What does AI and AGI mean?

This is the third and final post in this series on what got eliminated to produce the modern vocabulary of “AI” and “AGI.” The first traced the Dartmouth cut of 1956 — the institutional renaming of cybernetics out of cognitive science. The second traced the Turing cut of 1950 — the operational substitution of behavioral indistinguishability for the question of mind. The fork I want to look at here is the most recent and, in some ways, the most structurally constraining of the three, because it operates not at the level of ideas (which is what the first two cuts did) but at the level of the conditions under which alternative ideas can be entertained at all. The 2017–2026 consolidation in AI wasn’t a scientific result; it was a three-way resonance between a scaling law, unlimited capital, and a hardware monopoly that collapsed paradigm diversity, researcher diversity, geographic diversity, and time horizons into a single optimization loop — and taught a generation to mistake that loop for discovery.

The more the better

“More” stopped being a tactic and became a theory. Once the scaling hypothesis implicitly claimed that the question of cognition is exhausted by the question of scale, every other line of inquiry was reframed as either premature optimization or nostalgia. The technical history of how this happened is worth keeping in view because the scaling regime we now live inside is often misunderstood as an inevitable progression of ideas rather than the contingent product of a few specific papers, a few specific funding decisions, and a few specific hardware monopolies.

The architectural moment was Vaswani et al., “Attention Is All You Need” (NeurIPS 2017) — the transformer paper, which introduced an architecture that scaled better than the recurrent and convolutional alternatives that had dominated the prior decade. The empirical demonstration that the transformer architecture’s gains kept compounding with scale was Brown et al., “Language Models are Few-Shot Learners” (NeurIPS 2020) — the GPT-3 paper, which showed qualitatively new capabilities emerging with parameter count. The theoretical formalization came with Kaplan, McCandlish et al., “Scaling Laws for Neural Language Models” (arXiv 2001.08361, January 2020) — power-law relationships between compute, data, parameters, and loss, derived from a sweep across model sizes. The Chinchilla paper (Hoffmann et al., “Training Compute-Optimal Large Language Models,” arXiv 2203.15556, March 2022) recalibrated the compute-versus-data ratio Kaplan had proposed, sharpening the prescription. By early 2022, the field had a price-per-capability curve. Capital responds to predictable curves.

The capital lock-in followed. Microsoft’s $1B investment in OpenAI in 2019 was the opening move; the $10B follow-up in January 2023 was the consolidation; subsequent rounds (Google’s stake in Anthropic, Saudi and UAE sovereign wealth funds, SoftBank, the rise of GPU rental specialists like CoreWeave and Lambda Labs) extended the pattern. The Stargate Project, announced January 2025 by OpenAI, Oracle, SoftBank, and MGX, committed $500 billion over four years toward roughly 10 GW of AI data center capacity in the United States. As of early 2026, Stargate has the flagship Abilene, Texas campus partially online (first two buildings operational since September 2025) plus five additional US sites announced in September 2025 — Shackelford County (TX), Doña Ana County (NM), Lordstown (OH), Milam County (TX), and an unnamed midwestern site — with the combined Stargate footprint at roughly 7 GW of planned capacity and on a stated path toward the full 10 GW commitment. A custom AI chip — codenamed “Titan” — is being developed with Broadcom on TSMC’s 3 nm process for mass production in the second half of 2026. The story is not uniformly triumphant: the UK Stargate project was paused in April 2026 over regulatory and energy-cost concerns, and a planned Norway site was pulled back in favor of Microsoft assuming the offtake. OpenAI itself has reframed Stargate as “an umbrella term” for its compute strategy and is now emphasizing leased capacity from third-party providers rather than first-party data centers. Even with the partial retreats, the order-of-magnitude commitment is unprecedented.

Hardware: Nvidia’s near-monopoly position on the high-end training accelerators (H100, H200, the Blackwell-generation B200 and GB200) created a single bottleneck through which all frontier-scale training runs pass. Export controls compounded the bottleneck geographically. The Bureau of Industry and Security’s October 2022 rules restricted exports of high-end chips to China; expanded controls followed in October 2023; the Biden administration’s “Framework for Artificial Intelligence Diffusion” (the AI Diffusion Rule) was published January 15, 2025, just days before the administration left office. The Trump administration rescinded that framework in May 2025, calling it overly bureaucratic, and replaced it with a different approach: on December 8, 2025, the administration announced selective sales of some chips to approved Chinese customers, and on January 13, 2026, BIS revised export licensing to allow case-by-case review for entities in mainland China and Macau, replacing the prior presumption of denial. Huawei’s Ascend 910B, 910C, and 910D chips remained prohibited as end-use targets. Congress approved a 23% increase in BIS’s FY2026 budget, with several million dollars earmarked specifically for semiconductor enforcement. The geopolitical-technical regime around frontier AI is now an active policy artifact, revised every few months under each new political administration, and the underlying pattern — that frontier capability is gated by access to a few specific chip families produced by a few specific fabs — has not changed.

The result of all of this is a research environment in which the binding constraint on what can be investigated is access to roughly 10,000-GPU clusters, roughly 100 megawatts of power, and roughly $100 million to $1 billion per training run. Paradigm diversity got eliminated by investment selection, not by refutation. JEPA, active inference, neural cellular automata, neuromorphic substrates, predictive coding, classical symbolic reasoning, energy-based models — none disproven, all underfunded relative to what their proponents argue they need. Below I list a few of the things I believe got eliminated in the consolidation. The list is not exhaustive; it is the eliminations that compound enough to matter for the next decade.

The solo and small-lab researcher as frontier actor

The one-GPU paper that defined a subfield is extinct at the frontier. Geoffrey Hinton worked on backpropagation variants and Boltzmann machines from the late 1970s through the early 2010s — most of it in academic or small-lab settings, much of it considered marginal at the time. The Boltzmann machine work he and Terry Sejnowski developed between 1983 and 1985 used tools from statistical physics; the backpropagation algorithm he and Rumelhart and Williams demonstrated in 1986 (“Learning representations by back-propagating errors,” Nature, 1986) was the foundational technique that, three decades later, would underwrite the entire deep-learning revolution. Hinton shared the 2024 Nobel Prize in Physics with John Hopfield “for foundational discoveries and inventions that enable machine learning with artificial neural networks” — a prize that retrospectively credits the scaling era to the small-group long-horizon work that the scaling era has structurally eliminated. Hopfield’s 1982 paper introducing what is now called the Hopfield network (“Neural networks and physical systems with emergent collective computational abilities,” PNAS, 1982) was a single-author paper. Hochreiter and Schmidhuber’s LSTM paper (“Long Short-Term Memory,” Neural Computation, 1997) was a two-author paper. The kind of work the Nobel Committee recognized cannot be done at the frontier today.

The numbers are not subtle. Epoch AI’s amortized-hardware estimates put GPT-4’s final training run at roughly $40 million; Stanford’s 2025 AI Index Report, using a different methodology that includes cloud rental pricing, puts the same run at roughly $78 million, with Google’s Gemini Ultra at around $191 million and Meta’s Llama 3.1 405B at $170 million. Both methodologies show the same trend: training compute costs for the largest runs are doubling roughly every eight months. By Epoch AI’s projection, the largest training runs will exceed $1 billion by 2027. Frontier research stopped being a craft where individual intuition mattered and became an industrial process where compute access is the binding constraint on who gets to think. An entire mode of intellectual work — the lone theorist who can be wrong for twenty years and then right — was defunded out of existence, not because the lone-theorist mode failed, but because it cannot compete in a market structured around quarterly capability releases.

What dies first is the foundational-conceptual work that takes a long time to ripen. Friston’s free energy principle, which I have referenced across this trilogy, took roughly fifteen years (from the early 2000s to its current formulation) to develop in a small-group academic setting. The kind of synthesis it represents — drawing on thermodynamics, variational inference, neuroscience, and philosophy of mind — is not the kind of work a frontier lab can fund, because it doesn’t return a benchmark improvement on a quarterly timescale. What replaces it, structurally, is engineering refinement: better training pipelines, better data filtering, better post-training methods. These are valuable; they are not the same thing.

The academy as site of frontier inquiry

Hinton (Google Brain since 2013, then DeepMind, then independent), LeCun (Meta since 2013), Ilya Sutskever (OpenAI 2015 to mid-2024, then Safe Superintelligence), Andrej Karpathy (Tesla, then OpenAI, then independent education ventures), Ian Goodfellow (Google, then Apple), the Amodei siblings (OpenAI, then Anthropic), Demis Hassabis (DeepMind from founding through Google acquisition), countless others — defected to industry across the 2013–2024 window. Yoshua Bengio at Montreal made public statements as early as 2018 and 2019 about the difficulty of retaining graduate students who could double or triple their compensation by leaving for industry before defending. Universities became talent pipelines, not research institutions, at the frontier. PhD curricula retooled for industry employability — the Berkeley, Stanford, and CMU machine learning programs now look structurally indistinguishable from terminal-MS programs designed to feed FAANG and frontier-lab pipelines. Stanford’s HAI (Human-Centered AI Institute, founded 2019) was an institutional response — an attempt to keep a frontier-relevant AI institution under university auspices — but its annual budget is a rounding error against any single frontier training run.

The consequence is that the question “what is cognition” got structurally replaced by “what is achievable on this quarter’s cluster.” The university had been the one institution explicitly designed to pursue questions on geological timescales — that is what tenure structurally protects. That institution lost its grip on its own discipline. There are corners of academic ML that still operate in the older mode (Anthony Chemero’s enactive cognitive science work; the slow phenomenology research at Vienna and Copenhagen; the basal cognition work in Levin’s lab at Tufts), but they are not “frontier” by the field’s current definition because the field’s current definition requires frontier compute access.

Open science and reproducibility

Pre-2020, the deep-learning literature was a model of open science by the standards of empirical disciplines: papers came with code, model weights, full hyperparameter recipes, and dataset documentation; labs could and did replicate each other’s work routinely. The inflection point was OpenAI’s February 2019 staged release of GPT-2 — the first time a frontier lab withheld weights from a publication, citing dual-use risks. The argument was contested at the time and the weights were eventually released, but the precedent had been set: a frontier lab could, citing safety, decline to share what it had built. The full transition came with GPT-4 in March 2023, whose system card disclosed essentially nothing technical about the architecture, the training data, or the training procedure. Anthropic’s Claude system cards have been similarly opaque about their frontier models. Peer review became theater for frontier work — you cannot verify claims you cannot afford to replicate, and you cannot replicate what you cannot see.

The picture has been materially complicated, however, by the Chinese open-weight ecosystem that has emerged since DeepSeek-R1 (January 2025). DeepSeek released R1 with full open weights under the MIT license, then DeepSeek-R1-0528 (May 2025), and a preview of DeepSeek-V4 (V4-Pro and V4-Flash, both open-weight) in April 2026 — building on the V3 base model open-sourced at the end of December 2024. The Hugging Face one-year-after-DeepSeek-R1 retrospective documents what happened across the broader Chinese AI ecosystem in the same period: Baidu went from zero releases on Hugging Face in 2024 to over 100 in 2025; ByteDance and Tencent each increased open-weight releases by eight to nine times over the same window; Alibaba’s Qwen series, the 01.AI Yi series, and several smaller labs (Zhipu, Moonshot, MiniMax) all maintained aggressive open-weight release schedules throughout 2025 and into 2026. By any reasonable metric, the global open-weight model ecosystem is now dominated by Chinese labs, and the “frontier” gap between closed-weight Western labs and open-weight Chinese labs has narrowed substantially — DeepSeek-R1 reached benchmark parity with the best Western reasoning models at a small fraction of the training cost. This counter-development is real; it does not undo the structural Western shift toward closed-weight frontier work, but it means the open-science erosion is no longer a one-direction story. Closed-weight Western frontier labs no longer have a monopoly on frontier capability, and the open-weight tradition has migrated geographically rather than dying.

The Bender, Gebru, McMillan-Major, and Mitchell paper “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” (FAccT 2021) is worth naming for the institutional history. Timnit Gebru was forced out of Google in late 2020 over the paper; Margaret Mitchell was let go shortly afterward. The optics — a Black woman researcher pushed out for raising critical questions about scale — became iconic in subsequent debates about whose epistemic authority is welcome inside frontier labs. Joelle Pineau et al.’s reproducibility checklist (NeurIPS 2019) and the ML Reproducibility Challenge are honest community-level attempts to slow the erosion, but they cannot reach the frontier work whose internals are closed.

The honest record of failure

Training failures, loss spikes, mode collapses, dataset contamination, methods that didn’t work — systematically hidden by competitive pressure. Pre-scaling ML had a functional negative-results culture: ablation studies were standard, honest failure modes were reported, the field’s collective memory included what didn’t work. Post-scaling, that culture has substantially eroded. The field’s published memory is now biased toward success stories, which means its actual understanding of why things work is brittle.

Specific named exceptions exist and are worth holding in view. Meta’s withdrawal of Galactica (November 2022, three days after release, after the model produced confident scientific-sounding nonsense at scale) was a rare public accounting of a frontier release going badly. Google’s pause and recalibration of Gemini 1.0 image generation (February 2024, after the model produced historically inaccurate images that drew political attention) was another. Anthropic’s mechanistic-interpretability work — the Transformer Circuits Thread, ongoing since 2021, with major contributions by Chris Olah, Catherine Olsson, Nelson Elhage, and many others — represents the most consistent honest engagement with what the field’s models are actually doing internally. The March 2025 Anthropic introduction of “circuit tracing” via cross-layer transcoders, and the subsequent attribution-graph methodology, has produced empirical results that are reported with the kind of caveat-rich honesty that used to be standard across the field. The January 2025 collaborative paper “Open Problems in Mechanistic Interpretability” (29 researchers across 18 organizations, including Anthropic, Apollo Research, Google DeepMind, EleutherAI, and several universities) is itself a recovery of the older norm — a published consensus about what is unknown, written by the people closest to the work.

The exceptions matter. They are evidence that honest failure-reporting is not impossible inside frontier labs, only that it requires explicit institutional commitment against the default pressure not to publish what didn’t work.

Researcher cognitive diversity

Pre-2020 AI research drew from a wide range of disciplines: theoretical neuroscience (Hopfield, Tony Bell, Tony Zador), linguistics (Manning, the Chomskyans and their interlocutors), philosophy of mind (Pearl on causation, Dennett on intentionality, the Churchlands on neurophilosophy), physics (Hopfield himself coming from solid-state physics; Mehta and Schwab on the renormalization-group reading of deep learning; Hinton’s statistical-physics formation), complexity science (Kauffman, Wolfram), developmental psychology (Spelke, Carey, Tomasello), and anthropology (Hutchins, Lave). Post-2020, the research community has narrowed to a much more specific demographic profile: young, engineering-oriented, benchmark-driven, comfortable with PyTorch and distributed training, increasingly with industry-employment trajectories from undergraduate onward.

The narrowing is empirically visible in NeurIPS, ICML, and ICLR demographic surveys (analyzed in several recent papers, including the Stanford AI Index reports across 2023–2025); it is also visible in the kind of papers that get accepted versus the kind that don’t. Theoretical neuroscientists, developmental psychologists, philosophers of mind, contemplative scholars, complex-systems theorists — either marginalized or absorbed as “applied” labor on someone else’s agenda. Some of the most consequential recent ideas in machine learning came from outside the conventional ML pipeline: the renormalization-group analyses of deep networks (Mehta & Schwab, “An exact mapping between the Variational Renormalization Group and Deep Learning,” arXiv 2014); the connections between physics and learning that Hopfield’s Nobel re-validated; the predictive-processing/active-inference framework from Friston (computational neuroscience); the basal-cognition framework from Levin (developmental biology). The field still imports these ideas, but the people who originate them increasingly cannot get hired into the frontier labs because they don’t fit the engineering-track profile the labs select for.

The questions asked narrowed to what this demographic is interested in asking, which is a much smaller set than what’s worth asking. This compounds across hiring cycles: each generation of frontier-lab researchers is selected for compatibility with the current research culture, which then reinforces what the next generation of culture looks like. The monoculture is self-perpetuating, not as a conspiracy, but as the natural consequence of a sustained selection pressure on a single population.

Geographic and civilizational diversity

Frontier compute is now present effectively only in two locations: the United States (concentrated in a handful of hyperscalers and frontier labs) and China (with Tencent, Baidu, Alibaba, ByteDance, and the standalone DeepSeek and Moonshot labs). The United Kingdom retains a tail position through Google DeepMind, which is structurally a Google subsidiary now. The European Union has a tail position through Mistral and a few smaller efforts, with most of its policy energy directed at regulation (the AI Act, finalized 2024, in implementation 2025–2027) rather than at frontier research. Latin America, Africa, South and Southeast Asia, the Middle East — locked out by hardware costs and, increasingly, by export controls.

The export-control regime as it currently stands (after the Trump administration’s May 2025 rescission of the Biden Diffusion Rule and the January 2026 BIS revision) is more permissive on case-by-case Chinese exports than the Biden framework would have been, but it remains a structural exclusion mechanism for everyone outside the US-China duopoly. The chips that constitute the frontier (Nvidia H100/H200/B200/GB200) require export licenses for sale outside a small set of approved jurisdictions; the alternatives (AMD’s MI300/MI325/MI355, Intel’s Gaudi series, Cerebras and SambaNova specialty chips, Google’s TPUs) face similar control regimes or are not commercially available outside their parent ecosystems. A research lab in Lagos or São Paulo or Jakarta cannot purchase the equivalent of a frontier US lab’s compute access at any price short of multi-year sovereign-level commitments.

The DeepSeek phenomenon — and more broadly the Chinese open-weight ecosystem documented in the previous section — is evidence that the duopoly is not absolutely sealed. China’s open-weight models have given developers and researchers globally meaningful access to frontier-grade capability without the closed-weight Western labs as gatekeepers. But the broader exclusion of African, Latin American, MENA, and Southeast Asian research ecosystems from frontier model production (rather than consumption) is unchanged. The “universal” intelligence under construction is still the cognitive output of a narrow subset of humanity pretending at neutrality. Indigenous epistemologies, non-Western philosophies of mind (those not already absorbed into the Chinese training corpora), African oral-tradition approaches to distributed cognition, Indic theories of consciousness — zero structural influence on the artifacts being built. This is a more severe elimination than the Turing-era one because it is now structural (enforced by compute access and chip embargoes) rather than purely intellectual.

Research timescales

Hinton worked on devising backprop and neural networks across roughly thirty years of academic marginality. Post-scaling research operates on three- to six-month cycles. The current cadence is visible in the OpenAI release schedule for the GPT-5 family alone: GPT-5 launched August 7, 2025; GPT-5.2 December 11, 2025; GPT-5.3-Codex February 5, 2026; GPT-5.4 March 5, 2026; GPT-5.5 (with a Pro variant) April 24, 2026. That is five frontier releases in less than nine months from the same lab. Anthropic’s Claude release cadence and Google DeepMind’s Gemini release cadence are comparable. A result not scaled up within weeks is superseded by the next frontier release. Deep, slow, foundational work is systematically selected against. The publish-or-perish loop has compressed to lab-internal quarterly timescales. Nobody can afford to be wrong for ten years anymore, which means nobody can afford to think the kind of thought that takes ten years to mature.

The point is not that quick-iteration work is bad — much of it is valuable, and the cadence has produced real capability gains. The point is that quick-iteration work, by its nature, can only refine existing paradigms, not generate new ones. Generating new paradigms takes the kind of unprotected, marginal, long-horizon thinking that the current funding regime structurally cannot afford. This is a narrowing of the kinds of thoughts the field can produce, not just the rate at which it produces them.

The 20-watt brain as disciplining constraint

Biological brains do their thing at roughly 20 watts of metabolic power. This is not an engineering curiosity — it is a hard existence proof that some architecture achieves general intelligence at energy budgets five to nine orders of magnitude below current frontier training runs. The contrast is concrete and growing. The International Energy Agency reports global data center electricity consumption at approximately 415 TWh in 2024 (about 1.5% of global electricity), with projections approaching 1,050 TWh by the end of 2026 — roughly the consumption of Japan. AI-specific data center power consumption is projected to reach 90 TWh annually by 2026, a roughly tenfold increase from 2022. In the US specifically, total data center demand is expected to nearly double from 80 GW in 2025 to 150 GW by 2028, almost entirely driven by AI training and inference workloads. Individual frontier training runs are now measured in tens of megawatts of sustained power draw over months; one April 2025 estimate put a representative frontier training run at 25.3 MW peak draw, with another study estimating total energy consumption for a single training run at 50 GWh. The biological brain produces what we call intelligence, by some definitions, on roughly one ten-millionth of that.

Pre-scaling, energy efficiency was a serious research frontier. IBM’s TrueNorth chip (2014) and its successor NorthPole (2023) implemented neuromorphic architectures designed for the kind of energy efficiency the brain achieves; Intel’s Loihi (2017) and Loihi 2 (2021) pursued spiking neural network designs in commercial silicon; the Manchester SpiNNaker project (substantial deployments by 2018) built large-scale spiking-neural hardware. Carver Mead’s foundational Analog VLSI and Neural Systems (Addison-Wesley, 1989) — referenced from post 1 of this trilogy — established the engineering tradition this work descends from. Reservoir computing, sparse computation, conditional computation, and analog substrates all looked like serious frontiers as recently as 2017. Post-scaling, all of this got reframed as irrelevant. “Just add compute” became the dominant disposition. The question “why does a brain need so little to do what it does” — arguably one of the deepest open questions in science — is actively being defunded relative to its conceptual importance. If your theory of cognition is compatible with ten-gigawatt training runs producing performance comparable to a 20-watt brain, your theory has substantial chances to be wrong about cognition as it actually exists in nature.

This connects directly to the I=E framework I have been developing across these posts. Efficient elimination is tied to the substrate’s natural dynamics — a system that uses its own physical degrees of freedom to perform the elimination operation will be enormously more efficient than a system that simulates the elimination on a substrate built for general computation. The brain does the former; the transformer does the latter. The factor of ten million between them is, on this reading, evidence that they are doing very different things, even when their input-output behavior on certain narrow tasks looks similar. The scaling regime is structurally incapable of asking this question, because the question presupposes an alternative the regime has defunded.

Alignment as philosophy of mind

Pre-2020 alignment work was a philosophy-of-mind discipline. Eliezer Yudkowsky and the Machine Intelligence Research Institute (MIRI) had been working on decision-theoretic and reflective-stability questions since the mid-2000s. Nick Bostrom’s Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014) was the synthesis that brought the field into mainstream visibility. Stuart Russell’s Human Compatible: Artificial Intelligence and the Problem of Control (Viking, 2019) developed the inverse-reward-design and uncertainty-over-objectives framework. Paul Christiano’s early work on prosaic alignment — “Deep Reinforcement Learning from Human Preferences” (Christiano, Leike, Brown, Martic, Legg, Amodei, NeurIPS 2017) — was the technical antecedent of what would eventually become RLHF. Across this work, “alignment” meant: how do we create minds we can trust to pursue goals we endorse, given that we may not be able to fully specify those goals in advance? This was a problem in philosophy of mind, decision theory, metaethics, and metacognition, all at once.

Post-scaling, alignment collapsed to something narrower. RLHF (the InstructGPT formalization in Ouyang et al., “Training language models to follow instructions with human feedback,” NeurIPS 2022) became the canonical training-time alignment intervention. Red-teaming, evaluation suites (HELM, TruthfulQA, MMLU, the various capability-and-safety benchmark families), constitutional AI (Bai et al., Anthropic, 2022), preference-learning and reward-model training, and policy work absorbed the alignment-research budget. These are all real and useful — I am not dismissing them. But they are product-safety disciplines for current chatbots, not engagements with the deep version of the alignment problem. The deep problem — what does this system actually want, structurally, given the dynamics that produced it — got replaced by the surface problem because only the surface problem had a tractable metric and a quarterly deliverable.

The institutional history is telling. OpenAI’s Superalignment team, formed in mid-2023 under Ilya Sutskever and Jan Leike with a public commitment of 20% of OpenAI’s compute, was disbanded in May 2024. Sutskever and Leike resigned within hours of each other; Leike’s public statement was that “OpenAI’s safety culture and processes have taken a backseat to shiny products” and that the Superalignment team had been “struggling for compute,” with requests routinely rejected. Reporting in the months that followed established that the 20% compute commitment had never actually been delivered. Leike joined Anthropic the same month to continue the work; Sutskever founded Safe Superintelligence Inc. shortly afterward. In an October 2025 deposition, Sutskever alleged a pattern of dishonesty by Sam Altman, including memos he had written to the OpenAI board. OpenAI announced a “Mission Alignment Team” in late 2024 as a partial replacement for Superalignment; that team was itself disbanded in February 2026, after roughly sixteen months of operation. The pattern is consistent: alignment work that is structurally separable from product release tends to lose internal political battles inside frontier labs, regardless of how seriously the company nominally claims to take it.

The honest exceptions are narrow but real. Anthropic’s mechanistic-interpretability program continues to produce work that is recognizably philosophy-of-mind-adjacent, in the sense that it is trying to build empirical access to what the model is actually doing internally rather than just whether its outputs satisfy a preference function. The “circuit tracing” method introduced in March 2025, the cross-layer-transcoder formalism that followed, and the attribution-graph methodology developed across 2025 and into 2026 are technically deep work that could not have been done outside a frontier lab — and they are recognized as serious enough that MIT Technology Review named mechanistic interpretability one of its breakthrough technologies for 2026.

We are now, on net, less prepared for the hard version of alignment than we were in 2018. The intellectual apparatus for approaching it — the philosophical seriousness, the long-horizon framing, the willingness to ask what a system is rather than what it does — has been substantially diverted. My own self-eliminating-observer research program (active on this blog, as referenced in earlier posts) is structurally a philosophy-of-mind project rather than a product-safety project, and it would be effectively unfundable inside any current frontier lab even though it is squarely about what an aligned mind would be, structurally.

What scaling got right

I have been one-sided so far. As with the prior two posts in this series, honest steel-manning is owed.

Scaling-era AI got real things right. The most important is that emergent capabilities that were not predicted by the naïve view actually appeared. In 2019, the consensus position in mainstream NLP was that LLMs were stochastic-parrot statistical engines that would never produce coherent multi-step reasoning. By 2024, GPT-4-class models were producing legitimate multi-step reasoning across domains — including programming, mathematics, scientific writing, and law — to a degree that the 2019 consensus would have considered impossible. The empirical fact that this happened, regardless of the metaphysical interpretation of what is happening when these systems do it, is information about cognition that no other research program has produced. The transformer architecture (Vaswani et al., 2017) is better than what came before for the specific class of problems it solves, and the engineering achievements in scaling its training to billions of parameters and trillions of tokens are real engineering work that deserves engineering credit.

Industry frontier labs do produce real research, not just product work. The mechanistic-interpretability program at Anthropic is the clearest example, but there are many others: the scaling-laws work itself (Kaplan, Hoffmann, and the ongoing series of refinements) is a contribution to the science of learning systems regardless of how one feels about the consequences; RLHF as a technique (Christiano 2017, Ouyang 2022) is a substantive algorithmic advance; constitutional AI (Bai et al., 2022) is a thoughtful attempt to address one specific class of alignment problem; sparse autoencoders for feature interpretation (the Bricken et al. 2023 Anthropic work, “Towards Monosemanticity,” and its many descendants) opened a serious research direction. The distributed-training infrastructure that allows 10,000-GPU coordinated training runs is a non-trivial engineering achievement that reflects the kind of capability that only well-funded large organizations can produce.

Some of the pre-scaling AI research that the consolidation defunded was, in fact, not promising. The fragmentation of pre-2017 ML into many small subfields, each with its own benchmarks and its own dialect, was a real cost the field paid; the consolidation into a more uniform paradigm has made it easier to compare results and to communicate across research groups. Some criticism of the current monoculture reflects intellectual nostalgia for an earlier diversity that was, partly, a diversity of lower-quality work. The honest version of the critique should distinguish between paradigms that were defunded for good empirical reasons and paradigms that were defunded for capital-allocation reasons. The line between those is sometimes hard to draw.

So the honest position: the consolidation produced real progress and real losses. The engineering and empirical capability are real, including the surprising fact that scale alone produced what it did. The losses are mostly downstream of that same regime — the diversity of paradigms entertained, of researchers entering the field, of geographies where frontier work happens, of timescales over which questions can be asked, and of willingness to engage philosophy of mind as a live part of the project. The two ledgers don’t cancel.

Where does this leave us now?

There are other eliminations I could survey, but the point sits here: capital selection is itself a form of elimination, and it is the most dangerous form because it operates on predicted ROI rather than on truth-value. A paradigm can be correct and defunded; another can be wrong and flooded with money. The post-2020 regime didn’t eliminate alternative paradigms because they failed empirical tests; it eliminated them because they couldn’t promise linear capability scaling on quarterly timescales to investors.

The field as it stands in May 2026 has, by a reasonable accounting, narrowed along five dimensions simultaneously: paradigm (one architecture family, one training paradigm, one evaluation methodology); demographic (one educational profile, increasingly one career trajectory); geographic (two countries doing frontier work, with a handful of tail positions); temporal (three- to six-month iteration cycles displacing five- to thirty-year foundational work); and epistemic (one set of accepted methods for what counts as evidence). Within five to eight years at this pace, if no countervailing pressure intervenes, the field will collapse to a monoculture along all five dimensions at once. Some countervailing pressure is in fact intervening — the Chinese open-weight ecosystem has materially complicated the geographic and the open-science narrowing; export-control reversals have shifted but not eliminated the chip-access bottleneck; the small but persistent academic and small-lab work outside the frontier (Levin’s basal cognition, the active inference community, the neuromorphic engineering community, the contemplative-cognitive-science work I have referenced across this trilogy) keeps the alternative paradigms alive even if they cannot scale. Whether these countervailing pressures are strong enough to prevent the monoculture collapse is uncertain.

The critique of the capital filter is not a critique of the empirical results the filter has produced. The capabilities are there. It is a critique of what the filter has prevented from being investigated alongside.

What can we observe from our exploration arc as a whole?

The first elimination — Dartmouth, 1956 — removed a particular ontology of mind (cybernetic, embodied, observer-coupled, self-referential) from the space of what could be thought. The second elimination — the Turing test, 1950 — removed a particular set of questions about mind (what is it like, how does it come to be, how does it relate to its body and others) from the space of what could be operationally measured. The third elimination — the capital consolidation, 2017–2026 — removes the conditions under which alternatives to the dominant paradigm could even be entertained at scale. Each elimination is structurally distinct. The third is, in one sense, the most durable, because it operates not on ideas (which can be re-introduced) but on the institutional substrate that makes ideas pursuable.

Across all three: the most dangerous elimination is not any single paradigm that was defunded, or any single question that was sidelined, or any single institutional configuration that was lost. The most dangerous elimination is the loss of the metacognitive habit of the field — the ability to notice that it might be on a wrong path and redirect. Monocultures lose that habit because there is nobody left to argue from outside the consensus. The arguments from outside the consensus, when they come, get coded as adjacent disciplines (philosophy, systems theory, mysticism, niche neuroscience) rather than as live alternatives the field has to reckon with. The trilogy of cuts I have traced across these three posts has progressively reduced the field’s capacity to reckon with arguments from outside its own consensus, and that reduction is accelerating, not slowing.

The words “AI” and “AGI” did not appear ex nihilo. They are sediment from these three eliminations. “Artificial” carries the Dartmouth cut — the real thing is biological, what we build is imitation. “Intelligence” carries the Turing operationalization — it is whatever passes the test. “General” carries the scaling-era ceiling — as flexible as a human across our chosen domains, where “our chosen domains” is the residue of a particular cognitive monoculture. None of the three words has a definition independent of the elimination that produced it. That is why they don’t sit still — they cannot. Each successive frontier release shifts what the words refer to without shifting the words themselves, because the words have only histories, not referents.

What can be done? Naming the cuts is the prerequisite for choosing them deliberately rather than inheriting them by default. That is what these three posts have tried to do. The next move — what to build with the cuts named — is a different kind of work, longer and harder, and not entirely a writing project. A small portion of it is happening on this blog under the heading of the I=E research programs. A larger portion is happening on the margins of academic cognitive science, neuromorphic engineering, basal-cognition biology, philosophy of mind, contemplative science, and the small open-weight labs that have refused the consolidation’s ontology. The recovery, if there is one, will not come from the center; it will come from the periphery, slowly, against the structural pressure to converge. Whether that recovery materializes depends on choices being made now, by people who may or may not realize they are making them.