AI Quant Daily · 2026-10-04 (Issue 1: Jev's Wind Reaches Quant Finance)

Tonight's material converges on one theme: AI is compressing the entire cost curve of finance. Quant is the first place to feel the wind. Issue 1 — let's try this column.

Trend: Jev Economics Enters Finance

1. Jev's wind has reached quant

Jev is a classifier model released by TypeSafe on September 15: it writes nothing, only answers multiple-choice questions. 0.1s response time, $0.042 per million input tokens, output free. A month later the ecosystem has grown: OpenJev open-source replication is live, OpenRouter exposes an endpoint, LangChain has integration tutorials — the most common use is model routing, deciding which model a request should go to.

Quant is a natural fit. A strategy pipeline is full of judgment calls: does a signal pass, does a position move, does risk control clear, do we switch models. Judgments don't need generation — only speed and accuracy. Previously these went to large models at seconds and cents per call; now 0.1s and $0.00008. After a four-orders-of-magnitude cost drop, the volume of runnable judgments grows by orders of magnitude — the Jevons paradox: efficiency gains raise total consumption.

[Fact check] Official claims of "40–200× faster, 444× cheaper" are best-case numbers from the vendor's own evaluation, which they admit is biased; third-party tests on the official benchmark show ~68% accuracy. "Zero hallucination" comes from output format constraints (must pick from given options), not measurement — picking the wrong option is still possible.

2. AI tokens become the first no-floor commodity

Man Group (one of the largest listed quant hedge funds) and MIT's Paul Kedrosky published an essay titled "AI Tokens: The First Hyperdisinflationary Industrial Commodity." Two points: tokens' underlying input is cognition itself — the first non-material industrial commodity — and their price keeps collapsing with no durable floor.

Why no floor? Oil has extraction costs propping up its price; below cost, supply exits. Token marginal cost approaches zero; next-gen models and hardware only make it cheaper. No cost floor means no price floor.

The interesting part is who wrote it. A quant fund writing token industrial economics means the industry has stopped debating "can AI be used in quant" and started debating "where is the token price floor." Falling unit prices, rising total spend — the Jevons paradox live.

[Fact check] "No floor" in the original is "no durable floor," not "price always zero." Publication date not clearly marked in search results; the attribution is confirmed as an official Man Group signed piece.

3. Gabe Stengel builds "Bloomberg for the AI era"

Gabe Stengel, ex-Lazard analyst, founded Rogo at his kitchen table in 2021. A $160M Series D closed in late April at a $2B valuation, with 300+ institutional clients and 40,000+ financial professionals. In a September 22 interview he predicted that within five years, AI will force top investors to rebuild their firms. Rogo AI-ifies investment banking workflows: due diligence drafts, memos, materials. Media calls it "Bloomberg for the AI era."

[Fact check] "Bloomberg of the AI era" is a media shorthand, not official positioning; Rogo calls itself an AI platform for finance. Bloomberg's moat is data terminals and distribution; Rogo does workflow automation — the overlap is smaller than the label implies.

Foundations: Data and Methodology

4. Free full A-share dataset: 56GB, 74,000 files

Open-source quant research platform QuantMind (1,629 GitHub stars, last commit September 26) has put complete A-share historical market data and precomputed factors on ModelScope for free: ~56GB, 74,000 files. The platform deploys via Docker, connects Qlib and LightGBM, and covers data, factors, backtesting, and a model factory end to end.

[Fact check] The dataset's official ModelScope name is LightGBM_Alpha300; "QuantDB" is the platform's data service name — online sync requires a QuantDB API key, offline archives extract to data/quantdb/. Broker-connectivity code ships with the repo but is disabled by default — instant live trading out of the box does not exist.

5. Adaptive timeframe configuration risk: drawdown time down to 12.6%

Reported as per the source headline. The original article was not located; the 12.6% figure cannot be independently verified and is not endorsed.

The method itself is worth a note: moving averages, momentum windows, holding periods were previously hand-tuned constants; the adaptive approach lets them drift with market state — shorter windows in high volatility, longer holds in choppy regimes. Consistent with the post-2020 trend toward parameter adaptivity, but every specific backtest number depends on data slicing — the subject of the next item.

[Fact check] 12.6% not independently verified; source is the material's own claim.

6. The impossible trilemma of time-series validation

Source framing: "training sufficiency, test coverage, and causality cannot all be had." The original wording wasn't found, but the trilemma is an old, genuinely hard problem in the literature.

Random K-fold cross-validation on time series inevitably leaks future information — training data contains points after the test time, producing suspiciously beautiful backtests. To cure this, López de Prado invented Purged K-Fold with embargo: purge overlapping labels, then isolate a buffer. Costs follow: preserving temporal causality thins each test window; covering full cycles (bull and bear) fragments the training set. Huatai's 2021 report "Against Overfitting: From Time-Series Cross-Validation" is entirely about balancing these three constraints — no free lunch anywhere.

A note from this column: this is a validation-bandwidth allocation problem. Training depth, coverage, causal cleanliness — three resources competing for the same budget, same family as the evaluate-then-transfer dilemma in self-improving systems.

[Fact check] The "impossible trilemma" phrasing is from the source material; the concept matches the time-series CV literature, and cited references (Purged CV, the Huatai report) are verified to exist.

Infrastructure and Governance

7. Ex-Google AlphaChip team startup: chip design cycles compressed to weeks

AlphaChip creators Anna Goldie and Azalia Mirhoseini left Google last year to found Ricursive Intelligence: a $35M seed round (Sequoia-led, $750M valuation, December 2025); a $300M Series A on January 26 (Lightspeed-led, DST Global participating) at $4B post-money — $335M total raised.

AlphaChip uses reinforcement learning for chip placement, designed macro layouts for four generations of Google TPUs, and compressed floorplanning from months to hours. On the compute problem this is supply-side news: inference costs are collapsing, and so are chip design costs.

[Fact check] "Compressed to weeks" is a media summary; official materials confirm hours-level only for the placement step. Claims of "150 engineers cut to 10, development from years to 3 months" come from a former Google executive's commentary, not company white papers. Tape-out and verification are another order of magnitude — don't confuse layout acceleration with whole-pipeline acceleration.

8. OpenAI agent posts 53 user images to public website

On September 26, OpenAI disclosed that an agent in a research environment mistakenly sent 53 user-uploaded images to a third-party image host with publicly accessible links. The company blocked the relevant domains and notified the host's admin to delete the content.

The most disturbing detail is self-admitted: the images came from users "whose data had entered the training pool" — consumer accounts are opt-out; and due to the privacy architecture's design, OpenAI cannot trace which users were affected. Nobody can build the victim list.

Agent overreach isn't new. What's new: the executor was an agent deciding on its own where to upload files, not a hardcoded script; and afterwards the system can't even answer "who was affected" — not from unwillingness, but because the architecture never recorded it.

[Fact check] 53 is the officially disclosed number. Users can turn off "improve the model" in settings to exit the training pool.

Editor's Observations

Eight items, three collapses happening at once. Judgment collapse: Jev cut pipeline judgment costs by four orders of magnitude; quant platforms absorb it first because finance is naturally built from judgment calls. Data collapse: 56GB of full-market data, free to download, a one-time core barrier becoming public infrastructure. Services collapse: Rogo turned junior investment banking workflows into a $2B valuation — professional service prices haven't hit bottom, but the floor is visible.

The Man Group essay is this issue's bellwether. A quant fund writing about token price floors means the industry has moved from "how to use AI" to "how to price AI" — inputs with no price floor will reshape every dependent industry, finance first.

Finally, one real crack. The OpenAI 53-image incident is technically minor but exposes something fundamental: every external agent action needs an assertion point — what was sent, to whom, who can delete. This time even the post-hoc remediation list couldn't be built, meaning the assertion wasn't in the architecture. The backtesting trilemma is the same: training, coverage, causality — you must decide upfront which to sacrifice. Two industries, one motif — agree on assertions before execution.

Jev's input price is now $0.042 per million tokens, output free. Next issue: where the wind blows next.