×
Goodfire
AI Interpretability
Last verified August 15, 2026 · Updated daily
What Goodfire is building
Every neural network in existence today is, at its core, a black box. You put data in. You get output out. What happens in between – the billions of mathematical operations that transform a question into an answer, a prompt into a policy decision, a scan into a cancer diagnosis – is functionally invisible, even to the researchers who built it. The standard engineering response to this has been to treat it as an acceptable limitation: run more tests, add more safeguards, fine–tune on better data, and hope the outputs stay within acceptable bounds. That approach, naturally, is breaking down. Goodfire's thesis is that this opacity is not a fundamental property of neural networks but an engineering gap that can be closed. The science they're built on is called mechanistic interpretability: the practice of reverse–engineering neural networks to understand what is actually happening inside them, at the level of individual neurons, circuits, and features. The breakthrough that made this commercially viable is the Sparse Autoencoder (SAE): a technique for decomposing the tangled, overlapping activations of neural network neurons into individual, human–readable concepts. Goodfire's product, Ember, is the first hosted API that gives developers programmatic access to this capability at scale, i.e., developers can directly access and influence an AI model's internal workings to understand, debug, and safely steer its behavior. It ships as an API wrapper around large language models, currently supporting Llama 3 and other open models, that lets engineers inspect which internal features are activating during a model's reasoning, intervene on those features directly, and measure the precise behavioural effect of each intervention. The company has already demonstrated several real applications: cutting hallucinations in a large language model by nearly half by directly suppressing the internal features associated with confabulation; identifying a novel class of Alzheimer's biomarkers by reverse–engineering an epigenetic foundation model built by Prima Mente – the first major scientific finding obtained by interpreting a foundation model's internals rather than just its outputs; and improving safety benchmark scores for enterprise customers including Rakuten, Apollo Research, and Haize Labs. Their Series B funding, announced February 5, 2026 (less than 18 months after founding) will be used to build what they call a 'model design environment': a full platform for understanding, debugging, and intentionally designing AI models from the inside out. The phrase Eric Ho uses to describe the ambition: moving from growing AI 'like a wild tree' to shaping it 'like bonsai.' MIT Technology Review named mechanistic interpretability one of the 10 Breakthrough Technologies of 2026, and Goodfire is the company most directly commercialising that breakthrough.
Why this matters
The history of engineering is a history of fields that were transformed when practitioners stopped treating their medium as a black box and started understanding it from first principles. Steam engineers built more powerful engines for decades before thermodynamics gave them the science to understand why steam behaved the way it did and how to design engines rather than just iterate on them. Geneticists bred crops and mapped hereditary traits for a century before understanding DNA and once they understood DNA, the entire discipline of bioengineering became possible. Eric Ho makes this analogy explicitly and it is not rhetorical flourish: it is the actual structure of what is happening in AI right now. The black–box approach to AI development of training large models on vast data, evaluating outputs, adjusting training, repeat, has produced remarkable results and will continue to do so. But it is fundamentally a trial–and–error process applied to a system no one fully understands. The consequences of that opacity are becoming acute in three specific ways that are driving commercial demand for interpretability tools. First, enterprise deployment risk: 47% of organisations deploying AI in 2025 reported at least one negative consequence from that deployment, and the primary reason cited is unpredictable model behaviour that could not be diagnosed or fixed. When an AI model produces a harmful output, biased decision, or compliance violation, the current tools for diagnosing why it happened are rudimentary – the equivalent of trying to debug software by only looking at the program's final output. Ember gives enterprises a structured way to find the specific internal mechanism responsible and intervene on it directly. Second, regulatory pressure: the EU AI Act, which activates its highest–risk provisions in August 2026, requires that AI systems making consequential decisions — in credit, healthcare, hiring, law enforcement — be auditable. The current state of AI does not meet this requirement. Goodfire's interpretability layer is one of the few technically credible paths to AI systems that can be audited at the mechanistic level, not just evaluated on benchmarks. Third, and most consequentially in the long run: as AI models surpass human understanding in specific scientific domains – protein folding, drug discovery, materials science, genomics – the models themselves contain knowledge that cannot be extracted through normal interfaces. Asking an AI what it knows about Alzheimer's biomarkers only surfaces what it can articulate in language. Reverse–engineering what it has actually learned, for example, the internal representations it uses to make predictions that no human has thought of, extracts knowledge that is genuinely novel. Goodfire has done this once already, with the Alzheimer's biomarker discovery. If the technique generalises, the downstream implications for medicine and science are extraordinary.
Open roles at Goodfire
5 positions we're tracking. Roles are re-checked daily and removed when filled.
Interpretability Researchers
First seen 4 months ago
ML Engineers
First seen 4 months ago
Research Engineers
First seen 4 months ago
Applied Scientists
First seen 4 months ago
Enterprise Partnership leads
First seen 4 months ago
Hiring outlook
Raised $150M less than a year after Series A, scaling research team
Working at Goodfire
Goodfire is AI Interpretability, founded in and now people. What this means for you: opportunity to shape your role based on the company stage.
San Francisco is where most Goodfire positions are located.
Frequently asked questions
How many jobs does Goodfire have open?
We're tracking 5 active openings at Goodfire (verified February 2025).
Does Goodfire hire remotely?
Currently, Goodfire only has in-office roles in San Francisco.
What roles is Goodfire hiring for?
Goodfire is hiring across Other, Engineering, Data. The most recent opening is Interpretability Researchers.
How do I apply for a job at Goodfire?
Click through to apply, or see our detailed guide on landing a job at Goodfire.
Where is Goodfire based?
Goodfire is headquartered in San Francisco.
Get Goodfire roles before they're posted
Stay ahead of the AI job market. We track Goodfire and similar companies daily.
Get early access →