Opaque recurrence is the phrase that put a glossary in the news. On 7 September 2026 TechCrunch republished its running dictionary of artificial intelligence vocabulary under the headline “Opaque recurrence, and other AI terms that you should probably know.” The term had existed in public for five days.
That is a strange thing for a reference work to do, and it is worth taking seriously rather than mocking. Glossaries are supposed to lag the language. This one led it, promoting a word that had been reported on the previous Tuesday into the position normally reserved for the thing a reader most needs explained.
So we counted. We pulled the published glossary, split it into its individual entries, and measured every one: how many words each definition runs to, which entries point at which other entries, which companies get named, and where the links actually go. The findings are not in the coverage, and a few of them are awkward.
The headline term is one of the shortest definitions in the file. Three separate entries describe the same underlying technique. The article’s own web address still describes a different subject entirely. And the entry that points readers toward opaque recurrence contains a broken link. None of that makes the glossary bad — it makes it a snapshot of a vocabulary being assembled faster than it can settle, which is a more useful thing to understand than any single definition.
Table of contents
- What Opaque Recurrence Actually Is
- How Opaque Recurrence Went From Scoop to Glossary Entry in Six Days
- Opaque Recurrence Is One of the Shortest Entries in the Glossary
- Three Names, One Technique: Opaque Recurrence, Recurrent Depth and Neuralese
- The Glossary URL Still Says Hallucinations, Not Opaque Recurrence
- Two Broken Links, One Beside Opaque Recurrence
- What the Cross-References Say About Opaque Recurrence
- What Opaque Recurrence Means If You Are Buying AI
- The Other AI Terms Worth Your Time
- How to Read Any AI Glossary Without Being Sold a Word
- Frequently Asked Questions About Opaque Recurrence
- References
What Opaque Recurrence Actually Is
Before the counting, the substance. The term describes a real change in how a model can be built to think.
The mechanism in one paragraph
A conventional reasoning model works through a problem sequentially, in language. It writes out intermediate steps — the running commentary you see when a chatbot “thinks” before answering — and each step is a legible artefact. Opaque recurrence replaces part of that with a loop: the model passes the same query back through its own internal layers repeatedly, refining an answer in numeric representations rather than words.
Why anyone would want it
The appeal is efficiency. Looping through existing layers lets a smaller model reach a better answer without the compute cost of generating a long written chain of thought. In an industry where the cost of running a model is the cost of the product, that matters commercially. Opaque recurrence is a way of buying reasoning quality with fewer tokens.
Why safety researchers objected
The cost is visibility. Those written steps are not a perfect window into a model’s reasoning, but they are the best one anyone currently has, and they are the tool investigators reach for when a system misbehaves. A loop leaves far fewer readable traces. Buck Shlegeris, chief executive of Redwood Research, wrote that he was “extremely concerned by the reporting that Astra uses opaque recurrence,” and that pushing the technique further would give a lab “the option to massively increase the recurrence and totally destroy CoT monitorability.”
What OpenAI actually said
OpenAI has not accepted the framing. Chief scientist Jakub Pachocki wrote that the lab “has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models,” that it remains “a core goal of our current research program,” and that all models already perform some quantity of opaque reasoning. The company says Astra’s chain of thought stays legible and has pushed back on comparisons to a fully illegible model.
How Opaque Recurrence Went From Scoop to Glossary Entry in Six Days
The speed is the story. Set the dates side by side and the vocabulary’s arrival is unusually easy to time.
The timeline
The Information reported the technique on Tuesday 1 September 2026. TechCrunch covered the reaction on Wednesday 2 September at 13:19 UTC. OpenAI launched Astra on Thursday 3 September. The glossary carrying opaque recurrence in its headline published on Monday 7 September at 12:24 UTC. Six days, end to end.
Two names arrived at once
The Information’s reporting used “recurrent depth.” The safety commentary used “opaque recurrence.” Both entered the language in the same week, describing the same thing, and the glossary now carries both as separate entries.
| Date (2026) | Event | Term used |
|---|---|---|
| Tue 1 Sept | The Information reports the technique behind Astra | Recurrent depth |
| Wed 2 Sept | TechCrunch covers the safety reaction, 13:19 UTC | Both |
| Wed 2 Sept | Follow-up: Anthropic and Google DeepMind discussing it | Both |
| Thu 3 Sept | OpenAI launches Astra | — |
| Mon 7 Sept | Glossary republished, 12:24 UTC | Opaque recurrence, in the headline |
Which name won
The headline took the safety framing, not the engineering one. That is not our inference — the glossary says so itself, in the “Recurrent depth” entry: “recurrent depth” is the engineering term and “opaque recurrence” is “the framing that emphasizes the safety concern.” A reference work chose the more alarming of two synonyms for its title and then documented that it had done so.
Opaque Recurrence Is One of the Shortest Entries in the Glossary
Here is where the counting starts to bite. We measured every definition in the published article.
The baseline
The glossary contains 32 entries and 3,893 words of definition, plus a 148-word introduction. The average entry runs 121.7 words; the median is 122. That is a tight, consistent distribution — most entries sit within about 30 words of the middle.
Where the headline term lands
The entry for opaque recurrence runs 91 words. That is 25.2% below the average, and it is the sixth shortest of the 32. The entry for its synonym, recurrent depth, runs 70 words — the shortest definition in the entire glossary, 42.5% below average.
What gets more room
The longest entry is RAMageddon at 187 words, about a shortage of memory chips. Second is Weights at 182. Third is Chain of thought at 163 — appropriately, since it is the thing opaque recurrence displaces. RAMageddon is 2.06 times the length of the term in the headline.
The reasonable explanation, and its limit
A short definition is not a bad definition. New terms have less settled meaning, so there is less to say, and brevity is a virtue in a reference work. But the glossary is explicitly a living document, updated regularly. The asymmetry is worth noting precisely because it is the honest outcome of writing fast: the newest word gets the least writing, and it is still the word on the door.
Three Names, One Technique: Opaque Recurrence, Recurrent Depth and Neuralese
The glossary spends more of itself on this one idea than any headline suggests.
The three entries
Opaque recurrence (91 words) defines the loop. Recurrent depth (70 words) defines the same loop under its engineering name and says outright that media outlets use the two “almost interchangeably.” Neuralese (106 words) defines the endpoint — a model reasoning entirely in internal numeric representations — and exists mainly to say that opaque recurrence is a step toward it.
What that adds up to
Together the three run 267 words across 3 of 32 entries. That is 9.4% of the glossary’s entries and 6.9% of its definition words, all describing one axis: whether a model’s reasoning stays readable. No other concept in the file occupies three slots.
| Entry | Words | What it describes | Names Astra? | Points to |
|---|---|---|---|---|
| Opaque recurrence | 91 | The looping technique, safety framing | No | Chain of thought |
| Recurrent depth | 70 | The same technique, engineering name | No | Opaque recurrence, chain of thought (link broken) |
| Neuralese | 106 | The hypothetical endpoint; no model does this | Yes | Nothing |
| Combined | 267 | One axis: readable reasoning | 1 of 3 | — |
The entry that never names the model
This is the detail we did not expect. Every news story about opaque recurrence is a story about a specific product. The glossary’s definition of it names no company and no model at all — it is written as a general technique. The only entry in all 32 that names Astra is Neuralese, and it does so to say that Astra is not an example of it. The term arrived attached to a launch and was defined without one.
Why a glossary would include something that does not exist
Neuralese is the only entry in the file describing a thing no shipped system does. The glossary says so plainly: “No shipped model does this today.” Including it is a defensible editorial choice — you cannot explain why researchers are worried without naming what they are worried about — but it means one of 32 “terms you should probably know” is a scenario rather than a technology.
The Glossary URL Still Says Hallucinations, Not Opaque Recurrence
A living document leaves fossils. This one left an obvious one in plain sight.
What the address actually reads
The article’s web address ends in artificial-intelligence-definition-glossary-hallucinations-guide-to-common-ai-terms. The phrase “opaque recurrence” does not appear in it. Neither does anything about reasoning, loops or safety. The slug was minted when the glossary’s marquee term was hallucinations — the industry’s word for a model inventing facts — and it never changed, because changing a URL breaks every inbound link pointing at it.
What that tells you
The headline is a variable and the address is a constant. This is the same file that was, at some earlier point, “a guide to common AI terms” fronted by hallucinations. The vocabulary rotated; the document did not move. Anyone treating the headline term as a stable fixture of the field is reading a slot, not a word.
The same pattern will repeat
Hallucinations were the term everyone had to learn in 2023 and 2024. Opaque recurrence occupies that slot in September 2026. The slot will be filled again, and the address will still say hallucinations.
Two Broken Links, One Beside Opaque Recurrence
We checked every link in the published article. Two of them do not work, and where they sit is telling.
The malformed addresses
The heading anchor for Distillation points to http://distillation — a bare word that a browser will try to resolve as a hostname. And inside the Recurrent depth entry, the pointer that should send readers to the chain-of-thought definition points to http://@chain-of-thought. Both are the signature of a word typed into a link field where an anchor reference belonged.
Why it matters more than a typo
The second one sits in the entry a reader lands on when they are trying to understand opaque recurrence and want the contrast with ordinary step-by-step reasoning. It is the single most load-bearing cross-reference in the section, and it is the one that fails. Neither is a scandal. Both are evidence of the same thing the word counts show: a fast update to a large file.
The rest of the links are fine
To be fair to the work, the other outbound references check out — OpenAI’s charter, the Model Context Protocol documentation, a Hugging Face explainer on mixture-of-experts architectures, a Carnegie Endowment primer on compute. The glossary is well sourced. It is simply not, this week, carefully proofed.
What the Cross-References Say About Opaque Recurrence
The glossary’s “(See: …)” pointers form a small map of which ideas the editors treat as foundational.
How connected the file is
Only 12 of the 32 entries carry a pointer to another entry at all, and between them they make 14 pointers. That leaves 24 of 32 entries with nothing pointing at them — free-standing definitions in an alphabetical list rather than a connected web.
Which terms are treated as bedrock
Neural network and large language model receive three inbound pointers each, the most in the file. Chain of thought and inference receive two apiece. Opaque recurrence receives exactly one, from its own synonym.
Who gets named
Six of the 32 entries name OpenAI — more than any other company. Google appears in three, Meta, Microsoft and Mistral in two each, Anthropic in one (for donating the Model Context Protocol to the Linux Foundation). Nvidia, whose hardware underwrites the entire industry, is named in none.
What Opaque Recurrence Means If You Are Buying AI
Most readers of a glossary are not researchers. They are trying to work out what to do on Monday.
For almost every buyer, nothing changes yet
Astra’s use of the technique is reported as limited, its chain of thought is still expected to be legible, and no procurement decision turns on this today. If you are evaluating a model for a business workload, opaque recurrence is not currently a line item. Treating it as one would be reacting to a word rather than a capability.
The question worth writing down
What does change is a question you can now ask a vendor and expect a real answer to: can you show me the reasoning trace for a decision your system made, and will that still be true in twelve months? That is an auditability requirement, and it is the same requirement whether the model reasons in words, loops through its own layers, or does something not yet named. Any organisation with a serious AI strategy should already have it in a contract.
Where it becomes concrete
The places this lands first are regulated decisions and incident review — anywhere you may have to explain, after the fact, why a system did what it did. If your logs are the explanation, the format of those logs is a commercial term, not a technical detail. Firms building on artificial intelligence and machine learning platforms should ask what the provider commits to preserving.
Ignore the trio, learn the plumbing
If you have limited attention, spend it on the terms the glossary itself treats as foundational — the ones other definitions point back to. Inference is what you pay for. Tokens are the unit you are billed in. Fine-tuning and distillation are how a general model becomes yours. Those decide budgets this quarter. Opaque recurrence decides an argument about oversight that has barely started.
The Other AI Terms Worth Your Time
Beyond the three-entry cluster, the glossary’s most useful function is separating terms that cost you money from terms that describe research.
The cost words
Compute is processing power and the industry’s binding constraint. Tokens are the chunks of text a model is billed by. Token throughput measures how much work a system handles at once. Memory cache — key-value caching in particular — reduces repeat calculation. RAMageddon, the file’s longest entry, is the memory-chip shortage now raising prices on everything from consoles to data centre builds.
The architecture words
Mixture of Experts splits a model into specialised sub-networks and activates only a few per request, which is how very large models stay affordable to run. Distillation transfers behaviour from a large teacher model to a smaller student. Parallelization is doing many operations at once, the property that made graphics hardware the industry’s backbone.
| Term | Words in entry | What it decides for a buyer |
|---|---|---|
| Compute | 75 | Whether capacity exists at all — the binding constraint |
| Token | 121 | The unit on your invoice |
| Token throughput | 155 | How many users one deployment can serve |
| Memory cache | 138 | How much repeat work you stop paying for |
| Distillation | 132 | Whether a cheaper model can replace a costly one |
| Mixture of Experts | 134 | Why a very large model can still be affordable |
| RAMageddon | 187 | Why hardware quotes keep rising — longest entry in the file |
| Opaque recurrence | 91 | Nothing yet — an oversight question, not a line item |
The words that predate the boom
Several entries long predate any of this: deep learning, reinforcement learning, transfer learning, diffusion, GANs, weights, validation loss. A vendor selling natural language processing in 2018 was selling something real, and these terms have stable meanings. They are the ones least likely to be redefined by next week’s news cycle.
How to Read Any AI Glossary Without Being Sold a Word
The method we used here works on any reference of this kind, and takes about ten minutes.
Check the address against the headline
If the URL describes a different subject than the title, you are reading a rotating slot. That tells you how much permanence to assign the featured term.
Count the entries that say the same thing
Synonyms inflate a concept’s apparent importance. Three entries describing one technique is three times the shelf space, not three times the substance.
Follow the pointers
The definitions other definitions point back to are the ones the editors consider load-bearing. Learn those first. In this glossary those are neural network, large language model, chain of thought and inference — not the newest arrival.
Ask whether the thing exists
At least one entry here describes a scenario no shipped system implements. That is legitimate in a glossary explaining a debate, but it should be read as a position in an argument, not as a product category. Our analysis of OpenAI’s Astra and its critical cybersecurity designation covers what the model actually shipped with, which is the better test.
Watch what the definition omits
The opaque recurrence entry never names a vendor or a model. That is good lexicography and poor context — the reader learns the mechanism and not the fact that it entered the language through one company’s launch and one publication’s scoop. Both halves matter. Our AI models and tools hub tracks the products these words get attached to.
Frequently Asked Questions About Opaque Recurrence
Is opaque recurrence a new invention?
The technique of looping a query through a model’s layers is not new as an idea. What is new is a frontier lab reportedly shipping it in a flagship product, and the vocabulary that grew around that in the first week of September 2026.
Is opaque recurrence the same as recurrent depth?
Yes. TechCrunch’s own glossary states they describe “the same underlying method,” with recurrent depth as the engineering term and opaque recurrence as the framing that emphasises the safety concern.
Does Astra reason in a way nobody can read?
No. OpenAI says Astra’s chain of thought remains legible and has rejected comparisons to neuralese. The reported use of the technique is limited. The concern raised by safety researchers is about the direction of travel if the technique is scaled up.
Should this change what AI tools my business uses?
Not today. It should change one question you ask vendors — whether you can obtain a reasoning trace for a consequential decision, and whether that commitment survives their next model release.
Why does the glossary define a term that no model uses?
Neuralese is included to explain what safety researchers fear rather than what exists. The glossary is explicit that no shipped model reasons that way.
How many terms are in the glossary?
Thirty-two entries totalling 3,893 words of definition, averaging 121.7 words each, plus a 148-word introduction.
References
Opaque recurrence, and other AI terms that you should probably know — TechCrunch
OpenAI’s new reasoning technique alarms AI safety experts — TechCrunch
OpenAI launches Astra, its powerful (and controversial) new model — TechCrunch
Secret Technique Behind OpenAI’s Astra Model Sparks Security Concerns — The Information
Model Context Protocol — Introduction
Mixture of Experts Explained — Hugging Face
A Primer on Compute — Carnegie Endowment for International Peace
RSI is the new AGI, and it’s just as hard to pin down — TechCrunch
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.