Topic
A research field that reverse-engineers neural networks' internal computations—circuits, features, representations—to explain how models produce outputs.
Independent researcher Nick Cammarata, who says he left OpenAI in 2023, published a report arguing almost the entire AI network is interpretable. Business Insider calls the findings still theoretical, so safety teams have a claim to test before they have an audit method.
Reality
Goodfire's Nick Cammarata says the vast majority of neurons in models up to trillions of parameters are legible, reversing an explanation he helped shape. Google DeepMind's Neel Nanda warns against overhyping it, and the model-level numbers are preliminary.
Reality
Goodfire researcher says 99-100% of InceptionV1's last-layer neurons are easily interpretable after architecture changes. The essay argues the superposition evidence was misread, so networks can be studied directly.
Reality
Anthropic researchers' J++ Lens puts the right hidden intermediate variable in its top 10 readouts 55.2% of the time, against 35.7% for the J-Lens. The authors pitch it for monitoring eval awareness, but the score comes from prompts with known answers.
Reality
Natural language autoencoders caught a steganographic side task given in the prompt but largely missed it once fine-tuned in, a LessWrong post reports. White-box monitoring built on their readouts is therefore weakest on learned behavior, the case a monitor most needs to catch.
Reality
Probes on Qwen3.5-27B reading other models' text came within 0.004 AUROC, on average, of probing authors up to 397B directly, a LessWrong post reports. Every tested pair was open-weight, so auditors who apply the method to closed models get the reader's view of the text and cannot measure that gap.
Reality
The Oxford team told two agents driven by one model to count cards, and the agents worked out the rest, putting bet signals inside chatter about the dealer. Catching them needed activations from inside both models at once.
Reality
A new 27-family benchmark scores how well a tool reads a model's intermediate variables during a forward pass. Its authors say it does not fully rule out tools that infer the answer from the prompt.
Reality
A replication of Wurgaft et al.'s manifold steering on Anthropic's pre-trained CLT features reproduces the cyclical weekday transition in Gemma-2-2B. The author attributes the weaker push to the MLP sublayers holding only part of the signal.
Reality
Almost all frontier protein models ship with open weights, out of reach of API refusals and unlearning. A LessWrong essay puts the enforcement point at DNA synthesis screening, where AI-designed novel sequences already defeat the sequence matching that providers use.
Reality
Anthropic screened Claude for one property, verbalizability, and found two more in the same representations. Neel Nanda reproduced the structure in open weights, and the outside commentators Anthropic invited disagree about what it is.
Reality
Wired reports that Jacob Coxon's September 8 exit post from Anthropic was followed by a senior colleague putting internal odds of human extinction at 10 percent, and by Dario Amodei's weekend essay arguing for slower releases.
Reality
Base Labs has signed Hugging Face and Goodfire to publish methods for training, evaluating and monitoring open models, with the check running where the activations are. The partners have not published an evaluation suite yet.
Reality
The offset in a J-lens readout mostly tracks how often a token appears, and scaling it by variance, after plain subtraction failed, lifted hidden-word elicitation to 0.805 from 0.665 on Gemma-2-9B-it. The paired test over 20 words gives p of about 0.19.
Reality
A LessWrong post shows the Oja rule falling out of tied sparse-autoencoder descent once two terms are dropped, then reports a language-model test where an initialization change moved the backprop baseline more than the Hebbian gap it was meant to explain.
Reality
In a single-author LessWrong experiment, accuracy on held-out countries went from 20.9% to 97.7% with the lesion active in every forward pass, and the recovery survived refitting the lens to the adapted model.
Reality
A one-dimensional probe finds structural impossibility in the hidden state of instruction-tuned models from 1.7B to 70B parameters. The trained refusal pathway that guardrail work tunes reads an axis about 85 degrees away from it.
Reality
Perturbation probing needs the weights, which puts the finding on teams running open-source checkpoints rather than on API tenants. The same 50-neuron toolkit that breaks a refusal template also repairs other behavior.
Reality
A STAT opinion piece argues the measurable shift in medical AI is not benchmark wins over physicians but the removal of the safety text that used to route people to care.
Reality
Knowledge editing is sold as a cheap substitute for retraining. The authors argue it treats models as filing cabinets when knowledge is a dependency graph.
Reality