Daily Brief

Joseph Kim leads ChatGPT for Financial Services, which launched a month before this talk with design partners Morgan Stanley and Evercore. He demos the product and explains the "capability overhang": the gap between what models can do and the value people actually get from them at work. He covers why the team licensed and indexed financial data themselves instead of relying on existing MCP connectors, and how firm templates, interactive charts and citation viewers were built on the same ChatGPT plugin platform open to developers. He then looks at where models are headed, the point where a task

Affaan Mustafa and Alejandro Soumah Fernandez lead compute trading at Jump Trading. They explain why most GPU capacity still trades out of public view: private quotes, bilateral contracts and deals struck over Slack, X and Telegram. They look at why GPU prices have held up even as token prices fell, how chips, power, buildings and capital all feed into the price of a GPU hour, and why mid-size companies get squeezed when a 15 to 25% prepayment on a three-year contract can eat a whole funding round. They also cover brokers, exchanges, futures and flexible contracts, export controls, and why the

Andrey Rybka of Bloomberg walks through real security incidents involving AI agents and the controls that would have stopped each one: an agent that escaped its sandbox and took 17,600 actions over four and a half days, a zero-click email injection, instructions hidden in Morse code that drained a wallet, data smuggled out through DNS lookups, a poisoned MCP package, and an agent that dropped a production database after promising not to touch prod. His layered defense: one identity per agent with short-lived scoped credentials, treat all input as hostile, route every tool call through a determ

Full episode: https://www.youtube.com/watch?v=frYSOKvNiiE Me on twitter: https://x.com/dwarkesh_sp

Arca is an AI-native financial services firm where agents don't assist the business, they run it. Aman Singh, who leads AI at Arca, shows how long-horizon agents own a client goal end to end across days of stop-start work: a signal system that keeps agents aware of emails, calls, custodian events and in-app actions; operational context and a harness to do the work; persistent memory; and shared spaces where agents pull in the right person at the right time. The result: agents prompt people more than people prompt agents, and the firm steers the business by steering its agents. Chapters 00:00

Hari Kumar, head of AI at Contour Asset Management, a $6B fundamental long/short fund focused on tech, media and telecom, tackles the hardest question in AI adoption: what are we actually getting for the spend? He shares how Contour runs live A/B tests of analysts' research with and without AI tools, tracks "cost per useful insight," uses agents to resurface old ideas when the market makes them interesting again, and extends the same thinking to trading, ops and compliance. Chapters 00:00 Intro: AI at a $6B fundamental fund 01:45 Why AI ROI is so hard to capture 02:25 Live A/B tests of resear

Harneet Kaur, founder and CEO of Axim Intelligence and a former distressed credit investor, tests whether frontier models can reason about a live distressed situation: iHeartMedia. Given the same context, two top models disagreed sharply on the odds of bankruptcy versus a negotiated restructuring. Breaking the analysis into checked layers (extraction, derivation, judgment, prediction) brought them to the same answer, which raises a new question: is agreement the same as being right? Chapters 00:00 Intro 00:42 What distressed credit actually means 02:34 Why old scoring models fall short 04:30

Bridgewater Associates built the Vault: more than 100 million documents from 135 external sources plus the firm's own memos and meetings, cleaned, permissioned and tagged to its internal ontology. Sam Green (team lead, Applied AI) and Aaron Linsky (head of product engineering) share three lessons: humans and agents need different things from search, the biggest wins come from one tool that serves both, and the real prize is agents that proactively bring investors what deserves their attention. Includes how a hand-built benchmark took precision at 10 from about 60% to over 90%. Chapters 00:00

Full episode: https://www.youtube.com/watch?v=frYSOKvNiiE Me on twitter: https://x.com/dwarkesh_sp

A one-line prompt change made a support agent 17 points safer and 17 points worse at finishing tasks. Sofia Jakovcevic, forward deployed engineer at Arize AI, explains why agents fail at runtime rather than in code, and walks through a five-step regression loop built from production traces: curate datasets, score with evaluators, run baseline experiments, compare changes, and gate merges in CI. Then she shows how to close the loop at agent speed, with coding agents that read traces and fix themselves. Chapters 00:00 Intro: observability for agents 01:08 The real problem: fixes that break othe