ai
The main Latent Space podcast: AI engineering, Forward Deployed Engineering, AI for science, and ⚡ Lightning episodes. Other feeds: Engineering and FDE: https://podcast.latent.space/eng.xml; Science: https://podcast.latent.space/science.xml; Misc (Lightning and other shows): https://podcast.latent.space/misc.xml; Firehose (all available shows): https://podcast.latent.space/firehose.xml. Feed directory: https://latent.space/pods
Zach Lloyd explains that raw API usage for coding agents can be 5 to 50 times more expensive than subscription plans, noting the extreme cost gap developers exploit. He adds that this economic reality makes competing as a pure token-reselling coding agent exceptionally difficult.
Zach Lloyd discusses how Warp internally utilized automated tuning and testing features to dramatically lower the cost per pull request from around $80 down to between $15 and $20. He notes that during this time, their PR throughput increased by 40 percent without a rise in revert rates.
Zach Lloyd argues that the foundation of a self-improving software factory is defining workflows, repositories, and MCPs declaratively as code, comparing the approach to Terraform. This allows version control, branching, and automated pull requests from coding agents to improve the factory itself.
Jacob Lauridsen, CTO of Legora, discusses the evolution of legal technology and how their platform uses specialized agents, tabular reviews, and lists to automate complex M&A due diligence and legal workflows.
Jacob Lauridsen shares that Legora acquired the UK-based company Wexler to integrate fact intelligence features into their platform, primarily for litigation support.
Legora CTO Jacob Lauridsen talks about utilizing open-weight and frontier models depending on tasks, noting the challenges of model verbosity and the importance of localized evals.
Aman Singh shared that Arca recently raised sixty-four million dollars to build an AI-native wealth management firm. The company operates as the financial services firm itself, using a model-agnostic internal harness to run back-office operations and assist human advisors.
Aman Singh explained that Arca utilizes frontier models from labs like Anthropic and OpenAI, with Claude serving as their primary model. The firm maintains model-agnostic optionality, guiding model selection through internal benchmarks rather than relying on a single provider.
According to Aman Singh, while Arca's AI agents fully automate back-office tasks such as portfolio and tax analysis over a ten-day horizon, they are strictly barred from executing money movements, liquidations, or client communications without human approval.
Joseph Kim announced the release of OpenAI's new enterprise-focused product, ChatGPT for Financial Services, which integrates firm-specific templates and sources like S&P and Diligence, developed alongside design partners Morgan Stanley and Evercore.
Joseph Kim revealed that OpenAI experienced problems with pre-training and scaling models in the past, but the company has successfully resolved many of these issues in 2026.
Discussing product design decisions, Joseph Kim shared that OpenAI tested third-party MCP connectors and found error rates exceeding 40 percent, leading the company to bring financial data in-house for its finance product.
Joseph Kim stated that OpenAI is actively working to make the enterprise financial features more widely accessible to prosumers on the standard ChatGPT Pro plan.
Anthropic filed its S1 and expects an IPO in two weeks valued at two trillion dollars, highlighting massive revenue-per-employee growth and recent profitability.
Anthropic reported its first GAAP profitable quarter at 19 percent, marking a complete financial flip from losing 129 percent of revenue a year prior.
Anthropic implemented incremental API updates, such as enforcing preserved thinking blocks across turns, to make their models harder to distill from after a public research disclosure.
Anthropic successfully rolled out its Messages API rewritten entirely in Rust, achieving a 40 to 50 percent savings on CPU usage and a 10 to 20 percent latency improvement.
Anthropic discusses Claude Managed Agents, a suite of APIs that allows customers to define agents and run sessions without setting up complex custom infrastructure and harnesses.
Ian Quigley of LeashBio discusses the development of a full environment and eval for AI agents tackling structure-activity relationship problems in drug discovery. Using pre-measured experimental data, agents are given a budget of 100 molecules to find the best binder, allowing researchers to evaluate scientific decision-making at scale.
Ian Quigley shares that testing foundation models on drug discovery tasks revealed unexpected behavioral differences regarding resource allocation. More performant models are significantly quicker at recognizing bad trajectories and abandoning unproductive chemical neighborhoods, whereas weaker models tend to repeat unhelpful loops.
Alongside the podcast release, LeashBio is publishing a preprint paper, a dataset, and harbor packages containing thousands of transcripts from their AI model evaluations. This release aims to let the community analyze the failure modes and decision-making processes of different language models.
Discussing the roadmap for biological environments, Ian Quigley explains that LeashBio has already screened about 1,000 proteins against chemical libraries. The company is currently generating two environments a day and plans to scale that output to 20 environments a day.
Modal CEO Erik Bernhardsson shares that the company's revenue and usage have grown by at least 50 times over the past two years, moving from early struggles to rapid expansion driven by generative AI inference and sandboxes.
Erik Bernhardsson explains that to keep up with surging demand, Modal now works with roughly 25 different clouds, securing three-year deals with upfront payments and employing a dedicated four-person procurement desk.
Modal CEO Erik Bernhardsson notes that Modal Sandboxes, launched in 2023, now make up approximately one-third of the company's revenue due to massive growth from vibe coding companies, background agents, and reinforcement learning labs.
Erik Bernhardsson reveals that Modal has rapidly scaled its forward deployed engineering (FDE) team from four to twenty-four engineers this year to help customers with performance tuning, migrating code, and product optimization.
Erik Bernhardsson shares that Modal has grown to 200 employees, expects to hit 250 by the end of the year, and projects reaching approximately 700 employees by the end of next year, mostly consisting of systems engineers.