Writeups

A second image-generation pass became a model-composition tool—and made local inference, memory management and extensibility central to creative control.

Dagger’s approach to coding agents starts with their working environment: isolate parallel work, preserve reusable setup, and let humans inspect what happened.

A research agent’s growing tool history turns context into an engineering problem—one that reaches into retrieval, multi-agent coordination, memory, and the architecture around the model.

From CAD operations to code edits, Cursor’s early development showed why useful generation needs context, tools, and a workflow people can inspect and steer. This historical snapshot accompanies the original recording published August 22, 2023.

Configured cloud computers let agents run applications, demonstrate changes, and work in parallel. The harder handoff is turning that output into software people can confidently ship.

A materials search engine needs generators, progressively tougher evaluations, experimental feedback, and chemists who can recognize missing requirements.

How tokenization, filtering, data mixtures, and training rights shape a model’s knowledge, deployment costs, and benchmark results.

Generating tests is one step toward a larger goal: keeping specifications, tests, and code aligned, with developers deciding what the software should do.

Part 2 of the NVIDIA Brev and Dynamo conversation: adapting inference to changing workloads, giving agents access to compute, and distinguishing longer runs from useful autonomous work.

Apps SDK gives developers an interface inside ChatGPT. AgentKit brings workflow design, evaluation, and deployment together—with consequential limits at launch.

Part 1 of 4 follows developers’ first reactions to OpenAI’s announcements: less conversation plumbing, more tools, and practical questions about billing, context, control and reliable action.

Part 3 follows expense assistants, browser agents, code search, and open models through the practical questions behind DevDay: usable context, reliable outputs, and responsibility for automated actions.

Part 4 of 4 moves from the Keep AI Open party to Julius AI’s code-execution workflows: why latency matters, how parallel tools could shorten repairs, and where startups can build alongside a platform.

Part 2 follows the move from conversational configuration to connected workflows—and the consequences for retrieval control, hosting costs, and developers building on OpenAI.

How Genesis combines physics-guided co-folding, demanding evaluations, and laboratory feedback—and why useful drug-discovery agents depend on all three.

Why API traffic cannot easily reveal training intent, why the strongest model may be a poor teacher, and how coding benchmarks can reward knowledge their tasks never supplied.

An April 2023 conversation with Mike Conover about shaping model behavior, building open datasets, and evaluating what applications actually need.

Shared tools, coordinated agents, inspectable code, and the product judgment behind personal software.

Useful workplace agents depend on bounded instructions, carefully prepared company data, and a product employees can shape around their own work.

From oversight experiments to reusable notebooks, Elicit builds research workflows whose steps can be trained, checked, and scaled.

Chatbots for adult creators must coordinate persona, sales, memory, and boundaries. Convincing language is only one part of the system.

Glean’s workplace search problem opens a broader discussion of ranking, tool use, model costs, and knowing which work to automate.

Learning evolution’s constraints from protein sequences opens a route to structure prediction, molecular design, and discovery. Extending that approach to cells will require new data and experimental feedback.

Enterprise agents need useful context, bounded access, a workspace, and human oversight. Making those pieces fit an organization is as much a workflow problem as a model problem.

Understanding a query is only the start. Exa’s search ambitions depend on retrieving the right documents, clarifying intent and deciding where more computation is worth the cost.

Fireworks combines model serving, workload-specific optimization, and systems that connect models to tools and data. Its competitive bet is that inference engineering will matter more as open models improve.

FlashAttention preserves the attention computation while reducing memory traffic. Its development shows why mathematical ideas, hardware constraints and usable software have to advance together.

Part 2 moves from practical API integration to launch-day o1 questions: model stability, retrieval, batch processing, speech, and the cost and visibility of reasoning.

Part 3 of 4 follows builders from budget approvals and browser purchases to codebase search, guardrails, and open models—with usable context and reliable evaluation as recurring tests.

GPU marketplaces, image search, model training, and browser agents in Latent Space’s May 2023 conversation with Sharif Shameem.