Daily Brief

Full episode: https://www.youtube.com/watch?v=QbdbAhaJoCQ Me on twitter: https://x.com/dwarkesh_sp

Run terminal bench on Opus and on Haiku and Opus scores about three times better at a tenth of the cost, even though Haiku is far cheaper per token. Alex Atallah's point is that a small model pushed outside its training distribution thrashes, calling tools in loops until it costs more than the expensive model ever would. That inverts the obvious version of model routing, where you send each task to whichever model benchmarks best on it. Walden Yan calls that approach fragile for exactly the reason agents make it worse: a session starts as a question about a codebase, becomes a feature request,

Quantize a single number in a model and it gets 20% dumber. That finding, from the super weights paper, is why Daniel Han's claim is less absurd than it sounds: GLM 5.2 goes from 1.5 terabytes to 250 GB, 86% smaller, without being 86% dumber. Layers are wildly unequal. The first and last carry enormous weight, the middle ones barely matter, and a model trained on 30 trillion tokens never saturates its parameters, so many sit near zero and can simply be set there. Choosing which layers stay in high precision is a combinatorial search, not a setting. NVFP4 is the format the NVIDIA side leans on

When Fable was pulled back and access to frontier systems stopped looking guaranteed, Lucas Atkins watched enterprises move to Chinese open models, not because they scored better but because availability could be counted on. That is his working definition of trust, and he separates it hard from safety: an open model is a directory of files you can inspect, running on code you can read, while the same claim about a closed API is unverifiable by construction. Arcee's response was to reorient the whole company and pretrain a 400 billion parameter model in six months, which he says plenty of peopl

At our inaugural YCML at Startup School, YC Partner Ankit Gupta speaks with Jitesh Jain about building video agents that can adapt their reasoning to videos of different lengths. Current agents often struggle with long, open-ended video questions because temporal grounding is unreliable and training data is expensive. SAGE combines visual tools with transcripts and web search, then uses synthetic question-answer data, tool trajectories, and reinforcement learning to teach the model when each source of information is useful. As videos become longer, the agent takes more reasoning steps and imp

At our inaugural YCML at Startup School, YC Partner Ankit Gupta speaks with MIT PhD candidate Jovana Kondic about ChartNet, an open-source data generation pipeline and million-scale dataset for chart understanding. Charts require models to combine visual recognition, text understanding, and numerical reasoning. ChartNet generates diverse examples by translating charts into plotting code, augmenting that code, and rendering new images with corresponding tables, summaries, and reasoning traces. Training on ChartNet improved open-source models across a range of chart tasks and transferred to rea

At our inaugural YCML at Startup School, YC Partner Ankit Gupta speaks with Mark Žnidar about predictive models that work directly with relational databases without flattening their tables or manually engineering features. The method represents a relational database as a graph, samples the relevant neighborhood around an entity, and uses specialized attention mechanisms to capture columns, features, and relationships across tables. This preserves information that is often lost in traditional tabular machine learning pipelines. Despite having only 22 million parameters, the model outperformed

At our inaugural YCML at Startup School, YC Partner Ankit Gupta speaks with Surgan Jandial about evaluating how web agents plan, rather than only measuring whether they complete a task. The work breaks planning into individual skills such as temporal ordering, future-state prediction, action selection, and error correction. It then repurposes existing datasets to create inexpensive synthetic tests that measure each skill independently. These scores correlate with agents' performance on complete web tasks, providing a cheaper and more interpretable way to identify their weaknesses before runni