Daily incremental brief

Lucas Schuermann – Swapping Out Perpetual Futures (S7E34)

Reviewed as an industry signal only; its claims are not used as independently established facts.

Coverage window: 2026-09-01–2026-09-15 · publication dates shown on each item
01 / Podcast

Lucas Schuermann – Swapping Out Perpetual Futures (S7E34)

Reviewed as an industry signal only; its claims are not used as independently established facts.

02 / Company

Government of Canada welcomes major new investment in sovereign AI infrastructure in Saskatchewan

If executed at the stated scale, the project would be a major addition to Canadian sovereign-compute capacity and a large regional power-and-capital commitment. The figures describe a plan, not completed capacity or committed final spending, so approvals, customers, financing, grid readiness, and construction milestones remain the key diligence items.

03 / Research

The Measurement Revolution? Credible Measurement and Inference in the Age of AI

As AI-generated variables enter economic, financial, and business research, prediction accuracy alone is not enough for causal or descriptive claims. The proposed framework makes validation design an explicit part of the pipeline, but this is a conceptual NBER working paper rather than a completed empirical evaluation or peer-reviewed result.

04 / Industry

AI stocks drop, but the rest of Wall Street holds steadier after oil prices give back an early jump

The move links two valuation pressures—slower expectations for AI leaders and a higher long-duration discount rate—rather than establishing a change in AI demand by itself. It is a one-session market snapshot.

Primary releases

Only items selected by this edition’s manifest appear here. Company claims remain provider-reported unless independently verified.

Innovation, Science and Economic Development Canada Sep 14, 2026

Government of Canada welcomes major new investment in sovereign AI infrastructure in Saskatchewan

A Canadian government release says Bell plans to expand its Saskatchewan AI-infrastructure development by as much as 900 MW, creating a path to a 1.2 GW hub. The release associates the full concept with up to C$52.5 billion of capital investment and 4,500 jobs, but says community and Indigenous engagement is continuing and project-specific announcements will follow.

  • Bell plans up to 900 MW of additional Saskatchewan capacity, which the release says could bring the hub to 1.2 GW.
  • The release associates the full concept with up to C$52.5 billion in capital investment and 4,500 jobs, subject to further project development and announcements.
Why it mattersIf executed at the stated scale, the project would be a major addition to Canadian sovereign-compute capacity and a large regional power-and-capital commitment. The figures describe a plan, not completed capacity or committed final spending, so approvals, customers, financing, grid readiness, and construction milestones remain the key diligence items.
Primary Worth tracking
Open source

Research & policy

Academic papers, official research, regulatory material, patents, and standards are grouped together with their evidence labels intact.

National Bureau of Economic Research Sep 10, 2026

The Measurement Revolution? Credible Measurement and Inference in the Age of AI

Dell and Rambachan argue that AI changes empirical research by converting unstructured material into candidate variables at scale, shifting the constraint from extracting a measure to choosing and validating a credible one. They organize the workflow into construct definition, AI measurement, and statistical inference, and show how validation samples can support valid inference even when predictions are biased.

  • The paper proposes a three-stage workflow covering construct definition, AI-based measurement, and statistical inference.
  • The authors show that validation samples can permit valid inference even when AI predictions are biased.
Why it mattersAs AI-generated variables enter economic, financial, and business research, prediction accuracy alone is not enough for causal or descriptive claims. The proposed framework makes validation design an explicit part of the pipeline, but this is a conceptual NBER working paper rather than a completed empirical evaluation or peer-reviewed result.
Primary Worth tracking
Open source
arXiv cs.AI Sep 08, 2026

Harness or Model? Disentangling the Drivers of Agentic Coding Performance

A contamination-controlled study ran 792 of 800 planned coding-agent trials across 256 private tasks, comparing Claude Opus 4.6 and GPT-5.3-Codex in their native harnesses with a neutral harness. The reported average success differences were not statistically resolved: -1.25 percentage points for Opus and +1.25 points for GPT. Neutral operation used roughly 1.2–1.6 times more measured cost per success, although incomplete provider usage data prevents a definitive billed-cost comparison.

  • Across 792 completed runs, neither model showed a statistically resolved average success-rate advantage for its native harness over the neutral harness.
  • The neutral harness used more measured cost per successful task, but incomplete provider telemetry prevents a definitive billed-cost ranking.
Why it mattersAgent benchmark results can reflect the surrounding harness as well as the model. This preprint provides unusually controlled evidence that the average accuracy effect may be small, while efficiency differences remain economically relevant. It is a single-author, non-peer-reviewed study, and the revision corrects an earlier telemetry error.
Primary Context
Open source
arXiv cs.LG Sep 09, 2026

FINESSE: Financial Event Sequence Simulation and Evaluation

FINESSE is an agent-based simulator and benchmark for financial event sequences rather than isolated transactions. It generates temporally consistent activity and evaluates models on next-event prediction, cash-flow forecasting, fraud detection, and missed-payment prediction. The authors report releasing the framework and dataset to support reproducible testing.

  • The framework simulates longitudinal financial events and supplies benchmark tasks spanning forecasting, fraud, missed payments, and next-event prediction.
  • The authors report releasing the simulator and dataset for reproducible evaluation.
Why it mattersFinancial AI systems are often evaluated on static tables even though risk and behavior evolve over time. A controllable event-stream benchmark could improve stress testing and sequence-model comparison, but synthetic realism and transfer to live portfolios still need independent validation.
Primary Context
Open source
arXiv q-fin Sep 11, 2026

Favorite-Longshot Bias on Polymarket

Using 588 million trades from 2.48 million accounts, the authors report that purchases below 10 cents lost 19.3 cents per dollar on average, while purchases at or above 90 cents earned 0.83 cents. Yet the aggregate longshot result changes sign when contracts are grouped differently, from a 6.3-cent loss to a 4.1-cent gain; the pattern is reported as robust in crypto and politics but absent in sports.

  • The study analyzes 588 million Polymarket trades associated with 2.48 million accounts.
  • Reported longshot performance reverses sign under an alternative contract grouping, while favorite-side returns remain positive.
Why it mattersPrediction-market return patterns may depend as much on contract construction and aggregation as on headline probability buckets. The scale is notable, but the paper is a non-peer-reviewed observational study and the sign reversal makes taxonomy sensitivity central to interpretation.
Primary Context
Open source
arXiv cs.AI Sep 11, 2026

VRL-Bench: A Benchmark for Verbal Reinforcement Learning in Computer-Use Agents

VRL-Bench evaluates computer-use agents that convert experience into reusable verbal guidance across six settings. The authors report that ordinary verbal memory helps in some configurations but degrades others; their VEX² update rule is the only evaluated method reported to improve performance in all six.

  • The benchmark measures verbal reinforcement-learning methods across six computer-use settings.
  • The authors report that baseline verbal memory can hurt performance, while VEX² improved all six evaluated settings.
Why it mattersPersistent learning from interaction is a key step beyond one-shot computer agents. The benchmark highlights that naïvely storing lessons can hurt, although the positive VEX² result is author-reported, non-peer-reviewed, and still needs replication on broader tasks and production interfaces.
Primary Context
Open source
arXiv cs.AI Sep 04, 2026

Occamy-1.0: A Low-Cost Open Agent Model at the Pareto Knee

Occamy-1.0 is a 35-billion-parameter mixture-of-experts model with roughly 3 billion active parameters, post-trained from Qwen3.6 for agentic tool use. Across four benchmarks, the authors characterize it as a low-cost Pareto-knee model and report releasing weights plus a subset of training data.

  • The model has 35 billion total parameters and about 3 billion active parameters per token.
  • The authors report releasing model weights and a subset of post-training data.
Why it mattersAn open model optimized around active-parameter cost could broaden experimentation with tool-using agents and create a more useful efficiency comparison than headline parameter count alone. The performance and cost claims are provider-reported in a non-peer-reviewed preprint and require independent reproduction.
Primary Context
Open source

Industry desk

Independent reporting and specialist analysis that adds evidence beyond company announcements.

Associated Press Sep 14, 2026

AI stocks drop, but the rest of Wall Street holds steadier after oil prices give back an early jump

U.S. equities fell in a broad risk-off session led by AI-linked names: the AP reports the S&P 500 down 0.5%, the Nasdaq down 0.6%, and Nvidia down 3.4%, alongside a 10.7% fall in SoftBank in Tokyo. The U.S. 10-year Treasury yield briefly moved above 5% before ending near 4.98%, adding a discount-rate headwind to growth valuations.

  • The AP reported declines of 0.5% for the S&P 500, 0.6% for the Nasdaq, and 3.4% for Nvidia in the session.
  • The U.S. 10-year Treasury yield briefly exceeded 5% and ended near 4.98%, according to the report.
Why it mattersThe move links two valuation pressures—slower expectations for AI leaders and a higher long-duration discount rate—rather than establishing a change in AI demand by itself. It is a one-session market snapshot.
Verified source Context
Open source

Listen / read

Episode summaries use official descriptions or authorized transcripts. Timestamps appear only when they can be verified.

Latent Space Sep 14, 2026

Humanity’s Last Invention — Richard Socher of Recursive

Richard Socher outlines Recursive’s thesis that open-ended agents can help automate parts of AI research and discusses the growing overlap between AI systems and financial workflows; the technical and financing claims are founder commentary, not independent validation of recursive self-improvement.

Desk takeReviewed as an industry signal only; its claims are not used as independently established facts.
Listen / read
Odd Lots Sep 14, 2026

OpenAI President Greg Brockman on Doing Business in the Wake of Hugging Face

OpenAI president Greg Brockman discusses operational lessons from the Hugging Face security incident, coordination with rival labs, and the company’s approach to frontier-model risk; the episode is retained as executive interpretation, while incident facts remain grounded in the previously published primary disclosures.

Desk takeReviewed as an industry signal only; its claims are not used as independently established facts.
Listen / read
Flirting with Models Sep 14, 2026

Lucas Schuermann – Swapping Out Perpetual Futures (S7E34)

Variational co-founder Lucas Schuermann describes the firm’s request-for-quote liquidity model, flow segmentation, hedging, and proposed on-chain swaps for real-world-asset exposure; these are founder views on market structure and product design, not independently verified adoption or performance data.

Desk takeReviewed as an industry signal only; its claims are not used as independently established facts.
Listen / read

Podcast source watchlist

11 monitored sources

X signal wire

New post-level signals only. Earlier posts are not carried forward to fill a quiet edition.

Evidence rule:Each item below links to the original X post. Treat opinions and single-benchmark claims as provisional until replicated or corroborated by primary documentation.
No new source-linked X signal qualified for this edition.

Coverage & method

The publication layer follows a manifest-first, no-silent-repeat policy.

How to read this edition

Daily editions publish only first appearances and material updates.

Canonical links sit next to every item. Social posts remain separated from verified releases, and inaccessible sources are recorded as blocked rather than empty.

11published items
38sources checked
15blocked sources

Coverage run: 20260915T000005Z

Checked, no new relevant update

  • Adyen Knowledge Hub
  • Anthropic Research
  • BG2
  • BIS Innovation Hub
  • ECB research
  • FSB Financial Innovation
  • Google DeepMind Research
  • IMF FinTech Notes
  • Jane Street Engineering
  • Lex Fridman Podcast
  • Meta AI Research
  • Microsoft Research
  • NVIDIA Research
  • OECD AI and finance
  • OpenAI Research
  • Stanford AI Index
  • Stripe Engineering
  • Two Sigma Insights
  • arXiv cs.CL

Blocked or credential-limited

  • academic · 1 sources (OpenReview) — Official API returned HTTP 403; no complete dated listing could be verified.
  • academic · 1 sources (SSRN FEN) — Official FEN page returned HTTP 403; no complete dated listing could be verified.
  • academic · 1 sources (TMLR) — The journal index loaded, but its complete dated listing depends on the blocked OpenReview endpoint.
  • social · 12 sources (@AlexH_Johnson, @altcap, @bgurley, @demishassabis, @eladgil, @fchollet, @fintechjunkie, @karpathy, @patrickc, @saranormous, @simonw, @sytaylor) — X API account lookup failed: HTTP Error 402: Payment Required

Retrieval completed 2026-09-15T00:17:53Z. Links were verified against source pages where available.