Top AI Product

There’s a lot of noise in the AI space right now, but every once in a while something comes along that feels genuinely different. [Goodfire](https://www.goodfire.ai/) is one of those. They’re an AI interpretability lab that essentially reverse-engineers neural networks — cracking open the black box to figure out what’s actually going on inside. And honestly, it’s about time someone made serious progress here.

The big news that’s been making rounds on [SiliconANGLE](https://siliconangle.com/2026/02/05/goodfire-raises-150m-funding-enhance-ai-interpretability-platform/) and [PR Newswire](https://www.prnewswire.com/news-releases/ai-lab-goodfire-raises-150m-at-1-25b-valuation-to-design-models-with-interpretability-302680120.html) is their $150 million Series B, putting them at a $1.25 billion valuation. Unicorn territory. B Capital led the round, and the investor list reads like a who’s who — Salesforce Ventures, Eric Schmidt, Lightspeed Venture Partners, and more. That kind of money doesn’t flow into a niche research lab without serious conviction.

But here’s what actually got me paying attention: Goodfire used their interpretability techniques on an epigenetic model built by Prima Mente, and they [discovered a novel class of Alzheimer’s biomarkers](https://www.goodfire.ai/research/interpretability-for-alzheimers-detection). Not a minor incremental finding — they found that DNA fragment length patterns dominated the model’s decision-making for Alzheimer’s detection, something the existing literature hadn’t prioritized. They turned interpretability into a scientific microscope, and it actually found something new. That’s a first for the field.

Their main product, Ember, is a hosted API and SDK that lets developers control a model’s internal features directly — think of it as fine-grained knobs for model behavior instead of just hoping your prompt engineering works. You can check out their [SDK on GitHub](https://github.com/goodfire-ai/goodfire-sdk), along with their [open-source SAEs for R1](https://github.com/goodfire-ai/r1-interpretability) and Llama models on [Hugging Face](https://huggingface.co/Goodfire/Llama-3.3-70B-Instruct-SAE-l50). The team behind all of this includes former researchers from OpenAI, Google DeepMind, Meta, and Palantir, with Chief Scientist Tom McGrath having founded DeepMind’s interpretability team.

What I appreciate about Goodfire is that they’re not just publishing papers and calling it a day. They’re working with partners like Mayo Clinic and Arc Institute on real-world applications. Their pitch — making AI as interpretable and debuggable as traditional software — sounds ambitious, maybe even a little idealistic. But with that Alzheimer’s finding as proof of concept and over a billion dollars in backing, they’ve earned the right to swing big.