Daily Brief

The best early advice Chase Lochmiller got at Crusoe: most startups die from suicide, not homicide. It’s not competitors that kill you, it’s you not executing. If the opportunity is real, there’s room for multiple winners, and someone else chasing it is a compliment. #ai #infrastructure #leadership

Every new coding agent session starts like a new hire: it doesn't know your conventions, architecture rules or approved libraries, so it burns time and tokens searching your repo and repeats the issues already in it. Killian Carlsen-Phelan of Sonar introduces Sonar Vortex, a plugin that hands your agent a map of the codebase and your guidelines before it writes, then deterministically checks every edit. In a live session, Vortex catches a temp-file permission bug the agent's own tests missed, and rejects its first fix. Sonar reports 36% fewer tokens and 44% fewer production outages. Chapters

RAG gives agents a way to read, but memory gives them a way to write, and that's where your business value lives. Wojciech Pluta, head of technical marketing for Oracle AI Database, explains memory promotion: deciding what's worth keeping, checking it against existing records and admitting, rejecting or escalating it to a human. A mortgage-agent demo remembers an operator's preferences across sessions and refuses to store an "approve every application" rule. He also covers short-term, long-term and shared memory, why to store memory in a database, and the context card that keeps prompts lean.

Steven Latré leads AI at imec, the Belgian semiconductor R&D center, and his team ran thousands of benchmark runs to find out what moving from APIs to your own GPUs really costs for long-running coding agents. Testing open-weight models and three coding harnesses on SWE-Bench Pro, they found that hardware costs jump in steps rather than a smooth curve, that concurrency matters even more than accuracy, that break-even GPU utilization ranges from 89% on small systems to about 15% on 8x B200 machines, and that the right harness can cut GPU time in half for the same result. Chapters: 0:00 Intro 0

Bumble product manager Supriya Subudhi shares five lessons from A/B testing an AI photo feedback feature in a dating app with over 20 million monthly users. 78% of surveyed users said they wanted feedback, but only 10-15% used it; the average win hid odd outputs like telling smiling people to smile more; the best model cost about 10x more; switching to a cheaper model also shrank the results; and regulation meant the feature couldn't ship outside the US. She closes with five questions to ask before you call any AI test a win. Chapters: 0:00 Intro 1:31 Bumble's AI photo feedback feature 2:38 1

Full episode: https://www.youtube.com/watch?v=f0J_EeW42Lo Me on twitter: https://x.com/dwarkesh_sp

Richmond Alake was talking about agent memory before most people had heard of it, and now he argues the next layer after the control plane is an intelligence plane: what an agent knows, what it remembers, what it's allowed to see, and how it gets better. Using the WWII codebreakers who worked out that "AF" meant Midway, he shows why having the same information as everyone else isn't the same as intelligence. He then demos an agent built with an ontology and semantic layer that gets answers right 99% of the time using 32K tokens, versus 200K for a naive agent, and covers Oracle's agent memory p

Jeff Smith has built database tools at Oracle for 25 years, and he's here to make agents and databases get along safely. He explains why natural language to SQL is powerful but risky, since queries can run cleanly and still return the wrong data, and how Oracle's MCP server makes it safer: offering your team's vetted reports as tools instead of ad hoc SQL, pointing agents at a read-only standby for performance tuning, annotating schemas so agents stop brute-forcing your data dictionary, and deep data security that strips out rows a user isn't allowed to see before results ever reach the agent.

What does it take to build AI for finance on top of a general-purpose model? Joseph Kim, who leads work on ChatGPT for Financial Services at OpenAI, joins swyx and Richard at AI Engineer New York to discuss the decisions behind the product: training financial reasoning into a shared model, connecting market data, and building workflows financial professionals can review and trust. Joseph explains why OpenAI bundled financial data even when connectors already existed, what startups can build with the same plugin primitives, and why human taste and expert judgment are hard to measure. The conv

Arca head of AI Aman Singh joins Latent Space head of editorial Richard MacManus at AI Engineer New York, to discuss how Arca is building an AI-native wealth management firm. Singh explains Arca's architecture, which includes a custom harness and persistent memory. He also digs into how Arca runs agentic workflows that span days, not hours. Arca: https://arcawealth.ai/ AI Engineer: https://ai.engineer/ Latent Space: https://www.latent.space/ Chapters: 00:00 Introducing Arca’s AI-native wealth management firm 01:40 How AI supports human financial advisors 02:40 Choosing models with internal