Bottlenecks

Open the Products stock

1. #1Application development effort does not vanish. The "bitter lesson" applies to methods, not products. Real applications still need hand-built business logic, frontends, integrations, and evaluation. ANT, Slowing, Pivot
   Enterprise pilots with measurable P&L impact ●consistent with normal
2. #2Proof-of-concept to product gap. A demo that works 90% of the time is a capability, not a product. Executives see quick prototypes and miss the 90% of work needed to finish them. Pivot, SWE, Slowing
   Enterprise pilots with measurable P&L impact ●consistent with normalDeveloper productivity uplift (METR RCTs) ●consistent with normal
3. #3Capability-reliability gap. Accuracy has risen sharply while reliability has moved only a few points. A model that is right 70% of the time but fails unpredictably cannot remove the human from the loop, and reliability fixes are application-specific. ANT, Slowing, Left, OWE
   50%/80% horizon ratio ●consistent with normalMETR 80% time horizon ▲faster than normal
4. #4Compounding unreliability in agents. An agent making dozens of LLM calls with even a 2% error rate per step becomes useless end to end. AgentsMatter
   Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging50%/80% horizon ratio ●consistent with normal
5. #5Users expect software-like determinism. People expect AI products to behave like software. Stochastic outputs violate that expectation, and it is unclear whether determinism can be engineered in. Pivot, AGI
6. #6Missing interface paradigms. Current AI products are like PCs before the GUI. Higher-bandwidth interfaces that let users supervise without constant interruption have not been invented. Slowing, Pivot
7. #7Domain-specific product design. General capabilities have to be made useful one domain at a time, which requires deep domain knowledge to find the adoption hurdles (Cursor's code-verification UI is their example). AGI
   Enterprise pilots with measurable P&L impact ●consistent with normal
8. #8AI companies neglected product engineering. Labs assumed generality exempted them from UX and software engineering, and are rediscovering that both are hard. Slowing
9. #9Cost of inference. Cost, not capability, blocks many applications, especially agentic workflows that call models hundreds of times. Cost also sets accuracy, since retries improve success. Scaling, Pivot, AgentsMatter, OWE
   Price per unit of capability (cheapest METR-hour on OpenRouter) ○not yet measurable
10. #10Judgment tasks resist evaluation. The tasks whose automation would most transform a profession (legal filings, research) have no single correct answer, so they are the hardest to evaluate and improve. ANT, Legal
11. #11No verifiers in open domains. AI became superhuman at chess because of fast, accurate feedback. Law, medicine, and science have no such verifier, so progress is limited to automatically checkable tasks. Moravec, Agents26
12. #12Real-world knowledge caps reasoning. AI reasoning in medicine is limited by the medical knowledge that exists, the same limit humans face. Moravec
13. #13Brittleness outside closed domains. Systems that excel in narrow demos go off the rails in open-ended settings. Moravec, OWE
14. #14Agents lack judgment for open-ended work. In their shadow evaluations, frontier agents given real research questions lacked judgment, creativity, and the ability to backtrack or respond to feedback. Agents26, OWE
15. #15Non-functional requirements go unmet. Agent-built software sacrifices quality, maintainability, and security, and successes often rest on human-built test suites, thousands of lines of prompt, or memorized training data. OWE, Google
16. #16Benchmarks lack construct validity. Benchmarks overweight what models are good at, miss contextual reasoning, and encourage agents that score well without being useful. Vendor claims cannot be independently checked. Legal, AgentsMatter, Google
    Developer productivity uplift (METR RCTs) ●consistent with normal50%/80% horizon ratio ●consistent with normal
17. #17Real-world messiness cannot be simulated. Whether a system can automate a job is only knowable after diffusion, because no lab environment captures the world's complexity. AGI, OWE
    Developer productivity uplift (METR RCTs) ●consistent with normal

Open the Early adoption stock

1. #18Deployment is not diffusion. Instant availability to hundreds of millions says nothing about how many people use a capability, for how long, or for what. Guide, ANT
   Share of work hours assisted by generative AI ●consistent with normalUS firms using AI (Census BTOS) ●consistent with normal
2. #19Low intensity of use. Most users use generative AI infrequently. Nearly a year after release, under 1% of ChatGPT users touched thinking models on a given day. ANT, Guide
   Share of work hours assisted by generative AI ●consistent with normalAugmentation vs automation share (Anthropic Economic Index) ●consistent with normal
3. #20Human learning curves. Taking advantage of AI requires continual learning of new workflows and topics. The curve is steep and time-consuming, and it moves at human speed. Guide, Left, AGI
   Share of work hours assisted by generative AI ●consistent with normal
4. #21Workflow and habit change. The hard part of adoption is users adapting their workflows, which cannot even start in the first months after a launch. ANT, Guide, Legal
   US firms using AI (Census BTOS) ●consistent with normalUS businesses paying for AI (Ramp AI Index) ◐emerging
5. #22Most adoption decisions are "no.". Because deployment is instant, people constantly face adoption choices, and the vast majority of the time they decline, for rational and irrational reasons. Guide
6. #23Cost of supervising agents. Agentic work requires constant human supervision, which is time-consuming and mentally exhausting, so realized productivity lags. SWE, OWE
   Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging
7. #24Skill erosion and the dependence spiral. Using AI for tasks one has not mastered erodes skill and control, so responsible users limit offloading. Vendor-specific skills replace unaided ones. Left, Stack
8. #25Skilled judgment gates "democratization.". Past waves of no-code failed because the barrier is not syntax but the judgment and accountability needed to make good decisions. SWE
9. #26Verification needs domain expertise. Checking whether an agent actually succeeded at an open-ended task takes deep expertise and time, and long logs cannot be fully reviewed. OWE, Left
10. #27Privacy resistance. Useful assistants need access to sensitive personal and organizational data, which triggers outcry and regulatory limits on data sharing. Pivot, ANT

Open the Adaptation stock

1. #28Organizational change is slower than individual change. For past general-purpose technologies this stage took decades. It has barely begun even in software engineering. ANT, Guide, Left, Pivot
   US firms using AI (Census BTOS) ●consistent with normalShare of work hours assisted by generative AI ●consistent with normal
2. #29Drop-in replacement fails. Like electrification, benefits only arrive once work is reorganized: new layouts, processes, hiring, and training. The restructuring is discovered by experimentation, not designed in advance. ANT, Left
   Enterprise pilots with measurable P&L impact ●consistent with normalDeveloper productivity uplift (METR RCTs) ●consistent with normal
3. #30Tacit organizational knowledge. Much of what organizations know is unwritten and not in training data. It has to be made available to models gradually through adoption, sector by sector. ANT, Left, SWE
4. #31Integration into existing systems. Models must be wired into the many systems organizations already run, a downstream task no model release solves. Enterprise data is locked in Salesforce, Workday, and SAP. Left, Stack
5. #32Task specification is human labor. Unambiguously saying what is wanted is a large share of the work. Brooks's "essential complexity" of specification is untouched by AI. ANT, SWE
6. #33The "decide" and "deliver" layers do not compress. Understanding requirements and taking accountability for what ships remain human. Once a decision is delegable it stops being a competitive advantage, so human judgment migrates upward. SWE, Left
7. #34Staged trust. Organizations grant AI access to consequential decisions only after it proves reliable in less critical contexts. ANT
8. #35Human-in-the-loop erodes benefits. Requiring approval of every action mostly destroys the value of automation, so it degrades into rubber-stamping or is outcompeted. ANT
   Human interventions on 4–8 hour agent tasks (self-reported) ◐emergingDeveloper productivity uplift (METR RCTs) ●consistent with normal
9. #36Collective action and sclerotic institutions. Structural change requires solving coordination problems or reforming institutions, which is far less predictable than user behavior (air traffic control is their example). Guide, AGI
10. #37Underinvestment in complements. Diffusion depends on public goods the private sector underprovides: AI literacy, workforce training, digitization, open data. ANT, AGI
    US total factor productivity, private nonfarm business ●consistent with normal
11. #38National diffusion capacity. Following Jeffrey Ding, countries differ enormously in their ability to spread innovations through the economy, and this, not who reaches a milestone first, sets economic outcomes. ANT, AGI
    US firms using AI (Census BTOS) ●consistent with normal
12. #39Bespoke deployment. Transforming workplaces requires forward-deployed engineers and consulting partnerships, labor-intensive work outside the product. Stack, Guide
13. #40Enterprise resistance to lock-in. Embedded "digital workers" create extreme lock-in and target labor budgets, so enterprises resist, withhold data for training, and demand portability. Stack
14. #41Deskilling breaks the talent pipeline. Automating entry-level tasks removes the experiences through which juniors build expertise, a cost that emerges only over time. Legal, SWE
    Entry-level employment shortfall in AI-exposed occupations ◐emergingRecent-graduate unemployment rate (NY Fed) ●consistent with normal
15. #42Millions of small adaptations. Impact arrives through countless mundane process and policy tweaks, not a single leap. AGI, ANT
    US labour productivity, year on year ●consistent with normal

Open the Adaptation stock

1. #43Regulation in high-consequence domains. FDA, EU AI Act, and sector rules make deployment slow by design, and regulation often outright prohibits current AI use in productive settings. ANT, Left, SWE
2. #44Liability uncertainty. Unclear application of liability law deters adoption. Clear rules (FAA drone rules in 2016) spur it. ANT, SWE
3. #45Professional guardrails. Malpractice liability, professional codes, and device regulation keep doctors from delegating decisions to chatbots. Guide
4. #46Regulation that freezes categories. Rules insensitive to experimentation reify business models and organizational forms prematurely, and binary automated/not-automated rules discourage new oversight designs. ANT
5. #47Misplaced burdens on model developers. Regulation blind to the developer/deployer split saddles general-purpose model makers with context-specific obligations. ANT
6. #48Unauthorized practice of law. UPL rules and their state-by-state vagueness deter new entrants, and bar associations cite model unreliability as grounds for caution. Legal
7. #49Law firm ownership rules. Restrictions on who can own or share fees with law firms block outside capital and scaled business models. Legal
8. #50Incumbent resistance to reform. Lawyers and commerce groups lobby to narrow regulatory sandboxes despite scant evidence of harm. Legal
9. #51Human adjudication time. The speed of human judges, lawyers, and clients caps how fast legal processes can move. Adding judges is politically fraught, and Article III likely requires human judges. Legal
10. #52Courts may restrict access. Flooded courts respond by tightening doctrines to keep litigants out or by banning AI, counteracting the gains. Legal
11. #53Policy-mandated human steps. Platform policies require developer accounts, 2FA dialogs that block synthetic input, and a human pressing publish. OWE
12. #54State capacity sclerosis. Veto points and proceduralism hobble governments' ability to deploy AI, coordinate resilience, or even provide services they could. Gov, ANT
13. #55Policy backlash. Self-driving cars took 15+ years to deploy and now face bans because policymakers failed to prepare. Public anxiety and backlash from rushed deployments are headwinds. Moravec, ANT, Left, Pivot

Open the Adaptation stock

1. #56Clinical trials and real-world experiments. Treatments need trials with thousands of people over 10 to 15 years. Progress needs data from real experiments that faster AI cannot speed up, and society will not let AI run large-scale experiments on people. Left, AGI, Agents26
2. #57Costly, unsimulable errors. Where mistakes are expensive and the world cannot be simulated (driving), safety throttles each iteration of the improvement loop. ANT
3. #58AI targets non-bottleneck steps. In an already technological, regulated world, the workflow parts AI improves were optimized by earlier waves. The true bottlenecks resist for external reasons. Guide
4. #59Physical and social task bundles. Jobs bundle inspection, loading, paperwork, and negotiation with the "core" task. Bio-harm still depends on materials, equipment, and tacit know-how. ANT, Gov
5. #60Slow external review. Deployment depends on systems the developer does not control. An app took 45 minutes to build and ten days to pass App Store review. OWE
6. #61Infrastructure, energy, and compute. Grid capacity constrains training and inference. Token supply is scarce today. Breakthroughs deploy slowly when supporting infrastructure does not exist yet (steam to electric took 40 years). ANT, Stack, Agents26, Moravec
   Circular financing scale ▲concentratingCloud backlog (remaining performance obligations) ▲concentrating
7. #62Inherent limits on prediction. Many tasks resemble long-range weather forecasting, where mathematical limits are already reached. Forecasting and persuasion have high irreducible error. Left, ANT, SciTool
8. #63Unrecognized latent bottlenecks and Amdahl's law. Constraints like RL environments or energy only become visible once they bind, and if many hard steps remain, a hundredfold speedup on the AI-amenable parts yields a small overall gain. Agents26

Open the Adaptation stock

1. #64Safety limits in high-consequence tasks. They predict slow diffusion will remain the norm where errors matter, enforced by regulators and by organizations' own caution. ANT, AGI
2. #65Business incentives against unsupervised deployment. Poorly controlled AI is bad business. Unrecoverable failures like deleting production data make companies pull back from hasty automation. ANT, AGI, Left, Science
   Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging50%/80% horizon ratio ●consistent with normal
3. #66The general-purpose / high-stakes / automated trilemma. For now an agent can be only two of the three, which keeps AI a collaboration technology rather than an automation one. Left
4. #67Legibility and control. Removing humans from task boundaries to let AI run end to end reduces oversight, so autonomy is checked by the need to keep systems legible. ANT
5. #68Society's choice to keep humans accountable. Liability law and sector regulation are speed controls society can strengthen deliberately, regardless of capability. SWE, Left
6. #69Security risks of capable scaffolds. Prompt injection, data leakage, and agent worms must be solved before assistants get broad access; scaffolds capable enough to test the frontier carry risks that may prevent use. Pivot, OWE

Open the Adaptation stock

1. #70Automation devalues the automated task. Once a task is automated its value collapses and humans move to unautomated tasks, so the AGI goalpost keeps moving. ANT, AGI
   Price per unit of capability (cheapest METR-hour on OpenRouter) ○not yet measurableConsumer surplus from generative AI (willingness to accept) ▼dispersing
2. #71Competitive arms races dissipate gains. In law and science, competition is so intense that productivity gains fuel escalation rather than societal value. Guide, Legal
   Price per unit of capability (cheapest METR-hour on OpenRouter) ○not yet measurableConcentration of business AI spend across labs (Ramp, HHI) ○unmeasuredSemiconductors' share of stack operating income ▲concentrating
3. #72Credence goods. Buyers cannot verify the quality of legal work or judgment-heavy AI output even in hindsight, so they rely on prestige and trust, and normal price competition fails. Legal, Stack
4. #73Input-based incentives. Billable hours reward more hours and more output regardless of outcome, absorbing efficiency gains. Legal
5. #74Growth is capped by the slowest sector. AI's uneven sectoral effects mean long-run growth is bottlenecked by wherever it diffuses least. AGI
   US total factor productivity, private nonfarm business ●consistent with normalUS labour productivity, year on year ●consistent with normal
6. #75Commodity trap and recoupment. Undifferentiated models push inference to marginal cost, outcomes are hard to measure so value pricing stalls, and incumbents own distribution. This limits how far labs can fund diffusion. Stack
   Lab recoupment ratio ◐emergingPublic cloud software valuation multiples (Clouded Judgement) ◐emergingApplication layer's share of enterprise generative-AI spend (Menlo) ◐emerging
7. #76Fixed consumer attention. More apps do not increase usage, capping AI's impact on consumer software. SWE
8. #77Publisher distrust. Media organizations' justified wariness and inability to bargain collectively limit the use of journalistic content. ANT

Open the Adaptation stock

1. #78Production of findings is not the bottleneck. Cheap generation of results will not unlock progress because producing findings is not what limits science. AGI, Science
2. #79Overproduction drowns novel work. Attention is finite. As volume explodes, novel work is lost and AI search concentrates attention on already-famous papers. Science
3. #80Publish-or-perish incentives. Institutions reward measurable production, making researchers risk-averse and making AI a tool for chasing metrics. Reform is blocked by inertia and Goodhart's law. Science
4. #81Scientists lack software engineering practice. Most AI-for-science is software work, but fields have not adopted testing, version control, or code review, so leakage and other errors affect hundreds of papers. Science, SciTool
5. #82Science does not self-correct. Code and data go unshared, reviewers do not check code, retractions are near zero, and there is no incentive to debunk. Science, SciTool
6. #83Prediction without understanding. AI improves predictive accuracy without explanation, like epicycles, so wrong theories can persist and human understanding, which is the point of science, erodes. Science, Left, SciTool
7. #84Epistemology change exceeds field capacity. Adopting AI requires re-litigating validity per field and model type, which no field can do in a couple of years. SciTool
8. #85Misdirected tools and funding. AI-for-science tools chase headlines rather than real bottlenecks like error detection, and are evaluated on time saved, not on understanding. Science, SciTool

Open the Return arrow stock

1. #86External bottlenecks are immune to self-improvement. Limits on AI's power sit in deployment, not in the system's design, so improving the design cannot overcome them. Guide, ANT
   Agent-workdays per human workday in frontier research (self-reported) ◐emergingHuman interventions on 4–8 hour agent tasks (self-reported) ◐emerging
2. #87Herding, compute, and declining openness in methods research. Research herds around fashionable ideas, compute and cost constrain new paradigms, and the industry's culture of open sharing is fading. The AI community may be unusually bad at finding new paradigms. ANT, Guide
3. #88Data exhaustion. Readily available data is spent, and more data costs ever more in money, licensing, and reputational risk. Continued scaling is a business decision, not a technical given. Scaling, Slowing
4. #89Evaluation does not scale. Every capability increase creates new demand for domain-specific evaluation, which resists automation and absorbs human effort. Left