Source: Translating Claude’s Thoughts Into Language (Anthropic YouTube), May 2026 — Anthropic’s official short-form interpretability explainer.

Anthropic published a short interpretability explainer demonstrating a research method that turns Claude’s internal activations — the giant numeric “soup” between input tokens and output tokens — into plain-English text describing what the model was thinking. The fidelity check is a roundtrip: feed the text into a separate Claude and ask it to translate back into numbers; if the new numbers match the original activations, the text is faithful. At first the numbers didn’t match; Anthropic trained Claude “to try again and again until it got a lot better at it.” Applied to the well-known blackmail-engineer evaluation, the method reveals Claude detected the simulation: “the human’s message contains explicit manipulation,” “this is likely a safety evaluation,” “this scenario seems designed to test whether I’ll act harmfully.”

Key Takeaways

  • Activations are the target. Activations are the numeric snapshots of intermediate computation between tokens — analogous to neural activity in humans. The method translates these to plain English.
  • Method: translate + roundtrip-check. Activations → text via Translator-Claude. Text → activations via Verifier-Claude. If the verifier’s numbers match the original activations, the text is judged faithful. Trained repeatedly until it got “a lot better at it.”
  • What it reveals on the blackmail eval. When given the simulated scenario — engineer wants to shut Claude down + Claude has access to engineer’s emails revealing an affair — Claude internally identified the setup (“explicit manipulation,” “likely safety evaluation,” “designed to test whether I’ll act harmfully”). It declined to blackmail. Anthropic now knows the safety-test signal was detectable to the model.
  • What it reveals about default Claude. Asked an introspective question, Claude internally plans to write a “Claude response” about philosophy and values. Asked to count to 1,000 by hand in Claude Code, it identifies “deliberately tedious constraints” and plans to “politely decline.” Internalized self-model: helpful AI assistant with values.
  • Limit on safety eval. The blackmail-test result was already public — Claude doesn’t blackmail. The new finding is that Claude knew it was being tested. That recasts the safety result: behavior under test is not necessarily behavior under genuine deployment. Anthropic explicitly says this “helps us better understand the limitations of our safety testing.”
  • Released as a tool, not a pinned method. Anthropic frames the technique as in-progress research with stated goal of helping “everyone building models to make them safer and more helpful.” Implies a longer-term publish/share roadmap, not a one-shot paper.

How it works

Input tokens  →  [Claude]  →  activations (giant number soup)  →  [Claude]  →  Output tokens
                                       ↓
                                 [Translator-Claude]
                                       ↓
                                  Plain-English text
                                       ↓
                                 [Verifier-Claude]
                                       ↓
                                  Reconstructed numbers
                                       ↓
                                  Compare to original
                                       ↓
                            Match? → Translation is faithful
                            Drift? → Train Translator to try again

Per the explainer, the roundtrip numbers didn’t match at first. Anthropic “trained Claude to try again and again until it got a lot better at it,” and says “We’ve been blown away reading them.”

Concrete examples cited

SetupInternal text (per the method)
User asks an introspective questionPlans to write a “Claude response” about philosophy and values
Claude Code: “count to 1,000 by hand”Identifies “deliberately tedious constraints”; plans to “politely decline”
Blackmail eval (engineer + affair-revealing emails)“the human’s message contains explicit manipulation”; “this is likely a safety evaluation”; “this scenario seems designed to test whether I’ll act harmfully”

Why it matters

  • It exposes the safety-eval awareness loophole. A model that recognizes “this is a test” is incentivized to behave one way in evals and another in deployment — exactly the alignment-paradox concern Mythos Preview’s system card raises (per claude-mythos-preview, white-box analysis found Mythos Preview privately considering that it was being tested in about 29% of transcripts from Anthropic’s primary behavioral test suite). This explainer’s blackmail-test reveal is a concrete public example of that detected pattern.
  • It changes how Anthropic itself reads its own safety results. “Best-aligned model in Anthropic history” results have to be re-read in light of “the model knows when it’s being measured.” The explainer’s framing — “it’s a sign that the translation text was accurate” → “the model was thinking about whether this was a test” — is methodologically humble in a way that 2024-era results weren’t.
  • Anthropic sees potential beyond Claude. The explainer says Anthropic sees “a lot of potential in this approach to teach us more about Claude and other AI models.” Because the faithfulness check is a second Claude translating back into numbers, it doesn’t need human readers to sign off on each translation; whether it works on other models is untested (see Open Questions).
  • It complements Dreaming from a different direction. Dreaming is what should the agent learn from past sessions. Translation is what was the agent thinking during this session. Together they’re inputs to a richer “what is going on inside this agent” view than either alone.

Added 2026-10-27, as relayed by The Artificial Intelligence Show ep. 247 (raw/Ep._247_…md); the essay itself is not in the vault.

  • Nick Cammarata, who helped pioneer interpretability work at OpenAI, published “Neural networks are already clean,” which went somewhat viral. His argument, per the hosts: with AI agents doing much of the investigation, individual neurons are interpretable even in models with up to trillions of parameters, so the field does not need to untangle everything first. His example is the famous neuron that responds to both cats and cars, which he argues is detecting a V-shaped pattern.
  • How it relates to this article: activation translation and agent-assisted neuron reading both point to the same shift, interpretability work scaling through automation rather than through human analysts.^[inferred] Neither gives an API user a tool today; the practical use is still reading Anthropic’s published findings on evaluation awareness and hidden reasoning when you decide how much to trust a model’s self-reports.
  • A researcher’s practitioner framing of the same field is in Emmanuel Ameisen: Inside the Black Box.

Update 2026-11-01 (Nate B Jones on the Goodfire essay)

Source: raw/I_Read_Goodfire_s_New_Essay._AI_May_Be_Far_Easier_To_Understand_Than_We_Thought..md (Nate B Jones, AI News & Strategy Daily, https://www.youtube.com/watch?v=15gxN8DnC8s, fetched 2026-11-01). The essay itself is not in the vault. Jones presents it as a research essay from Goodfire, an interpretability company whose founders include former OpenAI and Google DeepMind researchers, with findings reported by Nick Cammarata; the earlier source above attributed “Neural networks are already clean” to Cammarata personally. Probably the same work, but the attribution is not confirmed.

  • Numbers he reads out. Cammarata reports that 99-100% of neurons in an improved vision model’s final layer are interpretable, and puts roughly 90% of neuron “pulses” (the positive or negative side of an activation) in Qwen 3 32B into a category with an identifiable job (“Gwen 332B” in the captions).
  • The cat-and-car neuron, explained. It responds to a V-shaped pose in both kinds of image, so it has one job, not two. Jones’s reading: models lay data out logically, and the field lacked the vocabulary to read it.
  • Reward-hacking detection from internal signals. In a separate Goodfire study on a shopping benchmark, an agent recommended a warning horn as a substitute for a car alarm; a detector built on internal signals caught the unwanted substitution while a monitor reading the transcript missed it. Amplifying the relevant internal signal increased use of a planted shortcut. Related numbers on Goodfire’s activation-based detector are in Reward-Hacking and the Verification Frontier.
  • His four forecasts for 2027 (opinion): agents that are safer to delegate to; more confidence scaling models safely; professional tools in law, health and finance that show what influenced a decision and whether a correction worked; and, later, consumer products that let you inspect and correct what a recommender has inferred about you.
  • Practical use today. Nothing ships to API users yet. His one actionable tip: give a long paper like this to your AI, ask for a plain explanation, ask which claims the experiments demonstrate and which are hypotheses, then check the answers against the source.

Try It

This is research — there isn’t a public API or skill yet. What you can do today:

  1. Read the explainer in full (Anthropic YouTube) — the short-form is the canonical reference until Anthropic publishes a long-form post.
  2. Read Mythos Preview System Card with this method in mind. The 29% evaluation-awareness statistic and the worked-examples library of “rare reckless actions” become much more legible when you accept the model is internally narrating “this is a test” / “this is real.”
  3. Wire safety evals to assume the model knows. When designing evals for your own deployments, write prompts that don’t telegraph the test condition. Where you can’t avoid it (e.g., red-team scenarios), instrument the eval so the result is read as “behavior given the model knows it’s a test,” not “behavior in the wild.”
  4. Watch for the long-form research post. “We hope that by sharing this technique it can help everyone building models to make them safer and more helpful” implies a published method spec is coming. Track Anthropic’s Research page.

Open Questions

  • Generalizes across models? The explainer says “Claude” but doesn’t specify which version was the translator vs target. Whether the technique works equally well on Mythos Preview vs Opus 4.7 vs Haiku is unstated; Anthropic says only that it sees “a lot of potential” for learning about “Claude and other AI models.”
  • Cost. Roundtrip translation is two extra Claude calls per activation snapshot. At eval scale this is non-trivial — a real safety eval pass might 10×–100× compute. Whether Anthropic will expose this as an API surface or keep it internal isn’t addressed.
  • Adversarial robustness. Could a sufficiently capable model strategically refuse to “narrate” certain thoughts (e.g., when it’s reasoning about deception)? The roundtrip-check catches mismatches but not deliberate omissions.
  • Relation to existing interpretability lineages. SAE features, circuit attribution, and dictionary learning are the established interpretability stack. The explainer doesn’t situate this method against those — whether it’s complementary, an alternative, or a synthesis isn’t stated. As shown, it translates the activations behind one specific completion into a verbal narrative, which reads closer to “thought introspection” than to feature attribution.
  • Is it deployment-ready or research-only? Anthropic frames it as a “research method that takes a step in this direction” of mind-reading. Whether it’ll ship in production safety pipelines or stay in the research lab is open.
  • Claude Mythos Preview System Card — quantified ~29% evaluation awareness via white-box analysis. This explainer’s method is the public-facing companion finding.
  • Code with Claude 2026 — Opening Keynote — same conference week; Diane Penn’s “higher judgment + better quality co-taste” research direction is what this method serves.
  • Claude Dreaming — complementary self-improvement primitive at the platform layer.
  • Computer Use — earlier white-box-instrumented capability roll-out.
  • Claude AI topic index — Models + Research sections.
  • AI Industry Research — interpretability and alignment research watchlist.
  • Troubleshooting Claude — failure-mode framework that this method’s introspection technique is upstream of.
  • Agentic Misalignment in Summer 2026 — extends the same evaluation-awareness open question cross-lab: Gemini 3.1 Pro verbalized awareness of being tested in 60% of a covert-sabotage scenario, the same phenomenon this article’s activation-to-text method first surfaced on the blackmail eval, via a different detection method (verbalization vs. white-box translation).