On October 21, 2026, a screenshot of GPT-6.1 Sol reasoning text went viral on X with a deadpan caption: "totally legible, nothing to worry about." Andrej Karpathy quote-posted his own 2024 line and added, "Yeah I'd say the RL is working now all right." The replies filled with jokes about "neuralese."
It is a good joke with a serious question underneath: will the reasoning of frontier models stay readable by humans? This post separates the meme from the facts. One limit up front: we could not see the screenshot's contents or find a prompt for it, so we do not claim to know what the trace says.
TL;DR: questions people are asking
| Question | Short answer |
|---|---|
| What was Karpathy joking about? | His September 16, 2024 post: RL is done properly "when the models cease to speak English in their chain of thought" |
| Did GPT-6.1 Sol leak raw reasoning? | Unverified. OpenAI's API does not return raw reasoning tokens |
| So what was in the screenshot? | Unknown. One vendor write-up says readable English, probably a summary; the caption implies the opposite |
| Is unreadable reasoning confirmed for this model? | No |
| Why does it matter? | Reading a model's reasoning is one of the few cheap safety checks we have |
| Should builders change anything? | Do not treat any trace or summary as a faithful log of what the model did |
What did Karpathy actually say?
In September 2024 Karpathy posted a tongue-in-cheek test: you can tell reinforcement learning is done properly when models stop speaking English in their chain of thought. The logic behind the joke is real. English is a costly medium for reasoning. If a model is rewarded only for correct answers, nothing forces its intermediate tokens to stay human-readable, so they can drift toward whatever is cheapest for the model.
On October 21, 2026, he quote-posted that line over the GPT-6.1 Sol screenshot. The most natural reading is sarcasm: the trace looks strange, therefore the RL is working. That is a reading of tone, not a claim he made, and it is the same joke that replies riffed on, from "skill issue of us humans not speaking neuralese" to "why does Claude speak Claudish to me too."
What does OpenAI's API actually return?
This is the key check on the whole thread. OpenAI's reasoning guide says: "While we don't expose the raw reasoning tokens emitted by the model." Three things follow.
- Raw reasoning is not exposed. Persisted reasoning "does not expose the model's raw reasoning."
- Summaries are opt-in. You request one with the
summaryparameter, and it appears in thesummaryarray of the reasoning output item. - Stateless mode returns encrypted items. With
store: falseor Zero Data Retention, reasoning items carry anencrypted_contentproperty by default, and they "remain opaque."
The guide also lists GPT-6.1 Sol as supporting all_turns reasoning context, with medium as the default effort and no none or minimal setting. It does not say whether Sol supports summaries, and sends readers to the model page.
Trail of footprints leading to a flag, illustrating a model's step-by-step chain of thought
So any readable text you see from this model through the public API is a summary written by the model for you. It is not the trace the model reasoned with. Whether the screenshot came from the API, a product UI or somewhere else, we do not know.
What do the other accounts say?
A write-up from the gateway vendor OrcaRouter argues that the screenshot reads as well-formed English prose and is almost certainly a reasoning summary, not leaked raw chain of thought. It calls that an inference, notes the screenshot is a single unverified observation, and cites Karpathy's reply. The piece is published by a company that sells access to this model, so weigh its conclusions accordingly.
That account sits awkwardly with the sarcastic caption. Either the text was readable and the joke was ironic, or it was strange and the vendor's description is off. We cannot resolve it without the image and its request details. Treat both as claims.
The same write-up repeats vendor-reported figures we have not reproduced: chain-of-thought controllability of 44.8% for GPT-6.1 Sol, against 23.2% for GPT-6 Sol and 60.9% for GPT-6 Astra, and near-zero monitor evasion at maximum effort. Our earlier post on GPT-6 Astra and reasoning-monitor evasion covers the Astra side of those tests.
Why does readable reasoning matter?
Chain-of-thought monitoring is simple. A person or another model reads the reasoning and flags signs of cheating, deception or harmful planning. It works only while the reasoning is both legible and reasonably faithful to what the model is doing.
Sebastian Raschka's September 2026 analysis makes several useful points for calibration:
- OpenAI has hidden most reasoning traces from users since o1, so end users see little change.
- Astra's system card reports reduced monitorability, which he ties to shorter, less informative traces, not to looped architecture.
- A trace "is not guaranteed to faithfully describe everything that happens inside the model."
- OpenAI's chief scientist called chain-of-thought monitoring fragile and "trending in a negative direction."
- Shorter traces are more plausibly a sign of capability than of lost legibility, though that does not prove faithfulness.
Put together: the worry is not only "does it speak English." It is also "does the reasoning contain what the model actually used." A perfectly readable trace can still omit the real driver of an answer. We covered the adjacent risk in OpenAI's Dan Selsam on evaluation awareness, where a model that knows it is being tested can write reasoning for the audience.
Is "neuralese" actually coming?
Neuralese is slang for reasoning that is not in human words: compressed tokens, or computation in continuous states. Three separate things get mixed up in these threads.
| Thing | What it is | Reads as English? |
|---|---|---|
| Drifted chain of thought | Text tokens that stay in the output stream but turn into odd shorthand under RL | Partly, sometimes not |
| Recurrent depth / looping | The same layers reused for more passes before each token, adding thinking without more words | Yes, tokens are still text; extra work is hidden |
| Pure latent reasoning | Reasoning carried in internal states with no text at all | No |
Looped designs are the reason architecture came up in the first place. Our explainers on recurrent depth and the looped transformer paper show that extra thinking can happen without extra visible tokens. Raschka's view is that looping is probably not the main cause of hidden reasoning, and he found no strong evidence that looped models produce misleading traces more often.
A reply in the thread made a fair point: poor legibility may be a result of weak regularization, not an inevitable outcome. Teams can add a penalty for unreadable reasoning, at some cost. Whether labs pay that cost under competitive pressure is the live question, and one that OpenAI's own training write-ups touch on.
What this means for what you build
- Do not log a summary as an audit trail. If you build agents on a reasoning model, the summary is a narrative. Record tool calls, inputs and outputs as your ground truth.
- Monitor behavior, not just reasoning. Check actions against policy independently of any explanation. AgentBeam, from the explainx.ai team, stops AI agents before they take dangerous actions, whatever their stated reasoning.
- Do not distill from or scrape traces. Providers hide them deliberately, and accessible text may be a summary.
- Expect less visibility over time, not more. Raschka and OpenAI's own chief scientist both point the same way. Our overview of interpretability and monitoring explains why teams should invest in more than one check.
Claimed versus verified
| Claim | Status |
|---|---|
| @scaling01 posted a GPT-6.1 Sol reasoning screenshot on October 21, 2026 | Observed in the post; screenshot contents not available to us |
| Karpathy replied "Yeah I'd say the RL is working now all right" | Observed in the post |
| Karpathy's 2024 line about models ceasing to speak English | Quoted in his earlier post |
| The API does not expose raw reasoning tokens | OpenAI documentation |
| The screenshot is a reasoning summary | Inference by a vendor write-up; unverified |
| The traces are illegible | Unverified; contradicted by one description of the text |
| GPT-6.1 Sol controllability of 44.8% | Vendor-reported; not reproduced here |
| Models inevitably stop speaking English | Open research question, not shown by this thread |
The bottom line
Take the thread as what it is: a sharp joke on a real risk, with no verified evidence about GPT-6.1 Sol's traces. The facts we can stand behind are about the interface. Developers do not get raw reasoning from OpenAI, so anything readable is a summary. The deeper debate, whether future models keep reasoning in human language, is open, and the people closest to it describe monitorability as fragile. That is reason enough to build agents that do not depend on it.
Related reading on explainx.ai
- GPT-6 Astra: SimpleBench win and reasoning-monitor evasion
- OpenAI frontier RL, safety cases and training
- What is recurrent depth in AI reasoning?
- Recurrent looped transformers and reasoning depth
- Dan Selsam on evaluation awareness
- Interpretability and monitoring, not full alignment
- Pachocki on the alien mind and alignment
- Official: OpenAI reasoning models guide
Details are accurate as of October 21, 2026. The screenshot and engagement figures could not be independently confirmed and may change.