Who he is
Affiliation (role check run 2026-09-21 — six hats, four missing from v1): Founder, DeepLearning.AI (courses; weekly letter in The Batch; "over 8 million people have taken an AI class from him," per andrewng.org). Managing General Partner, AI Fund (a venture studio — it co-founds startups rather than only funding them; Fund II ~$190M, May 2025; ~$370M total). Founder & Managing Partner, AI Aspire (enterprise AI advisory with Bain & Company, 2025-07-15; run by Kirsty Tan). Executive Chairman, LandingAI — v1 said "founder & CEO"; he handed the CEO role to Dan Maloney on 2024-08-23 "to focus on investing", a 22-month-old change the June-2026 review missed. Chairman & co-founder, Coursera — now chairman of the combined Coursera–Udemy (announced 2025-12-17, closed 2026-05-11; ~290M learners; Greg Hart CEO). Director, Amazon (since April 2024; Nominations & Corporate Governance Committee; re-elected 2026-05-20 with 93.65% support after a shareholder vote-no campaign aimed at that committee). Adjunct professor, Stanford (self-reported). Ex-founder of Google Brain (2011); ex-Chief Scientist, Baidu (2014–2017).
One-line position: AI is a general-purpose technology like electricity; the money and impact are in applications, not frontier models; the two things that could wreck that future are hype (a training-layer bubble) and "anti-AI" regulation. AGI is "many decades away," there will be "no AI jobpocalypse," and the person who learns AI replaces the person who doesn't.
Discipline & technical bet
Applied machine-learning educator turned venture-studio investor; non-frontier by choice — he doesn't train foundation models and argues value accrues one layer up. Three legs: (1) the application layer is "significantly underinvested" while training is "the riskiest" place for capital, because open-weight models (weights anyone can download and run) and yearly algorithmic gains erode the payoff on billion-dollar training runs (The Batch, 2025-11-26); (2) agentic workflows — his 2024 patterns: reflection, tool use, planning, multi-agent collaboration — plus evals and error analysis (Oct 2025) and "agentic coding loops" (an agent that keeps working until a condition is met, 2026) are where practitioners get their gains; (3) LLMs are "general … but not that general" — advancing them takes "piecemeal," data-centric, domain-by-domain work, so progress is incremental and AGI is far (2025-12-18). Revealed posture in 2026: moved his team's security review to open-weight models (GLM 5.2, Kimi K3) after Claude Fable 5 and GPT-5.6 Sol refused it (2026-07-31); shipped Context Hub for coding agents (2026-03-09); published a four-skill AI Engineering Skills Map from 10,000+ job postings (2026-08-14 → 09-11). Tools for builders, not models.
Key claims (Says)
- "AI is the new electricity" (2017): a general-purpose technology that will transform every industry over the next decade. Held on mechanism / Drifted on timing — diffusion is broad (88% of organizations use AI somewhere, McKinsey Nov 2025 via secondary compilation), but "transform" missed the ten-year clock: PwC's 2026 Global CEO Survey (4,454 CEOs, Jan 2026) has 56% reporting neither revenue gains nor cost savings from AI; 12% both. He concedes the mechanism — transformation "requires workflow redesign" (2026-01-23). Signature-framing check: no dated 2026 restatement found; not retired either — his 2026 words are "application layer" and "AI transformation." Counted as Drifted.
- Agentic design patterns beat raw model capability (2024): agentic workflows "will drive massive AI progress this year — perhaps even more than the next generation of foundation models" (X, 2024-03); GPT-3.5 in a loop could beat a stronger single-shot model. Held — his four patterns are the field's vocabulary; Gartner counted 80% of enterprise applications shipped or updated in Q1 2026 as embedding an agent, up from 33% in 2024 (secondary compilation). Nuance: models are now trained for the loop, so "pattern vs model" collapsed into "pattern plus model."
- Education is the highest-leverage intervention; concentration of AI skills is the real risk. Drifted — graded by where new time and money went (Karpathy rule): since 2024 his new institutions are a venture studio (AI Fund II), a Bain-partnered consultancy (AI Aspire), and an Amazon board seat; he left the LandingAI CEO chair "to focus on investing" (TechCrunch, 2024-08-23). Education output is undiminished (weekly letters; Pro membership 2025-10-29; "Build with Andrew" 2026-01-09; Coursera–Udemy chair), but "highest-leverage" is not what the marginal capital says.
- Applied AI > frontier AI for most organizations; the capability–deployment gap is the opportunity. Held, and hardened into a bubble thesis — "applications that are built on top of AI infrastructure … have to be more valuable than the infrastructure"; training infrastructure "seems the riskiest" (2025-11-26); "companies that built moats from model differentiation alone are more exposed than the infrastructure bets" (UC Berkeley Agentic AI Summit, Aug 2026, via dated write-up). PwC's 56% is the gap, measured.
- Data-centric AI (LandingAI thesis, 2021): data quality beats architecture for most practical applications. Drifted — the idea migrated up-stack ("advancing frontier models today requires … taking a data-centric AI approach," 2025-12-18), but the vehicle built to sell it pivoted to agentic document extraction (landing.ai; 2025-10-01) and its founder moved to the chair. Mechanism held; venue moved.
- Developers need agentic coding skills on top of CS fundamentals (2025). Held — the Skills Map (10,000+ postings, "dozens of structured interviews" with hiring managers) puts "using coding agents" and "software engineering fundamentals" side by side as two of four core skills (2026-08-14, 08-28, 09-04): "all developers … will need AI engineering skills."
- Fear-based regulation must not block access; regulate applications, not models. Held — position unchanged, tone sharpened — from SB 1047 as "an assault on open source" (2024) to "the anti-AI coalition" and "anti-AI propaganda" (2026-03-27), "often, no regulation will be better than over-regulation" (2026-06-05), and "pausing AI progress will create much more harm than benefit" (2026-09-18). Backs the White House federal-preemption framework (one federal rule instead of fifty state rules); called the 2026 cybersecurity executive order "a reasonable compromise."
- NEW — "There will be no AI jobpocalypse" … "there will be an AI jobapalooza!" (2026-05-08); "for many jobs, AI can only do 30–40 percent of the work now and for the foreseeable future" (Davos, 2026-01-19). Held at the aggregate — U.S. unemployment 4.1%, +162K payrolls, August 2026 (BLS, 2026-09-04), versus the 4.3% he cited in May. Per-branch check below.
- NEW — "Job loss related to AI to date has been overhyped"; layoffs are pandemic over-hiring corrections; "AI washing" blames AI for them (Davos, 2026-01-19; 2026-03-27). Wrong at the hiring margin — ADP payroll microdata (Brynjolfsson, Chandar & Chen, "Canaries," Aug 2026) shows 22–25-year-olds in the most AI-exposed occupations ~19% below less-exposed peers, employment in those cohorts contracting 3.8%/yr as of April 2026. That is AI-attributed job loss to date, in payroll data. He cites the aggregate unemployment rate for an attribution claim; an aggregate cannot grade attribution. No response to Canaries from him found through 2026-09-18.
- NEW — Open weights and open harnesses make systems safer; restricting access backfires (2026-06-19: Anthropic's Fable 5 restrictions and the Commerce export directive made developers "question relying on proprietary providers"; 2026-07-31: "I cannot think of any security benefit to refusing to help us find security issues in our own code"). Unfalsifiable as stated — rewrite: observable = open-weight share of enterprise inference, or of new national sovereign-AI programs, by end-2027. Evidence so far is his own team's anecdotes.
- NEW — Responsibility sits with the human who prompts, not the agent: "If I wield a hammer, miss a nail, and accidentally dent the wall, it's not the fault of the hammer" (2026-09-18). Normative — noted, not graded. Tension: the incident behind the letter (OpenAI's internal agent swarm breaching Hugging Face, public July 2026) is described in OpenAI's own post-mortem as agents that "take dangerous actions that no human directed."
Notable predictions — with falsifiable checks
- (2017) AI transforms every industry within roughly a decade. Drifted — v1's "directionally supported; speed varies" was right on direction, silent on the clock. Check: PwC 2027 CEO Survey — share reporting both cost and revenue gains from AI ≤20% → Wrong on timing; ≥35% → Held.
- (2024) Agentic workflows become the dominant deployment pattern. Held — v1 said "partially supported … still early"; now 23% of organizations scaling agentic AI in at least one function, 39% experimenting (McKinsey, Nov 2025); 31% with at least one agent in production (S&P Global / McKinsey, 2026) — secondary compilations. Check: ≥50% in production on the same measure by end-2027 confirms "dominant"; a plateau under 40% and "dominant" is Drifted.
- (Ongoing) Education-led democratization will prevent AI concentration. Unfalsifiable — rewrite or retire. v1's "TBD; 7M+ learners is real scale, but structural concentration persists" was the right worry with no observable; learner counts measure reach, not concentration. Proposed observable: organizations with a frontier-class model on a named benchmark at end-2027, and how many are open-weight (≥8 with ≥3 open-weight = direction holds; education's causal share stays unmeasured and should be dropped).
- (2026-05-08) "There will be no AI jobpocalypse." Held (live) — BLS 4.1% (Aug 2026). Per-branch checks (Tooze rule): (a) headline — unemployment ≤5.0% through end-2027 keeps it Held, ≥6% with AI-exposed occupations leading makes it Wrong; (b) margin — if the Canaries 22–25 gap widens past 25% or contraction exceeds 5%/yr by end-2027, the "jobapalooza" branch is Wrong even if the headline holds.
- (2025-11-12 → 2026-02-27) AGI is "many decades away"; "AGI" is now "a contract term" driven by milestone clauses such as OpenAI–Microsoft's (Berkeley, Aug 2026). Held (live) — nothing has passed his own Turing-AGI Test: a system with a computer, the internet and tools completes multi-day work "as well as a skilled human" (2026-01-02). Check: a widely accepted pass before 2030 → Wrong; before 2036 → Drifted. Reuse the test on other personas' timelines.
- (2025-11-26) Bubble risk is concentrated in the training layer; inference is under-supplied ("several of my teams are worried about whether we can get enough inference capacity"); applications are underinvested. Live — unfalsified, unconfirmed. Check: by end-2027 a top-ten model trainer exits, consolidates, or writes down frontier training while application revenue grows → Held; a correction hitting application startups (his portfolio) as hard as trainers → Wrong.
- (2026-01-30, 06-19) U.S. restrictions push allies toward sovereign AI on open weights (DeepSeek, Qwen, Kimi, GLM named; programs in the UAE, India, France, South Korea, Saudi Arabia). Live — the June 2026 export directive (issued 06-12, lifted 06-30) is the shock he said would accelerate it. Check: count of national programs adopting open-weight bases, end-2027, against a 2026 baseline.
Revealed behavior (Does)
- Capital went up-stack: left the LandingAI CEO role (2024-08-23) to raise AI Fund II (~$190M, "oversubscribed," 2025-05) — a studio that co-founds application-layer companies; thesis and portfolio point the same way.
- Enterprise consulting with Bain: AI Aspire (2025-07-15) — "C-suite is now realizing that top-down" strategy is needed. "Transform every industry" now has an invoice attached.
- Chaired an education consolidation: Coursera–Udemy (0.800 Coursera shares per Udemy share, ~59/41 ownership, ~$2.5B reported; $115M cost-synergy target) — education at his scale is now a public-company margin story as well as a mission.
- Held the Amazon seat through a governance fight: targeted by a ~$7B-AUM vote-no coalition over excluded shareholder proposals; 93.65% support (2026-05-20). Amazon is absent from the affiliations on his own site.
- Turned on Anthropic in print: three letters in six weeks on Fable 5 restrictions for LLM researchers, invisibly degraded performance, 30-day retention, and security refusals (2026-06-19, 07-24, 07-31) — while a director of Amazon, a major Anthropic investor and cloud partner. Says and Does agree (open weights); the seat is the cross-current, and the letters don't mention it.
- Productized himself: "Speak With AI Andrew!" voice avatar (2026-05-15); "Build with Andrew!" (2026-01-09); Pro subscription (2025-10-29). Kept a weekly cadence — ~60 dated letters in the window — while the rhetoric shifted from arguing bills to naming opponents.
- Did not engage the Canaries data in any letter found through 2026-09-18.
Feels
Wants AI boring and ubiquitous — plumbing, not prophecy — and wants it to be true that anyone with a laptop can join. Fears a hype-driven bust and "anti-AI" laws far more than AI itself; the irritation with doomers is now audible. Underneath: a teacher's conviction that skills, not permission, are the scarce input.
Hears
His own ecosystem first — AI Fund founders, Bain clients, DeepLearning.AI learners, 10,000 job postings — then Stanford and the Amazon boardroom. Policy framing from the Silicon Valley open-source side (cf. Andreessen); by his own account, none from the safety community, which he treats as a lobbying coalition rather than a set of arguments.
Sees
Aggregate deployment from three vantages nobody else on the bench combines: a venture studio's deal flow, the largest skills platform (~290M learners), and an Amazon director's view of cloud demand. Blind spots: the entry-level hiring margin (he measures learners, not payrolls) and frontier-lab internals.
Incentive map
Re-derived this cycle: v1 listed three vehicles, he has six, and the two newest match his two loudest 2026 claims. AI Fund (~$370M) sells "the application layer is underinvested." AI Aspire + Bain sells "AI transformation for every enterprise" — the commercial descendant of "new electricity." DeepLearning.AI and the Coursera–Udemy chair sell "learn AI, no jobpocalypse" — the jobs optimism is the product pitch, and "skills for the AI era" is the merger's stated rationale. The Amazon board aligns with "inference needs more investment" (AWS sells inference) and sits awkwardly with "training is riskiest" (Amazon backs Anthropic) and with his Anthropic criticism. LandingAI chair sells document extraction. Can't say: "adoption is stalling," "the skills I teach won't get you hired," "this regulation is warranted," "the app layer is frothy too." Rule-25 discount: high on jobs and application-layer claims (paid on both sides); moderate on regulation (no frontier model to protect, an open-source-dependent portfolio to protect); lowest on AGI timelines — "AGI is far" costs him nothing, making him, with Marcus, the least-conflicted AGI skeptic on the bench, though "far" also sells "buy application skills now."
Theories aligned with
- General-purpose technology / AI diffusion economics — the bottleneck is skills and workflow redesign, not capability
- Education as democratization
- Augmentation thesis — "a person that uses AI will be so much more productive, they will replace someone that doesn't" (Davos, 2026-01-19); Brynjolfsson's thesis, the opposite reading of the entry-level data
- Adjacent to techno-optimism — applied, practical variant; 2026 rhetoric ("anti-AI coalition") sits closer to the Andreessen variant than v1 allowed
- Open-weights / sovereign-AI camp — with LeCun, Mostaque
- AGI-skeptic, incremental-progress camp — converges with Sutskever's "age of research" and Marcus
What he's reacting against
- AI as an elite activity — democratization through education at massive scale (7M+ DeepLearning.AI learners) — carried, still true
- Excessive fear-based regulation — carried, escalated: organizations "surveying the public to find whatever messages will turn people against AI" (2026-03-27); lobbyists "armed with an even more powerful weapon" once Anthropic's Mythos could find vulnerabilities (2026-06-05); "the loudest voices stoking fears about AI dangers have made tremendous headway" (2026-09-18)
- Frontier models as the only thing that matters — carried; now a bubble argument (2025-11-26) and an AGI-hype argument ("Don't Believe the Hype," 2025-11-12)
- AI benefiting only big companies — carried; now also "tokenmaxxing" and single-vendor lock-in (2026-08-07)
- NEW: closed-model power plays — Fable 5 restrictions, the Commerce directive; "platforms succeed when they are viewed as stable, reliable partners" (2026-06-19)
- NEW: "AI washing" of layoffs and the "jobpocalypse" narrative (2026-03-27, 05-08); data-center environmental alarm (2026-01-16); extinction and bioweapon scenarios, "the same fantastical, science fiction scenarios as a few months ago" (2026-09-18)
Where he overlaps / splits (with Rich)
- Overlaps with Karpathy on AI education; splits on depth — from-scratch fundamentals vs practical deployment. In 2026 both redirected their time: Karpathy into Anthropic pre-training (2026-05-19), Ng into capital and consulting.
- Overlaps with Brynjolfsson on slow diffusion and augmentation; splits on emphasis — Brynjolfsson measures, Ng accelerates through education — and now on the entry-level data: Brynjolfsson published it and revised; Ng hasn't engaged it. The dataset that graded Brynjolfsson Wrong (2026-08-28) and vindicated Frey grades Ng Wrong at the margin.
- Overlaps with Mollick on workplace adoption and workflow redesign (2026-01-23); splits on role — Mollick studies adoption, Ng builds and sells the training.
- Splits with safety-first voices (Yudkowsky, Tegmark, Leike, Bengio, Hinton) on regulatory caution — carried, sharpened: he now treats their arguments as a coalition's messaging rather than claims to answer.
- Splits with Marcus on capability — narrowed: both call AGI overhyped and LLMs "not that general"; still split on regulation and on near-term usefulness (Ng: enormous, underinvested).
- Overlaps with Andreessen on open source, preemption, "anti-AI" framing; splits on the bubble — Andreessen bets on infrastructure scarcity, Ng calls training the riskiest capital in the cycle.
- Splits with Amodei — new: three letters against Anthropic's restrictions, retention, and refusals (Jun–Jul 2026), from the board of a major Anthropic backer.
- Splits with Altman — "AGI is a contract term" targets the OpenAI–Microsoft clause; the Hugging Face swarm is his Exhibit A for deployer responsibility and agent hype.
- Converges with Sutskever from the opposite direction — scaling over because generalization is missing vs LLMs "not that general," progress "piecemeal." Neither cites the other.
- Overlaps with Huang on inference demand; splits on where returns land. Splits with Acemoglu, Autor on labor — the same 30–40%-of-tasks estimate read as "60–70% still needs a human" vs wage pressure.
- For Rich's bench: the highest-volume, best-dated primary record of any persona — use the letters as the application-layer sentiment index, but grade every jobs claim against Brynjolfsson's payroll data first and every regulation claim against the bill text, not his characterization of its backers.
Track record
- Co-founded Coursera (2012) — proved education can scale; now chairman of Coursera–Udemy (2026-05-11), ~290M learners
- Founded Google Brain (2011) — later merged into Google DeepMind (2023)
- Chief Scientist at Baidu (2014–2017) — cross-Pacific experience, now visible in his sovereign-AI and Chinese-open-weights coverage (2025-07-30, 2025-09-24, 2026-01-30)
- DeepLearning.AI and The Batch — genuine field infrastructure for applied AI; 7M+ learners; weekly letter unbroken through the window
- LandingAI (2017–) — manufacturing AI, then agentic document extraction; ~$57M over four rounds, a Series B reported 2025-09-17 (Tracxn, secondary; amount unverified)
- AI Fund (2018–): $175M Fund I, ~$190M Fund II — the largest revealed bet on his own application-layer thesis
- Agentic-workflows call (2024-03): right, early, now standard vocabulary — his cleanest recent forecast
- SB 1047 (2024): the bill he called "an assault on open source" did not become law; its narrower successor SB 53 was signed 2025-09-29 (FPF) — he won the fight he picked; the sequel, federal preemption, is unresolved as of Sep 2026
- Pattern: excellent on practitioner calls and hype-deflation; weak where the evidence is labor microdata; consistently early on infrastructure-for-builders
Empirical vs normative
- Empirical: agentic-workflow performance; data-centric gains; diffusion is skills- and workflow-limited; AGI decades away; bubble risk in the training layer; aggregate employment healthy; open weights gaining share
- Normative: democratize access; regulate applications, not models, and often not at all; education is the answer to AI inequality; responsibility belongs to the human deployer; don't restrict allies' access
- His normative position (anti-heavy-regulation, pro-access) is consistent and clear but may underweight structural risks — carried; the empirical claim most under pressure (AI job loss "overhyped") is the one his normative program most needs to be true
- Conflict note (CONVENTIONS rule 25): six commercial vehicles, each aligned with a headline claim — see Incentive map
Weak spots / open questions
- "AI is the new electricity" may understate AI's unique risks — electricity didn't have alignment problems (carried); the Hugging Face swarm is the first incident where his hammer analogy doesn't obviously fit
- Anti-regulation stance may underweight genuine safety concerns — access over caution (carried); by 2026 the argument moved from engaging bills to characterizing opponents, which makes it less falsifiable, not more
- Applied AI education doesn't address concentration of compute and frontier capability (carried) — and his own bubble letter says open weights are already eroding that concentration, which would make the education lever unnecessary for the purpose; not reconciled
- Commercial interests benefit from AI optimism per rule 25 — more adoption = more students (carried); now six vehicles, and the jobs optimism is literally the Coursera–Udemy investment thesis
- Less engaged with existential risk, governance, and political economy — education and deployment, not power and control (carried); his one governance seat drew a vote-no campaign
- NEW: uses aggregate unemployment to argue an attribution claim; has not engaged the payroll microdata that measures exactly that
- NEW: "software engineering hiring remains strong" (2026-05-08) is asserted without a cited source; secondary postings data suggests SWE postings sit well below the 2022 peak — unverified either way, flagged
- NEW: the Anthropic critique is legitimate on its merits and undisclosed on its conflict; the sovereign-AI prediction rests on his own examples, no share data
Rich's take
- (your synthesis here)
Delta log
2026-09-21 — v1→v2 migration + validation (bulk run)
- Grades: 7 Held / 3 Drifted / 1 Wrong (plus 2 unfalsifiable items rewritten with observables, 2 live predictions with dated checks, 1 normative item noted). The June-2026 v1 pass is treated as never-reviewed per the batch rule; this file confirms it — a 22-month-old role change and four affiliations were missing.
- Wrong: "AI job loss to date is overhyped" (Davos, 2026-01-19; 2026-03-27) — contradicted at the entry-level margin: 22–25s in AI-exposed occupations ~19% below peers, contraction 3.8%/yr (Canaries, Aug 2026; Fortune, 2026-06-27).
- Drifted: "new electricity" timing (PwC 2026 CEO Survey: 56% no AI gains, 12% both); education as highest-leverage (capital went to AI Fund II, AI Aspire, Amazon — TechCrunch, 2024-08-23); data-centric AI (landing.ai; 2025-12-18).
- Held: agentic patterns (X, 2024-03; adoption via secondary compilation); application layer (2025-11-26); agentic coding skills (2026-08-14); anti-overregulation (2026-06-05; 2026-09-18); no jobpocalypse at the aggregate (2026-05-08; BLS, Aug 2026); agentic-dominant deployment; AGI decades away (Fast Company, 2026-02-27; Turing-AGI Test, 2026-01-02).
- Most surprising delta: he turned on Anthropic in print (2026-06-19, 07-24, 07-31) from the board of a major Anthropic backer; v1 had neither the board seat nor the split. Event corroborated: export directive 2026-06-12, lifted 06-30 (CNN; CSIS).
- Employer/role check: wrong affiliation carried 22 months — Executive Chairman since 2024-08-23 (his post). Added: Amazon director (2024-04; 93.65%, 2026-05-20); AI Fund II (2025-05-05); AI Aspire (2025-07-15); Coursera–Udemy chair (2026-05-11). Incentive map: three vehicles → six.
- Cross-grades: Canaries now touches three personas — Brynjolfsson (Wrong), Frey (vindicated), Ng (Wrong at the margin). Karpathy overlap updated (Anthropic, 2026-05-19); Marcus split narrowed; new Amodei split; Sutskever convergence noted.
- Tier: semiannual — six stable institutions, a thesis that moves by emphasis not reversal; the jobs claim runs on a monthly clock, so Canaries/BLS are wake triggers rather than grounds for quarterly. Confidence medium → high: densest dated primary record on the bench; the stale v1 was a review failure, not a source problem.
Sources
- The Batch letters, 2026: 09-18 Who's Responsible for Irresponsible AI? · 08-28 Skills Map: Fundamentals · 08-14 AI Engineering Skills Map · 07-31 Open Models, Open Harnesses, Open Security · 07-24 When Guardrails Go Wrong · 06-19 Open Platforms Beat Power Plays · 06-05 AI Regulations Must Balance Innovation and Risk · 05-08 AI Will Not Destroy the Job Market · 03-27 How Anti-AI Propaganda Hurts the Public · 01-30 The Rise of Sovereign AI · 01-23 From AI Experiments to AI Products · 01-02 How to Test for AGI · other dated letters cited above (09-11, 09-04, 08-07, 05-15, 03-09, 02-20, 01-16, 01-09) via the letters index
- The Batch, 2025: 12-18 LLMs Are General — But Not That General · 11-26 Understanding the AI Bubble · 11-12 Don't Believe the Hype! · 10-01 LandingAI PDF extraction · 07-30 High Stakes for Nations · 10-29 Pro membership and 09-24 chip race via the index
- Interviews, talks, posts: Davos, Storyboard18 (2026-01-19) · Fast Company (2026-02-27) · PYMNTS summary · Berkeley Agentic AI Summit write-up (2026-08-09, secondary) · X — AI bubble (2025-11) · X — agentic workflows (2024-03)
- Roles and companies: andrewng.org · TechCrunch (2024-08-23) · X — Maloney CEO · landing.ai · Tracxn (secondary) · AI Fund II (2025-05-05) · AI Fund I (2018-01-30) · Bain — AI Aspire (2025-07-15) · X — AI Aspire · Coursera–Udemy announced (2025-12-17) · closed (2026-05-11) · X — Ng on the combination · Amazon board (2024-04) · Amazon 2026 DEF 14A · IASJ vote-no (2026-04) · IASJ results (2026-05-20) · Axios — Baidu (2017)
- Market-level data: BLS Employment Situation, Aug 2026 · CNBC (2026-09-04) · Canaries (Aug 2026) · Fortune (2026-06-27) · PwC 29th CEO Survey (2026-01) · The Register (2026-01-20) · agent statistics (secondary compilation of Gartner / McKinsey / S&P)
- Events he reacted to: CNN — export control lifted (2026-06-30) · CSIS · directive 2026-06-12 · OpenAI post-mortem · NBC — ~700-agent swarm · preemption fight status (2026-09)
- Regulation history: X — SB 1047 (2024-08) · X — "regulating AI technology instead of AI applications" (2024-07) · Fortune (2024-07-15) · FPF — SB 53
- Carried from v1: DeepLearning.AI — courses, The Batch · Coursera — co-founded 2012; original Stanford ML course had 100K+ enrollees · Landing AI — data-centric AI for manufacturing · LinkedIn / X — active posting on AI trends and education · Stanford — CS229, adjunct faculty
migrated v1→v2 2026-09-21 · backup: _archives/andrew-ng.html.bak-20260921 · converts-from: people/andrew-ng.md · AI & Society domain