Source ledger: generic-finance-topical-2026-08-16-raw.md Episode URL: N/A — this is a cross-publisher discovery and transcript-capture note; individual recording URLs are cited inline. Credibility: MEDIUM overall; HIGH only for facts independently corroborated by first-party firm pages. Source status: Ten timestamped raw captures archived; 147 discovery candidates retained as leads. Method: 147 topical candidates across six lanes, followed by canonical-source resolution and timestamped caption capture for ten high-signal recordings. Search candidates are not treated as firm evidence until the transcript and source identity pass review.
The highest-value findings were not generic “AI is important” remarks. They were operating details that sit inside ordinary finance conversations:
- Jane Street links a large Texas GPU facility to research and trading, and its first-party page reports 4,032 GPUs across 56 liquid-cooled racks with roughly 8,000 km of fiber. Its ML page describes neural-network trading models, distributed training and inference, LLM/RL/CUDA research, and custom trading architectures. (data-center page; ML page)
- HRT’s follow-up Odd Lots discussion connects token use to GPU availability, data-center power, training scale, and the research cycle. It is a practical AI-FinOps and infrastructure signal, not a public model-card disclosure. (episode page; timestamped transcript)
- Numerai’s NumerCon recording points to MCP, open Skills, and agent-facing research/submission tooling. Numerai’s written recap remains the controlling source for the 8B Predictive LLM, its one-million-plus article corpus, and Faith feature generation. (recap; closing recording)
- A topical Bridgewater discussion led back to the first-party AIA Labs page, which lists the June 30, 2026 expert-judgment paper and describes models tuned for organizational tasks. Bridgewater also makes firm-reported claims about live-capital use and Pure Alpha integration; those claims are not independently verified here. (AIA Labs; recording)
- The CFA Institute and RavenPack recordings add research-method vocabulary: causal inference, factor construction, alternative-data transformation, and NLP-based signal extraction. They are source-discovery and method evidence, not proof of a specific hedge-fund deployment. (CFA; RavenPack)
- A title-blind Beryl Elites interview with RavenPack Chief Data Scientist Peter Hafez adds a distinct vendor/data-layer route. Hafez describes knowledge graphs, entity resolution, grounded retrieval, use-case-specific “micro data sets,” repeated LLM classification and voting, and mind-map convergence. The discussion is vendor methodology, not evidence of a named manager’s deployment, model weights, permissions, or performance. (episode; timestamped capture note)
Why this matters for the research process
The discovery unit should be the topical episode, not the AI-labeled episode. The transcript is then screened for compute, data, models, agents, automation, research platforms, signal generation, and evaluation language. Promotion still requires a named speaker, employer as-of the episode, timestamp, canonical URL, and an explicit statement of what the source does not show.
The researcher-name expansion found a three-episode Steve Hou lane that the hedge-fund-labeled queue had missed. Full Signal’s December 2025 episode identifies Hou as a Bloomberg quantitative researcher covering multi-asset strategy research and discusses AI’s effect on labor, markets, and a new stock index. Forward Guidance’s April 2026 episode supplies a deeper timestamped macro/AI agenda covering capex, agentic-AI compute demand, productivity, and physical bottlenecks. A February 2026 Full Signal episode adds a dated US–China AI-arms-race discussion. Bloomberg’s public 2026 index outlook and signal-research article provide a separate research-artifact trail. These are public market-research and media surfaces, not evidence of a hedge fund’s live AI deployment. See the capture note.
This catches infrastructure and workflow disclosures while avoiding the opposite error: treating a search result, host framing, sponsor claim, or auto-caption fragment as evidence of live investment authority or performance.
Captured transcripts
- Jane Street GPUs and trading
- Jane Street Texas data-center tour
- Jane Street machine learning
- HRT AI use
- HRT token burn
- NumerCon 2026 closing remarks
- Bridgewater / Thinking Machines discussion
- Systematic Investing and AI
- Causal AI and factor investing
- Alternative-data investment processes
- RavenPack / Peter Hafez interview
Researcher-name expansion — Steve Hou
- Full Signal — AI and the economy — publisher metadata and timestamps; transcript capture pending.
- Forward Guidance — The AI Bubble Is Widely Misunderstood — 29 April 2026; publisher timestamps; transcript capture pending.
- Full Signal — Wall Street MISSED this trade — 19 February 2026; publisher metadata and timestamps; transcript capture pending.
Evidence Boundaries
Auto-captions are not diarization and can misrecognize names, hardware, and acronyms. A podcast does not establish current employment, confidential methods, model weights, production deployment, investment authority, alpha, or returns. The 147-candidate count measures discovery coverage, not verified disclosures.