Writeups

Jeremy Howard’s case for accessible deep learning leads to a harder question: how can a language model learn a specialization while retaining its broader abilities?

Part 1 of this 2024 Latent Space recap follows the path from AI investment to useful systems: stronger models, synthetic data, cheaper inference, local execution and multimodal applications.

Part 1 of 3 follows an early-2024 debate over compute, data, model competition and reliable agents, ending with how faster inference could change the applications worth building.

Latent Space’s December 2023 recap follows the competition over training data, inference economics, multimodal products, and application tooling—then turns to coding agents and personal-context hardware.

AI economics, inference bandwidth and chip design in Latent Space’s November 2023 conversation with Dylan Patel—ending with the question of whether training can escape a single data center.

Part 2 of 3 follows Lila Sciences from improving drug-development odds to connecting instruments, shortening experimental cycles, and training scientific models—then asks where simulation still falls short.

From a bot over Notion and Slack to SQL agents and application traces, LangChain addresses the work surrounding a model call.

Tulu’s practical post-training recipe leads to harder questions: how models learn to use tools, choose plans, recognize success, and remain useful as rewards become more complex.

Radical AI connects hypothesis generation, robotic experiments and scientific judgment. The next challenge is carrying that feedback loop into manufacturing.

How examples, reasoning, automated optimization, and adversarial testing shape the behavior of language models—and where prompting becomes systems engineering.

Circuit tracing turns a model’s next-token prediction into a testable account of intermediate computation. Part 1 follows the tools, their limits, and the case for understanding models before relying on them.

Circuit tracing connects internal features to model outputs, then tests those connections by changing them. Part 2 explores shared concepts, advance planning and misleading explanations before turning to the tools’ limits and the trade-offs of publishing them.

Part 3 of 3 examines the work behind clear interpretability papers, the infrastructure that speeds up subsequent investigations, and the questions newcomers can pursue.

Vector search remains useful. The question is whether it needs a separate database—and how retrieval should fit the application’s data, quality requirements, and workload.

From Cicero’s unusual moves to coding agents and multi-agent learning: what extra reasoning buys, and why objectives, experience and elapsed time still matter.

Synthetic children’s stories make coherence, elementary reasoning and model internals accessible to small experiments. Which lessons transfer to larger models remains an open question.

Good-looking arrows, digital ink, and reusable canvas interactions led to a workspace for expressing intent—and revising what a model produces.

Together connects data selection, model architecture and GPU systems to lower the cost of building and serving models. Its goal of 5,000 tokens per second reaches beyond faster chat.

The engineering behind reliable clusters, predictable training runs, and evaluations that measure useful coding capabilities.

Gradient’s context-extension work combines positional scaling, distributed training and carefully constructed data. The harder test is whether a model can use the extra information.

How Hugging Face’s M4 team connected pretrained models, built interleaved web data, debugged unstable attention and planned a smaller successor.

How training budgets, numerical precision, memory and communication shape the practical work of running large language models.

AI-assisted scattering calculations show how a well-posed question, a simplifying assumption, and expert verification can turn an unwieldy expression into a research result.

Part 2 of 6: VideoPoet turns video generation into sequence modeling; Tali Dekel explores controllable editing through neural atlases and motion features; Sander Dieleman begins a geometric introduction to diffusion.

Recovering text from embeddings, aligning model representations, and measuring memorization reveal the gap between information stored and information we can use.

A practical guide to managed clusters, reliability and security from SemiAnalysis’s ClusterMAX 3.0 discussion.

A speaker-management contest tested AI-built software against the details of real event operations—and exposed the gap between a demo, an evaluation and an owned service.

Part 2 follows the choices behind useful model improvements: architecture, evaluation, training budgets, context, serving costs—and the research habits that sustain them.

Goal-reaching agents benefit from depth when the learning objective, architecture, and supply of experience scale together.

Shrinking model advantages put more weight on product design, integration and execution. The Unsupervised Learning × Latent Space crossover explores where that work creates value—and where technical limits still matter.