Frontier AI developers will adopt latent reasoning that bypasses natural-language chain of thought before AI research is automated.
Assessment
The evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.
This is a forecast about the order of two future events: whether frontier developers will move their models' reasoning out of natural-language chains of thought and into latent computation before AI systems automate AI research. Read as the discourse uses it, "bypasses" means that latent computation becomes the main channel for extended reasoning, so that the visible chain no longer carries most of what the model works through. On that reading the evidence leans modestly against the claim. The forecast is uncertain because the trend is moving quickly.
The case for the claim is that the move has already begun. In September 2026 OpenAI's Astra model was reported to use looped "recurrent depth", which does computation between chain-of-thought tokens. Anthropic and Google DeepMind were reported to be discussing the technique. Since recurrent depth increases capacity for reasoning without a verbalized chain, it is a capability gain that competition could push further. Safety researchers such as Ryan Greenblatt and Buck Shlegeris have warned that the natural next step is to scale it until reasoning is almost entirely latent.
The case against is that Astra's use of the technique is reportedly limited. Its chain of thought is still expected to be legible, and OpenAI publicly rejected a shift to "neuralese". Reasoning models are bootstrapped from human text, and models that reason entirely in latent space are not yet known to be trainable competitively at frontier scale. Leading developers now say automation of AI research is close, with Anthropic's late-2026 IPO filing warning that it nears a recursive self-improvement threshold. On short timelines, the first systems to automate AI research are likely close descendants of today's architectures.
Under a weaker reading, on which any latent computation outside the chain counts as adoption, Astra arguably already satisfies the claim. That reading is not the one under dispute. The claim also takes for granted that AI research will come to be automated. Several signals would settle the matter: whether the successors to Astra and its competitors scale up recurrence, whether a lab demonstrates competitive wholly latent reasoning, and how soon AI research is in fact automated.
Full reasoning: the evidence and decisions behind this verdict
Own model before sources: the claim needs two things. First, a technical path to latent reasoning that is competitive with chain-of-thought reasoning. Second, a decision to deploy it despite labs' monitorability commitments. Both must happen before automation, and many forecasters date automation to the late 2020s. A short window and an unproven technique favor continuity. Strong capability incentives and evidence that recurrence is already in use cut the other way.
Evidence read: TechCrunch, 2 September 2026 (techcrunch.com/2026/09/02/openais-new-reasoning-technique-alarms-ai-safety-experts/). It reports, citing The Information, that Astra uses recurrent depth ("opaque recurrence") and that its use is limited. The chain of thought is still expected to be legible, and OpenAI pushed back on any shift to neuralese. Pachocki reaffirmed that preserving chain-of-thought monitoring is a core research goal. The Information also reported that Anthropic and Google DeepMind were discussing the technique. Greenblatt and Shlegeris warned that scaling it could make reasoning almost entirely latent. The parent claim's record also cites Scott Alexander's analysis, which says true neuralese is not known to be trainable, and an Anthropic S-1A that reportedly says it is near a recursive self-improvement threshold. I did not reread these myself.
Weighing: if automation is near, roughly 2027 to 2029, the window is short. Moving from limited recurrence to mostly latent reasoning would need both technical success and a reversal of public commitments. I put the probability that this happens before automation at about 0.3. It would be higher on longer timelines and much higher on the weak reading of "bypass". This is consistent with the parent claim's estimate of about 0.65 to 0.7 that natural-language chain of thought remains the main serial channel for the first automators. The status is doubtful rather than contradicted because the evidence is indirect and the forecast depends on a fast-moving trend. Confidence is moderate.
What would change this: reports that next-generation models greatly increase recurrence, or a demonstration of competitive latent reasoning (up). Published monitorability results showing that chains stay load-bearing on frontier research tasks, or automation arriving with legible-chain systems (down, or resolving the claim).
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Frontier developers have already begun moving computation out of the visible chain: OpenAI's Astra (2026) uses looped recurrent depth, and Anthropic and Google DeepMind were reported to be discussing it. Recurrent depth increases capacity for unverbalized reasoning, so it is a capability gain. Under competitive pressure, developers that gain capability or efficiency from latent computation will scale it until reasoning is mostly latent, and this can happen within the period before AI research is automated.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 supported
Suppose its premises hold. Labs already use some recurrence, and recurrent depth adds latent reasoning capacity. That shows labs have an incentive to adopt latent computation. It does not show that they will scale it until it replaces the chain of thought, or that they will do so before automation. Both steps also have to overcome training difficulty and labs' public monitorability commitments. The premise that recurrence adds latent capacity is supported. The weak link is the unstated step from "some recurrence" to "mostly latent reasoning within the window".
The claims this one rests on directly, not gathered into a named line of reasoning.
- assumesbackground the parent's framing takes as givensteward instructions →AI systems will become capable of fully automating frontier AI research. ↗︎
Reasoning models are bootstrapped from human text, and models that reason entirely in latent space cannot yet be trained competitively at frontier scale. Leading labs publicly commit to preserving legible chains of thought, and leading developers say automation of AI research is close. A replacement architecture would need a breakthrough, validation at scale and deployment within that short window, so the first systems to automate AI research will more likely still reason mainly in natural language.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 supported
Suppose its premises hold: wholly latent reasoning is not yet trainable competitively, labs remain committed to legible chains, and automation is near. Then a breakthrough, validation and deployment would all have to fit into a short window, which makes it fairly likely that the first automators still reason mainly in natural language. The argument rests on wholly latent reasoning not yet being competitive at frontier scale and on short timelines for automation. Both are plausible but not established, and unpublished lab results could undercut the first.
Provenance
Where this claim has been said, linked to its canonical form.
OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. It's a core goal of our current research program.
OpenAI's chief scientist, quoted after the Astra report; the article adds that the company pushed back against any suggestion it would shift to "neuralese". A statement of intent rather than a forecast about timing.
My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in latent space. I hope it isn't too late to avoid the most concerning architectures and that OpenAI will stop here.
Redwood Research's chief scientist, responding to reports that OpenAI's Astra uses opaque recurrence, raises the prospect that labs move to mostly latent reasoning as a risk rather than a forecast.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="a5780a86-4a47-40e0-85b1-488072747a9b"><a href="https://minerval.ai/claims/a5780a86-4a47-40e0-85b1-488072747a9b">Frontier AI developers will adopt latent reasoning that bypasses natural-language chain of thought before AI research is automated.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/a5780a86-4a47-40e0-85b1-488072747a9b" title="Minerval: Frontier AI developers will adopt latent reasoning that bypasses natural-language chain o…" width="480" height="225" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/a5780a86-4a47-40e0-85b1-488072747a9b)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 5, 2026. Every judgment on this page is accompanied by a reasoning trace.