As of November 3, 2026, Level counts 102 live listings that mention interpretability, once Anthropic's boilerplate is removed. Most sit at Builds AI and many ask for research depth. A new MIT CSAIL post on reading model weights has brought fresh attention to the field.
What happened
MIT CSAIL researchers published a blog post on weight-space programming. It is dated November 2, 2026. The authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Their claim is that a language model can read and write neural network weights directly. The post reached the top of Hacker News, where it passed 2,500 points.
The authors gave the job to GPT-6 Astra. They report that it could tell what a network was trained to do from its raw weights. The main test was a small open model, Qwen3.5-0.8B, which often got stuck in reasoning loops.
Astra traced those loops to one direction written by two layers, then subtracted that direction from the weights. The edit changed 0.016% of the weights, with no retraining. The authors report that accuracy on the MATH-500 benchmark started at 7.9%. With 50 forward passes, it reached 48.1%. The share of runs hitting the token limit fell to 33.1%.
The result has a clear limit. Astra could not improve Gemma 3 270M, a model that rarely loops. The authors conclude that an edit can remove an obstacle hiding an ability, but cannot add one.
The authors say the work "marks the beginning of a dramatic shift" in their field. That is their claim. It comes from a blog post, and Level did not reproduce the results. They have published the run logs for readers who want to check.
Who is hiring for AI interpretability jobs?
The word appears in 776 live listings on Level, but most come from boilerplate. Anthropic accounts for 677 of them, because its standard text cites prior research on interpretability. We removed those unless the title itself names the field.
That leaves the 102 listings in our main count. They span 58 companies. AstraZeneca and LatchBio lead with 8 listings each. Amazon follows with 6. OpenAI and UPS post 5 each. Anthropic keeps 3, all with the word in the title.
The word covers two kinds of work. The first is research into how models work inside, sometimes called mechanistic interpretability. The second is explaining a model's outputs to a business, a regulator or a clinician.
Only 5 listings put the word in the title, and all come from Anthropic or OpenAI. That is the narrow core of the field. The wider group is mostly research, data and engineering work, where the word can be one topic among several.
Level's AI levels add a second filter. About 75% of these listings sit at Builds AI. Another 12 sit at Works on AI. Research is the largest category, with 33 listings you can compare with other AI research jobs.
What do AI interpretability jobs pay?
Pay data is partial, so read it as a guide. Of these listings, 46 publish a yearly salary in US dollars. Their median lower bound is $198,000+. Their median upper bound is $288,000+.
The 5 titled roles pay far more, though that sample is very small. Their median lower bound is $350,000+. Their median upper bound is $500,000+. Anthropic's Research Scientist, Interpretability listing has the highest posted top, at $850,000.
All 5 titled roles are based in San Francisco. Only 1 of them is flagged as remote, an OpenAI researcher post. In the wider group, 27% of listings are flagged as remote, and remote AI jobs has the full list.
For a wider pay comparison, see our guide to research scientist salaries.
What skills do AI interpretability roles ask for?
Python is named in 69% of these listings. PyTorch appears in 27%. A PhD is mentioned in 46%. A mention does not always mean a requirement.
OpenAI's Researcher, Interpretability listing shows this. It asks for a PhD or research experience in computer science, machine learning or a related field. It also asks for experience in AI safety or mechanistic interpretability.
Anthropic's Research Engineer, Interpretability listing offers a different entry point. It says candidates need not have worked on interpretability before, because the infrastructure expertise resembles general large model work. The duties include activation extraction and steering vector application.
That last duty is close to the MIT CSAIL experiment, since both act on a direction inside a model. One difference matters: Astra changed stored weights, while a steering vector is applied during a forward pass.
Senior and lead roles make up 59% of the group. Only 6 listings are labeled junior or intern.
Open roles right now
This list uses Level search for the word, filtered to Builds AI. It held 12 listings at the time of reading. Of those, 2 are flagged as remote: OpenAI's Researcher, Interpretability and Faculty's Senior Research Scientist, AI Safety.
