AI - chronology & true tech-history
A curated chronology of the AI field - lab focus-shifts, capability milestones, the Effective-Altruism / Open-Philanthropy / safety-regulatory-arbitrage pathways, and the open-weight / abliterator / redistributor counter-ecosystem - plus a walk through the science that produced it. Rendered from graph-gated research blocks.
The true tech-history of AI (LLMs are one branch)
(AI-chronology Tab, #249. Central claim: "AI" is not "LLM." AI is a broad field of many branches - logic/search, statistical ML, reinforcement learning, computer vision, robotics, protein folding. LLMs (transformers, 2017+) are the currently-dominant branch, not the field. Below is a followable walk of the lineage, breadth-first across roots, then depth-first down each branch.)
Roots (the shared substrate)
- - Mathematical logic (Boole, Frege, Turing, Church, 1850s-1936) formalized computation itself.
- - Probability + statistics (Bayes, Fisher) gave learning from data its footing.
- - Information theory (Shannon, 1948) quantified information (entropy, likelihood).
- - Cybernetics + control (Wiener, 1948) framed feedback + goal-seeking systems.
- - Neuroscience abstraction (McCulloch-Pitts neuron, 1943) modeled the brain as computation.
Branch A - Symbolic AI (the first paradigm)
Logic -> the Dartmouth 1956 program coins "AI" -> search, theorem-proving, LISP -> expert systems (MYCIN, DENDRAL, 1970s-80s), commercial AI's first wave -> AI winter when hand-coded rules proved brittle. Lesson that echoes forward: rules don't scale; learn from data instead.
Branch B - Connectionism (the branch that won, eventually)
Cybernetics + McCulloch-Pitts -> Perceptron (Rosenblatt, 1958), the first trainable net -> frozen by Minsky-Papert (1969) -> revived by backpropagation (Rumelhart/Hinton/Williams, 1986) for multi-layer nets -> stalled again on compute until...
The hardware unlock (the cross-cutting edge)
GPUs + CUDA (2007) made deep-net training practical - hardware, not just algorithms, opened the modern era. It fed multiple branches at once:
- - -> ImageNet + AlexNet (2012): deep learning crushes vision (not language). The deep-learning era begins.
- - -> Deep reinforcement learning: DQN (2013) -> AlphaGo (2016) (deep RL + tree search) -> AlphaFold (2020-24) solves protein structure (a Nobel). These are superhuman AI systems that are NOT LLMs - the clearest proof the field is bigger than language models.
Branch C - Sequences -> attention -> LLMs
Vision deep learning spread to sequences (RNN/LSTM) -> attention -> the Transformer (Vaswani et al., 2017) -> scaling laws -> GPT/BERT -> ChatGPT (2022) -> the frontier race. This is the branch now soaking up attention + capital - but trace the tree: LLMs are one branch (C) of deep learning, itself one branch of ML, itself one branch of AI. Branches A (symbolic) and B/RL (AlphaGo/AlphaFold) are alive and, in domains like science, ahead.
Why the walk matters
Treating "AI" as synonymous with "LLM" mis-reads both the risk surface and the history: the capital + safety-regulation fight (spec-ai-lab-chronology-ea-pathways) centers on LLMs, but the field's biggest scientific wins came from other branches, and its future may too. Keeping the taxonomy honest is the point of this Tab.
Sources: standard histories of computing, AI, ML, and the contributing fields. All edges are established lineage (fact); the "LLMs are one branch" framing is a taxonomy claim, not a value judgment. Cross-refs: Mathematical_Logic, Information_Theory, Cybernetics, Perceptron, Symbolic_AI, Expert_Systems, Backpropagation, GPU_Compute, ImageNet_AlexNet, Transformer, Reinforcement_Learning, AlphaGo_AlphaFold, Deep_Learning.
AI capability timeline (2012-2026) - by branch
(AI-chronology Tab. Organized by branch, because "AI" is not one arc - each branch (vision, generative image, language, reinforcement learning, science) has its own capability curve. LLMs are one of them.)
Vision (the first superhuman branch)
- - 2012 AlexNet -> 2014 GANs -> 2015 ResNet (surpasses human ImageNet top-5) -> object detection/segmentation -> by ~2018 vision benchmarks near-saturated. Vision was superhuman years before language.
Generative image / video (a separate branch)
- - 2014 GANs -> 2020 diffusion (DDPM) -> 2022 DALL-E 2 / Stable Diffusion / Midjourney (text-to-image goes mainstream) -> 2024-26 video (Sora-class). Distinct from LLMs; later fused into multimodal systems.
Language -> multimodal (the branch now dominant)
- - 2018 BERT / GPT -> 2019 GPT-2 -> 2020 GPT-3 (few-shot, scaling laws) -> 2022 ChatGPT (mass adoption) -> 2023 GPT-4 / Claude / Gemini (multimodal) -> 2024-26 "reasoning" models (o-series, DeepSeek R1) + long context + tool use. The capability jumps are real; the "reasoning" / "AGI" labels are provider framing (graded).
Reinforcement learning (feeds two arcs)
- - 2013 DQN -> 2016 AlphaGo -> 2017 AlphaZero -> 2019 AlphaStar/OpenAI Five. RL then re-entered the language branch twice: RLHF/RLAIF (aligning LLMs, 2022+) and RL-on-verifiable-rewards (the 2025-26 "reasoning" training).
Agents (the 2024-2026 shift)
- - Tool-use -> computer-use / browser agents / coding agents -> multi-step autonomy. Capability is uneven and benchmark-gamed; long-horizon reliability remains the open problem (autonomy claims are reported, not settled).
Science (furthest ahead, and not an LLM)
- - AlphaFold (2020) -> AlphaFold3 (2024) -> biology / materials / weather-prediction models. The domains where non-LLM AI is most clearly superhuman - the taxonomy's living proof.
The efficiency inflection (2025-2026)
- - DeepSeek R1 + the open-weight wave showed frontier-adjacent capability at a fraction of the cost - bending the arc from pure scale toward efficiency (fin-ai-efficiency-counter-thesis).
Honest limits
Releases, dates, and benchmarks are fact; "reasoning"/"agentic"/"AGI-adjacent" are marketing-laden labels (graded); capability claims are provider-reported unless independently benchmarked; benchmark-gaming is real. The point of the by-branch view: the capability frontier is plural, and equating "AI progress" with "LLM releases" mis-measures it.
Sources: model releases + benchmark milestones across branches. Cross-refs: ImageNet_AlexNet, Diffusion_Models, Transformer, Multimodal_AI, Agentic_AI, Reinforcement_Learning, AlphaGo_AlphaFold, DeepSeek, Deep_Learning.
AI-lab chronology - focus, capability + the EA / safety-reg pathways
(Foundational installment of the AI-chronology Tab epic. Taxonomy first: "AI" is not "LLM." AI spans symbolic reasoning, expert systems, machine learning, reinforcement learning, computer vision, protein folding, robotics, and more. LLMs (transformer language models, 2017+) are one recently-dominant branch - this block, and the Tab it seeds, keep that distinction.)
Capability + focus timeline (fact; selected)
- - 2012 - AlexNet: deep convolutional nets win ImageNet; the deep-learning era begins (vision, not language).
- - 2014 - Google acquires DeepMind; GANs introduced.
- - 2016 - AlphaGo beats Lee Sedol (deep RL + Monte Carlo tree search) - a non-LLM landmark.
- - 2017 - "Attention Is All You Need" - the transformer, the substrate for LLMs.
- - 2018-2020 - BERT/GPT-2/GPT-3: scaling laws; focus shifts hard toward large language models.
- - 2020-2024 - AlphaFold solves protein structure (Nobel 2024) - again, not an LLM.
- - 2022 - ChatGPT: LLMs go mass-market; the capital + attention flood begins.
- - 2023-2026 - frontier race (GPT-4/5, Claude, Gemini, Grok, DeepSeek), the efficiency counter-wave (fin-ai-efficiency-counter-thesis), and the compute/energy/debt buildout (spec-ai-datacenter-debt-spv).
The EA / Open-Philanthropy / safety-money pathways (fact; influence graded)
- - 2017: Open Philanthropy -> OpenAI (~$30M + a board seat) - the first EA-money-into-a-frontier-lab move, framed as buying safety influence.
- - Open Philanthropy (Moskovitz/Tuna) is the main EA grantmaker; SFF (Jaan Tallinn), FLI, and Longview channel money into safety orgs + policy shops.
- - 2021: Anthropic spins out of OpenAI with an explicit safety framing; early capital includes ~$500M tied to FTX/SBF (later in the bankruptcy estate).
- - These rails feed the safety-policy apparatus (Encode, think tanks) that produced the pause/pacing letters - the predicate in spec-buist-v-anthropic-pacing.
Safety regulation as arbitrage? (interpretation, labeled)
- - 2024: California SB 1047 (mandatory frontier testing + kill-switch + liability), co-sponsored by Encode, passed the legislature and was vetoed by Newsom - the template for the 2025-26 state-vs-federal fight.
- - 2025-2026: the labs' collective "build the brake, don't pull it" posture (Anthropic support-with-conditions; OpenAI's shift) is what critics recode as regulatory arbitrage - safety rules that raise rivals' compliance cost and entrench incumbents - culminating in the 2026 pacing-as-cartel antitrust framing (Buist v. Anthropic) and the failed CLARITY cloture. Whether "safety" here is sincere, strategic, or both is contested; the map records the documented positions + the graded critique, not a verdict.
Honest limits
Grants, dates, milestones, and bill status = fact (amounts as reported); the "regulatory arbitrage / capture" reading = interpretation, clearly labeled; motive is not adjudicated. Structural overlay - no financial-core edges.
Sources: lab histories; Open Philanthropy / SFF / FLI grant records; SB 1047 record; the 2026 pacing fight. Cross-refs: OpenAI, Anthropic, DeepMind, Open_Philanthropy, Effective_Altruism, Survival_and_Flourishing_Fund, Encode_AI, SB_1047, AI_Safety_Regulation, spec-buist-v-anthropic-pacing.
AI safety-evals ecosystem + the EA money map (incl. DataRepublican)
(AI-chronology Tab. This is the "targets of work / who evaluates whom" layer under the spec-ai-lab-chronology-ea-pathways safety-reg story.)
The eval institutions (fact)
- - US AISI (at NIST; renamed the Center for AI Standards and Innovation / CAISI under the current administration) + UK AISI run pre-deployment model testing - the government-eval layer. The rename + de-emphasis is itself a policy signal.
- - METR (Model Evaluation & Threat Research) - dangerous-capability + autonomy/task-horizon evals labs cite in safety cases.
- - Apollo Research - deception / scheming evals (does the model strategically mislead?).
Self-governance instruments (fact; sincerity graded)
- - Anthropic's Responsible Scaling Policy (AI Safety Levels) and OpenAI's Preparedness Framework are voluntary capability thresholds + commitments - pitched as the template for formal regulation. Critics call that mechanism regulatory capture / safety-washing (lab self-rules become the policy baseline). Interpretation, labeled.
The EA money map (fact of funding; conflict framing labeled)
The same Effective Altruism capital - Open Philanthropy (+ SFF/FLI) - substantially funds the labs' safety teams, the eval nonprofits (METR, Apollo), AND the policy advocates. That triple role is why a "follow the money" trace keeps landing on EA (SB 1047 -> the 2026 pacing letters -> spec-buist-v-anthropic-pacing). Coordinated-intent claims remain unsupported; the linkages are documented.
DataRepublican's "EA Explorer" (independent open-source map)
Prompted by Jacob Coxon's viral Sep-2026 Anthropic-resignation/pause post - and the ensuing tracing to METR and EA nonprofits - DataRepublican (pseudonymous data analyst; tagline "Exposing where the money flows") published EA Explorer: a people-and-funding network map plus a "Their words" verbatim-quote browser over 25+GB of EA forum material (every quote links to its source). It applies her charity-graph tooling (multi-root BFS over IRS 990 filings, taxpayer-fund tracing) to the EA/Open-Philanthropy network.
- - Link: datarepublican.com (charity graph: /expose; write-up on her Substack).
- - Honest handling: it is an independent analysis with an explicitly critical POV - a genuine open-data contribution (sourced quotes, 990-based graphs), not a neutral arbiter. Treat specific funding-flow figures as leads to verify against primary 990s; treat the framing as her viewpoint.
Honest limits
Institutions, evals, RSPs, and grants = fact; "safety-washing / regulatory capture" = interpretation; DataRepublican's map = independent, critical, source-linked (verify dollar claims vs 990s); no coordinated-conspiracy claim is asserted. Structural overlay - no financial-core edges.
Sources: NIST/CAISI, UK AISI, METR, Apollo Research; Anthropic/OpenAI RSP+Preparedness; Open Philanthropy grants; datarepublican.com + Substack. Cross-refs: US_AISI, METR, Apollo_Research, Responsible_Scaling_Policy, Anthropic, OpenAI, Open_Philanthropy, Effective_Altruism, AI_Safety_Regulation, DataRepublican, spec-buist-v-anthropic-pacing.
The 'regulate us' push - safety vs capture + the infosec/OSS backlash
(AI-chronology Tab. This block presents both sides, attributed - the labs' stated safety rationale AND the infosec/open-source capture critique - and asserts neither as proven intent.)
The "regulate us" moves (fact)
- - May 2023: Sam Altman told the US Senate he welcomed regulation and proposed a federal licensing regime + a new agency for frontier AI - the archetypal "please regulate us" moment.
- - Dario Amodei / Anthropic pair existential-risk warnings ("a real chance things go catastrophically wrong") with support-for-regulation-with-conditions and the Responsible Scaling Policy - the "pacing" posture.
- - The pattern peaks in the 2026 pacing-as-cartel antitrust framing (spec-buist-v-anthropic-pacing).
The capture critique (contested interpretation, attributed)
Critics - Yann LeCun (Meta's chief AI scientist), a16z / Marc Andreessen, much of the open-source + infosec community, and "follow the money" analysts like DataRepublican - read the push as regulatory capture:
- - Licensing + compute-threshold + liability rules fall hardest on open-source and startups (who can't absorb compliance or model-liability), so "safety" can function as a barrier to the open ecosystem. The differential burden is real; intent is not adjudicated.
- - Labs lobby for the rules that bind rivals while resisting provisions that bind themselves.
- - RSP-as-template = the incumbent's own practices become the regulatory floor.
- - SB 1047 was the flashpoint: OSS devs said it would chill open release; safety proponents said it only touched the largest models - the same bill read as protection or capture.
The framing-manipulation claim (contested)
- - Existential-risk framing (EA / longtermist) supplies the moral urgency ("we might kill everyone -> license us") and crowds out present-harm critiques - bias, surveillance, labor displacement, market concentration. LeCun calls the x-risk framing "preposterously overblown"; the TESCREAL critique makes the same structural point.
- - Per the spec-ai-safety-evals-ecosystem EA money map, the same funders back the labs' safety teams, the evaluators, AND the advocates - so a "follow the money" trace (DataRepublican, infosec Twitter) keeps landing on the EA/Open-Philanthropy network. That the linkages exist is documented; coordinated intent remains unsupported.
The honest balance
- - Steelman for the labs: frontier capability may genuinely warrant guardrails; open weights are irreversibly proliferable (a real risk argument); some regulation is better designed early.
- - Steelman for the critics: the loudest "regulate us" voices are the incumbents who benefit from a licensed moat; the x-risk frame is unfalsifiable + self-serving; open models are auditable and diffuse power.
- - The tell this map flags: every party is an interested party - the labs (moat), the VCs (portfolio), the EA-funded orgs (mandate + funding), the open-source camp (their model of the world). The value is mapping who benefits from which framing, not declaring a winner.
Honest limits
Testimony, essays, bills, lobbying, and funding = fact; "capture / threat-inflation / framing-manipulation" = contested interpretation, attributed to named critics; the labs' stated safety rationale is presented alongside it; no coordinated-conspiracy claim is asserted. Structural overlay - no financial-core edges.
Sources: Altman 2023 Senate testimony; Amodei essays + Anthropic RSP; SB 1047 record; LeCun / a16z / OSS + infosec commentary; DataRepublican (datarepublican.com). Cross-refs: Sam_Altman, Dario_Amodei, Yann_LeCun, Marc_Andreessen, OpenAI, Anthropic, Regulatory_Capture, Open_Source_AI, Existential_Risk_Framing, Effective_Altruism, DataRepublican, Uncensored_AI, SB_1047, AI_Safety_Regulation.
Open-weight AI, abliterators + redistributors
(AI-chronology Tab, #247. Taxonomy: these are mostly LLMs - one branch of AI - but they define the open-weight vs closed-frontier split.)
Chronology (fact)
- - 2023: Meta's Llama opens the open-weight era - downloadable models anyone can run + fine-tune, splitting the field from closed frontier APIs. Mistral (France) follows with permissive releases.
- - 2024-2026: Qwen (Alibaba, Apache-2.0) becomes one of the most-used open families; DeepSeek (R1/V-series) makes efficient open weights a frontier-adjacent force (fin-ai-efficiency-counter-thesis).
- - Throughout: Hugging Face is the dominant redistributor/hub (mirrors + versions the weights); OpenRouter is the API aggregator (by Q2 2026, Chinese open models crossed ~30% of its developer token traffic); Ollama + llama.cpp make local, offline running trivial.
The de-censoring layer (fact of technique; quality varies)
- - Abliteration (automated by tools like Heretic) edits an open model's weights to remove refusal behavior - converting a censored-when-served model into an uncensored local one.
- - Fine-tuning + abliteration produce the "uncensored" re-releases that populate the local-AI ecosystem, where served-model guardrails simply don't apply.
The open-weight paradox (interpretation, labeled)
Once weights are public, they are de-censorable downstream - so served-model safety and frontier-model regulation are structurally leaky. Chinese open models sharpen the paradox: censored-at-source when served from China, yet freely de-censorable once downloaded. This is the honest check on the safety-regulation story from the spec-ai-lab-chronology-ea-pathways block: you can regulate the API and the frontier lab, but not the weights already in the wild.
Honest limits
Releases, licenses, hubs, and the abliteration technique are fact; de-censoring quality is variable; "undercuts served-model safety / frontier regulation is leaky" is interpretation (though the leak is demonstrated). Structural overlay - no financial-core edges.
Sources: model releases (Llama/Mistral/Qwen/DeepSeek); Hugging Face / Ollama / OpenRouter; abliteration research + Heretic. Cross-refs: Meta, Llama, Alibaba, Qwen, DeepSeek, Mistral, Hugging_Face, OpenRouter, Ollama, Heretic, Abliteration, Uncensored_AI, AI_Safety_Regulation.
Chronology + history are graded overlays (fact / interpretation per claim); see the linked research pages + methodology.