Writeups

AI’s accumulated breakthroughs, the Unix architecture behind personal agents, and the institutional work between technical capability and widespread adoption.

Part 2 follows the work of making AI useful: turning intent into tests, choosing long context or retrieval, and separating fast responses from sustainable inference.

The final part of Latent Space’s 2024 review examines what keeps agents from everyday use, revisits the year’s lasting technical lessons, and sets out the production test facing 2025.

Part 3 of this ICML 2024 reading edition explains diffusion’s training and sampling mechanics, follows DreamFusion through CAT3D, and introduces Ricky TQ Chen’s transport-based account of flow matching.

Part 1 follows Brev’s approach to GPU access, NVIDIA’s “speed of light” planning, and Dynamo’s controls for serving models across machines.

What Glean, Windsurf, and Roadrunner reveal about building enterprise AI: permissions, customer trust, sales leadership, and the commercial rules behind every quote.

Part 3 closes the hardware discussion with cloud economics, chip competition and overordering, then follows Milind’s work from video understanding to applied AI at Capital One.

Marble turns text and images into editable 3D scenes. Its camera control makes that useful today; learning how worlds behave remains a larger challenge.

A July 2023 conversation between Latent Space and Practical AI explores what turns model capabilities into useful systems: operations, evaluation, local data, customization and user experience.

How model size, data quality, human preferences, and tool use shaped Llama—and the research still needed to make agents more capable.

Agents can assemble financial research faster than an analyst can, but judgment, market regimes, and trusted data still matter. Part 2 of 3 follows a forecasting experiment into the economics of AI infrastructure and Microsoft’s software dilemma.

Fast experiments, reproducible trials, changing inputs, and production feedback belong in the same engineering workflow.

Part 3 of 3 follows the closing Q&A, research AMA recap, and two demos: reasoning limits, immigration drafting, and a prototype that turns o1’s plans into executable code.

Learning from rewarded attempts, building a competition checkpoint, and confronting the gaps in planning, data efficiency, and research judgment.

Sora explores what video generation can learn about the world. Genie adds frame-by-frame actions. Part 1 follows their mechanisms, demonstrations, and limits on the path toward agent training.

Part 4 follows flow matching into crystals and discrete tokens, then examines efficient image synthesis, controllable speech, synthetic pretraining data, and DeCAF’s legacy in computer vision.

Why reinforcement learning’s uneven generalization makes workflow context, product design, and learning from mistakes central to Cursor’s ambitions.

Part 5 of 6 follows DeCAF’s transfer-learning legacy, Lucas Beyer’s case for captioning and reward tuning, Ashley Edwards’s path from robot imitation to Genie, and Seung Jae Lee’s introduction to tokenized behavior generation.

David Luan’s account of OpenAI, Google and Adept connects research organization to a practical challenge: making models operate software reliably and learn from the people they help.

Part 5 of 5 examines why reversing a learned fact differs from reversing one supplied in a prompt, then explores how DSPy optimizes language-model programs—and why it still needs a useful evaluation metric.

Part 1 of Latent Space’s conversation with Yi Tay follows the path from general-purpose model research to Reka, where research judgment meets the difficulty of keeping rented compute running.

Building video models, preserving history, and combining generation with reasoning and editing tools: a reading edition of the Latent Space conversation with Ethan He.

Controlling reasoning is one part of building with Gemini. Live voice adds session limits, tool execution, conversational timing and the decision to stay silent.

Medal’s clips pair what players saw with what they did. General Intuition aims to use that experience to teach agents to act, then reconstruct situations where they can learn through interaction.

From reconstructing a workweek to redesigning Copilot, GitHub’s agent ambitions depend on shared context, credible review, and infrastructure built for changing workloads.

Cowork combines a coding agent, a controlled execution environment, browser access, and reusable instructions. Its central challenge is making work easy to delegate—and safe enough to leave running.

Production monitoring, live steering, and scientific discovery expose both the promise of mechanistic interpretability and the distance between finding a feature and controlling a capability.

Long-context reasoning, controlled coding agents, practical prompting, and workload-dependent costs in the April 2025 conversation.

Output specifications, validation and repair give developers more control over model behavior. SQL makes the mechanism concrete; an email assistant exposes the harder problem of specifying what the user wants.

Part 2 follows coding agents from demonstrating a change to working inside an enterprise—and asks which harness investments will remain useful as models improve.