Videos and podcasts worth watching in full.

HRT continues to hire interns; the guest describes AI tooling as amplifying roles rather than replacing the current hiring plan.

Early adoption is framed as giving people time and room to experiment with their workflows.

The discussion connects model productivity to utilization and the economics of data-center capacity.

The guest separates model capability from the harder question of reliably predicting markets.

The guest says large-model training is technically conceivable for HRT but describes frontier development as highly capital intensive.

Reported GenAI uses include coding, experiment ideation, experiment monitoring, and agents proposing signals for comparison with human researchers.

The reported safety case depends on short horizons, automated risk checks, and a well-posed problem.

A team-level estimate places ordinary daily token use in the hundreds of dollars per employee, with occasional bursts higher.

Low-level market events are described as a key data substrate for intraday prediction.

The guest declines to generalize short-horizon market-data evidence to month-ahead fundamental prediction.

The neural model is described as producing a plan or prediction, with audited and risk-checked layers acting on the output.

The conversation flags contamination risk when a general model has already seen historical speeches or news outcomes.

The speakers distinguish research and training workloads from the nanosecond path of trading decisions.

The conversation describes prediction targets such as fair value rather than assuming a single next-order-book prediction.

Model diversity and faster experimentation are presented as design choices across research applications.

The speakers describe hiring across machine-learning architectures, custom models, and the LLM training lifecycle.

The tour distinguishes LLM training from custom architectures adapted to other research problems.

The site is described as holding 4,032 GPUs across 56 racks, with power balancing as an operating concern.

The speakers connect the cluster to quantitative research on trading strategies.

The infrastructure story is framed as an evolution from an earlier small operation to a data-center model.

The opening describes effort invested in a data and modeling pipeline built with software-engineering principles.

The conversation connects macro research to systematic information gathering and repeatable modeling workflows.

The source is useful for mapping the research process; it does not disclose a named proprietary model.

The interview names an investment-committee AI agent called Maya and references a separate model provider.

The guest describes automated agents running thousands of tasks daily across the business.

A deeper research agent is described as taking hours on more complicated questions.

The host identifies Balyasny’s founder, managing partner, and chief investment officer.

The interview is most useful as organizational context for a multi-strategy platform.

AI references should be reconciled with direct technical or firm-published evidence before promotion.

The guest describes using language models for coding and idea generation, while warning that models can suggest incorrect alpha.

Reinforcement learning is discussed as a way to tune models for constrained tasks.

An agent is connected to a controlled MCP server rather than given unrestricted notebook access.

Feature engineering, model selection, and hyperparameter tuning are described as possible constrained search spaces.

The publisher description frames agents as junior analysts inside research and trading workflows.

The episode description also names proprietary model stacks, alternative data, and strategy/risk use cases.

The episode introduces the data-science function as part of the investment process.

The source is valuable for historical organizational context rather than current deployment claims.

Any current AI inference should be treated as a follow-up research question, not as evidence from this older episode.

The episode identifies Mike Schuster as a leader of Two Sigma’s AI Core team.

Because this is a 2021 source, it is retained as historical evidence rather than a current-state claim.

The introduction identifies Bouchaud as CFM’s chief scientist, head of research, chairman, and co-founder.

CFM’s research division is described as an academic-style department with many PhD researchers.

The discussion places machine learning, text analysis, and large language models alongside longer-running quantitative methods.

The guest emphasizes the need to understand whether a model’s behavior makes sense before production deployment.

The interview discusses meta-models that could help assess whether a backtest is credible enough for production.

The discussion lists discussion boards, Discord, Reddit, and podcast transcripts as possible research inputs.

Expert interviews are described as a fundamental research method that language models can help scale.

An agent conducting interviews is presented as a possible decomposition of the research process.

The conversation distinguishes a language-model research layer from a data-science layer that must validate the information.

The episode opens by separating the desire to use AI for trading from the practical problem of financial data exposure.

Historical market data is described as coming from regimes that may differ materially from the present.

The source is retained as a methodological counterpoint on where GenAI may fit around a strategy.

The guest identifies sentiment analytics from news reports as a possible signal-generation use case.

The conversation frames AI as a tool that should work symbiotically with quantitative researchers.

The source is relevant to QIS and systematic investing, but the speaker is from Morgan Stanley rather than a hedge fund.

Governance is described as part of requirements, testing, and the full AI life cycle rather than an end-stage add-on.

The discussion connects delegated authority to documented decision boundaries and human approval thresholds.

Agent decisions are connected to human approvals through risk-based thresholds, with the institution retaining responsibility.

Responsible-AI principles are placed inside the agent development life cycle, with risk analysis before lower-risk fast paths.

The conversation treats an agent as software whose apparent autonomy is created by human design and control.

Monitoring decisions, disparate impacts, transparency, and corrective action are described as continuing obligations.

Large language models are described as useful for multidimensional synthesis across large information sets.

A language model is framed as an interpreter between a researcher and a larger analytical library.

The conversation emphasizes verification and testing when LLM output is used in a research workflow.

Agentic systems are discussed as possible research assistants, not as evidence of autonomous trading authority.

Primary research through expert conversations is presented as a workflow that language models can help scale.

An agent conducting an interview is discussed as a possible decomposition of research work.

A transcript library is described as a possible training or customization substrate for question generation.

The publisher identifies McAuliffe as Voleon’s co-founder and chief investment officer.

The transcript places the guest’s background across Harvard, D. E. Shaw, Amazon, and Berkeley.

The interview describes a systematic process built around computer horsepower, databases, machine learning, and a predictive engine.

The episode frames Numerai as a quantitative fund built around collective intelligence rather than one internal research team.

The transcript describes a model-contribution interface in which external researchers submit signals.

Ensembling is described as a way to combine models whose errors differ, rather than relying on a single forecast.

The episode makes strategy time horizon a central condition for applying deep-learning methods to market data.

The discussion treats regime change and overfitting as failure modes when models fit one market environment too closely.

GPUs are described as useful for the calculus-heavy training workloads behind gradient descent and backpropagation.

The speakers describe quant investing as systematic implementation of investment ideas with risk controls, not a black-box substitute for insight.

NLP is described as a way to analyze speeches, earnings-call transcripts, and other unstructured company information.

The discussion connects AI and machine learning to research access, alpha ideas, portfolio construction, risk, and implementation.

The speakers retain a role for human investment ideas and judgment while describing possible future agentic trading workflows.

The quantitative team describes processing millions of data points and using optimization to control factor exposures, risk, and trading costs.

NLP is described as extending analysis beyond numeric statements into financial-statement text and earnings-call transcripts.

The speakers describe factor selection and model construction as human activities that introduce judgment and potential bias.

The source is a traditional asset-manager comparison and does not identify a named GenAI system.

The publisher chapter map labels a new economic regime discussion.

The publisher chapter map labels AI impact on the labor market.

The publisher chapter map labels a new investing strategy.

The publisher chapter map labels index rebalancing.

The publisher chapter map labels AI’s macro impact.

The publisher chapter map labels AI capex as a growth driver.

The publisher chapter map labels agentic-AI compute demand.

The publisher chapter map labels a productivity discussion.

The publisher chapter map labels physical-world bottlenecks.

The publisher chapter map labels geopolitics becoming macro.

The publisher chapter map labels US versus China risk and camp exposure.

The publisher chapter map labels currencies and returns.

The publisher chapter map labels the AI arms race.