Writeups

Part 1 follows developers from simpler conversation APIs and custom GPTs to tool execution, expressive speech, and the unresolved engineering questions behind vision.

Part 2 of Latent Space’s DevDay recording explores conversational programming, managed retrieval’s limits, plugin migration, and Zapier’s move from chat actions to scheduled assistants.

A June 2023 developer roundtable on native function calling, cheaper inference, and the decisions application code still needs to control.

The March 2025 launch conversation explains why Responses became a new API, how its tools compose, and how handoffs and traces could support better agents.

Alex Atallah and Anjney Midha on model distribution, inference competition, fusion, and the security challenges of a growing token economy.

Anj Midha’s case for a compute grid begins with idle accelerators and expands into a theory of frontier labs: pool resources, reduce coordination losses, and let the mission determine what to build.

Part 3 moves from writing before the experiments are settled to the environments that shape research judgment, then closes with an argument for building AI ecosystems around experienced technical contributors.

Liam Fedus and Ekin Doğuş Çubuk explain how uncertain measurements, simulation and lab automation fit into materials discovery.

Better context and stronger models turned a search product into a coding assistant. Specialized training promised more control over cost and capability; consistent correctness remained the harder problem.

Applied Intuition connects autonomy models to the engineering that makes them usable: sensor supervision, reliable operating systems, validated simulation and inference that fits the machine.

Better images are only part of the problem. Playground’s ambition connects model quality, precise editing, visual control, and feedback from people making things.

Josh McGrath on the engineering behind reinforcement learning, efficient agents, and useful long context, in a conversation at NeurIPS 2025.

Linear attention, state-space models and RWKV reduce the cost of processing sequences. Their harder challenge is preserving the information a task needs—and making that computation fast on real hardware.

From incompatible datasets to efficient training: how Axolotl turns model-building experience into reusable configuration.

Braintrust connects production logs, comparative evaluation, and interactive debugging so teams can improve AI behavior—and decide which work belongs in ordinary code.

LlamaIndex’s early evolution shows why connecting language models to data requires more than a larger context window.

Part 2 of the Railway conversation follows self-deploying workloads, durable workflows and cloneable environments to a practical question: how does faster code generation become safe production software?

Railway’s closing ambition is an instantaneous path from idea to shareable production deployment, pursued at every necessary layer of the stack.

Deploying an application is only the beginning. Part 1 follows Railway’s approach to evolving it through environment versioning, owned hardware, agent interfaces and safer production experiments.

Andon Labs tests whether agents can sustain a business, coordinate with people and act reliably in the physical world. Earnings matter—but so do the decisions that produce them.

Gray Swan connects adversarial testing, specialized guard models and system permissions to make agents safer without stripping away their usefulness.

How a tool for sharing research became a platform for running and customizing models—and why packaging, GPU economics and hacker culture all mattered.

From financial text analysis to Ghostwriter: a reading edition of the May 2023 conversation with Reza Shabani on training code models, evaluating their feel, and moving beyond autocomplete.

Faster code generation leaves verification work behind. Retool’s bet is on AI inside business processes, with developers controlling the steps and people reviewing the results.

An unaffordable recommendation feature led to Turbopuffer’s storage architecture. Agents issuing many searches at once are now changing its workload and economics.

How instruction tuning, reward models, and DPO turn feedback into model updates—and why the resulting behavior remains difficult to predict.

A browser-testing problem leads to cheaper token processing, multilingual models, and a crucial distinction: processing a long input is not the same as remembering it.

Selecting an object is the beginning. SAM 2 connects tracking, correction and training data through an interface designed alongside the model.

Concept prompts let users name what they want to see. SAM 3’s architecture, data engine and deployment examples explain how that request becomes masks and tracks—and where human intent still matters.

A March 2024 conversation about combining model reasoning with software workflows—from a stubborn output format to a thousand-call textbook generator.