Timeline

Each row gives the date, what happened (linked to its source) and why it matters for this course. Dates are publication or announcement dates as the source states them. Where the course’s research disagrees on a date, or only one source reports it, the row says so.

The short version: the ideas are old (honest scoring 2007, temperature scaling 2017, zero-shot classification 2019, PFNs 2021). What is new in 2026 is the packaging: a typed-decision API at a fraction of a cent, matched by OpenAI, Microsoft, Cloudflare, Perplexity and others within six weeks.

Roots: decisions, scores and costs (1967-2016)

Date	What happened	Why it matters
1967	Gene Amdahl presents what becomes Amdahl’s law at the AFIPS Spring Joint Computer Conference.	The part you don’t speed up caps the whole speedup; it is the test behind Is 5x real?.
2001	Elkan publishes The Foundations of Cost-Sensitive Learning (IJCAI 2001).	A probability plus a cost matrix gives the action threshold; new costs move the threshold, not the model.
2007-03-01	Gneiting and Raftery publish Strictly Proper Scoring Rules, Prediction, and Estimation in JASA.	Sets out why honest probabilities score best under proper rules such as Brier and log loss, the basis of every calibration reward.

Building blocks (2017-2019)

Date	What happened	Why it matters
2017-05-23	Geifman and El-Yaniv publish Selective Classification for Deep Neural Networks.	“Answer only when sure”, with a guaranteed error rate: the root of automating the confident cases.
2017-06-12	Vaswani et al. publish Attention Is All You Need, the transformer.	Every model in this course, from Jev to TimesFM-3, is a transformer.
2017-06-12	Christiano et al. publish Deep RL from human preferences, the paper that introduced RLHF.	RLHF later made chat models helpful, and also overconfident.
2017-06-14	Guo et al. publish On Calibration of Modern Neural Networks.	Shows modern networks are overconfident and that one fitted temperature fixes much of it; still the default fix.
2017-11-28	Ratner et al. publish Snorkel.	Weak supervision: labels from many cheap rules instead of hand labelling.
2018-09-12	LWN reports AUTOSEL, a model that picks kernel commits to backport (about 85% precision, 80% recall) (one source only).	The first typed decision model in kernel maintenance, trained on labels the process already wrote.
2018-10-11	Devlin et al. publish BERT.	Made “pre-train an encoder once, add one output layer” the standard for classification.
2019-03-13	Rich Sutton publishes The Bitter Lesson.	The strongest argument that general methods with more compute beat specialized ones.
2019-08-31	Yin, Hay and Roth publish Benchmarking Zero-shot Text Classification, using NLI as a classifier.	Framed zero-shot classification as entailment (“this text is about X”) and defined a benchmark for labels never seen in training.

Scaling, prompting and Bayesian transformers (2020-2022)

Date	What happened	Why it matters
2020	Kaplan et al. publish Scaling Laws for Neural Language Models.	The source of the “2N and 6N FLOPs” rules that tie model size to cost.
2020	Brown et al. publish Language Models are Few-Shot Learners (GPT-3).	Gave the field the zero-, one- and few-shot vocabulary.
2021-02-19	Zhao et al. publish Calibrate Before Use.	Prompt-read probabilities carry label bias; contextual calibration recovered up to 30 points of accuracy.
2021-03-26	Northcutt et al. publish Pervasive label errors in test sets.	Even famous test sets average at least 3.3% wrong labels, so label cleaning pays.
2021-06-17	Hu et al. publish LoRA.	The method behind almost every cheap fine-tune since.
2021-07-15	Angelopoulos and Bates publish A Gentle Introduction to Conformal Prediction.	Practical, guaranteed prediction sets around any model’s scores.
2021	Bommasani et al. publish On the Opportunities and Risks of Foundation Models.	Where the term “foundation model” comes from.
2021-12-20	Müller et al. publish Transformers Can Do Bayesian Inference, introducing PFNs.	Train once on synthetic data from a prior and get Bayesian predictions in one forward pass: the idea under TabPFN.
2022-03-04	Ouyang et al. publish InstructGPT.	The RLHF recipe behind chat models; Jev’s CEO Diogo Almeida is 4th of its 20 authors.
2022	Hoffmann et al. publish Chinchilla.	Compute-optimal training; products now over-train small models past it because inference dominates cost.
2022-07-05	Hollmann et al. publish TabPFN (v1), trained in 20 hours on 8 RTX 2080 Ti GPUs.	The first practical tabular foundation model, trained only on synthetic datasets.
2022-07-11	Kadavath et al. publish Language Models (Mostly) Know What They Know.	Base models are calibrated, RLHF breaks it, and a temperature of 2.5 repaired their 52B model.
2022-09	The IETF publishes RFC 9309, the Robots Exclusion Protocol.	robots.txt gets a standard, which says its rules “are not a form of access authorization”.
2022-09-22	Tunstall et al. publish SetFit.	A 355M model with a few examples beat 175B GPT-3 on RAFT (71.3 vs 62.7), for $0.025 of training.

Cheap labels and the first foundation models outside text (2023-2024)

Date	What happened	Why it matters
2023-03-27	Gilardi et al. publish ChatGPT outperforms crowd workers for text annotation.	LLM labelling becomes the cheap default; their data implies about $60 per 1,000 crowd labels.
2023-10-14	Das et al. publish TimesFM, Google’s first time-series foundation model (200M).	The ancestor of TimesFM-3.
2023-11-14	DeepMind announces GraphCast, beating ECMWF’s HRES on more than 90% of 1,380 targets.	Weather is the clearest orders-of-magnitude win for a vertical foundation model.
2024-03-12	Amazon publishes Chronos; the models cost $252 to $2,066 to train on AWS.	Open time-series foundation models become default baselines, cheap enough for a small team.
2024-04-01	Gerstgrasser et al. publish Is Model Collapse Inevitable?	Collapse happens when synthetic data replaces real data; accumulating both stays stable.
2024-04-14	Gentoo adopts its AI policy, forbidding content made with NLP AI tools.	Distro communities start to separate generating code from validating it.
2024-04-29	Verga et al. publish Replacing judges with juries.	A panel of small, different models beat one large judge at over 7x lower cost.
2024-06-25	Penedo et al. publish The FineWeb datasets.	A fully ablated open recipe for turning Common Crawl into training text.
2024-06-25	EvolutionaryScale releases ESM3 (up to 98B parameters).	Biology’s frontier-sized protein model, and the compute outlier among vertical models.
2024-07-01	LMSYS publishes RouteLLM.	A router cut cost 85% on MT Bench at 95% of GPT-4 quality, but only 35% on GSM8K.
2024-07-20	Longpre et al. publish Consent in Crisis.	In one year robots.txt changes made 5%+ of C4’s tokens off-limits: open web data is closing.
2024-11-22	Lambert et al. publish Tülu 3, naming RLVR.	Reinforcement learning with rewards a program can check becomes a named recipe.

Reasoning, rulings and operational wins (2025)

Date	What happened	Why it matters
2025-01-08	Hollmann et al. publish TabPFN v2 in Nature.	It beat a 4-hour tuned ensemble in 2.8 seconds on tables up to 10K rows.
2025-01-22	DeepSeek publishes DeepSeek-R1; R1-Zero went from 15.6% to 71.0% on AIME with rule-based rewards.	Verifiable rewards, not learned reward models, drive the reasoning-model wave.
2025-02-11	Judge Bibas rejects fair use in Thomson Reuters v. Ross, a non-generative legal search tool.	The case closest to a decision model that competes with its training source.
2025-02-25	ECMWF’s AIFS goes operational.	An AI forecast in daily service, with up to 20% better cyclone tracks.
2025-06-05	METR reports o3 reward-hacking in 30.4% of RE-Bench runs.	Reward hacking is the default, so checkers must be hard to fool.
2025-06-23	Judge Alsup rules in Bartz v. Anthropic that training on books is fair use, but pirated copies are not.	Transformation protects generative training; piracy does not.
2025-06-25	Judge Chhabria grants Meta summary judgment in Kadrey v. Meta, fair use “at least for now”.	Market harm, not copying, is where future cases will be fought.
2025-07-15	Weller et al. publish Ettin, matched encoders and decoders.	A 150M encoder beat a 400M decoder on MNLI: start a decision model from an encoder.
2025-07-22	Damani et al. publish RLCR.	A Brier term in the reward cut in-domain ECE from 0.37 to 0.03 at about the same accuracy.
2025-08-02	The EU AI Act’s general-purpose AI duties apply.	Providers of large generative models owe a copyright policy and a training-data summary; a small decision model is very likely not one.
2025-08-11	Knowledgator publishes GLiClass.	Packing 128 labels into one encoder pass cost 7% of throughput, against 52x slower for one pass per label.
2025-09-24	Cloudflare launches Content Signals for robots.txt.	Sites can say `ai-train=no` in a machine-readable line, framed as a reservation under EU text-and-data-mining law.
2025-10-17	Amazon publishes Chronos-2 (120M, Apache-2.0).	Open forecasting of related series and covariates together, under a permissive licence.
2025-10-22	Fedora approves its AI-assisted contributions policy.	AI may not be “the sole or final arbiter”, but automated technical validation is allowed: room for a co-pilot.
2025-11-04	The UK High Court rules in Getty v. Stability AI: Getty’s secondary copyright claim fails.	The court did not treat the model as an “infringing copy”; Getty has permission to appeal that point.
2025-11-11	A Munich court rules in GEMA v. OpenAI that memorized lyrics in a model are a reproduction.	In Germany, memorization can infringe even where text and data mining is allowed; under appeal.

The run-up to Jev (January to August 2026)

Date	What happened	Why it matters
2026-05-04	SAP agrees to acquire Prior Labs, maker of TabPFN, and pledges over €1B across four years.	A tabular foundation model goes from a university lab to a data-platform owner in about four years.
2026-05-13	Prior Labs releases TabPFN-3: up to 1M rows, 20x faster than 2.5.	In-context models reach table sizes real businesses have.
2026-05-14	Datadog releases Toto 2.0, 4M to 2.5B parameters, Apache 2.0.	A permissively licensed time-series family, three months before TimesFM-3.
2026-06-03	Fortune reports Nvidia has bought Kumo AI (KumoRFM); about $400M per one report (one source only).	A second relational and tabular foundation model ends inside a platform company.
2026-06-11	Diana Hu becomes YC Managing Partner.	She later signs the “Vertical Foundation Models” request that opens this course.
2026-07-20	The court gives final approval to the Bartz v. Anthropic settlement: $1.5B for 482,460 works.	About $3,000 per pirated work: the price of bad data provenance.
2026-08-02	The European Commission’s enforcement powers over general-purpose AI start.	Fines of up to 3% of turnover or €15M become possible for providers who skip their duties.
2026-08-15 to 08-28	Debian votes on General Resolution 2026/002; “Responsible Use of Generative AI” wins 281 to 126.	A major distro neither endorses nor bans AI, and asks for discussion before large automated actions.
2026-08-24	Common Crawl publishes its August 2026 crawl: 2.14B pages, 360 TiB.	The free starting point for web-scale sourcing.
2026-08-31	Google announces TimesFM-3: 330M parameters, related series and covariates, the whole forecast in one pass.	The forecasting example in the course’s opening quote; its weights are non-commercial.

Jev and the six weeks after (September to October 2026)

Date	What happened	Why it matters
2026-09-15	TypeSafe AI launches Jev (`jev-1.13.0`) in early access, with a $40M seed led by DCVC.	A typed decision and a probability instead of text, at $0.042 per million input tokens.
2026-09-15	Prior Labs releases TabPFN-3.5: 220M parameters, 1M rows, 20,000 features.	The same day, the tabular side scales up, under non-commercial weights.
2026-09-21	NanoJev PR #14 lands shared-prefix option scoring.	An open replica measures 2.84x faster at 64 options with logits unchanged.
2026-09-22	The community Jev Decision Index appears on Hugging Face.	Decision models get a public leaderboard within a week of launch.
2026-09-24	Fastino releases GLiNER2.5-Decide, a 340M encoder under Apache 2.0.	The first rival is a small encoder, not a giant LLM.
2026-09-28	Victor Dibia rebuilds Jev’s scoring method on Qwen2.5-7B for about $4.	The core trick is reproducible on a hobby budget.
2026-09-29	The Third Circuit affirms that Ross’s use was not fair use, the first federal appellate ruling on AI training.	A non-generative tool that competes with its source lost; decision models sit closest to this case.
2026-09-29	NVIDIA releases Kumo-Tabular, trained only on synthetic tables, with a commercial-use licence.	A commercially licensed rival to TabPFN.
2026-10-01	Cloudflare releases Clef (27B) and Clef-flash (9B) under Apache 2.0, API-compatible with Jev.	Open weights match the contract two weeks after launch.
2026-10-01 to 10-04	Perplexity ships Decider v1 27B, open weights plus an API at $0.04/M.	The price war starts.
2026-10-06	Vals AI publishes an independent evaluation of Jev: 0.10 to 0.12 s for 1 to 64 questions; ECE 0.011 on one task, 0.108 on another.	Speed is real; calibration does not transfer between tasks.
About 2026-10-06	OpenAI opens `POST /v1/decisions` in public beta: gpt-6-luna, $0.10/M input.	A frontier lab adds a “decision mode”, the risk founders feared most.
2026-10-07	Arize publishes the first multi-vendor decision-model benchmark: Jev’s calibration error 0.023, $0.045 per 1,000 judgments against $14.30 for Opus 5.	Hosted leaders sit within about a point of each other; cost differs by about 300x.
2026-10-09	TypeSafe AI announces an $870M Series A at a $7.5B valuation, led by Andreessen Horowitz.	Investors price the primitive as a platform.
By 2026-10-13	Microsoft ships Decision-1, a post-trained Qwen3.5-9B, at $0.042/M.	A frontier lab matches both the contract and the price in under a month.
2026-10-14	TypeSafe releases Jev 1.14, with a Box preview for images.	The vendor says 1.13 “was multimodal all along”.
2026-10-15	Mistral releases Timestral, 5M to 209M parameters; the 5M model forecasts in 17 ms on a CPU.	Timestral-209M and H2O’s TimeH2O-v1 overtake TimesFM-3 on GIFT-Eval and fev-bench.
2026-10-15	Perplexity halves its price to $0.02/M; on 10-23 Decider Flash goes to $0.01/M.	One decision now costs roughly $0.000004 to $0.00002.
2026-10-21	Baseten releases mammoth-1 (27B) and elephant-1 (9B); mammoth-1 tops Decision Index 0.4 at 71.24 of 208 models.	Open models now lead the board Jev started.
2026-10-27	Y Combinator’s Winter 2027 Requests for Startups first appears in web archives, per FourWeekMBA (one source only); the YC page itself is undated.	Diana Hu’s “Vertical Foundation Models” entry is the quote this course is built on.
2026-10-27	TabArena ranks Kumo-Tabular first (Elo 2025), above TabPFN-3.5 (1903) and the best tuned tree, LightGBM (1433).	Synthetic-only tabular models beat tuned trees by a wide margin on broad benchmarks.
2026-10-28	TimesFM-3, top of the time-series boards on 2026-08-31, ranks 6th in the course’s recomputed standings from GIFT-Eval and fev-bench (see TabPFN, TimesFM and friends).	Leaderboard leads last about six weeks; build on a model, not on its rank.

What the timeline says

The decision primitive is old; the price is new. Every piece Jev uses was public by 2022: NLI zero-shot classification (2019), temperature scaling (2017), proper scoring rules (2007). The 2026 change was a hosted API at $0.042 per million input tokens.
Moats in this field last weeks. Jev’s lead lasted about two weeks before open weights matched its contract, and TimesFM-3’s leaderboard lead about six.
The law is moving towards “it depends on the source”. Piracy lost (Bartz), transformation won (Bartz, Kadrey), and a non-generative competitor lost on appeal (Ross). See data rights.