Yoko Li

Most of us in engineering have accepted that AI models are black boxes. We optimize prompts, fine-tune with RLHF, add safety filters on top — but we rarely ask what's actually happening inside the model. It works well enough, so we move on. Goodfire is betting that "well enough" isn't good enough. Founded by Tom McGrath, who built the interpretability team at DeepMind, and Nick Cammarata from OpenAI's interpretability research, they're building tools to decompose what a model learns into individual understandable components — and then selectively keep or discard them. They used this to cut hallucinations by 50%, not through external filtering but by editing the model's internal mechanisms directly. The result that caught my attention: they reverse-engineered a biological AI model trained on DNA sequences to understand how it detects Alzheimer's from blood. The scientific community expected the model relied on DNA methylation. It didn't. It had found an entirely different class of biomarkers — fragment length patterns in cell-free DNA — that researchers weren't even looking for. They're calling it the first major scientific discovery made by reverse-engineering a foundation model. Validated with Mayo Clinic and Arc Institute. They just raised $150M at a $1.25B valuation. Their product Ember gives engineers direct control over model internals, with Microsoft and Mayo Clinic among their clients. Interpretability went from academic curiosity to unicorn territory in under a year. I find myself in two minds about this. On one hand, understanding what's inside the black box feels obviously more valuable than just making it bigger. On the other, their own blog acknowledges these techniques currently work for "relatively simple behaviors" — scaling to complex compositional reasoning is unsolved. The question is whether interpretability can keep pace with model complexity, or if we're opening boxes that keep getting deeper. https://lnkd.in/d7iH_43B

2

Head of Design Ryo Lu helped transform Cursor from a feature-layer on top of VS Code into one of the world's leading AI code editors. He joins YC's Aaron Epstein on Design Review to talk about the path that brought him to Cursor, how rapid prototyping reshaped the core product, and how he's on a mission to break down the barriers that once separated designers and coders. https://lnkd.in/gXPCpsZ3

508

14 Comments

Every engineering leader experimenting with AI eventually runs into the same problem: You can’t improve what you can’t measure. To help, we’re launching the next generation of Jellyfish AI Impact featuring: - Autonomous Agent Insights - Developer Insights - Report Builder - Deeper AI Impact Telemetry in Metrics Explorer This update brings together the full picture across AI-DLC including usage patterns, delivery outcomes, developer experience, and strategic insights so engineering leaders can finally see what’s working, why, and what to do next. Learn more about our newest update to AI Impact here: https://lnkd.in/etSVpyK5

107

1 Comment

💭 Y Combinator CEO says YC startups launch with 95% of their code written by AI 😮. Harness Field CTO Martin R. brings valuable insight to the conversation with LeadDev around this trend and its impact on software developers. "Having a human in the loop will continue to be vital. The future of coding will involve a combination of human creativity and AI-driven efficiency." Read more on how AI is reshaping engineering roles ⬇️. https://lnkd.in/gnRmrWxS

2