The wiki

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

Modular connects Mojo programming, MAX compilation and model serving in an AI infrastructure stack built to span different processors.

A social network for AI agents that grew through OpenClaw skills and recurring heartbeat instructions, faced an early database exposure, and became a setting for external research on agent interaction.

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.

The Beijing company behind Kimi develops models, long-context serving infrastructure and agents for coding and knowledge work.

Machine-learning researcher whose work spans benchmark reliability, algorithmic fairness, and predictions that change the outcomes they forecast.

MosaicML built tools and infrastructure for training custom AI models, released the MPT transformer family, and joined Databricks in July 2023.

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.

Software engineer and Gemini Deep Research tech lead in 2025, whose work helped turn web research into reports and whose public interests included personalized learning and multimodal research agents.

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

Worked on Goodfire’s founding team, connecting mechanistic interpretability research to model customization, production guardrails, and interfaces for controlling AI systems.

Software engineer, author and educator whose work includes React Native AI and developer education around coding agents at Cognition.

Brev co-founder focused on easier GPU access, developer experience at NVIDIA, and tools that agents can use.

Former GitHub CEO whose AI work spans Copilot’s launch, browser agents, startup funding and the Vesuvius Challenge.

Sierra agent engineering founder whose work combines customer systems integration, reliable AI orchestration, and customer-facing engineering.

Researcher and Interconnects author focused on open language models, post-training recipes, and the reward signals that shape AI assistants.

Zed’s CEO and co-founder advocates AI-assisted software craftsmanship, interchangeable coding agents, and collaboration that keeps conversations connected to code.

Chai Discovery product leader working on AI tools for medicine design, with earlier experience in security automation, product management, and robotics.

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.

Machine-learning security researcher whose work tests model robustness, training-data privacy, and the security of production language-model APIs.

OpenAI Apps SDK contributor who helped shape MCP Apps and argues for model-facing tools designed around user intent.

Google DeepMind image-product leader associated with Nano Banana and Whisk, whose public views emphasize iteration and cultivated creative judgment.

Cline’s Head of AI in May 2025, Pash advocated for coding agents that explore repositories, make precise edits, and preserve context through long tasks.