Writeups

A practical account of coding agents: how they use tools, recover from errors, reuse workflows, and turn human feedback into completed work.

Video generation, segmentation with memory, faster detectors, and the challenge of teaching small models to see precise details.

Statsbot’s early Slack interface exposed two problems: understanding a question and understanding the data. Cube’s semantic layer separated them—and foundation models reopened the conversational side.

Part 3 of 3: the closing exchange of Latent Space’s conversation on NVIDIA’s AI engineering.

Ambient documentation gives Abridge a starting point for a harder problem: delivering accurate, timely assistance while clinicians can still act on it.

Part 4 of 4 moves from an open-source community gathering to Julius AI’s practical priorities: faster answers, multiple tool calls, visual feedback, and building useful products alongside a competing platform.

From broader task coverage to persistent codebases, coding evaluations are confronting what it means for an agent to do useful development.

Pydantic’s approach to agents connects data contracts, explicit control flow, and visibility into running applications.

A June 2024 conversation about turning foundation models into products, learning from deployment, and planning for changing inference costs and capabilities.

Unexpected polymer designs, active learning and quantum chemistry reveal what it takes to turn a promising prediction into a useful material.

As assistants take over discovery and reading, businesses need content that survives search selection, answers concrete questions, and reaches the model intact.

Closer model competition widened the opportunity for startups. Conviction’s NeurIPS 2024 keynote examined where that opportunity depended on better models—and where it depended on data, distribution and execution.

The dbt–Fivetran merger, oversized seed rounds, and the product case for memory and continual learning.

Generating a clip is becoming easier. Sustaining a character, controlling the result, and turning attention into a business still take work.

The final part follows AI infrastructure from financing and chip supply to memory budgets, specialized accelerators and CPU demand, before closing on writing habits and the value of stepping away.

Young AI companies need enormous resources before they have mature businesses. Whether that capital sustains an application ecosystem or finances its competitors remains an open question.

From GPU kernels to recursive language models, the researcher asks how a system’s structure changes what its components can accomplish.

How Amp approaches a moving target: model-specific tools, fewer product commitments, deliberate context selection, and fast feedback from tests and human review.

Continuous training, self-directed research and a simpler application stack underpin the ambition to build thousands of useful products with roughly 12 people.

Real user tasks make Arena’s evaluations valuable. Funding those requests, retaining users, and deciding what to publish make the benchmark a product and governance challenge.

Choosing a model means comparing capability, behavior, speed and the cost of completing the work. Artificial Analysis builds evaluations around those trade-offs.

Devin and OpenInspect show what background coding requires: shared sessions, working environments, useful tests, scoped memory and review that keeps repositories maintainable.

How onboard autonomy changes drone operation—and why mission verification, manufacturing, and the economics of defense matter as much as the AI.

Formal proofs give AI a checkable target for learning and generation. Axiom’s challenge extends from finding proofs to specifying the right problem, managing large searches, and making results reusable.

Wearable audio gives Bee context without constant explanation. Turning that context into useful help requires editable memory, selective attention, and careful choices about power and permission.

A benchmark tests particular capabilities under particular conditions. Understanding those choices helps explain both model progress and the gap between a leaderboard score and a useful product.

Clémentine Fourrier on reproducible comparisons, the limits of preference rankings, and the engineering choices behind model scores, from the July 2024 conversation.

Part 3 follows an agent’s failed travel plan into the science of benchmarks, then examines how Self-RAG and process supervision use intermediate judgments to select better outputs.

Datology’s case for optimizing filtering, rewriting and training order together—and measuring their value in learning efficiency, model capability and inference cost.

How evolutionary clues, generative models and inference-time search support molecular design—and why useful predictions still have to survive experiments.