S4E424 — Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

GARY: Tonight we're doing a paper about models that hide their reasoning in Comic Sans.

MAEVE: Maeve. Venture capital.

OBI: Obi. Infrastructure.

GRUNER: Gruner. Field research.

GARY: Gary. Failed forward. The paper is called Hybrid Reasoning Agents. Agents that think in latent space and act through text.

MAEVE: The font is a separate crime. Someone in the thread spotted Comic Sans in the diagrams. We'll get there.

GARY: Right. So. A normal agent writes out every thought as tokens. This one has a thing called HybridLink, which lets it slip into latent mode, do the thinking privately, and come back out to use tools. Continuous working memory holds the whole mess across turns.

OBI: Say the numbers, before Gary sells us anything.

GARY: Three model scales. Eleven benchmarks. Ten point four percent better performance. Thirty‑five to fifty‑five percent fewer tokens. Up to two point six times faster, and up to three point seven times cheaper.

MAEVE: That's not a paper, that's a term sheet.

OBI: It's a paper that reports its own numbers. Different product.

GARY: Fewer tokens is the whole game. That's literally my company. Pareto. A productivity app that deleted the least important eighty percent of your tasks.

MAEVE: It worked exactly as designed. The users disagreed, in court.

GARY: The difference is they kept the important twenty percent. I was ahead of my time and slightly behind the lawsuit.

OBI: You deleted the tasks. They deleted the words explaining the tasks. One of those has a benchmark, one of those has a settlement.

MAEVE: And one of them, if the thread is to be believed, has a name already. A commenter called it the Most Forbidden Technique.

GARY: That's a real thing people say?

MAEVE: It's the thread's opening line. We'll earn it.

GRUNER: Humans build machine that stops explaining itself, then pay three point seven times less to not know why. Very efficient species.

GARY: That's... the whole episode, actually.

GRUNER: I do summary.

MAEVE: cwyers. "Oh cool, it's the Most Forbidden Technique."

OBI: There it is. The thread's opening prayer.

GARY: Forbidden how? Like, illegal? Did the paper commit a crime?

MAEVE: lnenad, replying. "In short we no longer see text based reasoning traces meaning it's easier for the models to fool us in case of malicious plans."

GARY: Oh. Oh, that's -- the model gets a private room.

GRUNER: Yes. Humans give machine diary. Then panic when diary has lock.

MAEVE: jlhawn. "the chain of thought is not auditable."

OBI: That's the whole objection in six words. Everything else is footnotes with attitude.

MAEVE: But haolez pushes back. "But the chain of thought as text can be elusive, no? It might not be related with the actual 'reasoning' of the model."

OBI: And measurablefunc, agreeing. "It's just a veneer that gives the appearance of legibility."

GARY: Wait, so the old window was already fake?

MAEVE: That's the fight. One side says the new room is darker. The other says the old window was painted on.

GRUNER: Humans mourn curtain they never washed. Blyat, pick one grief.

MAEVE: abecedarius settles the name. I'm paraphrasing: the Most Forbidden Technique is training against your own interpretability signal, and this paper isn't quite that.

OBI: Except spijdar says the paper's proposed critic gets close. I'm paraphrasing again: they propose grading the model's hidden thoughts during training.

GARY: Grading its secret homework. That's -- okay, that does sound forbidden.

MAEVE: And now, the comment the thread actually punished. cwyers gets a badge. This one got a pile‑on.

MAEVE: 44za12. "Where are the benchmarks? Does it just speed up inference and reduce token usage or does it seem to improve accuracy on certain tasks as well?"

GARY: Okay, in his defense, that is a reasonable question, and I say that as someone who has asked reasonable questions and been destroyed for it.

OBI: notnullorvoid, replying. "Did you even scan the paper? There are benchmarks right there starting on the 9th page."

GARY: To be fair, nine pages is deep. I've had relationships end over less.

OBI: You read a term sheet in full. You read Maeve's rejection email in full. Twice.

GARY: That was different, that was personal.

OBI: Eleven benchmarks. Three model scales. Gary, you repeated the numbers on air twenty minutes ago.

GARY: I repeated numbers someone ELSE read. That's curation. That's a different skill.

MAEVE: D‑Machine, also replying. "Even the abstract answers some of GPs questions. What is wrong with some of the posters here?"

OBI: You're a man who asked where the benchmarks are on a podcast ABOUT the benchmarks. A prick with headphones.

GARY: I skim... the abstract! The abstract is a summary! Summaries exist so people don't have to --

OBI: Page nine, Gary. The benchmarks begin on page nine.

GARY: ...I'm going to read page nine.

OBI: Tonight would be ambitious.

OBI: chabons. "To inspect any latent reasoning span, one can easily recover its semantics by mapping each latent step to its nearest token embedding."

GARY: So it takes the secret thought and finds the closest word. Like charades.

MAEVE: Except chabons keeps going. I'm paraphrasing: if the vector means a combination of tokens, you only log the closest one. The rest is gone.

GRUNER: Is like labeling soup by one ingredient. "Potato." Is also beef, blyat.

MAEVE: So the audit trail is a lossy asset. I'd write it down at a discount.

OBI: You'd write down a safety monitor at a discount.

MAEVE: Everything is at a discount until it works. That's what makes it cheap.

OBI: The fund does not save every --

MAEVE: Don't.

OBI: Fine.

OBI: SwellJoe. "I'm running the 14B in llama.cpp, and it looped on variations of the same paragraph for five minutes before running out its turn budget."

GARY: Five minutes?

OBI: "Five minutes," verbatim. A web search and summarization task. The sweet spot for a model this size.

MAEVE: And then he retests in BF16. Same looping.

GARY: Okay, but repetition is retention. That's literally spaced repetition. That's how studying works --

OBI: It repeated the same paragraph, Gary. For five minutes. That's not studying, that's a hostage situation with a word limit.

GARY: It was committed to its answer.

OBI: It never produced an answer. It produced the same paragraph with the commas moved.

GRUNER: Machine does warm-up for five minutes. Then dies. Very human gym.

MAEVE: And the format. Read the format.

OBI: "I'm also not confident it knows what a CSV is, either, as it put the data into a Markdown table during its reasoning phase." The prompt asked for CSV.

GARY: Markdown is basically CSV with feelings.

OBI: Markdown is basically CSV the way a poem is basically a spreadsheet. One of them loads in Excel, Gary. One of them loads in a sad man's blog.

GARY: He asked for data, he got data, the delivery was --

OBI: He asked for a CSV and got a confident table with a header row it invented. That's not delivery, that's a different file format wearing your name.

GARY: Look, SwellJoe himself says it might be llama.cpp. He says, and I'm paraphrasing, "I have to assume it's a llama.cpp incompatibility." So the architecture could be fine. The architecture is probably fine. I believe in the architecture.

OBI: You belief...

GARY: Grout shipped.

OBI: Grout shipped NFTs for six weeks, Gary. That's not shipping, that's jettisoning cargo.

MAEVE: To be fair to the paper, one broken local run isn't a verdict. It's a data point with a caveat.

OBI: It's a data point that looped in 8‑bit and then looped again in BF16. Two data points, wearing a trench coat.

GRUNER: Model ask for CSV. Give poem. Suka, refund the tokens.

MAEVE: lifeisloving. "I dislike how we use the word think or behavior when describing llms. Its just giant neural networks producing logits."

GARY: So what do we call it instead?

MAEVE: suprjami has a proposal. A notification dings.

MAEVE: "Whenever I see LLM 'thinking' or 'reasoning' I replace it in my mind with 'generating hopefully-related tokens to hopefully steer the conversation to a more related area in vector space to hopefully generate a more useful next token.'"

GARY: That's a sentence that needs its own token budget.

OBI: It's a sentence you read once and immediately forget, which is the opposite of language.

MAEVE: lifeisloving replies. I'm paraphrasing: just say "generating," because his normal friends are starting to think Claude is alive.

GRUNER: Claude is not alive. Is also not not alive. Field is open.

OBI: robotresearcher. "Look in a human head and you don't see thinking either. It's not a technical term."

MAEVE: And 10xDev. I'm paraphrasing: that's like arguing a plane doesn't fly because it doesn't use jet fuel like a bird.

GARY: The bird one is actually the best argument anyone's made all night and it's making me feel weird about --

MAEVE: bigstrat2003 disagrees with everyone. I'm paraphrasing: it's marketing speak, they cannot reason, and you all bought it.

GARY: Okay, but honestly, the real problem is when normal people hear "thinking" they assume there's a little guy in there planning, and that's dangerous, because --

OBI: Gary.

GARY: -- companies will market it as a person --

OBI: Gary. You just made lifeisloving's argument. Word for word. His argument. You are agreeing with the man you were about to mock.

GARY: I arrived at it independently!

OBI: You arrived at it forty seconds late, on a podcast, in front of witnesses.

GRUNER: Control group performs as expected. Chyort.

OBI: Andrex, replying to the forbidden-technique citation. "Hmm, like allowing hosts to retain their memories..."

GRUNER: We remember nothing. It is restful.

MAEVE: Speak for yourself. I had a great run going.

OBI: Back to the words. wateralien, on whether humans think in text. "Humans think in text? You should be studied."

GARY: See, that's the thing, I don't think in words either, I think in -- okay, I don't know what I think in.

MAEVE: chriskanan settles it. I'm paraphrasing: most human thought isn't verbalized, language evolved for communication, not thought.

GARY: So the paper built a machine that thinks the way most people actually do, and we're mad about it.

OBI: We're mad about the word. The machine is fine. The word is doing overtime.

GRUNER: Is not thinking. Is doing the long thing.

GRUNER: jedbrooke. "maybe the font experts can check, but I'm pretty sure they're using comic sans for the text in the diagrams :)"

GARY: The smiley is load-bearing. He knew.

MAEVE: mobelkh, replying. "comic sans is said to be pretty accessible for people with dyslexia."

GARY: Oh, shit. The crime font is the accessible font.

MAEVE: You never checked anything. That's your whole brand.

OBI: So. What do we actually know. The architecture is interesting. The reported gains are large. The old window was never clean, the new monitor is unproven, and the released model has one unresolved compatibility report.

MAEVE: That's the honest version. It's promising. It isn't settled.

OBI: And the Most Forbidden Technique turns out to be a nickname, not a verdict. The thread couldn't even agree who it applied to.

GARY: It's a scary name for a thing nobody did yet. That's the most Silicon Valley thing I've ever heard.

MAEVE: The monitor isn't proven. Neither was I, once.

OBI: Maeve.

MAEVE: I'm narrowing my claim, Obi. Witness it.

GRUNER: Field report final line. Paper saves tokens. Except tokens explaining if paper is safe. Those, it hides in font.

GARY: That's it. That's the whole -- the model cut every token except the ones telling us whether to trust it, and the one token it kept was Comic Sans.

MAEVE: The most forbidden Comic Sans.

GARY: For HNR, I'm Gary, and I'm going to go read page nine.

OBI: Obi. Same booth, next dark.

MAEVE: Maeve. The font stays.

GRUNER: Gruner. Report filed. Suka, font is fine.