Writeups

Lindy’s redesign makes required actions explicit while leaving models room to interpret information—a division of labor that shapes its workflows, memory, evaluations, and product strategy.

A July 2023 launch-day conversation about commercially usable models, the engineering behind human feedback, and the applications still waiting to be tested.

From XGBoost to browser-based language models, MLC tackles the engineering cost of making models useful across hardware.

From laptop models to camera conversations, 2024 expanded what people could do with LLMs—and exposed the reliability, interface and trust problems that builders still needed to solve.

Part 2 of the January–February 2024 recap: Gemini’s million-token context, prompt transparency, anniversary demos of Lindy, RWKV, Pixee, and Julius, and listeners on evaluation and technical community.

Part 3 of 3 of Latent Space’s January–February 2024 recap closes with listener questions about research coverage, AI user experience, and organizing a community around working demos.

Learning to represent sound lets Suno reuse transformer machinery for music. The harder questions are how to generate fast enough for listening, give people useful control, and judge whether a song moves them.

Why Model Context Protocol separates tools, resources, and prompts—and how client support, transport, authorization, and governance shape what developers can build.

A large model needs more than a fast inference engine: it needs capacity that scales, predictable latency and workflows that keep the product running.

Voxtral TTS generates speech one frame at a time. The same conversation connects specialized models, enterprise training, formal proofs, and learning from agents that work for hours.

How a runtime for changing workloads grew into infrastructure for inference, training and agents—with programmable configuration and observable execution at its center.

Moonlake combines reasoning models, simulation tools, and learned rendering to create interactive worlds with persistent state—and give agents somewhere to learn from their actions.

Affordable models depend on frontier research. This conversation with Jeff Dean follows that connection through search, chip design, model specialization, and coding agents.

How MosaicML combined attention implementations, data choices, and automated recovery—and why cheaper experiments mattered as much as a finished model.

From coding agents that do too much to tools that get called but ignored, the training problem is deciding which actions deserve credit.

Part 2 follows QLoRA’s four-bit fine-tuning results, DataComp’s controlled dataset experiments, LLaVA’s visual instruction pipeline, and Tree of Thoughts’ search over alternative reasoning moves.

Part 3 moves from an arithmetic puzzle to Minecraft agents, planning tests, and sequence models—examining how capabilities are built and where apparent success breaks down.

Part 1 follows MosaicML, Fireworks and Cursor through model specialization, production inference, synthetic evaluation data and the value of reusing computation.

Part 2 of 4 follows Cursor’s careful hiring, Perplexity’s move into model serving, Metaphor’s search architecture, and Answer.ai’s plans to make AI useful and discover overlooked talent.

Part 3 moves from conference conversations to Cerebras’ sparse training research, FiftyOne’s dataset debugger, and Atlas’s maps of unstructured data—ending with the question of who can afford to run models.

Word2Vec’s asynchronous updates, metric-induced emergence, DPO’s algebra, repeated-data scaling, and QLoRA’s memory savings—with their mechanisms and limits.

Nomic on sustaining open source, Lightning on reducing training and infrastructure mistakes, and Jay Alammar on generation, model knowledge, and Transformer ideas worth teaching.

Open models, local inference, LoRA, and Mojo reshape the question of where lasting engineering advantages come from.

From human tumor datasets to spatial transformers: how Noetik aims to turn patient biology into better-defined clinical trials.

Sourcegraph combines search, code graphs and language models to supply useful context—and treats reliable individual changes as the foundation for greater automation.

How Hex Magic embeds AI in analytical work: small, reviewable tasks, warehouse-aware retrieval, and evaluations that turn recurring failures into engineering improvements.

Background work requires understandable permissions, dependable tools, useful evaluations, and context that extends beyond a chat.

MCP’s first year exposed the hard parts of shared agent infrastructure: enterprise identity, bidirectional transport, trusted discovery, long-running work, and neutral governance.

From PyTorch’s hardware tradeoffs to shared feedback and home robots: a reading edition of the March 2024 conversation with Soumith Chintala.

Open models gave builders more control. Shared training pipelines let researchers reuse one another’s work—but compute, data access and commercial incentives still limited participation.