Writeups

Part 4 of Latent Space’s NeurIPS 2023 recap closes with the common ingredients of attention-free models, their remaining bottlenecks, and the different demands of interactive responses and bulk processing.

Part 2 of Latent Space’s Q2 2024 recap examines the infrastructure agents need, the work customers will pay for, and the possibilities and limits of ChatGPT’s early advanced voice mode.

Brushes, database columns, and inspectable programs show how interface design can turn language-model capabilities into useful, controllable work.

Stronger code generation made Bolt possible. A browser runtime, an observable repair loop, and integrated deployment made it useful. Enterprise agents face the same challenge under very different constraints.

Selecting evidence, preserving its meaning, and turning it into investment questions—with expert judgment guiding the system.

Claude’s SWE-bench reference agent shows how much depends on tool design. Computer use and robotics extend the lesson to integration cost, reliability, and trust.

Part 2 of the OpenAI DevDay 2024 conversation moves from voice interfaces and application-specific evaluation to Cosine’s software-engineering models, then to planning products around unpredictable research advances.

Part 3 of Latent Space’s OpenAI DevDay 2024 coverage explores deployment safety, agent reliability, product discovery, internal workflows, and the gap between generated code and useful software.

Part 1 follows realtime voice from database lookups to a 400-strawberry order, then explores evaluation-driven model improvement and o1-assisted application development.

Modular connects asynchronous execution, compiler fusion and Mojo’s Python-oriented language design to a common goal: making new computations easier to express without giving up hardware performance.

Singapore’s AI strategy connects workplace adoption, practical skills and compute access—with distinct roles for government, employers and engineers.

Superhuman connects deep search, focused tools and demanding evaluations with a harder product question: how much work can an assistant remove without creating more review?

A reading edition of the Microsoft Build 2026 conversation on company-specific learning, reusable software assets, and AI whose value appears in completed work.

Google Labs’ coding agent puts independent execution, simpler orchestration, and persistent context in service of a harder problem: specifying and verifying useful software.

From capability discovery and conversational interfaces to product differentiation, model evaluation, and personalized AI.

Saving an insight takes a triple-tap. Making it searchable and useful later takes audio alignment, careful model choices and an interface built around listening habits. This conversation was published [March 14, 2025](https://www.latent.space/p/snipd).

From calendar classifiers to company-wide search: useful AI, the infrastructure around it, and the judgment needed to rethink work.

Designing an antibody means specifying where it should bind, what it should avoid, and how it should behave. Chai’s models and visual tools aim to make that process programmable, while experiments remain the test.

From unwanted utility bots to social AI: how Chai lets users shape the experience, evaluates models in hours, and spends inference compute on selecting better replies.

How community comparisons became a model benchmark—and what style, voter expertise, private testing and inference cost mean for its scores.

Shared tools, persistent files and inspectable outputs extend Codex beyond software development. Making that useful requires helping people discover tasks, preserve context and distinguish activity from progress.

A launch-day discussion moves from video editing and plugin distribution to personal assistants, context limits and the permissions models need to act.

AI can organize research without knowing what matters. Part 1 follows Doug O’Laughlin’s path into semiconductors, then explores coding agents for financial analysis, evaluation rubrics, context management and shared research infrastructure.

How planning, visible execution, direct model access, and persistent task state shape an open-source coding agent.

CloudChef combines cooking-state perception, reusable robot skills and structured recipes to automate commercial kitchen work. Its single-demonstration promise depends on skills the robot already has.

Part 2 of 3 follows practitioners discovering useful capabilities, testing the limits of an isolated workspace, and explaining why code that runs and revises itself changes what they choose to build.

Part 3 of 3 examines missing charts in shared sessions, requests for local code execution, and a central tension in business analysis: how do you choose an upload before you know what data you need?

Part 1 follows the early beta from uploaded databases and iterative debugging to sandbox limits, interactive exports, and the distinction between files a program can execute and information a model has read.

How GPU efficiency, completion design and workload-specific evaluation turn model capability into everyday utility.

Long-running coding agents depend on familiar tool interfaces, context management, and working habits developers can inspect and evaluate.