46 thoughts on the near future

bayes · tweet · 2026-06-30 Reader 🖥️

theoretical · mixed — A numbered list of 46 discrete predictions organized under eight thematic headers (Intelligence; The intelligence supply chain; Robotics; Progress; Capital and Production; Culture and Psychology; Coordination; Power, violence, security, liberty). Individual items range from bare assertion ('X will happen') to occasional explicit chains of reasoning (the algorithmic-depth moat argument, the MAD argument, the jobs/disintermediation argument). There is no single connected essay-level argument; the unity is a shared worldview and timeline applied across many independent domains, not a step-by-step case for one conclusion.

Structure

Key terms

• algorithmic OOMs (orders of magnitude)
  Discrete, multiplicative jumps in the efficiency of converting compute into 'intelligence output' from algorithmic (not hardware) progress alone; the author estimates 4-10 such jumps remain before the universe's physical/computational limits are reached.
• early takeoff
  The claim that AI-improving-AI (recursive self-improvement) has already begun and is accelerating, distinct from a future hypothetical threshold.
• algorithmic depth
  How much unexploited efficiency remains latent in deep-learning practice and theory itself, as distinct from raw compute scale — the variable that determines whether labs' research advantages are durable or catch-up-able via distillation.
• Move 37 moment
  A generalized reference to AlphaGo's famous move 37 — a superhuman, initially incomprehensible insight — extended to predict that every domain of technical human endeavor will have such a moment, and that these will rapidly become unremarkable.
• vulnerable world (Bostromian)
  Invokes Nick Bostrom's 'vulnerable world hypothesis' — that some technology levels enable catastrophic misuse by default — as a live, unresolved possibility for the current AI moment, requiring extreme surveillance/control ('panopticon') absent another solution.
• permanent underclass
  A durable segment of humanity locked out of contributing economically to AI-automated production — defined by the author specifically as restricted agency rather than necessarily restricted income, distinguishing it from ordinary poverty.

Propositions

• explicit⚡ disputed & corroboratedThere are many algorithmic orders-of-magnitude (OOMs) of gains left in the production of intelligence — plausibly four to ten, with going beyond ten implausible but not impossible — and realizing them would make an already-underway 'early takeoff' look far weirder than markets, governments, and even frontier AI labs are currently pricing in.conf: low
  contradictsSorry, Sam and Elon, we have not reached the Singularity · p7 — Bayes asserts we're already in an early takeoff with a big unanticipated jump coming; Marcus argues current AI is nowhere close to I.J. Good's intelligence-explosion threshold, managing at most 2 of 10 benchmark tasks in his own bet.
  supports ←Automatisert KI-koding vil endre alt ("Automated AI coding will change everything") · p1 — Both anticipate substantial further AI capability growth (this piece via LessWrong's 'superhuman coder' threshold, the other via remaining algorithmic orders-of-magnitude), converging on the same general forward trajectory from different source material.
• explicitcontestedWe are already in an early phase of AI recursively improving AI (RSI), driven by falling researcher-time opportunity costs since agents can now be sent to explore speculative research directions autonomously.conf: medium
  contradictsArvind Narayanan on Why AI Isn't All That Revolutionary · p1 — Bayes claims we are already in early recursive self-improvement/takeoff; Narayanan argues AI's real-world transformation will unfold over decades, not months or years, due to adoption bottlenecks even as capability keeps increasing.
• explicitThere will be 'Move 37' moments — surprising, superhuman insights beyond current human conception — across every domain of technical human endeavor, arriving so rapidly in succession that they will quickly stop feeling remarkable.conf: low
• explicitLong-horizon agentic tasks do not require correspondingly long-horizon training, because error-correction generalizes across scales — this is offered as a rebuttal of Yann LeCun's (1-e)^n error-accumulation argument and as the explanation for recent gains on METR's agent-task-length benchmark.conf: low
• explicitAn engineering-grade science of deep learning (e.g., a science of scale-invariance letting small-scale experiments predict large-scale outcomes) is imminent and will accelerate progress toward algorithmic maturity faster than currently expected.conf: low
• explicitWhether frontier AI labs retain a durable competitive moat depends on how algorithmically 'deep' deep learning turns out to be: if practice stays shallow, secrets become cheap to discover and distillation/open source catches up over time; if it deepens with scale, each increment of automation buys an increasingly unreachable advantage — and which branch we're on is currently unknown.conf: medium
• explicitCompute will remain a highly contested resource for at least a few years but will progressively commoditize as capital, fabs, and energy production scale up.conf: medium
• explicitThe intelligence supply chain's current centralization around labs will shift toward advantages in capital, compute access, proprietary data and product as labs automate away their researcher-driven algorithmic edge — assuming labs don't lock down their internal AI-researcher models.conf: low
• explicitAcademic and open-source AI research may languish even with capital available, because the opportunity cost of directing scarce compute toward non-frontier research (versus renting it to labs) will keep them priced out of the resources they'd need.conf: medium
• explicitRobotics will undergo two step-change moments analogous to ChatGPT (Nov 2022) and Opus 4.5 (Nov 2025) for language models, with the gap between them likely shorter than the roughly three years those language-model moments took.conf: low
• explicitPhysical scale-up to on the order of 100 million robots produced per year by 2030 is a reasonable expectation if capital and algorithms move quickly, reasoning by analogy to ~100M cars/year and ~1B smartphones/year already manufactured, with 10 million/year considered clearly achievable per existing drone-market precedent.conf: low
• explicitCurrent apparent hard limits on robotics (poor sample efficiency, hardware cost for hands/motors, the fractal complexity of the physical world, unrecorded tacit knowledge like plumbing) will dissolve as research scaling laws are exploited to their limit, the same way comparable limits fell for language and code.conf: low
• explicitGlobal demand for robots is easily in the tens of billions of units once all form factors are summed, because there is enormous physical work worth automating and markets will pursue this largely unimpeded.conf: low
• explicitScientific progress will increasingly come from automated and virtualized labs — combinations of neural models, explicit simulations, and real-world experiments — that remove bottlenecks in biology, materials science and other domains, rather than from human-run wet-lab work alone.conf: medium
• explicitCivilizational progress itself follows an S-curve-like 'progress law,' and whether we are near a saturation point (technological maturity) or still early on a much deeper, largely unexplored tech tree is currently unknown — 'most likely' saturating, but the timing could be close or far.conf: medium
• explicitIntensified capitalism driven by more capital and more intelligence will drive prices for most important goods (AI, food, housing, medicine, electronics, entertainment, travel) toward near-zero marginal cost, absent people getting in the way of markets.conf: low
• explicitPhysical supply chains — mining, shipping, factories, factory labor, distribution, and their own maintenance and scaling — will be comprehensively automated.conf: medium
• explicitcorroboratedWhat fraction of humanity retains paid employment long-term is a genuinely open question that both maximalist ('high') and zero forecasts get overconfidently wrong; a given job's survival will depend on how disintermediated, outcompetable, or fungible it is (illustrated by contrasting a TikTok influencer and a doctor, both with zero intermediate layers, against a factory worker with many) — and the whole scenario assumes no precipitous demand-side collapse if productivity and UBI/UHI can't keep pace with job loss.conf: medium
  supports ←SWLW #711: The laws of this world (game theory), Should you still become a Software Engineer in 2026, and more. · p4 — Ellenbogen's 'where's your margin once it goes to zero' is the same open question the source piece poses about what fraction of humanity retains paid employment as AI absorbs knowledge work.
• explicitA 'permanent underclass' — best understood as durably restricted agency rather than necessarily restricted income — is a real possible outcome; for most people this could ultimately be tolerable given how restricted agency already is under modern society, though it will require psychological adaptation that may be painful.conf: low
  extendsArvind Narayanan on Why AI Isn't All That Revolutionary · p14 — Both flag genuine uncertainty about how AI-era wealth gets distributed to displaced humans; bayes extends Narayanan's income-focused doubt by reframing the risk specifically as restricted agency rather than restricted income.
• explicitAbundant intelligence and automation will let humans engineer durably well-adjusted psychologies within decades — compressing what would otherwise be a thousand years of psychiatric and psychological innovation — making crude wireheading an overrated risk relative to more skillful mind-engineering options that will become available.conf: low
• explicitThere is a specific hypocrisy ('doublespeak') among the very wealthy who claim AI will benefit everyone while being unwilling to trade their own position for a randomly-selected person's life even years from now; this pattern risks building an unjust world and deserves more public discussion than American politics currently provides — a critique the author explicitly applies to himself as well.conf: medium
• explicitElon Musk is likely to be the first quadrillionaire, given plausibly >>1000x future demand for chips, robots, and spaceships, a large share of which he can capture.conf: low
• explicitAt least some international coordination on AI (e.g., treaties, GPU counting) is likely valuable for slowing adversarial military/governmental power accumulation with minimal impact on beneficial research, but may not be achievable because GPUs are too broadly and generally useful — unlike nuclear weapons, which almost no rational actor actually wants to use.conf: medium
  supportsArvind Narayanan on Why AI Isn't All That Revolutionary · p9 — Both argue that durably controlling AI capability diffusion via coordination is structurally hard — bayes because GPUs are too broadly useful to restrict by treaty, Narayanan because doing so would require sustained authoritarian-level global control.
• explicitA coordinated pause or slowdown of frontier AI production is more justifiable today than it was in 2023, because the 'it will be squandered' objection is weaker now that automated engineering (though not yet automated research) exists — but the author personally opposes a pause at this time because it would disrupt other necessary parts of navigating the transition and adversaries are real.conf: medium
• explicitOur world may be 'vulnerable' in Bostrom's sense — possessing degrees of freedom too dangerous to coordinate around fast enough while preserving current governance and liberty norms, absent something like a panopticon — and worlds fitting this pattern tend to end up bad for most people via power-accumulation slippery slopes.conf: low
• explicitAI diffusion will happen at some speed greater than zero regardless of rate-limiting attempts, because there are already many computers in the world and the FLOPs-to-intelligence exchange rate will never again be as unfavorable as it is today.conf: medium
• explicitA 'permanent overclass' is the logical mirror of a permanent underclass, and its historical justification has always ultimately rested on violence-backed domination; advanced AI could produce a world where humans have no agreed, justifiable basis to govern each other, though this will never be fully true.conf: low
• explicitInstitutions will face transformation pressure from many directions, some of which could lead to tyranny — via safety-justified power grabs or benign power-creep toward powerful-AI-plus-automated-military-supply-chains-plus-automated-weapons.conf: medium
• explicitcorroboratedRobotics capability at scale creates real coup/takeover risk beyond what purely computer-based AI models present, plus new cyberattack surface and vectors, and these risks warrant serious mitigation work.conf: medium
  supports ←This Week in AI: When agents outnumber people · p6 — Both flag that AI-driven physical/robotic action raises the stakes of error and risk beyond what purely digital/computer-based AI presents.
• explicitMutually assured destruction (MAD) is not a structural given going forward, because it is based on 20th/21st-century technology and a compressed, roughly millennium-scale wave of technological change is coming that could produce decisive strategic advantage; this is one of the most serious topics available for discussion and deserves serious treatment rather than the dismissiveness it has received.conf: medium
• explicitMilitary, police, and core law-enforcement functions will be automated and become smarter than humans.conf: medium
• explicitFrontier AI labs could plausibly be nationalized in a strong sense despite this being incompatible with the current American political system in the author's view; private-company control differs in kind from state control because companies cannot legally enact violence directly, while unilateral federal power over this capability is itself extremely risky.conf: low

Arguments

Narratives

• Superhuman AI doctors available for $20/month plus a la carte testing, offering significantly improved medical technology, contrasted against the possibility that medical cartelization keeps 'being a doctor' a great profession regardless.
  → Makes the abstract disintermediation/fungibility argument about job survival concrete for a specific, high-status profession, illustrating how demand for even elite cognitive labor could collapse — or be artificially propped up by institutional capture rather than technical necessity.
• A rhetorical question posed as a resource-allocation dilemma: are GB300 GPUs more valuable serving GLM5.2 or Fable, and is it more valuable doing non-frontier research in an academic lab or building 'Mythos 2' inside Anthropic?
  → Illustrates the opportunity-cost argument for why academic and open-source AI research may remain compute-starved even with capital in hand — the market allocates compute to wherever demand (and pay) is greatest, which currently means the labs.
• A three-way comparison: a TikTok influencer has zero intermediate layers between them and their audience, a doctor also has zero layers, but a factory worker sits behind many layers of supply chain.
  → Supplies the concrete typology behind the disintermediation heuristic used to predict which categories of job survive AI-driven automation.

Questions addressed

• ✅ How accurately do current institutions, markets, and governments understand the trajectory of AI capability?
  The author asserts they are systematically extrapolating from the wrong (too-conservative) baseline and will be blindsided, but this is stated as a claim rather than demonstrated with evidence.
• ⭕ What happens to human employment as AI and robots automate both cognitive and physical labor?
  The author explicitly frames this as an open question and declines to give a confident percentage, offering a disintermediation heuristic instead of a resolved answer.
• ⭕ Can international coordination (treaties, GPU counting) durably control the spread and risks of advanced AI capability?
  The author says such coordination is 'likely a good idea' but explicitly doubts it will be achieved, because GPUs (unlike nuclear weapons) are too broadly useful to restrict by consensus.
• ✅ Is a pause or slowdown of frontier AI development currently advisable?
  The author states a personal position against a pause at this time, citing disruption to other necessary parts of the transition and the reality of non-complying adversaries — reasons given, but not developed into a full argument.
• ⭕ Could AI-driven change produce a stable but unjust world (a 'permanent underclass' and its 'overclass' mirror), and is this being adequately discussed?
  The author explicitly flags this as under-discussed, criticizes American politics for handling it poorly, and self-implicates in the wealth-doublespeak critique — but does not resolve what should be done about it.
• ⭕ Will humans retain a justifiable basis to govern one another in a world with superhuman AI?
  The author raises the possibility that humans may have 'no justifiable rights to govern' but explicitly declines to resolve it, framing it as something that 'might become more important to think about.'
• ⭕ Does mutually assured destruction (MAD) remain a stable strategic equilibrium as AI accelerates military and technological capability?
  The author asserts MAD 'is not a given' and calls it one of the most serious topics available, but offers no concrete mechanism, timeline, or path to a resolution.
• ⭕ Will frontier AI labs end up under direct state (nationalized) control?
  The author offers a personal lean (skeptical of nationalization's compatibility with the American system) but explicitly says paths to it 'don't seem off-limits,' leaving the question open.

Gaps

• uninformedThe central quantitative claim — 'many algorithmic OOMs left... maybe up to ten, with four to seven seeming more likely' — is stated with false numeric precision but sourced to nothing beyond the author's own intuition ('my guess,' 'I suspect'). No historical algorithmic-efficiency growth-rate data (e.g., the widely reported compute-efficiency doubling times tracked by groups like Epoch AI) is cited to calibrate whether four to ten further OOMs is plausible on any observed trend, and the piece never distinguishes measured compute-efficiency OOMs from the much fuzzier, unmeasured notion of 'intelligence output' OOMs it actually claims. This directly fails the definitional-rigor bar this corpus's Gary Marcus material applies to comparably sweeping claims (a claim this consequential 'carries a burden of definitional precision before it can be meaningfully evaluated at all').
  Read next: Compare directly against extract-gary-marcus-not-reached-singularity, which insists on defining terms before evaluating claims of this magnitude, and against any corpus item with actual scaling-law or benchmark data on measured algorithmic-efficiency trends.
• illogicalThought 5 dismisses the concern that long-horizon agentic tasks require correspondingly long-horizon training by re-labeling the same empirical phenomenon (the METR agent-task-length curve) as 'error correction' rather than the 'error accumulation' framing it is rebutting, without offering an independent benchmark, ablation, or mechanism to distinguish the two readings. No alternative explanation (e.g., that recent gains reflect narrow task/tooling improvements rather than a generalized error-correction capability) is considered or ruled out before the stronger claim ('generalization in time exists... long tasks are not made of longness') is asserted.
  Read next: Cross-reference with any corpus material directly engaging LeCun's error-accumulation argument or presenting independent analysis of the METR time-horizon benchmark's drivers.
• incompleteThe robotics-scale-up claim ('order of 100M robots/yr by 2030') leans on ~100M cars/year and ~1B smartphones/year as manufacturing-scale precedents, and separately treats the ~10M/year drone market as proof that 10M/yr humanoid robots is 'definitely achievable' — but drones (largely small, non-autonomous, low-complexity hardware) and humanoid robots (explicitly the subject under discussion) are not remotely comparable in actuator/sensor cost, manufacturing complexity, or software maturity. No current humanoid-robot production figures are cited to show how many orders of magnitude of scale-up '100M/yr by 2030' actually implies, leaving a several-orders-of-magnitude, ~4-year production ramp asserted with no supply-chain or capital-feasibility analysis behind it.
  Read next: No corpus material on robotics manufacturing currently exists to compare against; this is flagged as a standalone gap pending future extraction of robotics-specific sources.
• incompleteThe job-survival heuristic (disintermediation/outcompetition/fungibility) — despite the author's own admission that 'this analysis is quite subtle and this paragraph is not going to do it justice' — is tested against only three anecdotal examples (TikTok influencer, doctor, factory worker) with no reference to labor-economics literature on historical automation waves (e.g., agriculture's employment share falling from roughly 40% to 2% of the US workforce over a century, or manufacturing employment trends since the 1970s) that would let a reader calibrate plausible timeframes or magnitudes for AI-specific labor disruption. This corpus's own extract-arvind-narayanan-ai-isnt-all-that-revolutionary reaches a more developed (though still incomplete, per an existing open gap) decide/execute/deliver decomposition of the same problem; bayes doesn't reach that level of mechanism, let alone cite data.
  Read next: Compare against extract-arvind-narayanan-ai-isnt-all-that-revolutionary#p4 and #p14, and the existing corpus gap noting Narayanan himself doesn't explain the wage mechanism by which scarce human time would command a livable wage.
• incompleteThought 4 in Culture and Psychology diagnoses a wealth 'doublespeak' (people who say AI benefits everyone but wouldn't trade their own position for a random person's) and unusually applies this critique to the author himself. But the same reflexive scrutiny is not extended to the piece's own epistemics: nearly every uncertain judgment call across the other 45 points (how many OOMs remain, how fast robots scale, whether a pause is wise, whether nationalization happens) resolves toward 'faster/bigger/more capable than consensus expects' — also the reading most flattering to, and most aligned with the visible incentives of, someone with insider familiarity with frontier AI (the piece name-drops GB300s, Anthropic, GLM5.2, Fable, and 'Mythos 2' with specificity). The self-awareness about wealth-doublespeak is not turned on the analogous possibility that consistently forecasting acceleration is itself a predictable output of the author's own position and incentives.
  Read next: No direct corpus parallel exists; flag as a standalone critical gap about author-incentive disclosure in AI-maximalist forecasting content generally.
• illogicalThe thread simultaneously asserts (a) that institutions, markets, and governments are systematically blindsided and unable to correctly track AI's current trajectory (thought 1), and (b) confidently prescribes fairly sophisticated institutional responses to a faster future — international treaties with GPU counting, a CERN-model-style pause calculus, and judgments about the feasibility of lab nationalization — without reconciling why institutions incapable of even tracking current progress would suddenly become capable of designing and enforcing these more demanding coordination mechanisms in time.
  Read next: Compare against extract-arvind-narayanan-ai-isnt-all-that-revolutionary#p9, which argues durably preventing dangerous-capability diffusion would require sustained, effectively authoritarian global control precisely because institutional coordination at this level is so hard to sustain.

Related documents

• What OpenAI and Anthropic Think Happens Next With AI
  AI Alignment & Safety AI Regulation & Governance (External) Recursive Self-Improvement
• How the 4 New AI Models Change How You Work
  AI Labor Disruption AI Regulation & Governance (External) Riding Model Improvements
• A big week for AI denialism
  AI Alignment & Safety AI Regulation & Governance (External)
• AI Companies Are Hiring More
  AI Labor Disruption AI Regulation & Governance (External)
• AI Sovereignty: Bargaining with Big Tech and the Promise of Full Stack Open Source AI
  AI Regulation & Governance (External) Compute & Energy Infrastructure
• Anthropic Can Now Read Claude's Mind
  AI Alignment & Safety AI Regulation & Governance (External)
• Presentation: A Few Predicted Talks From QConAI 2030
  AI Labor Disruption AI Regulation & Governance (External)
• Arvind Narayanan on Why AI Isn't All That Revolutionary
  AI Labor Disruption AI Regulation & Governance (External)
• Cybersecurity and the Gap Between Skill and Ability
  AI Alignment & Safety AI Regulation & Governance (External)
• Don't train the model, evolve the harness
  Recursive Self-Improvement Riding Model Improvements