Writeups

Sarah Nagy’s case for natural-language analytics starts with interrupted analysts and waiting colleagues. The challenge is making generated queries dependable enough to guide business decisions.

From camera-based board games to interactive segmentation, Roboflow’s story shows how better models change annotation, deployment, and the economics of computer vision.

Part 4 of 4 of Latent Space’s ICLR 2024 recap examines how pre-training on task data changes the comparison between transformers and recurrent models—and lets simpler architectures close the gap.

Modal connects fast container startup, Python orchestration and GPU utilization to make custom cloud workloads easier to run—and economical to scale.

Browserbase combines hosted browsers, natural-language automation and human takeover to handle the state, access and reliability problems beneath an agent’s clicks.

Short experiments need flexible compute. Hardware financing needs long commitments. SF Compute connects them through resale, short reservations, and standardized delivery.

Durable workflows, streaming interfaces, operational context, and external security controls turn an agent loop into a usable application.

Shared experiments, deliberate code review, simulated shoppers and specialized models show what it takes to turn abundant generation into useful work.

Part 3 of 3 examines why simulated materials data can miss experimental needs, how training accommodates laboratory timescales, and why better GPU utilization could free money for lab work.

Human simulation combines personal histories, observed choices, and controlled experiments to model how people respond—and help decision makers explore what to do next.

Foundation models let developers begin with trained capabilities, then build the data connections, infrastructure and interfaces that make them useful. A 2023 view of the engineering role emerging around that change.

Fast speaking drills, adaptive curricula, and patient turn detection show what it takes to turn capable models into a language tutor.

Part 3 of the ChatGPT voice preview tests the gap between conversational control and precise audio behavior: recognizing an illusion, producing a tuning note, copying pitch, and thinking before speaking.

Coding agents move the work toward asking questions, supervising execution, and integrating changes. The promise comes with a learning curve—and a bottleneck that more code cannot solve.

Part 1 follows the path from production outages to dependable model interfaces: constrained decoding, tool execution, refusals, schema design, evaluation, and model selection.

Code execution, open weights and long context expanded what developers could build. Evaluation, deployment control and reliable behavior determined whether those capabilities could become useful products.

How generated textbooks, rewritten web pages and synthetic labels support capable models that cost less to run and can stay on the device.

AI can generate a starting interface. Choosing a direction, making its behavior coherent, and developing a distinctive product still demand design judgment.

Cosine’s software agent learns from reconstructed development work: finding relevant code, planning changes and recovering from mistakes. Its experience shows why better patches depend on more than code generation.

Part 6 of 6 of Latent Space’s ICML 2024 recording: faster robot policies, Chelsea Finn’s approaches to learning from language and experience, and Younghyo Park’s case for automatic environment shaping.

How requirements, mutable harnesses and agent security fit together in a conversation about the next era of coding agents.

A harness gives a model tools, context and boundaries. As models learn more of the coordination once supplied by software, harness design increasingly turns toward permissions, decisions and human attention.

An AIE Europe debrief on agent infrastructure, model specialization, and what changes when software generation outruns human review.

From queued business-software requests to specialized software teammates, Part 1 explores agent composition, knowledge graphs, engineering tradeoffs, and the discovery infrastructure a network of agents would need.

Part 2 follows the practical problems of delegating to agents: generating interfaces, testing candidates, choosing models, pricing work, and sharing context—with a closing turn to domain markets and long-term conviction.

Part 3 of 3 moves from accessible video tools and controlled image edits to reusable expertise—and the personal and community priorities that keep useful work sustainable.

Canvas and Tasks connect interface design with model training: choosing when to act, understanding what the user wants, and building trust through collaboration.

A conversation with Bret Taylor about building at technology’s limits, defining what agents should deliver, and making their decisions inspectable and correctable.

How Factory combines enterprise context, behavioral evaluations, retrieval, and human planning to turn coding tasks into delegated work.

An October 2023 conversation about applying models, delegating work to agents, and making useful AI software economically sustainable.