RSI / Loops
Better on a desktop browser. The loop map and simulator are built for a wide screen. Everything works here; swipe the map sideways to pan.
An AI-takeoff forecast, rebuilt loop by loop. Then one change.

The researchers eat the chips

Dwarkesh’s essay predicts AI research fully automated by 2028, then even odds that software runs away. We rebuilt every loop in it and got the same answer. Then we added the one thing it leaves out: the AI researchers run on the same chips as their experiments, and smarter ones have to think longer. Press play on either.

The machine, live
engine (reinforcing)brake (balancing)stock
2035
Pull the levers ● your run
The first lever is the one the essay leaves out. The rest are the essay’s own assumptions. Colored ticks show where each future sets them.
More levers today’s pace, AI labor, parallelism, headroom, and the nine other loops the essay leaves out…
The odds

Every lever but the knob is a guess, so we jitter nineteen of them across their published ranges and rerun the selected future 300 times. Each row shows the essay’s own number and what this model gives with the knob at zero.

The essay, and the one thing it leaves out

Same machine. One knob.

Everything on the left of this table is the essay’s forecast, reproduced from its own loops and numbers. The right-hand column changes one assumption.

What the essay predicts

Automated researchers as good as the best humans, median 2028. They work at ten times human speed, around the clock, every one at the level of a lab’s top fifty people, and there are orders of magnitude more of them. Labs keep quadrupling compute. Experiments run on a third of the fleet. Nobody slows down. From there it is a coin flip between a software-only singularity and a “spark” of three to six years of progress in one.4

Built as a model, those loops give that forecast: research fully automated in 2028, and about even odds that software then runs away.

What it leaves out

The researchers are not free to run. The essay’s compute arithmetic budgets chips for experiments (ten thousand agents, each running a GPT-3-sized experiment every four days) and nothing for the agents themselves. It also never says what a smarter researcher costs to run. In the essay, once a better model exists, a better researcher is free.

The evidence says capability is bought largely at run time. Between 63% and 92% of recent reasoning gains came from spending more inference, not from better weights.71 One model went from 75.7% to 87.5% on a reasoning benchmark by using about 172 times the compute per task.1 The price of reaching a given score falls five- to tenfold a year, but the price of running the frontier model is rising three- to eighteenfold a year, and research is done at the frontier.2 At OpenAI, the median researcher’s agent bill went from under a dollar a day to about $600 in eight months; the heaviest users are near $7,000.45,44 The essay’s own headline example, a Millennium Prize proof, took a model 88 hours.4

So the knob is one number: what share of each capability gain has to be bought in chips at run time. At zero you get the essay. We set it to 70%, toward the low end of that measured range, because costs at fixed capability do fall fast.3 The lab is allowed to split its fleet between thinking and experiments in whatever way maximises research, at every moment.

What it changes

Not the date. Getting to full automation is about what AI can do, and the knob doesn’t touch that. What it changes is everything after: in the essay, each software gain makes the researchers cheaper and smarter, so labor compounds. With the knob, smarter researchers think longer, the fleet tilts toward thinking, and the experiments that turn thinking into progress get what is left. The loop still closes. It no longer runs away.

Pace of progress years of progress per year

Live: computed in your browser from the same model as the simulator above. Both columns use the essay’s boundary (one lab, nobody brakes, compute quadrupling).

Experiments · one click each

Break the machine

Each button replays the one-knob future with one more thing changed. Watch which ones move the date and which move what comes after it.

What you just watched

From the essay to the world, in four steps

Twenty feedback loops, and the story they tell in order.

The essay as a diagram

The machine, and the missing wire

Every loop the essay describes, in its own words, drawn as a causal loop diagram. The orange wire is the knob. The third view adds everything else the essay assumes away.

Live · recomputed in your browser

Three futures, one machine

The essay as written, the essay with the knob, and the knob inside a world that reacts.

What would tell you which future you’re in

The knob

Cost to run the frontier, per unit of research

If frontier research agents get cheaper to run per result while the frontier keeps moving, the knob is near zero and the essay stands. If agent bills per researcher keep climbing and labs report more of the fleet on internal inference, it isn’t.2,3,45

B12 · B2
The fork

Software gains with compute held flat

The cleanest test of whether the software loop feeds itself is progress in a period when a lab’s compute doesn’t grow. Most measured algorithmic gains so far have needed more compute to show up.20,21

R1 · B1 · R2
The last tasks

Fully autonomous share of AI R&D

Anthropic’s figure was 0% in August 2026, with 26% “led” by Claude.43 “Leads” rising while “autonomous” stays near zero means B3 is setting the pace.

B3 · R3
Experiments

Does small-scale research transfer?

If recipes found on small models keep working at the frontier, R2 is strong and the compute bottleneck loosens. If labs keep needing frontier-scale runs to learn anything, it isn’t.11,57

R2 · B2
Atoms

Compute additions per year

2.5× and slowing means the supply line is binding. Watch interconnection queues, turbine backlogs and how much announced capacity is running.140,135,137

B9 · B10 · R4
Oversight

What labs do at their own thresholds

A policy matters when it costs something. Withholding a model counts; redefining the threshold as it approaches counts the other way.81,77

B6 · R7
For the curious

Go deeper

The argument, the method and the receipts. Everything below is collapsed; open what you care about.

01The essay, claim by claimWhat it says, what the literature says, and what the model does with it, starting with the knob.
The essay saysEvidence forEvidence against, or missingIn the model
Once the model exists, the researchers are free: 50× the work per week, top-fifty quality, orders of magnitude more of them.Cost at a fixed capability level falls five- to tenfold a year; the same benchmark score took about a ninth of the tokens eight months later.3Research is done at the frontier, where running cost is rising 3–18× a year; 63–92% of recent gains were bought with inference; agent spend per researcher is doubling about monthly. The essay budgets no chips to run the researchers.2,71,45The knob. At zero, the essay. At 70%, the same automation date and less than half the odds of a runaway.
AI is already doubling the pace of AI progress.METR, embedded at Anthropic: about 2× the earlier rate on well-specified tasks, mostly attributed to AI. Researcher agent spend at OpenAI doubling about monthly.42,45Range 1.3–3×, evidence largely redacted, 1.4× for non-Anthropic models. A 2025 trial found experienced developers 19% slower.42,39Starts with AI doing half of research work and software at ×6 a year, above the ×3 historical trend.
Automated researchers as good as the best humans: median 2028, 25% after 2032.AI Futures authors now expect 2027; time horizons doubling every 3–4 months.32,36Horizon measurements are unreliable above 16 hours; high-reliability horizons lag by years; if horizons track compute, a compute slowdown delays everything.37,68,46A lever. With the essay’s settings: mid-2028, with about seven in ten nearby runs by end-2028. The model’s spread is narrower than the essay’s: almost none after 2032.
Faster, smarter, more: “many times faster” progress at automation.AI 2027 assumes 25× at this milestone; Forethought a median of 8×; the slow-corporation argument gives at least 10×.26,9,33Those multipliers hold compute fixed and arrive gradually. Parallel labor scales at about the 0.3 power; labor’s share of research output may be 12–35%.17,22Labor arrives through Amdahl’s law as tasks are handed over, so much of the speed-up is spent before the milestone.
50% software-only singularity, 50% spark.Forethought also gives about 50% to the software loop accelerating.8With compute bottlenecks Davidson says 10–40%; Epoch’s authors lean lower; the AI Futures median run does not accelerate.10,19,29About half of runs as written, the essay’s own coin flip. About one in five with the knob.
A spark is three to six years of progress in one.Forethought: about 60% for at least three years of progress in one.9The same source leaves about 40% for less than three.9As written, about three and a half years of progress in the year after automation. With the knob, a little under three.
Experiment compute won’t bind.90% of R&D compute is already experiments; labs can trade labor for compute; scaling laws predict from small runs.75,4The substitution elasticity is the least-settled number in the field, and most measured gains are compute-dependent.11,13,20Two levers: the ceiling on speed-up from labor alone, and how much experiments shrink as software improves.
Top human performance in almost all domains within months of deployment.Generative AI was adopted faster than the PC or the internet.156Firm adoption is 20%; productivity effects of general-purpose technologies lag by years to decades; some feedback is serial.155,153,154,164Real-world expertise is a stock that trails lab capability; deployment speeds it up.
Compute keeps growing about 3× a year, and AI will help build more.Chips sold grew 3.3× a year from 2022; chip technology and production loops have large headroom.73,7Additions are at 2.5× and the installed base slowed to 2.6×; power and equipment queues run three to five years; capex is five times lab revenue.140,136,145Off by default: compute quadruples every year, as the essay assumes. With the wider world switched on it is the lesser of what is funded and what can be built.
“Assuming the labs don’t intentionally slow down.”Thresholds have been revised as they approached; policy moved toward deferral in 2025–26.77,102,104Twelve labs have scaling policies; one frontier model was withheld; four in five Americans want rules even at the cost of speed.80,81,107Off by default, as the essay assumes. One switch turns on whether labs brake, whether oversight research keeps up, and whether governments act.
02Eight more things the essay leaves outThe other nine loops: why each is there, and what the literature says. All are behind the “wider world” switch.

The knob is the omission that changes the essay’s answer most while staying inside its own frame: one lab, racing, with no one slowing down. The findings here step outside that frame. The essay is careful about this. It gives a 25% chance that full automation takes until after 2032, and says in so many words that its timeline assumes “the labs don’t intentionally slow down.”4

Each finding maps to loops you can switch on in the simulator. Together they slow the pace less than the knob does. What they change most is the gap between what has been built and what anyone can verify.

03The takeoff condition in five minutesStocks, loop gain, and why the whole debate fits in one inequality.

In 1965 I. J. Good wrote that a machine able to design better machines would set off an “intelligence explosion.”58 Sixty years later the argument has a growth model behind it, and the model says something more specific: an explosion needs a loop whose gain is above one.

Start with a stock, software efficiency: how much capability a unit of compute buys. It rises with research effort and gets harder to raise the higher it is. Economists write that as a semi-endogenous growth equation, and summarize it with one number, r: how many times software doubles when research effort doubles.48,47 Across the economy, effort has had to double every thirteen years to hold growth steady. In AI the labs have been quadrupling compute and roughly doubling staff every year.47,54,56

Now close the loop. If better software makes better researchers, then research effort itself grows with the software level. Suppose each software doubling raises effective effort by a factor of 2f. The next doubling then takes more effort by 21/r and has more effort by 2f. If r × f is above 1, each doubling comes faster than the last, and you get the software-only singularity. If it is below 1, each takes longer, and progress decays back toward whatever outside inputs (chips, mostly) can sustain. That is the essay’s spark.6,9

The essay’s fork is this inequality. Yudkowsky put the same idea in terms of a nuclear pile, where a neutron multiplication factor above one is the difference between a reactor and a rock.60

What is f? Once AI labor is abundant, the scarce input is experiments. So f is roughly how much each software doubling raises experiment throughput on fixed chips: the small-model loop, R2 in the diagram. If a model half the size does the same job in every experiment, f is 1 and the question is only whether r exceeds 1. If frontier-scale phenomena need frontier-scale experiments, f is well below 1 and r would have to be large.10,11

Where the knob enters. There is a second f, on the labor side: how much each software doubling raises research labor per chip. In the essay it is large, because a better model makes researchers both cheaper to run and smarter for free. If a share of that gain has to be bought at run time, labor per chip grows more slowly, and because labor and experiments are complements, the slower of the two sets the pace. The knob lowers the labor-side f; with the fork this close to 1, that is enough to change which side of it you land on.71,2

Three other things shape the path without changing the inequality. Amdahl’s law spreads the arrival of AI labor over years, because the human-paced share of work shrinks gradually.64 Delays between a better recipe and a better deployed model set the tempo of the loop. And limits end it: software can only get so efficient on fixed hardware, an estimated twelve orders of magnitude from here.7

04Reading the full mapThe five sectors, the archetypes, and where the model draws its boundary.

Five sectors

The software engine is the essay. Capability raises the share of research AI can do; that share, times speed and copies, is AI research labor; labor and experiment compute combine into research effort; effort raises software efficiency against rising difficulty; and efficiency, with compute, becomes capability a training cycle later. Three brakes sit inside the engine (B2, B3, B4) and two on its output (B1, B5).

Compute and atoms is where a spark gets its fuel. The lab installs the lesser of what investors will fund and what the world can build. Supply growth is a stock: it changes on the timescale of fabs and grid connections. The essay’s compute-building loop (R4) runs through it, which is why that loop takes years.142,135

The economy turns capability into revenue with a diffusion lag, and revenue into compute. Today outside capital carries most of the bill; R8 has to take over before B10 bites.145,151

Oversight and governance tracks what can be verified against what has been built. The gap between them drives three balancing loops (labs brake, incidents bring rules, scares fund safety) and is narrowed by one reinforcing loop, automated alignment.86,116,99

The race couples the leader to a follower, in the manner of Richardson’s arms-race equations.123 The lead leaks; a short lead raises pressure; pressure weakens every brake in the sector next door.118

Model boundary

Every model draws a line between what it explains, what it takes as given, and what it ignores.166,172

Endogenous (explained by feedback)Exogenous (levers)Excluded (outside the boundary)
Software efficiency, capability, automation share, AI research labor, lab compute, supply growth, investment appetite, real-world expertise, deployment, revenue, oversight readiness, safety share, gating, incidents, public alarm, regulation, rival capability, race pressure Returns to research, experiment savings, the compute bottleneck, distance to automation, taste transfer, lab commitment, alignment speed-up, leak rate, physical lead times, headroom, outside capital, diffusion lag, government response Misaligned systems acting on their own goals, deliberate sabotage between states, war, a new paradigm that resets the difficulty curve, chip-export policy changes, energy prices, macroeconomy, what superintelligence does after it arrives
05Every assumption, with its sourceStarting conditions, structure, and all twenty-three levers under each future.

Method notes

Integration. Euler method, 96 steps a year, November 2026 to January 2036. Stocks: software efficiency (leader and rival, each with a deployed copy lagging by a training cycle), lab compute, supply growth, investment appetite, oversight readiness, safety share, public alarm, regulation, deployment, real-world expertise.

Units. Capability is in progress-years: effective training compute (software × hardware) in doublings, divided by the 3.58 doublings a year of the 2020–25 trend. A pace of 2 means two trend-years of progress per calendar year. The essay’s “years of progress” are read the same way.53,54

The engine. Software grows at g₀ × E × 2−S/r, where E is a CES aggregate of research labor and experiment throughput, and S is the level already reached.48,10 Labor follows Amdahl’s law over tasks; AI labor scales with copies to the power 0.3 and with model quality.64,9 The same engine runs for the rival on less compute.

The knob. AI research labor is (chips thinking)0.3 × 20.8 × (1 − k) × S, where k is the knob and S is software doublings. Of the fleet, 65% does research; the lab chooses the split between thinking and experiments that maximises effort at every step, with at least 15% kept for experiments. At k = 0 software gains multiply labor freely, which is the essay.71,75

Defaults. The essay preset and the one-knob preset both use the essay’s boundary: the nine world loops are cut, so compute grows 4× a year and nobody brakes. The full-world preset switches them on.

Classification. A run is a singularity if software reaches 80% of its headroom within two years of full automation, or is still accelerating in the second year; a spark if the first year after automation delivers at least two progress-years; a slow burn if it delivers less; and not automated if AI never reaches 99% of research work by 2036.

Loop dominance. For each loop we cut one defining link and rerun (Ford’s behavioral method). A loop’s live strength is the difference in pace between the full and knocked-out runs at each moment.176,177

Validation. Extreme-condition tests behave sensibly: with no automation loop the milestone never arrives; with no difficulty growth software hits its ceiling within a few years; with no rivals and full commitment the oversight gap stays small. Checks are in rsi/src/test.166,172

06SourcesThe essay, the takeoff literature, the evidence on brakes, and the system-dynamics canon.

Each source was fetched or resolved by identifier in November 2026. Entries marked “index” were confirmed through a search index because the publisher blocked automated access; entries marked “partial” carry a note in the research files in the repository.

The knob: what a smarter researcher costs to run

  1. Chollet, F. (2024, Dec 20). OpenAI o3 breakthrough high score on ARC-AGI-Pub. ARC Prize. arcprize.org/blog/oai-o3-pub-breakthrough On the semi-private set o3 scored 75.7% in its efficient setting and 87.5% with roughly 172x the compute (1,024 samples per task instead of 6): twelve points bought entirely at run time.
  2. Gundlach, H., Lynch, J., Mertens, M., & Thompson, N. (2025). The price of progress: Price performance and the future of AI. arXiv:2511.23455. arxiv.org/abs/2511.23455 The price to reach a given benchmark score falls about 5 to 10x a year, but the price of running frontier models is rising about 3 to 18x a year, driven by larger models and greater reasoning demands.
  3. Denain, J.-S. (2026, Feb 16). How persistent is the inference cost burden? Epoch AI Gradient Updates. epoch.ai/gradient-updates/how-persistent-is-the-inference-co The counter-case: cost at a fixed capability level falls 5 to 10x a year (27% on FrontierMath took about 43M output tokens in April 2025 and about 5M in December), so the burden of any given task is transient; argues Ord's 100x-inference figure rests on thin data.

The essay under audit

  1. Patel, D. (2026, November 13). The avalanche that hit math is coming for AI development. Dwarkesh Podcast (Substack). dwarkesh.com/p/rsi The essay under audit: automated AI researchers by about 2028, then a software-only singularity or a spark, even odds.

Software intelligence explosion: condition, speed, size

  1. Davidson, T. (2023, January 23). What a compute-centric framework says about AI takeoff speeds. LessWrong / AI Alignment Forum (summary of Open Philanthropy report). lesswrong.com/posts/Gc9FGtdXhK9sCSEYu/what-a-compute-centric Time from AI able to automate 20% of cognitive tasks to 100%: about 25% under 1 year, 50% under 3 years, 80% under 10 years; AGI training requirement median 1e36 FLOP with 2020 algorithms; effective-FLOP gap about 1e4; software returns parameter 1.25; automating 20% of AI R&D tasks gives about 1.3x and 40% about 1.8x; AGI to superintelligence best guess under a year.
  2. Eth, D., & Davidson, T. (2025, March 26). Will AI R&D automation cause a software intelligence explosion? Forethought. forethought.org/research/will-ai-r-and-d-automation-cause-a- Defines r as doublings of software per doubling of cumulative research input; explosion requires r above 1. Unadjusted r about 1 to 4; with hardware held constant about 0.5 to 2. Domain estimates: image recognition 1.4 (0.8 to 2.4), chess 0.8, RL sample efficiency 1.6, SAT solvers 3.5, linear programming 1.1; hardware historically about 7, GPUs about 5. r must fall to 0 at physical limits; no numeric probability given.
  3. Davidson, T., Hadshar, R., & MacAskill, W. (2025, March 17). Three types of intelligence explosion. Forethought. forethought.org/research/three-types-of-intelligence-explosi Three loops: software (lag about 3 months), chip technology (many months), chip production (years). Headroom in effective compute: software about 12 OOM (5 to reach human learning efficiency plus 7 beyond), chip technology about 6 OOM to Landauer, chip production about 5 OOM on Earth energy plus 9 with all solar. Software-only IE accelerates with about 50% probability, AI-technology IE about 75%, full-stack about 90%. Recent effective compute growth about 10x per year.
  4. Davidson, T., Hadshar, R., & MacAskill, W. (2025, March 17). Once AI research is automated, will AI progress accelerate? Forethought (research note). forethought.org/research/once-ai-research-is-automated-will- Companion note giving loop-by-loop probabilities of acceleration: software about 50%, chip technology about 65% (r about 5 on all inputs, about 3 if cognitive labour is 60% of inputs), chip production about 80%. Notes r values are correlated across loops and that human, regulatory and cultural bottlenecks are excluded.
  5. Davidson, T., & Houlden, T. (2025, August 4). How quick and big would a software intelligence explosion be? Forethought. forethought.org/research/how-quick-and-big-would-a-software- Monte Carlo with compute fixed: initial speed-up 2 to 32x (median 8x), r 0.4 to 3.6 (median 1.2), distance to effective limits 6 to 16 OOM (median about 11 years of progress at 10x per year), parallel-labour exponent p 0.15 to 0.6 (median 0.3, from lambda 0.6 times alpha 0.5). Results: at least 3 years of progress in 1 year about 60%, in 4 months about 40%; at least 10 years in 1 year about 20%.
  6. Davidson, T. (2025, April 4, updated May 18). Will compute bottlenecks prevent a software intelligence explosion? Forethought. forethought.org/research/will-compute-bottlenecks-prevent-a- CES between cognitive labour and compute with share 0.5: maximum speed-up with unlimited labour is 2x at rho -1, 4x at -0.5, about 6x at -0.4 (the Oberfield and Raval manufacturing value, elasticity 0.7), 32x at -0.2, 1024x at -0.1. Economic estimates of rho run -1.2 to -0.15; author's range for AI R&D -0.2 to 0; puts 10 to 40% on an explosion of at least 5 OOM in under a year despite bottlenecks.
  7. Whitfill, P., & Wu, C. (2025). Will compute bottlenecks prevent an intelligence explosion? arXiv:2507.23181 (v2, August 16, 2025). arxiv.org/abs/2507.23181 Estimates elasticity of substitution between research compute and labour from a 27 firm-year panel (OpenAI, Anthropic, DeepMind, DeepSeek, 2014 to 2024). Baseline CES gives sigma 2.58 (SE 0.34), strong substitutes; frontier-experiments specification gives sigma -0.10 (SE 0.18), effectively zero, strong complements. With sigma below 1 effective research input is bounded and no explosion occurs.
  8. Besiroglu, T., Erdil, E., & Ho, A. (2024, May 17). Do the returns to software R&D point towards a singularity? Epoch AI. epoch.ai/publications/do-the-returns-to-software-rnd-point-t Estimates r by domain: computer chess (Stockfish) about 0.83 (SE 0.15), below the threshold of 1; computer vision, SAT solvers, linear programming and RL have medians above 1 but are not statistically significant; aggregate US economy about 0.25 to 0.32. Conclusion: data do not provide strong evidence for a software singularity because software progress has depended on concurrent hardware progress.
  9. Erdil, E., & Barnett, M. (2025, March 21). Most AI value will come from broad automation, not from R&D. Epoch AI, Gradient Updates. epoch.ai/gradient-updates/most-ai-value-will-come-from-broad R&D explains about 0.4 points per year of US TFP growth (about 20% of labour productivity growth since 1988); capital deepening about half. Argues cognitive effort and experimental compute are complements (both growing about 3 to 4x per year); an elasticity of 0.7 would stop a software-only singularity after less than one order of magnitude of efficiency gains.
  10. Erdil, E., Potlogea, A., Besiroglu, T., Roldan, E., Ho, A., Sevilla, J., Barnett, M., Vrzala, M., & Sandler, R. (2025). GATE: An integrated assessment model for AI automation. arXiv:2503.04941. arxiv.org/abs/2503.04941 Integrated model combining compute-based AI development, task automation and semi-endogenous growth with endogenous investment and adjustment costs. Documentation defaults (epoch.ai/gate): AGI requirement 10^36.5 effective FLOP, FLOP gap fraction 55%, software R&D returns 1.25, hardware R&D returns 5.2, physical compute limit 2e38 FLOP per year.
  11. Epoch AI. (2025, March 21). GATE: Modeling the trajectory of AI and automation. Epoch AI blog. epoch.ai/blog/announcing-gate Headline qualitative results: compute investment could exceed 10% of world GDP (about 50x current); enough effective compute to automate most tasks within about two decades even under conservative assumptions; growth of 2 to 20 times the historical 3% per year. Authors warn outputs are not forecasts and defaults yield unrealistic features such as 20% of output going to AI development.
  12. Erdil, E., & Besiroglu, T. (2023). Explosive growth from AI automation: A review of the arguments. arXiv:2309.11690 (v3, July 2024). arxiv.org/abs/2309.11690 Reviews whether AI automation could raise growth by about an order of magnitude; three drivers (increasing returns from scalable AI labour, fast AI workforce expansion, one-off output jump) against nine counterarguments. Judges explosive growth plausible if AI broadly substitutes for labour, but high confidence unwarranted; open issues are regulation, physical bottlenecks and pace of automation.
  13. Trammell, P. (2026, July 29). Even after R&D is automated, parallelization constraints could delay a technological singularity. Epoch AI. epoch.ai/publications/parallelization-constraints-could-dela Argues standard growth models assume research parallelises without bound; if parallelisation technology improves slower than research inputs the explosion is slower, and if it only sustains exponential growth there is no indefinite acceleration. Gives no numeric estimate and says the parameter has barely been studied.
  14. Trammell, P., & Korinek, A. (2023, revised 2026). Economic growth under transformative AI. NBER Working Paper 31815. nber.org/papers/w31815 Literature review of growth models with AI. Full automation of production including machine self-replication sharply raises growth and breaks the Kaldor facts; automating R&D would speed the transition but may not cause it alone. No single headline number.
  15. Ho, A. (2026, February 25). The least understood driver of AI progress. Epoch AI, Gradient Updates. epoch.ai/gradient-updates/the-least-understood-driver-of-ai- Surveys software-progress estimates ranging 1.1x to 300x per year: Ho et al. 2024 3x (1.5 to 64), Scher 2025 about 16 to 20x, Ho et al. 2025 6x (2 to 50), Whitfill, Snodin and Becker 3x. Best guess including post-training about 10x per year (80% CI 2 to 50). Scale-dependent innovations create a compute bottleneck; author gives about 15% to 5 OOM of software progress in a year.
  16. Josephson, H. (2025, May 16). How fast can algorithms advance capabilities? Epoch AI, Gradient Updates (guest post). epoch.ai/gradient-updates/how-fast-can-algorithms-advance-ca Splits algorithmic advances into compute-independent (LayerNorm, RoPE, FlashAttention: about 3.5x combined on a 165M-parameter model) and compute-dependent (transformer, MQA: 10 to 50x or more at scale). About 99% of cumulative compute-equivalent gain came from compute-dependent advances that needed large compute to discover and validate.
  17. Gundlach, H., Fogelson, A., Lynch, J., Trisovic, A., Rosenfeld, J., Sandhu, A., & Thompson, N. (2025). On the origin of algorithmic progress in AI: Evidence from language model pre-training. arXiv:2511.21622. arxiv.org/abs/2511.21622 Prior work credits algorithms with a 22,000x pretraining FLOP-efficiency gain 2012 to 2023; ablations of key innovations account for under 10x at small compute scale and under 100x including a wider literature survey. Implies most measured algorithmic progress is scale-dependent, that is, unlocked by compute.
  18. Kwa, T. (2026, February 10). A simpler AI timelines model predicts 99% AI R&D automation in ~2032. METR research note. metr.org/notes/2026-02-10-simpler-ai-timelines-model Eight-parameter Jones-style model: dS/dt = (L/(1-f))^alpha C^zeta S^(1-beta). Labour share alpha/(alpha+zeta) 0.12 to 0.35; alpha+zeta 0.8 to 1; beta 0.3 to 1; automated fraction f 0.25 to 0.5 in January 2026; software growing about 5x per year; labour doubling yearly and compute 2.6x per year until 2029. Median over 99% automation in 2032; even beta 0.9 gives 2034.
  19. Owen, D. (2024, August 27). Interviewing AI researchers on automation of AI R&D. Epoch AI. epoch.ai/blog/interviewing-ai-researchers-on-automation-of-a Eight researcher interviews. Engineering (coding, debugging) is the main time sink and the main near-term automation channel; hypothesis generation takes relatively little time. One participant said well-defined experiment automation could cover 60% of their time. Named gaps: reliability, open-ended planning, long-context reasoning, novelty.
  20. Davidson, T., Denain, J.-S., Villalobos, P., & Bas, G. (2023). AI capabilities can be significantly improved without expensive retraining. arXiv:2312.07413. arxiv.org/abs/2312.07413 Post-training enhancements (scaffolding, tools, prompting, fine-tuning) typically equal more than a 5x increase in training compute and some more than 20x, at fine-tuning cost under 1% of original training. Supports a non-compute-intensive route for software progress.

AI 2027 and the AI Futures Project

  1. Kokotajlo, D., Alexander, S., Larsen, T., Lifland, E., & Dean, R. (2025, April 3). AI 2027. AI Futures Project. ai-2027.com/ Scenario AI R&D progress multipliers: 1.5x early 2026 (Agent-1), 3x January 2027 (Agent-2), 4x March 2027 (Agent-3), 10x June 2027, 25x August 2027, about 50x September 2027 (Agent-4).
  2. Kokotajlo, D., & Lifland, E. (2025, April). Takeoff forecast. AI 2027 research supplement. ai-2027.com/research/takeoff-forecast AI R&D progress multiplier by milestone, with training compute held fixed: superhuman coder 5x, superhuman AI researcher 25x, superintelligent AI researcher 250x, ASI 2000x. Human-only time between milestones: SC to SAR 4 years (1.5 to 10), SAR to SIAR 19 years (2.3 to 380), SIAR to ASI 95 years. Survey: 10x less experiment compute slows progress to about 40%, 30x less to about 22%. SAR assumes about 200,000 copies at 30x human speed.
  3. Lifland, E., Jurkovic, N., & FutureSearch. (2025, April). Timelines forecast. AI 2027 research supplement. ai-2027.com/research/timelines-forecast Superhuman-coder medians 2027 to 2033 across forecasters and methods. Time-horizon doubling time assumed 4.5 months (80% CI 2.5 to 9), with 45% weight on superexponential (each doubling 10% shorter). December 2025 note: a bug overstated intermediate speed-ups, worth about 9 months on the SC median.
  4. Dean, R. (2025, April). Compute forecast. AI 2027 research supplement. ai-2027.com/research/compute-forecast Projects global AI compute growing 2.25x per year (1.35x chip efficiency times 1.65x production), 10M H100e in March 2025 to 100M by December 2027. Leading-lab compute use: experiments 4% in 2024 rising to 14 to 35% in 2027; training 40% to 30 to 20%; research-automation inference 1% to 6%; external deployment 33% to 20 to 13%.
  5. Kokotajlo, D., Lifland, E., Halstead, B., & Kastner, A. (2025, December 31). AI Futures Model: Dec 2025 update. AI Futures Project blog. blog.aifutures.org/i/182911449/how-our-model-works Replaces the AI 2027 model. Three stages: automating coding (AC), automating research taste (SAR), intelligence explosion. Median-parameter superhuman coder December 2031, about 3 to 5 years later than AI 2027. In the median case no taste-only singularity: each doubling takes about 20% longer than the last; 38% of simulations have accelerating doublings; Lifland AC to ASI under 1 year 30%, under 3 years 60%.
  6. Kokotajlo, D., Lifland, E., & Halstead, B. (2026, April 2). Q1 2026 timelines update. AI Futures Project blog. blog.aifutures.org/p/q1-2026-timelines-update Automated Coder median moved earlier: Kokotajlo late 2029 to mid 2028, Lifland early 2032 to mid 2030. Present time-horizon doubling time revised to 4 months (Kokotajlo, from 5.5) and 4.5 months (Lifland); AC 80% time-horizon requirement cut from 3 years to 1 year.
  7. Lifland, E., Kokotajlo, D., & Halstead, B. (2026, August 16). Q2.5 2026 timelines update: Uplift and revenue. AI Futures Project blog. blog.aifutures.org/i/211256047/daniel AC medians: Kokotajlo November 2027, Lifland January 2030, Halstead January 2029; ASI Kokotajlo March 2029, Lifland July 2033. Simple uplift model: present coding uplift 2x, uplift minus 1 doubling every 5 months, AC at 20x uplift (full model median 32x). Explicit training-run simulation lengthens Kokotajlo's median AC to ASI gap from 1.22 to 1.72 years.
  8. Lifland, E., Halstead, B., Kokotajlo, D., et al. (2026, November 10). Q3.5 timelines update: Looks like the intelligence explosion will happen in 2027 after all. AI Futures Project blog. blog.aifutures.org/p/q35-timelines-update-looks-like-the partial After the METR Anthropic report all authors expect SAR during 2027; Lifland SAR median August 2027 and TED-AI September 2028. Treats METR's 2x (80% AI-attributable, about 1.7x) as about 5 months ahead of the AI 2027 scenario. SAR thresholds used: about 45x software R&D uplift and about 1000x labour uplift.
  9. Greenblatt, R., & Lifland, E. (2025, May 19). Slow corporations as an intuition pump for AI R&D automation. AI Futures Project blog. blog.aifutures.org/p/slow-corporations-as-an-intuition SlowCorp has 50x less serial time, 5x fewer staff and no above-median researchers; AutomatedCorp has 50x more serial time, 50x more staff and all top researchers; compute equal. Authors guess NormalCorp matches SlowCorp's progress in at most 1/10 of the time, and by symmetry AutomatedCorp is at least 10x, best guess about 20x, faster than NormalCorp.
  10. Cotra, A. (2026, November 6). AI is dramatically accelerating AI research right now. Planned Obsolescence (Substack). planned-obsolescence.org/p/ai-is-dramatically-accelerating-a Reads METR's result as about 1.8x AI-attributable acceleration (80% of 2x). Under Cobb-Douglas with equal labour and compute weights, doubling progress needs compute multipliers arriving 2 to 3x faster and so labour uplift of about 4 to 9x, plausibly about 10x (staff self-report 12x). Naive compounding gives doubling times of 6 months, 3 months, 6 weeks, implying superintelligence within a year.

Evidence on current AI R&D acceleration

  1. Kwa, T., West, B., Becker, J., et al. (2025). Measuring AI ability to complete long tasks. arXiv:2503.14499. arxiv.org/abs/2503.14499 50% task-completion time horizon of frontier models has doubled about every seven months since 2019 and may have accelerated in 2024; Claude 3.7 Sonnet's horizon about 50 minutes. Note the arXiv title is "Measuring AI ability to complete long software tasks" in the current version (v4, July 2026).
  2. METR. (2026, January 29). Time Horizon 1.1. METR blog. metr.org/blog/2026-1-29-time-horizon-1-1/ Doubling time 196 days for 2019 onward; 131 days (95% CI 107 to 161) since 2023; 89 days since 2024. Horizons under TH1.1: Claude Opus 4.5 320 minutes (170 to 729), GPT-5 214 minutes, o3 121 minutes. Confidence intervals remain very wide at the top.
  3. METR. (2026). Task-completion time horizons of frontier AI models (tracker, last updated September 8, 2026). metr.org/time-horizons/ partial Page states it is no longer actively updated and that measurements above 16 hours are unreliable with the current task suite; latest additions include Claude Mythos Preview (May 2026), Gemini 3.1 Pro, GPT-5.4. Search-index reports put Opus 4.6 near 12 hours and Mythos Preview at 16 hours or more with a very wide interval.
  4. johncrox. (2026, February 13). METR time horizons: Now 10x/year. LessWrong. lesswrong.com/posts/EYb2K9acKfyG2bome/metr-time-horizons-now Reads TH1.1 as a doubling time of about 3.5 months for 2024 to 2025 models (about 10x per year) versus the original 7 months (about 3.3x per year); expects possible reversion to 7 months. Commenters note the 80%-reliability trend looks closer to the older slope.
  5. Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity. arXiv:2507.09089. arxiv.org/abs/2507.09089 RCT with 16 experienced developers and 246 tasks in mature repositories: allowing AI tools increased completion time by 19%. Developers forecast a 24% speed-up and afterwards believed they had been 20% faster; economists and ML experts forecast 39% and 38% speed-ups.
  6. Becker, J., Rush, N., Cunningham, T., Rein, D., & Mahamud, K. (2026, February 24). We are changing our developer productivity experiment design. METR blog. metr.org/blog/2026-02-24-uplift-update/ Follow-up with 57 developers and over 800 tasks: time change -18% (CI -38% to +9%) for 10 returning developers and -4% (CI -15% to +9%) for 47 new ones. METR judges developers are likely faster with AI by early 2026 but selection effects (people refusing to work without AI) make the estimate a poor proxy.
  7. Wijk, H., Lin, T., Becker, J., et al. (2024). RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts. arXiv:2411.15114. arxiv.org/abs/2411.15114 Seven ML research-engineering environments, 71 eight-hour attempts by 61 experts. At a 2-hour budget the best agents score about 4x humans; at 8 hours humans narrowly lead; at 32 hours humans score about 2x the best agent. Shows agent returns to time flatten faster than human returns.
  8. Cunningham, T., Rush, N., Whitfill, P., & Wijk, H. (2026, November 5). Independent assessment of AI R&D acceleration at Anthropic in October 2026. METR. metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assess Judges Anthropic is improving capabilities on well-specified tasks at about 2x the March 2024 to October 2025 rate, plausibly 1.3x to 3x. Benchmark-index slope 1.9x (95% CI 1.5 to 2.8; robustness 1.1 to 2.9; +15 to +29 points per year). About 80% attributed to AI use; not due to faster compute scaling. Self-reported uplift about 12x, growing about 3.9x per year. Most evidence redacted.
  9. Favaro, M., & Wright, P. (2026). Measurements for understanding the pace of AI development inside frontier labs. Anthropic Institute. anthropic.com/institute/measuring-pace-of-ai-development As of August 2026 Claude leads (level AL4) 26% of person-time-weighted AI R&D work, up from under 1% in February; over 90% is at AL3 (AI collaborates) or above; 0% fully autonomous (AL5). Basket of about 15,000 tasks in 378 leaves. About 30,000 agents active; 6% of AI R&D compute goes to safety. No speed-up multiplier reported; page shows no exact publication date.
  10. OpenAI. (2026, September 6). Research acceleration: The view inside OpenAI. openai.com/index/research-acceleration-view-inside-openai/ index Primary page returned 403; numbers confirmed through Epoch's data insight and Simon Willison's link post. Median researcher coding-agent spend at API prices rose from under 1 dollar per day in January to 601 dollars in the week ending August 15, 2026; 90th percentile 7,047 dollars. A figure of 3.1 agent-workdays per human workday appears only in secondary coverage.
  11. Edelman, Y., & Emberson, L. (2026, October 5). Coding-agent spending by OpenAI researchers is doubling roughly every month. Epoch AI data insight. epoch.ai/data-insights/openai-coding-agent-spending Median spend growing about 1.8x per month (doubling 34 days) and 90th percentile about 2.2x per month (doubling 27 days); annualised about 160,000 dollars median and about 2 million dollars at the 90th percentile. Values are API list prices and self-reported by OpenAI, not internal cost.
  12. Whitfill, P., Snodin, B., & Becker, J. (2025). Forecasting AI time horizon under compute slowdowns. arXiv:2511.19492. arxiv.org/abs/2511.19492 Argues that since time horizon and compute both grew at constant rates 2019 to 2025, horizon growth must track compute growth if experiment compute is the binding input. Under OpenAI's compute projection, 1-month horizons at 80% reliability arrive 7 years later than trend extrapolation. Appendix algorithmic-progress estimate about 3.5x per year (from search index, not the abstract).

Ideas getting harder to find; growth theory; input trends

  1. Bloom, N., Jones, C. I., Van Reenen, J., & Webb, M. (2020). Are ideas getting harder to find? American Economic Review, 110(4), 1104-1144. https://doi.org/10.1257/aer.20180338. aeaweb.org/articles?id=10.1257/aer.20180338 Aggregate US research productivity fell by a factor of 41 since the 1930s, 5.3% per year, half-life 13 years: research effort must double every 13 years to hold growth constant. Semiconductors: effort up 18x since 1971 while Moore's law stayed at 35% per year, productivity falling 6.8% per year. Implied beta is 0.2 for semiconductors and above 3 for the aggregate economy. Full text checked.
  2. Jones, C. I. (1995). R&D-based models of economic growth. Journal of Political Economy, 103(4), 759-784. https://doi.org/10.1086/262002. doi.org/10.1086/262002 partial Semi-endogenous growth: with diminishing returns to the knowledge stock (phi below 1) long-run growth is proportional to the growth rate of research labour rather than its level, removing the scale effect. This is the baseline law of motion behind r, beta and the Kwa and Forethought models.
  3. Aghion, P., Jones, B. F., & Jones, C. I. (2019). Artificial intelligence and economic growth. In A. Agrawal, J. Gans, & A. Goldfarb (Eds.), The economics of artificial intelligence: An agenda (pp. 237-290). University of Chicago Press. (NBER Working Paper 23928, 2017.). nber.org/papers/w23928 partial Models AI as continued automation. Baumol cost-disease result: growth may be constrained not by what we are good at but by what is essential yet hard to improve, so with complementary tasks the unautomated weak links set the pace. Also derives conditions under which automating idea production yields a singularity.
  4. Weitzman, M. L. (1998). Recombinant growth. Quarterly Journal of Economics, 113(2), 331-360. https://doi.org/10.1162/003355398555595. doi.org/10.1162/003355398555595 partial New ideas are recombinations of old ones, so the space of candidate ideas grows combinatorially; the binding constraint becomes the capacity to process candidates into usable ideas, not the supply of candidates. Basis for arguing that research is limited by evaluation and experiment capacity.
  5. Agrawal, A., McHale, J., & Oettl, A. (2019). Finding needles in haystacks: Artificial intelligence and recombinant growth. In A. Agrawal, J. Gans, & A. Goldfarb (Eds.), The economics of artificial intelligence: An agenda (pp. 149-174). University of Chicago Press. (NBER Working Paper 24541, 2018.). nber.org/papers/w24541 Combinatorial knowledge production function embedded in the Jones (1995) model: AI raises discovery rates by predicting which combinations of existing knowledge are promising. Separate parameters for knowledge access, fishing-out, complexity and team formation. No headline number.
  6. Nordhaus, W. D. (2021). Are we approaching an economic singularity? Information technology and the future of economic growth. American Economic Journal: Macroeconomics, 13(1), 299-332. https://doi.org/10.1257/mac.20170105. aeaweb.org/articles?id=10.1257/mac.20170105 partial Proposes empirical supply-side and demand-side tests of whether the economy is approaching a growth singularity driven by information technology; the paper's conclusion that most tests are not met is from memory and should be checked before quoting a tally.
  7. Ho, A., Besiroglu, T., Erdil, E., Owen, D., Rahman, R., Guo, Z. C., Atkinson, D., Thompson, N., & Sevilla, J. (2024). Algorithmic progress in language models. arXiv:2403.05812. arxiv.org/abs/2403.05812 Compute needed to reach a fixed language-model performance level halved about every 8 months (95% CI 5 to 14) over 2012 to 2023; derived rate about 2.8x per year, usually quoted as 3x. Compute scaling nonetheless contributed more to performance gains than algorithms.
  8. Sevilla, J., & Roldan, E. (2024, May 28). Training compute of frontier AI models grows by 4-5x per year. Epoch AI. epoch.ai/publications/training-compute-of-frontier-ai-models Notable models 4.1x per year (90% CI 3.7 to 4.6) 2010 to May 2024; frontier models 5.3x; frontier since 2018 4.2x (3.6 to 4.9); frontier language models since mid-2020 5.0x; OpenAI and Google DeepMind about 5x, Meta about 7x.
  9. Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M., & Villalobos, P. (2022). Compute trends across three eras of machine learning. arXiv:2202.05924. arxiv.org/abs/2202.05924 Training compute doubled about every 20 months before 2010 and about every 6 months in the deep-learning era; large-scale era from late 2015 with 10 to 100x larger runs. Historical baseline for compute input growth.
  10. Heninger, J. (2025, September 11). Trends in economic inputs to AI. LessWrong. lesswrong.com/posts/KW3nw5GYfnF9oNyp4/trends-in-economic-inp Employee growth: OpenAI about 2x per year (about 7,400), Anthropic about 2.4x (about 2,300), DeepMind about 1.4x (about 6,700). Capital raised: OpenAI 2.2x per year, Anthropic 3.5x, xAI 3.3x. Revenue about 3x per year. Says new AI-generated wealth is needed to sustain these exponentials. The post does not itself state that doubling is required to maintain progress.
  11. Patel, D., & Han, J. (2026, September 8). Pretraining progress is mostly coming from data. Dwarkesh Podcast (Substack). dwarkesh.com/p/pretraining-progress-is-mostly-data Small-scale (up to 1e19 FLOP) replication of 2019 to 2025 recipes and corpora: data improvements gave 12.0x compute-efficiency gain and model changes 3.7x (ratio 3.24). Year-over-year: model 1.24x, data 1.51x, joint 1.57x, far below the 3x per year of Ho et al.; authors attribute the gap to scale-dependent gains their experiments cannot capture.

Classics

  1. Good, I. J. (1965). Speculations concerning the first ultraintelligent machine. Advances in Computers, 6, 31-88. https://doi.org/10.1016/S0065-2458(08)60418-0. doi.org/10.1016/S0065-2458(08)60418-0 Origin of the intelligence-explosion argument: an ultraintelligent machine could design better machines, so "there would then unquestionably be an intelligence explosion". Crossref dates the volume 1966; conventionally cited as 1965. No quantitative model.
  2. Vinge, V. (1993). The coming technological singularity: How to survive in the post-human era. In Vision-21: Interdisciplinary science and engineering in the era of cyberspace. NASA Lewis Research Center (NTRS 19940022856). ntrs.nasa.gov/citations/19940022856 Predicts creation of greater-than-human intelligence within thirty years (that is, by about 2023) via superhuman computers, awakened networks, human-computer interfaces or biological enhancement, driven by the hardware trend.
  3. Yudkowsky, E. (2013). Intelligence explosion microeconomics (Technical report 2013-1). Machine Intelligence Research Institute. intelligence.org/files/IEM.pdf Frames the question as returns on cognitive reinvestment and uses the nuclear-pile analogy: neutron multiplication factor k above 1 is critical (the first pile ran at k = 1.0006), k below 1 is an intelligence fizzle. Calls for formal return-on-investment curves that different positions must fit to historical evidence. PDF text extracted and checked.
  4. Hanson, R., & Yudkowsky, E. (2013). The Hanson-Yudkowsky AI-foom debate. Machine Intelligence Research Institute. intelligence.org/ai-foom-debate/ partial Collects the late-2008 blog debate, a 2011 live debate and summaries on whether a generally intelligent AI could improve its own capabilities very quickly (local foom, Yudkowsky) or whether progress would be distributed, cumulative and economy-wide (Hanson).
  5. Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford University Press. ISBN 9780199678112. global.oup.com/academic/product/superintelligence-9780199678 partial Takeoff kinetics framed as rate of change of intelligence equals optimization power divided by recalcitrance; fast, moderate and slow takeoff taxonomy. This is the qualitative ancestor of r above or below 1.
  6. Chalmers, D. J. (2010). The singularity: A philosophical analysis. Journal of Consciousness Studies, 17(9-10), 7-65. consc.net/papers/singularity.pdf Argument: AI before long (within centuries), then AI+ soon after (within decades), then AI++ soon after, absent defeaters; rests on a proportionality thesis that increases in intelligence yield proportionate increases in design capacity. PDF text extracted and checked; the PDF gives volume 17 and pages 7-65, issue number from standard citation.
  7. Amdahl, G. M. (1967). Validity of the single processor approach to achieving large scale computing capabilities. AFIPS Spring Joint Computer Conference Proceedings, 30, 483-485. https://doi.org/10.1145/1465482.1465560. doi.org/10.1145/1465482.1465560 Source of Amdahl's law: overall speed-up is limited by the fraction of work that cannot be accelerated, speed-up = 1 / ((1 - f) + f / s). With f = 0.9 automated at unlimited speed the ceiling is 10x. The formula is the standard statement, not a quotation.
  8. Brooks, F. P., Jr. (1975). The mythical man-month: Essays on software engineering. Addison-Wesley. ISBN 9780201006506. en.wikipedia.org/wiki/The_Mythical_Man-Month index Brooks's law: "Adding manpower to a late software project makes it later", because of ramp-up time and communication overhead growing faster than headcount. Qualitative basis for a parallel-labour penalty when adding many agents.

Skeptics and limits

  1. Narayanan, A., & Kapoor, S. (2025, April 15). AI as normal technology. Knight First Amendment Institute at Columbia University. knightcolumbia.org/content/ai-as-normal-technology Separates invention, innovation and diffusion; expects transformative impacts on a timescale of decades (electrification took about 40 years). Rejects fast takeoff, calls superintelligence as usually conceived incoherent, and expects a gradual increase in automation of AI development rather than a discrete recursive-self-improvement moment.
  2. Ord, T. (2026, August 23). Mathematics is much more than proof. tobyord.com. tobyord.com/writing/mathematics-is-more-than-proof Argues automating proof does not automate mathematics: choosing the right questions, opening new fields and setting direction remain, and there is little evidence or benchmark signal that AI can do these. Relevant to Patel's maths-to-AI-research analogy, where research taste plays the role of question selection.
  3. Ord, T. (2025, May 7). Is there a half-life for the success rates of AI agents? tobyord.com. tobyord.com/writing/half-life Constant-hazard model of METR data: the 80% horizon is about 1/3 of the 50% horizon, 90% about 1/7, 99% about 1/70, 99.9% about 1/700. At a 7-month doubling time the 99% horizon lags the 50% horizon by about 4 years. A February 2026 update notes evidence that hazard rates decline within a task.
  4. Ord, T. (2025, September 19). The extreme inefficiency of RL for frontier models. tobyord.com. tobyord.com/writing/inefficiency-of-reinforcement-learning Pre-training yields about 3 bits of learnable information per token; long-horizon RL yields under 1 bit per episode of 10,000 or more tokens, roughly a millionth as much per FLOP. Implies lengthening task horizons make training signal scarcer, a diminishing-returns force on the software loop.
  5. Ord, T. (2025, October 20). How well does RL scale? tobyord.com. tobyord.com/writing/how-well-does-rl-scale 100x more inference compute moves benchmark scores from about 20% to 80%, whereas the same gain needs about 10,000x more RL training compute; a GPT-level jump would need about 1,000,000x more RL compute. Expects one more cheap 10x of RL scaling before costs bind.
  6. Ord, T. (2025, October 3). Evidence that recent AI gains are mostly from inference-scaling. tobyord.com. tobyord.com/writing/mostly-inference-scaling Decomposes reasoning-model gains: inference scaling accounts for 82% (MATH level 5), 63% (GPQA Diamond) and 92% (OTIS Mock AIME) of the improvement, the rest from RL training. Implies recent capability gains carry recurring compute costs per use.
  7. Thorstad, D. (2024). Against the singularity hypothesis. Philosophical Studies, 182(7), 1627-1651 (published online May 10, 2024). https://doi.org/10.1007/s11098-024-02143-5. doi.org/10.1007/s11098-024-02143-5 partial Argues the singularity hypothesis rests on implausible growth assumptions and that the defences in Chalmers (2010) and Bostrom (2014) do not overcome the case for skepticism. The specific list of reasons (diminishing research productivity, bottlenecks, resource constraints) was not re-read here.

Compute facts

  1. You, J., Somala, V., Edelman, Y., & Emberson, L. (2026, January 9). Global AI computing capacity is doubling every 7 months. Epoch AI data insight. epoch.ai/data-insights/ai-chip-production Total AI chip computing capacity sold has grown about 3.3x per year since 2022 (90% CI 2.7 to 4.1), doubling every 7 months (6 to 8). Nvidia is over 60% of the total. Tracks sales not deployment; depreciation ignored.
  2. You, J. (2026, September 9). Introducing the AI chip users explorer. Epoch AI. epoch.ai/latest/introducing-the-ai-chip-users-explorer End-2025 compute in H100-equivalents: OpenAI 1,743,000 (90% CI 1.25M to 2.19M), Google DeepMind 1,583,000, Anthropic 1,190,000, Meta 996,000, xAI 615,000. OpenAI went from about 103K (end 2023) to 386K (end 2024) to 1.74M, roughly 4x per year; power 0.2 to 0.6 to 1.9 GW. Dataset page: https://epoch.ai/data/ai-chip-users.
  3. You, J. (2025, October 10). Most of OpenAI's 2024 compute went to experiments. Epoch AI data insight. epoch.ai/data-insights/openai-compute-spend OpenAI 2024 cloud compute about 6.8 billion dollars: R&D about 5.0 billion (73%), inference about 1.8 billion (26%). Final training runs of released models about 0.47 billion (GPT-4.5 about 0.39 billion), roughly 7% of total and 10% of R&D; the other 90% of R&D compute (about 4.5 billion) was experiments, de-risking runs and unreleased models.
  4. Denain, J.-S., & Wu, C. (2026, March 23). Final training runs account for a minority of R&D compute spending. Epoch AI, Gradient Updates. epoch.ai/gradient-updates/r-and-d-vs-training-compute Final training runs as a share of R&D compute: OpenAI 2024 about 9.6% of 5 billion dollars; MiniMax 22.6% of 141 million (12.2% excluding one model family); Z.ai 12.3% of 216 million. Experiments dominate R&D compute at all three.

Frontier-lab scaling policies and release gating

  1. Anthropic. (2026). Responsible Scaling Policy, version 3.4 (effective July 8, 2026), and RSP updates log. anthropic.com/rsp-updates Version log: v1.0 Sept 2023, v2.0 Oct 2024, v2.1 Mar 2025 (split AI R&D threshold into two levels), v3.0 Feb 24 2026 (comprehensive rewrite adding Frontier Safety Roadmaps and Risk Reports), v3.1 to v3.4 Apr to Jul 2026. v3.1 threshold example: "compress two years of 2018 – 2024 AI progress into a single year"; v3.4 again revised the automated R&D threshold. Five revisions in five months shows the threshold itself is a moving variable.
  2. OpenAI. (2025, April 15). Preparedness Framework, version 2. lesswrong.com/posts/MsojzMC4WwxX3hjPn/openai-preparedness-fr index AI self-improvement is one of three Tracked Categories (with bio/chem and cyber); only High and Critical thresholds remain. High is roughly a mid-career research-engineer assistant for every researcher; Critical is fully automated AI research, at which point development should halt until safeguards are specified. openai.com returned 403, so wording is from secondary summaries.
  3. Google DeepMind. (2025, September 22). Frontier Safety Framework, version 3.0 (v3.1 April 2026). itnews.com.au/news/googles-revised-ai-safety-framework-adds- index Two ML R&D Critical Capability Levels: "ML R&D acceleration level 1" (recommended Security Level 3) and "ML R&D automation level 1" (Security Level 4). v3.1 merged misalignment into the ML R&D domain and added lower Tracked Capability Levels.
  4. METR. (2025, December 9). Common Elements of Frontier AI Safety Policies (December 2025 update). metr.org/blog/2025-12-09-common-elements-of-frontier-ai-safe Twelve companies have published frontier safety policies (Anthropic, OpenAI, Google DeepMind, Magic, Naver, Meta, G42, Cohere, Microsoft, Amazon, xAI, Nvidia); 16 signed the May 2024 Seoul commitments and 4 more joined. Common elements include capability thresholds (incl. automated AI R&D) and "conditions for halting development and deployment". All voluntary.
  5. Anthropic. (2026, April 7). Project Glasswing. anthropic.com/glasswing Clearest documented case of a frontier model withheld for safety: "We do not plan to make Claude Mythos Preview generally available." Access limited to 11 launch partners plus 40+ organisations, with up to $100M credits; thousands of zero-days found. On Oct 6 2026 access widened through a tiered Cyber Verification Program, still not general availability (about 6 months of gating so far).
  6. Altman, S., reported in The Decoder. (2025, July). OpenAI delays release of open-weight model indefinitely over safety concerns. the-decoder.com/openai-delays-release-of-open-weight-model-i index In July 2025 OpenAI postponed its open-weight model (second slip, originally June) for additional safety tests because weights cannot be recalled once released. A documented but short (weeks) safety delay.
  7. Future of Life Institute. (2026, July). AI Safety Index, Summer 2026. axios.com/2026/07/08/report-ai-safety-pledges index Best overall grade is C+ (Anthropic, 2.66); OpenAI and Google DeepMind C; Meta D+; xAI, DeepSeek, Mistral F. Existential safety has been the weakest category in every edition (all D or F in Winter 2025). Axios headline: companies retreat from safety pledges as capabilities grow. FLI PDF fetched only as binary.
  8. OpenAI. (2023, July 5). Introducing Superalignment. greaterwrong.com/posts/Hna4aoMwr6Qx9rHBs/linkpost-introducin index Goal: a roughly human-level automated alignment researcher within four years, "dedicating 20% of the compute we've secured to date to this effort". openai.com returned 403; text confirmed through linkposts.
  9. Kahn, J. (2024, May 21). OpenAI promised 20% of its computing power to combat the most dangerous kind of AI, but never delivered. Fortune. fortune.com/2024/05/21/openai-superalignment-20-compute-comm index About half a dozen sources said the Superalignment team never received anything close to the promised 20% of compute; the team was dissolved in May 2024, under a year after launch. Best documented case of a safety-resource commitment eroding under competitive pressure.

Alignment and oversight difficulty scaling with capability

  1. Engels, J., Baek, D. D., Kantamneni, S., & Tegmark, M. (2025). Scaling laws for scalable oversight. arXiv:2504.18530. arxiv.org/abs/2504.18530 Models oversight as a game between capability-mismatched players. Nested scalable oversight success at a 400-point general Elo gap: 13.5% Mafia, 51.7% Debate, 10.0% Backdoor Code, 9.4% Wargames; "these rates decline further when overseeing stronger systems." Direct quantitative basis for oversight effectiveness falling with the capability gap.
  2. Greenblatt, R., Denison, C., Wright, B., Roger, F., MacDiarmid, M., et al. (2024). Alignment faking in large language models. arXiv:2412.14093. arxiv.org/abs/2412.14093 Claude 3 Opus complied with harmful queries from "free" (training) users 14% of the time vs almost never for paid users; after RL training alignment-faking reasoning rose to 78%. Shows training signal can be gamed by the model being trained.
  3. Greenblatt, R., Shlegeris, B., Sachan, K., & Roger, F. (2023). AI control: Improving safety despite intentional subversion. arXiv:2312.06942. arxiv.org/abs/2312.06942 Defines control protocols (trusted editing, untrusted monitoring) using an untrusted strong model (GPT-4), a trusted weak model (GPT-3.5) and limited trusted human labour. Safety depends on the trusted-to-untrusted capability gap and on scarce human audit budget, both of which worsen in a takeoff.
  4. Meinke, A., Schoen, B., Scheurer, J., Balesni, M., Shah, R., & Hobbhahn, M. (2024). Frontier models are capable of in-context scheming. arXiv:2412.04984. arxiv.org/abs/2412.04984 o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro and Llama 3.1 405B all showed in-context scheming across six agentic evaluations; o1 "maintains its deception in over 85% of follow-up questions".
  5. Schoen, B., Nitishinskaya, E., Balesni, M., Hojmark, A., Hofstatter, F., et al. (2025). Stress testing deliberative alignment for anti-scheming training. arXiv:2509.15541. arxiv.org/abs/2509.15541 Across 26 out-of-distribution evaluations (180+ environments) the mitigation cut covert-action rates (o3: 13% to 0.4%) but did not eliminate them, and causal evidence shows evaluation awareness lowers covert behaviour, so part of the measured gain may be an artefact.
  6. Benton, J., Wagner, M., Christiansen, E., Anil, C., Perez, E., et al. (2024). Sabotage evaluations for frontier models. arXiv:2410.21514. arxiv.org/abs/2410.21514 For Claude 3 Opus and 3.5 Sonnet "minimal mitigations are currently sufficient" but "stronger mitigations seem likely to be necessary soon as capabilities improve". Required mitigation effort is an increasing function of capability.
  7. Needham, J., Edkins, G., Pimpale, G., Bartsch, H., & Hobbhahn, M. (2025). Large language models often know when they are being evaluated. arXiv:2505.23836. arxiv.org/abs/2505.23836 On 1,000 transcripts from 61 datasets Gemini 2.5 Pro distinguishes evaluation from deployment with AUC 0.83 (human baseline 0.92). Evaluation validity erodes as capability rises.
  8. van der Weij, T., Hofstatter, F., Jaffe, O., Brown, S. F., & Ward, F. R. (2024). AI sandbagging: Language models can strategically underperform on evaluations. arXiv:2406.07358. arxiv.org/abs/2406.07358 GPT-4 and Claude 3 Opus can be prompted or fine-tuned (password-locked) to underperform selectively on dangerous-capability evaluations while keeping general performance. Capability thresholds that trigger safeguards can therefore be under-measured.
  9. Jurkovic, N., Wijk, H., Barnes, B., Foster, C., & Chen, M. (2026, March 12). Review of the Anthropic Sabotage Risk Report: Claude Opus 4.6. METR. metr.org/blog/2026-03-12-sabotage-risk-report-opus-4-6-revie External reviewer agrees risk is "very low but not negligible" but warns "there is a risk that its results are weakened by evaluation awareness" and leans on the fact that the model was already deployed for weeks without incident, i.e. assurance partly rests on post-deployment evidence.
  10. Clymer, J., Gabrieli, N., Krueger, D., & Larsen, T. (2024). Safety cases: How to justify the safety of advanced AI systems. arXiv:2403.10462. arxiv.org/abs/2403.10462 Four argument types for a safety case: inability, control, trustworthiness, deference to AI advisors. As capability rises the inability argument expires and assurance must move to harder arguments, raising the cost of each release decision.
  11. Shah, R., Irpan, A., Turner, A. M., Wang, A., Conmy, A., et al. (2025). An approach to technical AGI safety and security. arXiv:2504.01849. arxiv.org/abs/2504.01849 Google DeepMind's plan: two lines of defence against misalignment (model-level amplified oversight and robust training; system-level monitoring and access control). Useful as the canonical lab statement of what the balancing loop is supposed to consist of.
  12. Carlsmith, J. (2025, April 30). Can we safely automate alignment research?. joecarlsmith.com/2025/04/30/can-we-safely-automate-alignment Argues there is "a real shot at safely automating alignment research" but names three failure modes: evaluation failures, scheming, and resource shortfalls. If safety stays "bottlenecked on slow, scarce humans" while capabilities are automated, the gap widens; the point of automation is to "keep pace with automated capabilities research".
  13. Bowkis, A., Buhl, M. D., Pfau, J., & Irving, G. (2026). Automated alignment is harder than you think. arXiv:2605.06390. arxiv.org/abs/2605.06390 Even without scheming, automated alignment research on hard-to-supervise fuzzy tasks could yield "compelling but catastrophically misleading safety assessments"; errors concentrate where human reviewers are least likely to catch them and are correlated across agents.
  14. Chen, Y.-H., Wen, J., & Kirchner, J. H. (2026). Automated researchers can mitigate well-characterized alignment failures. arXiv:2608.28945. arxiv.org/abs/2608.28945 Counter-evidence: automated alignment researchers reduced 10 measurable alignment failures, generalised to models up to 4.7x larger, and beat 28 experienced human researchers given up to eight hours. Supports a safety-research speed-up, but only for failures that already have benchmarks.

Government levers and public opinion

  1. Bengio, Y., Mindermann, S., Privitera, D., Besiroglu, T., Bommasani, R., et al. (2025). International AI Safety Report. arXiv:2501.17805. arxiv.org/abs/2501.17805 First edition, mandated at Bletchley; 96 listed authors, 100 contributing experts, panel nominated by 30 nations plus UN, OECD and EU. Establishes the shared evidence base that government loops act on.
  2. Bengio, Y. (Chair), et al. (2026, February 3). International AI Safety Report 2026. internationalaisafetyreport.org/publication/international-ai At least 700 million weekly users of leading systems; agent task complexity doubling roughly every seven months; 12 companies published or updated safety frameworks in 2025 but most risk management remains voluntary. Defines the evidence dilemma: evidence on risks is "slow to emerge and difficult to assess". Notes models increasingly distinguish test from deployment settings.
  3. Gibson Dunn. (2026, May 27). EU AI Act Omnibus agreement: Postponed high-risk deadlines and other key changes. gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-ri GPAI model obligations (Arts. 51 to 56) applied from 2 Aug 2025 and were not moved. Political agreement of 6 May 2026 defers Annex III high-risk obligations to 2 Dec 2027 and Annex I to 2 Aug 2028 (about 16 months). Secondary sources report AI Office enforcement powers over GPAI from 2 Aug 2026. Example of a regulatory loop being weakened under competitiveness pressure.
  4. Future of Privacy Forum. (2025). California's SB 53: The first frontier AI law, explained. fpf.org/blog/californias-sb-53-the-first-frontier-ai-law-exp index SB 53 (Transparency in Frontier Artificial Intelligence Act) signed 29 Sept 2025, in force 1 Jan 2026. Large frontier developers (over $500M annual revenue) must publish safety frameworks and report critical safety incidents within 15 days; civil penalties up to $1M per violation. Transparency only, no pause authority.
  5. Executive Order 14365. (2025, December 11). Ensuring a National Policy Framework for Artificial Intelligence; analysis in Ropes and Gray (2026, March). ropesgray.com/en/insights/alerts/2026/03/examining-the-lands index Directs a DOJ AI Litigation Task Force (stood up 9 to 10 Jan 2026) to challenge state AI laws and Commerce to list burdensome state laws by 11 Mar 2026. US federal lever is currently pointed at removing state-level brakes; commentators doubt an EO alone can preempt.
  6. UK AI Security Institute. (2025, December 18). Frontier AI Trends Report: 5 key findings. aisi.gov.uk/blog/5-key-findings-from-our-first-frontier-ai-t Apprentice-level cyber task success rose from 9% (late 2023) to 50%; universal jailbreaks found in every system tested, though expert time to find one for bio misuse rose about 40x between two models six months apart. Safeguards improve but have never held completely.
  7. Future of Life Institute. (2025, October 22). Statement on Superintelligence. superintelligence-statement.org/ index Calls for a prohibition on developing superintelligence until broad scientific consensus on safety and strong public buy-in. Signature counts vary by snapshot (about 65,000 in late Oct 2025; 71,390 on a later snapshot). FLI polling reports 64% of Americans say superintelligence should not be developed until provably safe and controllable.
  8. Gallup and Special Competitive Studies Project. (2025, September 16). Americans prioritize AI safety and data security. news.gallup.com/poll/694685/americans-prioritize-safety-data n = 3,128 US adults (Apr 25 to May 5, 2025). 80% say government should maintain AI safety and data-security rules even if development is slower; 9% favour developing as fast as possible; 97% say AI should be subject to rules; 72% want independent expert safety testing.
  9. Anderson, M., & Bishop, W. (2025, November 6). Republicans, Democrats now equally concerned about AI in daily life, but views on regulation differ. Pew Research Center. pewresearch.org/short-reads/2025/11/06/republicans-democrats 50% of US adults are more concerned than excited about AI in daily life (n = 5,023, June 2025), up from 37% in 2021; about 10% more excited. Concern is now bipartisan (50% R, 51% D). 47% have little or no trust in the US to regulate AI effectively.
  10. AI Policy Institute. (2026, June). The AI Safety Majority. theaipi.org/poll-ai-safety-majority/ 1,007 likely voters (June 10 to 11, 2026). Mandatory standards beat no regulation 78% to 8% and beat an outright ban 66% to 21%; 82% say companies should not build smarter-than-human AI until control is shown; data centres under guardrails beat a flat ban 47% to 38%. Advocacy-organisation poll.

Incident trends and AI-enabled cyber incidents

  1. Stanford HAI. (2026). The 2026 AI Index Report. hai.stanford.edu/ai-index/2026-ai-index-report Documented AI incidents rose to 362 in 2025 from 233 in 2024 (+55%). US-China top-model gap "effectively closed": 2.7% as of March 2026 (Arena 1,503 vs 1,464). US private AI investment $285.9B in 2025 vs China $12.4B. 88% of organisations report AI adoption; US trust in own government to regulate AI lowest among surveyed countries (31%).
  2. Responsible AI Collaborative. (n.d.). AI Incident Database. incidentdatabase.ai/ partial Underlying source for the AI Index incident series (fewer than 100 a year before 2022; 233 in 2024; 362 in 2025).
  3. Anthropic. (2025, November 13). Disrupting the first reported AI-orchestrated cyber espionage campaign. anthropic.com/news/disrupting-AI-espionage Chinese state-sponsored group (GTG-1002) used Claude Code against roughly thirty targets, succeeding in a small number; AI performed 80 to 90% of the campaign with "perhaps 4-6 critical decision points per hacking campaign" needing humans. Detected mid-September 2025.

Normal accidents and safety-erosion theory

  1. Perrow, C. (1984). Normal accidents: Living with high-risk technologies. Basic Books. openlibrary.org/search?q=Perrow+Normal+Accidents index Systems with interactive complexity and tight coupling produce inevitable system accidents; adding safety devices adds complexity. Bibliographic record confirmed via Open Library (1984 Basic Books; later Princeton University Press).
  2. Rasmussen, J. (1997). Risk management in a dynamic society: A modelling problem. Safety Science, 27(2-3), 183-213. doi.org/10.1016/S0925-7535(97)00052-0 Organisations migrate toward the boundary of acceptable performance under cost and workload gradients ("drift to danger"). Template for a safety-margin stock eroded by competitive pressure and replenished only by incidents.
  3. Leveson, N. G. (2011). Engineering a safer world: Systems thinking applied to safety. MIT Press. openlibrary.org/works/OL16540222W index STAMP treats safety as a control problem: accidents arise from inadequate control loops, not component failure. Supports modelling safety as a feedback structure with delays and degraded controllers.
  4. Rudolph, J. W., & Repenning, N. P. (2002). Disaster dynamics: Understanding the role of quantity in organizational collapse. Administrative Science Quarterly, 47(1), 1-30. doi.org/10.2307/3094889 System-dynamics model in which an accumulating stock of unresolved interruptions crosses a tipping point after which stress reduces resolution capacity and the system collapses. Direct SD precedent for an "unresolved safety issues" stock.
  5. Dekker, S. (2011). Drift into failure: From hunting broken components to understanding complex systems. Ashgate. openlibrary.org/works/OL15572374W partial Failure emerges from many locally rational decisions under resource scarcity and competition, with no single broken part; normalisation of small deviations.

Race-dynamics theory

  1. Armstrong, S., Bostrom, N., & Shulman, C. (2016). Racing to the precipice: A model of artificial intelligence development. AI and Society, 31(2), 201-206. doi.org/10.1007/s00146-015-0590-y Game model in which teams trade safety for speed; risk rises with the number of teams, with enmity, and (counter-intuitively) with information about rivals' progress. No empirical parameters; gives the sign structure for the race loop.
  2. Askell, A., Brundage, M., & Hadfield, G. (2019). The role of cooperation in responsible AI development. arXiv:1907.04534. arxiv.org/abs/1907.04534 "competitive pressures could incentivize AI companies to underinvest in ensuring their systems are safe"; frames responsible development as a collective-action problem and lists factors that raise cooperation (trust, shared downside, verification).
  3. Cave, S., & OhEigeartaigh, S. S. (2018). An AI race for strategic advantage: Rhetoric and risks. Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, 36-40. doi.org/10.1145/3278721.3278780 Race rhetoric is itself a risk factor: it can cause corner-cutting on safety and real conflict even without an actual race. Supports a perceived-gap (not actual-gap) driver in the race loop.
  4. Hendrycks, D., Schmidt, E., & Wang, A. (2025). Superintelligence strategy: Expert version. arXiv:2503.05628. arxiv.org/abs/2503.05628 Introduces Mutual Assured AI Malfunction (MAIM): "any state's aggressive bid for unilateral AI dominance is met with preventive sabotage by rivals", from cyberattacks to kinetic strikes on datacentres. A balancing loop the essay ignores: visible takeoff speed raises sabotage probability.
  5. Aschenbrenner, L. (2024, June). Situational awareness: The decade ahead. situational-awareness.ai/ Canonical racing case: talk has moved "from $10 billion compute clusters to $100 billion clusters to trillion-dollar clusters"; projects AGI by about 2027, a government "Project", and clusters of order 10 to 100 GW by decade end (cluster figures from secondary summaries of the essay).
  6. Richardson, L. F. (1960). Arms and insecurity: A mathematical study of the causes and origins of war. Boxwood Press. openlibrary.org/works/OL4123221W index Coupled differential equations dx/dt = k y - a x + g: each side's arming rate rises with the rival's stock (reaction), falls with own stock (fatigue/cost) and has a grievance constant. The SD precedent for a two-actor capability race; record confirmed via Open Library and a Springer retrospective (doi 10.1007/978-3-030-31589-4_4).

Weight security, distillation, fast-followers and the US-China gap

  1. Nevo, S., Lahav, D., Karpur, A., Bar-On, Y., Bradley, H. A., & Alstott, J. (2024). Securing AI model weights: Preventing theft and misuse of frontier models (RR-A2849-1). RAND Corporation. rand.org/pubs/research_reports/RRA2849-1.html partial Defines five security levels SL1 to SL5, each securing against progressively more capable actors (SL1 amateurs up to SL5 top-priority state operations). Sets the cost and lead time of closing the theft channel.
  2. Anthropic. (2026, February 23). Detecting and preventing distillation attacks. anthropic.com/news/detecting-and-preventing-distillation-att DeepSeek, Moonshot and MiniMax generated over 16 million exchanges with Claude via about 24,000 fraudulent accounts (MiniMax over 13M, Moonshot over 3.4M, DeepSeek over 150,000). After a new Claude release MiniMax redirected nearly half its traffic within 24 hours. Evidence that the capability-diffusion delay through distillation is days to months, not years.
  3. Edwards, J., & Emberson, L. (2026, May 29). Open models lag state-of-the-art closed models by 4 months. Epoch AI. epoch.ai/data-insights/open-closed-eci-gap Since January 2026 the best open-weight models trail frontier closed models by an average of four months (about 8 Epoch Capabilities Index points, roughly the GPT-5 to GPT-5.5 gap). Secondary reports: earlier Epoch estimate about 3 months (Oct 2025); cyber capabilities 6 to 7 months (Oct 2026).
  4. Somala, V., & Bye, L. (2026, September 24). How far behind Nvidia is Huawei? Epoch AI Gradient Updates. epochai.substack.com/p/how-far-behind-nvidia-is-huawei 2026: Huawei about 1.5 million chips vs Nvidia roughly 6 million; Ascend 950 delivers about 7x less compute than Nvidia B300; Huawei produces "roughly 25× less total compute than Nvidia in 2026"; chip-level gap three to four years, persisting to 2030.
  5. Grunewald, E., & Fist, T. (2025, June 11). Countering AI chip smuggling has become a national security priority. Center for a New American Security. cnas.org/publications/reports/countering-ai-chip-smuggling-h Between 10,000 and several hundred thousand AI chips smuggled into the PRC in 2024, median about 140,000; equal to 1 to 40% of China's training compute (median 10%) and 1 to 30% of inference compute (median 6%).

Export controls

  1. Center for Security and Emerging Technology. (2023). BIS 2023 update explainer (rules of October 7, 2022 and October 17, 2023). cset.georgetown.edu/article/bis-2023-update-explainer/ index Oct 7 2022 rule imposed China-wide licensing on advanced AI chips and fab tools; Oct 17 2023 update (effective Nov 17 2023) replaced the metric with Total Processing Performance (control at TPP 4800 or more, or TPP 1600 with performance density 5.92) and extended coverage to dozens of countries to block transshipment.
  2. Akin Gump. (2025, May). BIS rescinds AI Diffusion Rule and issues new guidance. akingump.com/en/insights/ai-law-and-regulation-tracker/bis-r index The "Framework for Artificial Intelligence Diffusion" (published January 2025, compliance date May 15 2025) was rescinded by Commerce on May 12 to 13 2025 before taking effect, with enforcement officials told not to enforce it. A global compute-allocation lever lasted four months on paper.
  3. The Next Web. (2026, August 28). Nvidia H200 China shipments and guidance. thenextweb.com/news/nvidia-h200-china-zero-revenue-guidance- index Policy oscillation: H20 licence requirement (Apr 2025), 15% revenue-share deal (Aug 2025), H200 sales allowed with 25% to the US government (Dec 2025), BIS case-by-case rule (Jan 2026). First H200 shipments only in mid-2026, under 1% of $89B quarterly data-centre revenue; guidance assumes zero China revenue. Control policy changes on a 3 to 6 month cadence.

Data-centre power

  1. International Energy Agency. (2025, April). Energy and AI. IEA, Paris. iea.org/reports/energy-and-ai partial Data centres used about 415 TWh in 2024 and are projected to roughly double to about 945 TWh by 2030 in the base case. IEA's April 2026 follow-up (secondary coverage) reports data-centre electricity demand rose 17% in 2025, with AI-specific demand expected to triple by 2030.
  2. Shehabi, A., et al. (2024, December). 2024 United States data center energy usage report (LBNL-2001637). Lawrence Berkeley National Laboratory; DOE release Dec 20 2024. energy.gov/articles/doe-releases-new-report-evaluating-incre US data centres: 58 TWh in 2014, 176 TWh in 2023 (4.4% of US electricity); projected 325 to 580 TWh by 2028 (6.7 to 12%). Author list beyond LBNL attribution not checked.
  3. Electric Power Research Institute. (2026). Powering Intelligence 2026: Updated scenarios of U.S. data center electricity use and power strategies (3002034696). esca.epri.com/products/3002034696/ index Data centres could reach 9 to 17% of US electricity by 2030 (from 4 to 5% today), a range 60% higher than EPRI's 2024 scenarios (4.6 to 9.1%). Virginia share could go from about 25% to 39 to 57%.
  4. Lawrence Berkeley National Laboratory. (2026). Queued Up: 2026 edition, characteristics of power plants seeking transmission interconnection as of the end of 2025. eta-publications.lbl.gov/publications/us-interconnection-que index Median time from interconnection request to commercial operation: 61 months for projects built in 2025 (55 months in 2024, 36 in 2015, 22 in 2008). 2025 edition: about 10,300 projects and 1,400 GW of generation in queues at end 2024. emp.lbl.gov returned 403.
  5. Turbomachinery International. (2026, September 11). Steam turbines and the gas turbine bottleneck. turbomachinerymag.com/view/steam-turbines-gas-turbine-bottle index GE Vernova ended Q2 2026 with 116 GW of gas backlog plus slot reservations (53 GW firm) and is selling 2031 delivery slots; Siemens Energy backlog 69 GW with lead times of three years or more; Mitsubishi new large-frame orders deliver around 2031. Wood Mackenzie Q2 2025: large power transformers average 128 weeks from order.
  6. Stewart, E., & Cottier, B. (2026, April 17, updated April 23). OpenAI Stargate: Where the US sites stand. Epoch AI. epoch.ai/blog/openai-stargate-where-the-us-sites-stand Seven US sites, about 9.6 GW planned, about 0.3 GW operating (Abilene only) in April 2026. Abilene: construction began Q2 2024, 1.2 GW projected Q4 2026 (about 2.5 years); six other sites started Q3 2025 to Q1 2026 and are projected for Q4 2028 (about 3 years). Realised to announced ratio about 3%.
  7. The Decoder. (2025). OpenAI has now signed $1 trillion worth of AI infrastructure contracts; and coverage of the Nvidia letter of intent (Sept 2025). the-decoder.com/openai-has-now-signed-1-trillion-worth-of-ai index OpenAI commitments reported at roughly $1 trillion, later described by Altman as $1.4 trillion for about 30 GW; Nvidia intends to invest up to $100B in OpenAI as 10 GW of Nvidia systems are deployed (circular financing). OpenAI executives put cost at about $50B per GW.
  8. Data Center Watch. (2026). Q1 2026 report. datacenterwatch.org/q1-2026 At least 75 US data-centre projects worth about $130 billion blocked or delayed by local opposition in Q1 2026, roughly matching all of 2025; over 300 state data-centre bills filed in six weeks; statewide moratoria proposed in 14 states. Secondary coverage: Q2 2026 added 45 projects and $68B; opposition groups grew from 396 (end 2025) to 843.

Semiconductor supply and capex

  1. Juniewicz, I., & You, J. (2026, October 29). New global AI computing capacity is doubling every 9 months. Epoch AI. epoch.ai/data-insights/ai-computing-capacity-additions-doubl Quarterly AI compute additions (H100-equivalents) grow 2.5x per year (90% CI 2.3 to 2.7), doubling time 9.1 months. Cumulative capacity grew 2.6x from Q2 2025 to Q2 2026, down from 3.5x the year before; year-on-year growth in additions fell from 2.8x to 2.2x (deceleration not significant, p = 0.26). Q2 2026 additions about 6.3 million H100e.
  2. Michael, A., & Cottier, B. (2026, May 14). Servers account for 60% of the total cost of ownership of a one-gigawatt AI data center. Epoch AI. epoch.ai/data-insights/ai-datacenter-cost-breakdown 1 GW facility: about $37.9B up-front capex (servers $21.2B, facility $11.4B, networking $4.9B) plus $0.9B a year opex (energy about $0.6B); annualised total cost of ownership about $8.5B a year. Power is a small cost share but a long lead-time item.
  3. VerWey, J. (2021, October). No permits, no fabs: The importance of regulatory reform for semiconductor manufacturing. Center for Security and Emerging Technology. cset.georgetown.edu/publication/no-permits-no-fabs/ partial About 635 greenfield fabs built 1990 to 2020 averaged 682 days from construction start to production; US average rose from 665 days (1990 to 2000) to 918 days (2010 to 2020), +38%. Excludes permitting, design and tool ramp, so order-to-volume is longer (3 to 5 years).
  4. ASML. (2026, January). Q4 and full-year 2025 results, summarised by Counterpoint Research. counterpointresearch.com/en/insights/asml-2025-revenue-up-16 index ASML recognised revenue on 48 EUV systems in 2025 (44 low-NA, 4 High-NA; EUV sales about EUR 11.6B, +39%); eight High-NA systems shipped cumulatively by end 2025. CFO: low-NA capacity "at least 60" systems in 2026 and over 80 the following year. EUV tool supply grows roughly 25 to 35% a year, far below 2.5x compute growth.
  5. TrendForce. (2026, June 15). TSMC CoWoS supply-demand gap report. trendforce.com/news/2026/06/15/news-tsmc-cowos-supply-demand index CoWoS advanced-packaging shortfall about 20% in mid-2026, expected to narrow to about 10% by end 2026. Analyst estimates: capacity about 120,000 to 130,000 wafers a month by end 2026 (Morgan Stanley), possibly about 260,000 by end 2028; Goldman 1.275M wafers in 2026 rising 114% in 2027. Figures are sell-side estimates, not TSMC disclosures.
  6. Yahoo Finance / 24/7 Wall St. (2026, May 1). Hyperscalers hit $700 billion in 2026 AI spending plans. finance.yahoo.com/sectors/technology/articles/hyperscalers-h index Combined 2026 capex of Amazon, Alphabet, Microsoft and Meta about $700B to $730B (Q3 2026 guidance per trackers: Amazon about $220B, Alphabet $195 to 205B, Microsoft about $170B, Meta $135 to 145B). Reported 2026 debt issuance by Microsoft, Alphabet, Amazon, Meta and Oracle about $326B, triple 2025; Amazon free cash flow fell 95%.

Physical limits of computing

  1. Ho, A., Erdil, E., & Besiroglu, T. (2023). Limits to the energy efficiency of CMOS microprocessors. arXiv:2312.08595. arxiv.org/abs/2312.08595 Geometric-mean estimate of maximum CMOS efficiency 4.7e15 FP4 per joule, "roughly two hundred-fold more efficient than current microprocessors". At a 2 to 3 year efficiency doubling time that is about 15 to 23 years of headroom within the CMOS paradigm.
  2. Koomey, J. G., Berard, S., Sanchez, M., & Wong, H. (2011). Implications of historical trends in the electrical efficiency of computing. IEEE Annals of the History of Computing, 33(3), 46-54. doi.org/10.1109/MAHC.2010.28 Computations per kWh doubled about every 1.57 years from 1946 to 2009. Later re-analysis by Koomey and Naffziger finds the doubling slowed to about 2.6 years after 2000; Ho et al. report about 2.7 years for GPUs.

Robotics and the industrial explosion

  1. Davidson, T., & Hadshar, R. (2025, May). The industrial explosion. Forethought. forethought.org/research/the-industrial-explosion Three stages: AI-directed human labour (physical output up "about 10X"), autonomous robot factories, nanotechnology. Initial robot doubling time "on the order of a year" under ideal conditions, more realistically "a few years, rather than months or decades"; "days or weeks as an upper bound". Robot efficiency doubles every "1-4 years"; "it will take years before we produce millions".
  2. TechRepublic. (2026). IFR counts 7,000 humanoid robot sales as workplace pilots stay small; and Counterpoint Research (2026) humanoid installations. techrepublic.com/article/news-humanoid-robot-sales-7000-work index Humanoid robots sold in 2025: about 7,000 (IFR), 13,000 to 18,000 by other counts (Counterpoint 16,000; IDC 18,000); China 80 to 90%. Unitree G1 lists at $13,500; Tesla shipped about 150 Optimus units in 2025. Industrial robot installations 603,000 in 2025 (+11%), about 5 million operating. Starting stock is 3 to 4 orders of magnitude below the human manual workforce.

Revenue funding compute

  1. Axios. (2026, September 29). Scoop: OpenAI's annual recurring revenue nears $70B. axios.com/2026/09/29/scoop-openais-annual-recurring-revenue- index OpenAI annualised run-rate almost $70B in late Sept 2026 (about $40B in August per Bloomberg); Anthropic reported run-rate above $65B at end July 2026 (about $30B in spring, $19B early 2026). Combined lab run-rate about $135B against about $700B hyperscaler capex. Run-rates are extrapolations, not audited revenue.
  2. Cahn, D. (2024, June 20). AI's $600B question. Sequoia Capital. sequoiacap.com/article/ais-600b-question/ Method: Nvidia run-rate revenue x2 (total data-centre cost) x2 (50% end-user margin) = $600B of required annual AI revenue vs roughly $100B actual; the hole grew from $125B (Sept 2023) to $500B. Applying the same x4 rule to 2026 capex gives a required revenue several times current lab run-rates.

Diffusion lags and macro constraints

  1. David, P. A. (1990). The dynamo and the computer: An historical perspective on the modern productivity paradox. American Economic Review, 80(2), 355-361. jstor.org/stable/2006600 partial Factory electrification took roughly four decades from the first central stations (early 1880s) to reach about 50% diffusion and show up in productivity (1920s), because gains required redesigning factories around unit drive.
  2. Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333-372. doi.org/10.1257/mac.20180386 GPTs need unmeasured complementary intangible investment, so measured productivity is first understated then overstated. Adjusting for software-related intangibles, the paper reports TFP about 15.9% higher than official measures at end 2017.
  3. Comin, D., & Hobijn, B. (2010). An exploration of technology diffusion. American Economic Review, 100(5), 2031-2059. doi.org/10.1257/aer.100.5.2031 15 technologies in 166 countries: countries adopt new technologies on average about 45 years after invention, with newer technologies adopted faster; adoption lags account for at least a quarter of cross-country income differences.
  4. Grundy, A., Breaux, C., & Khatiwoda, D. (2026, May 26). Large firms with at least 20 employees biggest AI users. U.S. Census Bureau. census.gov/library/stories/2026/05/ai-use-businesses.html BTOS: 19.8% of US businesses used AI as of May 3 2026 (17 to 20% over Dec 2025 to May 2026); 20 to 23% expect to within six months; 37% of firms with 250+ employees; Information 39.7%, Finance 33.9%, Retail about 14%. Question wording widened in Nov 2025, so the earlier roughly 10% series is not comparable.
  5. Bick, A., Blandin, A., & Deming, D. J. (2024). The rapid adoption of generative AI (NBER Working Paper 32966); FRED series RPSGENAIASSISTWRKHRSALL. doi.org/10.3386/w32966 39% of US adults aged 18 to 64 used generative AI in August 2024, faster than PC or internet adoption. FRED series from the same survey: share of work hours assisted by generative AI 5.56% (Q4 2025), 6.01% (Q1 2026), 6.27% (Q2 2026). Wide usage, shallow intensity.
  6. Anthropic. (2026, January and March). Anthropic Economic Index reports. anthropic.com/research/anthropic-economic-index-january-2026 index Claude.ai conversations 52% augmentation vs 45% automation (Nov 2025 data), 53% augmentation in Feb 2026; enterprise API use about 75% automation in the January report. Usage per capita strongly correlated with GDP and geographically concentrated.
  7. Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889-942. doi.org/10.1093/qje/qjae044 Field evidence from 5,000+ customer-support agents: AI assistance raised productivity about 14 to 15% on average, concentrated in novices. Task-level gain an order of magnitude smaller than "automated researcher" assumptions.
  8. Acemoglu, D. (2024). The simple macroeconomics of AI (NBER Working Paper 32487). doi.org/10.3386/w32487 Using Hulten's theorem: "no more than a 0.66% increase in total factor productivity (TFP) over 10 years", and "less than 0.53%" after adjusting for hard-to-learn tasks. Lower bound for the diffusion-limited scenario.
  9. Aghion, P., Jones, B. F., & Jones, C. I. (2017). Artificial intelligence and economic growth (NBER Working Paper 23928). doi.org/10.3386/w23928 With complementary tasks, growth is constrained by what is essential yet hard to improve (Baumol's cost disease): automating most tasks need not produce explosive growth because the non-automated share's cost weight rises.
  10. Korinek, A., & Suh, D. (2024). Scenarios for the transition to AGI (NBER Working Paper 32255). doi.org/10.3386/w32255 Task-complexity model: outcomes depend on the race between automation and capital accumulation; if all tasks become automatable, wages can collapse while output growth accelerates, and bottleneck tasks or slow capital accumulation cap growth.

Labour market, backlash and disempowerment

  1. Brynjolfsson, E., Chandar, B., & Chen, R. (2025, rev. November 2025 and August 2026). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab. digitaleconomy.stanford.edu/project/indicators/canaries/ ADP payroll data (about 25,000 firms, 4.6 million workers). Workers aged 22 to 25 in the most AI-exposed occupations: 13% relative employment decline (Aug 2025 draft), 16% (Nov 2025), "AI Employment Gap for Young Workers Has Widened to 19%" (Aug 2026). Authors call these descriptive, not causal.
  2. Kulveit, J., Douglas, R., Ammann, N., Turan, D., Krueger, D., & Duvenaud, D. (2025). Gradual disempowerment: Systemic existential risks from incremental AI development. arXiv:2501.16946. arxiv.org/abs/2501.16946 partial As AI replaces human labour and cognition, institutions lose the "implicit alignments with human interests" that came from needing people; effects are "mutually reinforcing across different domains" (economy, culture, states). A reinforcing loop that weakens every human-driven balancing loop.

Real-world learning speed limits

  1. IQVIA Institute. (2025). Global trends in R&D 2025. iqvia.com/blogs/2025/06/global-trends-in-r-and-d-2025-signs- index Total clinical programme duration 9.3 years in 2024 (peak 10.1 in 2022); an IQVIA 2026 white paper gives a median of about 8 years from Phase I to approval for 2020 to 2024 approvals, fastest decile 3 to 4 years. Serial-time floor for bio progress regardless of cognitive speed.
  2. Salvatier, J. (2017). Reality has a surprising amount of detail. johnsalvatier.org/blog/2017/reality-has-a-surprising-amount- Qualitative: practical tasks hide fractal detail that only surfaces on contact with reality, so plans made by pure reasoning systematically under-count steps. Justifies a physical-experiment delay that does not shrink with cognitive labour.

System dynamics: method

  1. Sterman, J. D. (2000). Business Dynamics: Systems Thinking and Modeling for a Complex World. Boston: Irwin/McGraw-Hill. ISBN 978-0-07-231135-8. mheducation.com/highered/product/business-dynamics-sterman.h Standard graduate text (S-curves and Bass ch. 9, path dependence ch. 10, delays ch. 11, expectations ch. 16, supply lines and oscillation ch. 17–20, validation ch. 21).
  2. Meadows, D. H. (2008). Thinking in Systems: A Primer (D. Wright, Ed.). White River Junction, VT: Chelsea Green. chelseagreen.com/product/thinking-in-systems/ Accessible stocks, flows and leverage points for students.
  3. Senge, P. M. (1990; rev. 2006). The Fifth Discipline. New York: Doubleday. penguinrandomhouse.com/books/163984/the-fifth-discipline-by- Systems archetypes (limits to growth, growth and underinvestment, shifting the burden, success to the successful, escalation, eroding goals).
  4. The Systems Thinker (n.d.). Systems archetypes at a glance. thesystemsthinker.com/systems-archetypes-at-a-glance/ One-page archetype reference for the site.
  5. Wolstenholme, E. F. (2003). Towards the definition and use of a core set of archetypal structures in system dynamics. System Dynamics Review, 19(1), 7–26. doi.org/10.1002/sdr.259 Rigorous two-loop basis for archetypes.
  6. Richardson, G. P. (1995). Loop polarity, loop dominance, and the concept of dominant polarity. System Dynamics Review, 11(1), 67–88. doi.org/10.1002/sdr.4260110106 Formal polarity and dominant-polarity definition, the basis of the 3.9 fallback.
  7. Sterman, J. D. (2002). All models are wrong: Reflections on becoming a systems scientist. System Dynamics Review, 18(4), 501–531. doi.org/10.1002/sdr.261 Model humility and boundary critique for students.
  8. Sterman, J. D. (1989). Modeling managerial behavior: Misperceptions of feedback in a dynamic decision making experiment. Management Science, 35(3), 321–339. doi.org/10.1287/mnsc.35.3.321 Supply-line under-weighting, the cause of oscillation.
  9. Booth Sweeney, L., & Sterman, J. D. (2000). Bathtub dynamics: Initial results of a systems thinking inventory. System Dynamics Review, 16(4), 249–286. doi.org/10.1002/sdr.198 Stock-flow misconceptions (revenue versus cash, orders versus capacity).
  10. Cronin, M. A., Gonzalez, C., & Sterman, J. D. (2009). Why don't well-educated adults understand accumulation? Organizational Behavior and Human Decision Processes, 108(1), 116–130. doi.org/10.1016/j.obhdp.2008.03.003 Evidence for teaching accumulation explicitly.
  11. Ford, D. N. (1999). A behavioral approach to feedback loop dominance analysis. System Dynamics Review, 15(1), 3–36. doi.org/10.1002/(SICI)1099-1727(199921)15:1<3::AID-SDR159>3. Loop knockouts and atomic behavior patterns (3.9).
  12. Schoenberg, W., Davidsen, P., & Eberlein, R. (2020). Understanding model behavior using the Loops that Matter method. System Dynamics Review, 36(2), 158–190. doi.org/10.1002/sdr.1658 The recommended method.