LEON

COEFFECT / JOURNAL

Markets. Machines.
The city after hours.

Latest edition · 21 November 2026

The complete edition

20 November 2026.

37 notes, arranged by topic.

Originally published 20 November 2026. Reporting, forecasts, prices, odds and event availability reflect the original edition.

In this edition 37 notes

The Day

Quant & Markets

AI & Engineering

Business & Institutions

Art, Design & City

Restaurants & Food

Menswear

Sports

Music

Film & TV

Words & Trivia

Balyasny reportedly closes to new money

Balyasny is telling investors it is closed to new capital for the foreseeable future, according to Business Insider's Thursday report, citing a person familiar with the matter. The firm declined to comment. The report puts assets at thirty-seven billion dollars, compared with twenty billion at the end of twenty twenty-four. It doesn't establish which vehicles or existing-investor additions are covered, or why the firm is closing. There's no reliable flow breakdown behind that change in assets.

Source 1

Back to contents ↑

Jane Street’s larger reported IREN interest

Jane Street's latest IREN filing reports beneficial ownership of about twenty-seven point three five million shares, or six point five percent, as of November twelfth. Its December-end disclosure was roughly fifteen point four five million, or four point seven percent. The earlier number included shares obtainable through convertible bonds, and the share denominator has changed. Those filings establish a larger reported beneficial interest; subtracting them would not give you net purchases or net trading exposure.

Source 1 · Source 2

Back to contents ↑

A fellowship that leaves the research with the researcher

Susquehanna, meanwhile, has launched a doctoral fellowship. In the United States, it provides a hundred thousand dollars a year for two years as an unrestricted gift to the recipient's department, plus a senior quantitative-research mentor and computing resources. Fellows keep their research intellectual property and publication freedom. The ten-week summer visit is optional. It's a structured relationship with outside research in areas including statistical inference and sequential decision-making, without a requirement that the student join the firm. Nominations, rather than direct applications, govern entry; terms outside the United States vary.

Source 1 · Source 2

Back to contents ↑

KaiConvex’s choice among prices of protection

KaiConvex plans a hedged-equity fund and a separate tail-protection fund. In a Thursday Opalesque interview, its managers describe obtaining equity beta through calls and futures, using more than forty quantitative signals, and adding options on volatility, long bonds and gold. The tail strategy funds protection partly from Treasury-bill and box-spread yield, with an additional budget. These are descriptions of the intended portfolios, not results from the new funds.

The protection is divided among sleeves using variance futures and options tied to the S and P, the VIX, high-yield credit, long Treasuries and gold. They can switch a sleeve off when its protection becomes expensive. That leaves the strategy choosing among different prices of protection, rather than maintaining the same option position through every volatility regime.

And they plan to realize hedge profits in tranches. Staged selling lets them take some gains while retaining some protection, instead of choosing a single exit for the entire hedge. The interview doesn't give us the strikes, maturities or leverage needed to reconstruct the payoff.

Source 1 · Source 2

Back to contents ↑

When a company-metric option stops trading

Twenty-four X's proposed company-metric options add some consequential timing rules to the company-metric product design. Issuer reporting practice determines whether a series is treated as morning or afternoon expiration. Morning series stop trading the day before expiry; afternoon series trade on the expiration day. A placeholder date can be replaced when the issuer announces its earnings date. This remains a proposal, not an approved launch.

If the company reports the metric early and the exchange confirms it, trading stops and expiration accelerates. If disclosure is late, trading ends at expiry and processing waits for the number under clearing rules. Once settlement has happened, a subsequent restatement doesn't change the cash outcome. A corrected company account and a settled contract can therefore preserve different numbers.

Twenty-four X also proposes excluding these binaries from its standard obvious-error rules, arguing that the theoretical-value test doesn't fit their payout. It says listing will wait for a class-level maximum execution-price control. That is a specific combination: an upfront price control, without the usual error procedure afterward.

Source 1 · Source 2

Back to contents ↑

State Street joins a shared securities-loan record

State Street has signed on to Equilend's OneSource, according to Thursday's announcement. The service puts agreed securities-loan terms in a shared smart-contract record alongside firms' existing books. Marks, rate changes, returns and recalls are recorded once for both counterparties. The design addresses divergence during a loan's life, rather than just matching its opening terms. Equilend names Goldman, BNY and National Bank of Canada as already executing live trades; State Street's announcement doesn't establish that its trades are live. The provider's claimed savings haven't been independently measured here.

Source 1 · Source 2 · Source 3

Back to contents ↑

The liquidity that immediate settlement still needs

A separate settlement question came from the New York Fed's Anna Nordstrom on Monday. In her remarks on foreign exchange, she raised the possibility that atomic settlement could reduce netting and thin liquidity outside normal hours. A trade settling individually and immediately can lose the funding economy of offsetting obligations before settlement. The timing improvement comes with a change in how much liquidity has to be available along the way.

Her observations also included stable spreads, depth and market impact through the stress periods she discussed, and AI's use in pre-trade work and transaction-cost analysis. The FX Global Code's next strategic review begins in twenty twenty-seven. These are Nordstrom's personal views on the market's development, rather than an announced Federal Reserve policy change.

Source 1

Back to contents ↑

Rumble 2 Flash: downloadable speech recognition

Rumble is a half-billion-parameter model with downloadable Apache two-point-zero weights, based on Qwen Three ASR. Its model card lists forty-five languages. The developers train the smaller model on transcripts it generates itself, using a larger teacher to judge those outputs. Their GOLD method handles the mismatch between the teacher's and student's tokenizers.

Reson8 reports seven point eight six percent word error on an AppTek test, against seven point seven eight for Nemotron and seven point four for the larger Qwen model. AppTek's material consists of role-played customer-and-agent conversations, not recordings of actual customers. That gives the comparison a specific conversational setting, without establishing equality across languages or recording conditions.

You can run the weights through Transformers or vLLM. Baseten separately offers a real-time endpoint. Its very large throughput claims come from saturated batches of recorded files, so they don't tell you how long a caller waits for the next partial transcript. The open model and the hosted streaming product are separate deployment choices.

Source 1 · Source 2 · Source 3 · Source 4

Back to contents ↑

AssemblyAI’s endpointing choices—and its running meter

AssemblyAI divides STT three point seven into Recorded, Realtime and Realtime-Agent. Realtime supplies partial transcripts and turn-end labels. The agent version supplies formatted finals and uses punctuation and silence in endpointing. Its LiveKit integration can add another configurable delay. That means a conversational agent's response timing depends on the endpointing path as well as recognition speed.

And the meter keeps running during silence. AssemblyAI bills streaming for the whole time the WebSocket session stays open, with concurrent sessions charged separately. The listed hourly rates are twenty-one cents for recorded audio, thirty cents for realtime and forty-five cents for the agent version, before optional extras. An application that leaves sessions open between conversations pays for those gaps.

Realtime supports thirty-two languages. Recorded audio can use an optional fallback covering ninety-nine. Choosing between AssemblyAI's interfaces changes the available languages, the billing unit and how a turn is finalized; choosing Rumble's weights means taking on the serving configuration yourself.

Source 1 · Source 2 · Source 3 · Source 4 · Source 5 · Source 6

Back to contents ↑

One model, a very different harness result

Google's Wednesday Gemini Four Neon launch includes a striking spread in its own SWE-sweep results: twelve point eight percent with mini-SWE, thirty-nine percent with Antigravity, using the same model. The comparison doesn't isolate which harness changes produced the gain. Neon has a million-token context window. Standard Developer API pricing is three dollars per million input tokens and fifteen per million output tokens, including thinking. Implicitly cached input is fifteen cents per million; explicit caching isn't supported, and there is no free tier for this model.

Source 1 · Source 2

Back to contents ↑

Paper2Agent exposes the work before the benchmark

Paper2Agent tackles a different source of friction: turning a research paper's software into something an agent can actually use. A September Nature study, revisited in an author post yesterday, converted seventy-four of a hundred biology papers. The other twenty-six exposed missing code, artifacts or dependencies. Those failures occurred before the resulting tool could be judged on scientific questions.

The test questions came from the papers' tutorials: three hundred questions, with Sonnet Four scoring ninety-one point two percent through the converted agents, versus eighty point three using the repositories. Average query cost fell from thirty-eight cents to twenty, and time from four point three minutes to one point six. The initial conversion cost isn't included in those per-query figures. A lab repeating related work could amortize that preparation; the study doesn't establish where the break-even point lies.

Source 1

Back to contents ↑

What Frontier RSI’s score actually measures

Frontier RSI, from Abundant, instead gives models GPU research and engineering assignments. Its twenty-seven tasks cover pretraining, agent post-training and harnesses, serving, and kernels. The public repository currently contains eight, including optimizer work on one H-one-hundred and bulk inference that must preserve exact outputs. Each model gets two trials per task, under a fixed hardware allocation and budget.

Its score is relative to the best valid trial. That trial gets one; nearby results receive partial credit, while failures of correctness or resource limits get zero. The reported average-at-two averages those normalized rewards. Reading it as the percentage of tasks completed would change the meaning of the table. Different models also use different harnesses, and most of the task set isn't public yet. The eight available tasks are the part another team can inspect and try.

Source 1 · Source 2

Back to contents ↑

Move the irregular lookup, not the whole workload

LinkedIn's retrieval system found a bottleneck in a seemingly convenient place: document-ID lookups on the GPU. Its September paper describes narrowing a billion-plus-document corpus to tens of thousands of candidates, then assembling features for a neural pre-ranker. GPU binary searches took about five milliseconds at batch one and forty-seven at batch eight in the reported tests. Moving those irregular lookups to a parallel CPU hash map brought them to roughly one millisecond across tested batches.

The GPU still gathers dense document features, joins request-specific viewer-to-author edges, and scores candidates. That graph-edge join is separate from resolving document IDs. Their progressive throughput table goes from a tuned single-request baseline to one point seven five times with batching, then ten times with CPU lookups and a larger batch, under a hundred-and-twenty-millisecond ninety-ninth-percentile latency budget. Ten times measures the cumulative redesign, not the isolated lookup change.

They also preallocated cuBLAS workspaces after variable batches fragmented the remaining GPU memory. The document index occupied most of it. Larger batches had made the lookup slower and the spare memory less predictable; the authors changed both the lookup location and workspace allocation.

Source 1

Back to contents ↑

Deleting history can make inference more expensive

A November seventh preprint, Think or Remember, examines a related serving interaction: deleting old reasoning can break an exact-prefix cache. The authors forked a hundred and three matched Qwen Three eight-billion-parameter agent states, then let Keep and Drop versions continue independently. Under modeled generated-state reuse, Drop cost one point seven four times Keep in operations. Under no reuse, the ratio fell to zero point three four.

Those continuations also generated different behavior, so that comparison mixes cache reconstruction with the model's response to losing its history. A separate mechanical replay held continuations fixed. Keeping only the latest five reasoning blocks cost about four point five four times retaining everything, despite nearly equal prompt volume, under its specified unlimited exact-prefix cache. Removing an early block forced later material to be processed again.

These are modeled operation costs, not elapsed times or API bills. Quality also depended on the setting: dropping reasoning lost nine to twenty-three percentage points on live HotpotQA runs, while the gpt-oss ScienceWorld result favored dropping. A history policy changes both the reusable computation and what the agent subsequently does.

Source 1

Back to contents ↑

Moving prefix state beyond GPU memory

iFlytek's November second HiCache account concerns a different part of reuse: where cached state lives. The project and operator report extending it from GPU memory into host memory and external storage, taking prefix-cache hits from around thirty percent to ninety-five. Typical inputs ran from thirty thousand to two hundred thousand tokens. Median time to first token was under three seconds afterward, but they don't establish the previous production latency or hardware cost. The hit-rate change is the reported before-and-after comparison.

Source 1

Back to contents ↑

An accelerated operator is only part of decoding

Nvidia's November seventeenth sparse-attention discussion adds another physical cost. Selecting only some keys doesn't eliminate the indexer's work over the long history. A related April GVR paper exploits overlap between adjacent decoding steps to perform exact top-k selection. Under its DeepSeek V-three-point-two, Blackwell and TensorRT-LLM setup, the operator averaged a one point eight eight times speedup, while the end-to-end time per output token improved by up to seven point five two percent. The smaller end-to-end gain reflects how much of decoding remained outside that accelerated operation.

Source 1 · Source 2

Back to contents ↑

The cost of proving a calculation

Anthropic's Flashprove prototype pays a much larger computational premium for a different property: proving that an inference calculation was performed with a committed model. Its integer-quantized approach uses a cryptographic proof, a zk-SNARK. The reported single-H-one-hundred rate is about a hundred and ten tokens a second, with a nineteen-megabyte proof per thousand and twenty-four tokens. GPU time is seventy to ninety-five times the vLLM baseline. That proof concerns the computation used to produce the tokens; it doesn't certify that the answer is true. It is an older research prototype, with a substantial proving cost still attached.

Source 1

Back to contents ↑

Carronade arrives during Brookdale’s property reshaping

Carronade's new Brookdale disclosure puts a different portfolio decision in play. Carronade and Dan Gropper report a five point eight percent beneficial interest in the senior-living company. The thirteen point seven nine million shares include three million underlying ten-dollar calls expiring next April. The filing names capital allocation, portfolio changes, mergers, governance and a possible REIT conversion as subjects for discussions with management and the board. None is an agreed transaction.

Brookdale was already buying properties it used to lease. In September it completed the purchase of seventeen communities for a hundred and fifty-seven point four million dollars, plus transaction costs. A hundred and eight million of non-recourse mortgage debt helped finance it, at SOFR plus two hundred and fifteen basis points.

A separate agreement to buy thirty more leased communities for two hundred and ninety million would bring the owned share of units to nearly eighty percent. It remains conditional. Brookdale has also sold twenty-four owned communities this year. So Carronade is entering an active reshaping of the property portfolio; those purchases and disposals preceded its November nineteenth filing.

And occupancy complicates a pure property-value argument. Third-quarter consolidated occupancy improved to eighty-three point one percent, but management said growth missed its goal. It lowered full-year adjusted EBITDA guidance to four hundred and ninety-five to five hundred and two million dollars, also citing marketing costs and a community fire. Carronade calls the assets undervalued. Brookdale's operating shortfall is already visible alongside them.

Source 1 · Source 2 · Source 3

Back to contents ↑

Palliser asks Chemring to narrow further

At Chemring, Palliser wants a strategic review that could separate or sell Roke, the electronic-warfare business, while accelerating investment in energetic materials. Its Wednesday letter reports an economic interest equivalent to about three percent. Palliser argues that the businesses compete for capital and management attention despite having different needs. That is the shareholder's case; Chemring hasn't agreed to dispose of Roke.

Chemring had announced a disposal plan six days earlier, but for a different business: US Sensors. Management's November twelfth update retained both Roke and Countermeasures and Energetics as its focus. Roke's year was below expectations because UK Ministry of Defence and international orders were delayed, even while order intake reached a record. Selling US Sensors already narrows the group; Palliser is asking it to narrow further.

Roke also has a five-year miniature-radar-altimeter agreement with a US prime contractor, expected to be worth at least a hundred and seventy million dollars, including an initial sixteen-million-dollar production order. The expected agreement value is much larger than that first order. It gives the business a prospective production stream alongside the delayed awards, without making all five years of expected value current revenue.

Energetics requires patience of a different kind. Chemring expects its existing Scottish and Norwegian expansion projects to enter production in twenty twenty-seven and twenty twenty-eight. Its UK business has an initial twenty-three-million-pound, three-year production contract under a longer Ministry of Defence framework. Those projects put actual dates and orders behind the capital-allocation debate. They were company plans before Palliser's letter, not concessions made in response to it.

Source 1 · Source 2 · Source 3 · Source 4

Back to contents ↑

A father on Earth, a daughter on Mars

BAM lists La Distance on Friday and Saturday, November 20–21 at 7:30 p.m., at the Harvey Theater on Fulton Street in Fort Greene. Tiago Rodrigues sets the two-person drama in twenty seventy-seven: a daughter leaves for Mars while her father stays on Earth. Alison Dechamps and Adama Diop perform on a rotating stage, in French with English supertitles. It's eighty-five minutes about separation and survival.

Source 1

Back to contents ↑

Ghost Factory brings an industrial town’s memories to 92NY

Ghost Factory plays 92NY’s Buttenwieser Hall on Friday and Saturday, November 20–21 at 7 p.m. Bridgman and Packer combine two live dancers with projections drawn from Johnson City's abandoned factories and its residents. The companion installation, Places With Hidden Stories, is open for an hour before and after. That's the Upper East Side option for dance and an industrial town's memories. Current ticket availability is unconfirmed; these are the venue’s scheduled performances.

Source 1

Back to contents ↑

Kuroneko: ramen, squash and crab roe in the East Village

Kuroneko is a new discovery at 218 East 10th Street in the East Village, in the former Rai Rai Ken space. Josh Reisner's twenty-five-seat noodle bar combines house-made ramen with Chinese and Southeast Asian influences. The menu's butternut tantanmen brings together squash, king oyster mushroom, toasted sesame and whole-wheat noodles; the crab fried rice includes snow crab, crab roe, seafood XO and charred alliums. Those are reasons to investigate the menu, not a review of a meal we've eaten.

Source 1 · Source 2 · Source 3 · Source 4

Friday tables have not been confirmed.

Back to contents ↑

Huntsman broadens its New York offer

Huntsman's West Fifty-seventh Street atelier is adding made-to-measure and ready-to-wear, plus a client lounge. The atelier itself dates to twenty seventeen. The maker quotes a starting price of four thousand seven hundred and fifty dollars for a made-to-measure two-piece, with eight to twelve weeks from a single appointment. Its bespoke service uses an individual paper pattern and at least three fittings. Permanent Style's historical description of the house cut emphasizes a strong shoulder and close chest and waist; that's context for the silhouette, not a review of this new service.

Source 1 · Source 2 · Source 3

Back to contents ↑

Shelton reaches his first ATP Finals semifinal

Ben Shelton reached his first ATP Finals semifinal after beating Jodar seven-six, seven-five on Thursday, saving two set points in the first-set tiebreak. His qualification became certain when Alexander Zverev took the opening set against Arthur Fils. Zverev went on to win six-four, seven-six, finishing the group three and oh without losing a set; Shelton finished two and one. Shelton faces the other group's winner. His opponent and semifinal start time weren't yet established in the tournament report.

Source 1 · Source 2

Back to contents ↑

Giannis faces Milwaukee—and Miami wins again

In Wednesday's NBA result, Giannis Antetokounmpo faced his former Bucks for the first time and finished one assist short of a triple-double: twenty-seven points, fourteen rebounds and nine assists. Miami won a hundred and thirty-seven to a hundred and thirty. That was the Heat's sixth straight win, while Milwaukee fell to one and fourteen, its worst fifteen-game start. Miami had led by thirty; Milwaukee's late thirty-to-nine run made the final score much closer.

Source 1

Back to contents ↑

Lael Neale’s kitchen-recorded Friday release

Scheduled for Friday, November 20, Lael Neale's As Killers Do was made with Guy Blakeslee on a Tascam cassette eight-track in an Echo Park kitchen, combining Omnichord and psychedelic guitar. In The Skinny, Zoë White describes the recording's softened, intimate sound and contrasts the wildness of Joan of Arc with the restraint of Picasso's Wife. Slip Away was available ahead of the album, with a stop-motion video made from Neale's paintings.

Source 1 · Source 2

Back to contents ↑

What to stream: La Grazia

For a film at home, MUBI's US catalog lists Paolo Sorrentino's La Grazia. This is an older streaming arrival, not a new November premiere. Toni Servillo plays an Italian president nearing the end of his term, grieving his wife while considering an assisted-dying bill and clemency cases. Time Out's Phil de Semlyen describes wit and surreal comedy moving through the palace's formal spaces. The president also turns to his daughter for counsel on the law. It's a hundred and thirty-three minutes, with serious grief and end-of-life questions beneath the humor.

Source 1 · Source 2 · Source 3

Back to contents ↑

What to stream: CQ, before November 30

Criterion Channel has a much campier alternative: Roman Coppola's two-thousand-one film CQ, about an American editor in nineteen-sixties Paris who inherits the director's chair on a science-fiction production. Criterion compares its world to Barbarella and Danger: Diabolik, and lists it as leaving November thirtieth. Availability depends on subscription access.

Source 1

Back to contents ↑

Two questions for Friday

Question 1

Which element was identified in the Sun’s spectrum decades before it was found on Earth?

Reveal answer to question 1Hide answer to question 1

Helium. Its spectral line was recognized in 1868; identification on Earth followed almost 27 years later. The name comes from Helios, the Sun. NASA

Question 2

The Curonian Spit separates the Baltic Sea from the Curonian Lagoon. It spans which two countries?

Reveal answer to question 2Hide answer to question 2

Lithuania and Russia. The 98-kilometer sand-dune peninsula separates the lagoon from the Baltic Sea. UNESCO

Back to contents ↑

Copy this link

Select the address below and copy it to share.

Open this link →