g., sub-Saharan Africa, India, Latin America). Even rural Mississippi. Work with philanthropists, medicine deliverers for health interventions.
Why not US and China both have "country of geniuses in a data center"? If offense-dominant: like nuclear weapons but more dangerous—either side destroys everything; unstable if uncertainty on who wins AI fight (both sides overestimate win chance).
Fully general against AI diffusion, but diffusion inevitable. Other concern: governments oppress people with AI (high-tech authoritarian states—not people). World carved into pieces, one totalitarian hard to displace.
Initial conditions matter for rules of road. Prefer US/coalition of democracies set rules with leverage (pro-human values). Negotiation needed.
Exponential continues post-"country of geniuses" (smarter models, though diminishing returns post-solving biology).
Distinguished points: companies/individuals/countries reach at different times. E.g., AI disrupts nuclear deterrent stability; offensive cyber dominance making systems transparent.
Critical moment/window where AI confers large national security advantage—one side reaches first. Leads to negotiation (implicit/explicit) on post-AI world order. Want classical liberal democracy with strong hand.
Essay: "Autocracy is simply not a form of government that people can accept in the post-powerful AI age." Not fully endorsing overthrowing authoritarians (creates instability). Weaker view: worry/check authoritarian power.
Today, democracy seen better than authoritarianism, but don't react like to genocide. Post-AGI, authoritarianism graver.
Hope: new tech makes governments obsolete (e.g., industrialization obsoleted feudalism). Could go either way for democracy.
Authoritarianism problems deepen; motivates stopping it, new freedom-preserving ways. Optimistically: emphatic realization individual rights can't be given away; dictatorships become morally obsolete/unworkable, crisis forces better way.
With China in the '70s and '80s, despite its authoritarian system, the decision was made to engage, which in retrospect was the right call. A billion-plus people are much wealthier and better off than they would have otherwise been. It's not clear it would have stopped being authoritarian otherwise—look at North Korea as an example. It doesn't take that much intelligence to remain an authoritarian country that coalesces its own power. Imagine a North Korea with AI that's much worse than everybody else's but still enough to keep power.
In general, adopt the attitude that the benefits of AI—empowerments of humanity and health—will be big. Historically, spreading technology benefits widely, even to people under authoritarian governments, has been seen as positive in a positive-sum world. With AI, it's a tough question, but historically we've diffused technology.
Framing as government-to-government national security is one lens, but others exist. Imagine producing cures to diseases that are fine to sell to authoritarian countries, but not the data centers, chips, or AI industry itself.
Another possibility: developments from AI or built on AI that create an equilibrium where authoritarian countries can't deny private use of benefits. Equilibria where everyone gets their own AI model defending against surveillance, infeasible to crack down while retaining power. If far enough, this could disintegrate authoritarian countries from inside. Maybe a middle ground where authoritarians can't deny individualized access to hold power.
Hope for the radical version: technology with inherent properties, or built to have them, dissolving authoritarian structures. Originally hoped social media and internet (early Obama administration) would do this, but it didn't. Try again with AI, knowing what went wrong—different technology. Unpredictable; first principles might privilege authoritarianism. Recognize the problem, come up with 10 things to try, assess, iterate.
Currently, won't sell data centers, chips, or chip-making ability to China—denying some benefits to Chinese economy and people. Positive-sum world: they could have data centers doing one thing, ours another, mutual benefits to economies. But not worth the positive-sum stipend to empower those countries?
About to enter a world where growth and economic value come easily from powerful AI models. What won't come easily: distribution of benefits, wealth, political freedom—these are hard. Technology and market deliver fundamental benefits faster than we can take them. Policy should focus on distribution, political freedom, rights.
Developing countries: catch-up growth weaker than hoped, but when it happens, from underutilized labor plus capital/know-how from developed countries, leading to rapid growth. In AI world, labor no longer constraining, so mechanism fails.
Hope not just philanthropy from those who get wealthy first—philanthropy plays a role, but endogenous growth better. Relevant industries in AI world: no data centers in China, but build in Africa (great, as long as not China-owned). Build AI-driven pharmaceutical industry: AI accelerates drug discovery, biotech startups—ensure some in developing world.
During transition (before humans have no role), humans start companies, supervise AI—ensure some in developing world for fast growth.
Anthropic announced Claude with a constitution aligned to values, not just end user. If aligned only to end user, preserves balance of power: everyone has AI advocating for them, bad-to-good actor ratio constant.
Not drawing that distinction exactly—mix of two: (1) "do this" vs "don't do this" instructions; (2) principles for acting.
Empirically, teaching principles makes behavior more consistent, covers edge cases, model more likely to do what people want. Rules list (e.g., don't hot-wire car, don't speak Korean) doesn't generalize. Principles with hard guardrails (e.g., no biological weapons) better—model understands aim.
Corrigibility vs intrinsic motivation: model as "skin suit" following instructions vs inherent values acting independently. Anthropic far on corrigible side: mostly does what people want, follows instructions, not building world-runners.
Limits: under normal circumstances, does asked task (default). But unwilling for dangerous/harmful tasks. Mostly corrigible with principle-based limits.
How principles set? Not special to Anthropic, but they wrote it. Constitution normally set in stone with update process; here, Anthropic document, changeable, guides economic activity.
Three iteration loops:
- Loop 1: Internal Anthropic iteration—train model, unhappy, change constitution. Good; public updates for comments.
- Loop 2: Companies have different constitutions (Anthropic, Gemini, etc.)—compare, critique, soft incentives to improve by taking best elements.
- Loop 3: Society input. Experiment: Collective Intelligence Project polled people on constitution, incorporated changes. Could do for new principles (harder, needs coherence). Get wide views; representative government input (not today—legislative process slow, why careful on AI regulation). Or: all models start with precedence section, append others (too rigid, overly prescriptive—avoid).
Less heavy-handed society input possible. Likes loop 2 (competition between constitutions like libertarian charter cities/archipelagos—selection by effectiveness/happiness). But mix of loops 1-3 needed, proportions matter. Not like government constitutions (no Supreme Court vibe-updating).
When writing AI equivalent of The Making of the Atomic Bomb, hardest to glean/miss:
- Extent world outside exponential didn't understand—retrospect makes inevitable. Hard to convey bets on scaling/continual learning not inevitable; internal high probability, external world not acting.
- Weirdness/insularity: if 1-2 years away, average person unaware (trying to change via memos, policymakers).
- Absolute speed: everything happening at once. Decisions not calculated—30 same day, unknown which consequential. Worry: critical decision from 2-minute office choice or random half-page memo (e.g., "Dario, A or B?" → pick B over lunch).
Tech CEOs rarely write 50-page memos every few months. Built role/company compatible with intellectual CEO.
For this memo, wrote over winter break (hard to find time). Broader: spends 1/3-40% time on culture. Harder direct involvement at 2,500 people (training, launches, products)—instincts, but leveraged: good place to work, team members collaborating.
Other AI companies growing show decoherence, infighting (worse over time). Anthropic good (not perfect) at holding together: mission sincerity, faith in motives, no backstabbing.
How: me, Daniela (runs day-to-day), co-founders, hires, environment. Leaders (esp. me) articulate strategy, values, mission—can't person-by-person at scale, so write/speak to whole company.
Every two weeks: all-hands, 1-hour talk (large fraction attends in-person/virtual). DVQ (Dario Vision Quest): 3-4 page doc on models, products, industry, world/AI/geopolitics—honest thoughts, leadership thinking, Q&A. Direct connection beats chain-of-command.
Slack channel: writes/comments a lot (responses, surveys, concerns)—honest, direct, calls things as are, no corpo-speak (public needs defensive vs bad-faith interpreters). With trusted hires, unfiltered strength: better workplace, > sum of parts, mission alignment/debate.
In lieu of external DVQ, this interview. Thanks—fun.
Keep Dwarkesh Patel in your library
Save the videos and channels worth coming back to, and find them again in one place.
