If Your Agent Is Right 80% of the Time, Ten Steps Leave You With 10%

Imbue is still pre-product, has 27 full-time employees, and has just raised $200 million from investors including NVIDIA, with a significant slice earmarked for a cluster of 10,000 H100 GPUs. The company, formerly Generally Intelligent, is led by founder and CTO Josh Albrecht and his co-founder Kanjun Qiu. Host Nathan Labenz admits he struggled, even after the papers and other interviews, to see the firm as one thing. What clicked on a second listen was an investor thesis: this is a bet on a fork in the road for AI.

One path is that raw compute scale keeps pushing the frontier, and today's leaders — OpenAI, DeepMind, Anthropic, Meta — keep leading. Another is that ever more pre-training becomes uneconomical, and the way to make agents robust, reliable, and restrained enough for real work is painstaking fine-tuning plus complementary systems: better retrieval, supervision, and novel interfaces. Labenz thinks Imbue pays off hardest in that second world, because it may take a string of Eureka moments, and those often come from small, highly aligned teams that dogfood their own half-working tools.

Albrecht's own pitch is more arithmetic than manifesto. If you are right 80% of the time and you do ten things in a row, you are fairly likely to fail; 0.8 to the tenth is about 11%. To take much longer sequences you have to get that up to 90, 99, even 99.9 percent. He thinks the next year or two will bring a visible jump on those longer-horizon tasks, and that the result will be a pretty interesting, weird world in which the systems actually work the way people already expect them to.

Albrecht's path back into research is a little old-fashioned. He did a master's when support vector machines were the cool thing and deep learning was not yet a thing, and he watched from the sidelines until around 2019. Self-supervised learning was starting to work in vision, audio, and text, and he could finally see a way off the treadmill of huge human labeling teams stuffing in the right answer. He and Qiu stepped back from what they were doing and founded Imbue to tackle the larger problems that now looked tractable.

They have insisted, from day one, on calling themselves a research company — not a research lab, not a startup. The agents they want, systems that can actually reason and do what we want computers to do, remain an open research problem. At the same time they are close enough to useful products that this latest raise is funding things they hope to talk about sometime next year. The two sides will not separate cleanly, because how you should interact with an agent is as open as how you should train one.

If something goes off for days, taking actions and spending hundreds or thousands of dollars on your behalf, you really want to know what it is doing, and you do not want that in a giant text log or a chat box. There are UI questions, annoyance questions, and drift questions: how do you let it generalize without wandering off what you meant? For now they are building tools for themselves, so they can feel the pain. Using today's model APIs to assemble a larger system, Albrecht says, is an exercise in frustration — prompts, components that go off the rails, evaluations that are expensive and slow.

Labenz asks why GPT-4 is, of everything he has tried, the best reasoning language model. Albrecht does not claim to know OpenAI's full playbook, but he is blunt that garbage in, garbage out still holds. He is sure they have teams cleaning Common Crawl and mixing in books and code, and pretty sure they run a large labeling operation that writes back-and-forth coding dialogues — one person asking how to invert a matrix in NumPy, another writing a polished answer — which makes in-distribution questions look astonishing. He also notes that GPT-4's habit of unrolling a problem step by step is now baked into the training data, partly as a side effect of RLHF and partly as an intentional shove onto trajectories that are more likely to be right.

Beyond chain of thought, the academic menu includes trees of thought, graphs of thought, and sampling for answer consistency. What excites Imbue most is spending more compute at inference time. In a chat app, users will not wait around for a 10 or 20 percent better answer; if the system is working overnight, you do not care how long it takes. Labenz hears an echo of Imbue's earlier conversation with Noam Brown on Cicero, the Diplomacy agent whose lesson was that runtime compute can buy a qualitative jump. Curriculum tricks in pre-training, Albrecht adds, are more like a 2x on compute than a better final model, because you still run out of data; the leverage is in later stages, in generated fine-tuning data, and in asking a different question rather than a more expensive one.

Labenz asks whether "pretty good features" from pre-training means a robust world model, the Ilya Sutskever claim that next-token prediction learns a representation of the world that generated the token. Albrecht grants that it learns a representation, then insists there are many representations, some better than others, and that most of what transformers learn does not correspond to reality. Chris Olah's monosemanticity work at Anthropic is his exhibit: features like "does this look like base64," "is this Hebrew," "the in a math context versus a physics context." Swap `print` and `length` in Python and a person can follow the weirdness; the language model collapses, because it has never seen `print` in the length slot. Transformers, in his telling, use statistical co-occurrence of words to solve tasks that happen to be solvable that way.

Addition makes the gap provable: transformers at best learn modular addition in a strange trigonometric space and cannot generalize to longer sums the way people can. Hattie Zhou, a friend of the company, recently showed that prefix markers in the input let the model learn the actual carry algorithm, which does generalize. Multimodality, especially first-person causal video plus good text, plays a similar role by forcing representations toward something more shared with the world, though Albrecht does not expect that within a year. Internally, ridiculously small transformers crush "Einstein puzzles" and are trash at everything else, which is why he treats reasoning as a set of heuristics for transforming information — useless if there is no information to transform. A later Imbue paper on stepwise self-supervised image learning, born when an intern noticed notches in a CIFAR-10 loss curve, showed features coming online one at a time, first color, then contrast, then brightness: "dumb" features, kernel learning versus feature learning, and a conserved learnability budget that colleague Jamie proved, like momentum in physics, has to be spent on the most important directions first.

About a year and a half before the interview, Imbue built a procedurally generated virtual world so reinforcement-learning agents could practice, and they made it on the order of 100 times faster than Minecraft, work later published at NeurIPS. The agents did learn to open doors, even ones with complicated locks and bars. Watch the video, though, and they bounce and jiggle beside the door until something works, then slowly grind toward a more efficient wiggle. That is not how people open doors. People turn the handle and pull because they understand the mechanism.

The lesson sent them back to humans. You do not get a PhD by trying 10,000 times and reinforcing the run that worked; you plan, anticipate, and reason. So Imbue shifted toward coding, reasoning, browser, and desktop tasks where planning is the point. Reinforcement learning is not discarded: it is still good for which button to click, a lower layer complementary to the planner. Labenz is wary of smuggling confusion in with human analogies, then admits the split is hard to ignore, including Jim Fan's Eureka result, where GPT-4 writes denser reward functions that teach a robot hand to twirl a pencil.

Fan's Voyager paper, which stores Minecraft skills as code, is the other half of why code is the inner loop. In both Eureka and Voyager, writing code makes a better agent, and Albrecht thinks most new code will be written inside agents — optimizing their own prompts, picking few-shot examples, pulling regularities out of experience. Real tasks look like Stripe: "all you do is take a payment online," and somehow that employs an army, because the details never end. Today's self-driving stacks are rigid because people pre-decomposed the problem; future agents should decompose dynamically, notice that a language model is doing addition the stupid way, and swap in Wolfram Alpha or a calculator. That kind of self-improvement is what huge human-designed systems cannot do.

The fundraising demo was a suite of automated decompositions on academic datasets, mixing Imbue's own models, open-source checkpoints, GPT-4, and Claude, and GPT-4 was not always best when you needed speed or parallelism. One focus was Meta's ANLI, adversarial natural language inference built to trick models. A passage says a song was written by X and Y; asked whether it was written by just X, the model says yes because the word "just" is sitting there. Ask who all the authors are, reassemble the answer, and the trick evaporates. Tree-of-thought papers that appeared later looked a lot like what they had been doing, which told Albrecht the method was general, not a one-off. Chain of thought, in this view, is just one way of spelling information so attention can force the right tokens; generating sub-questions is another.

Labenz's back-of-envelope on 10,000 H100s at retail looks like $300 million, more than the whole round, so something else is going on. Albrecht says they are working with a new provider, Voltage Park, which actually bought the chips. Imbue uses the cluster for less than the chips' full lifetime, which lets Voltage Park stand up a larger cluster and lets Imbue avoid locking the entire raise into hardware. They had already trained hundreds or thousands of language models on a smaller cluster as part of CARBS, their hyperparameter optimizer, so mixing owned, open, and API models was already the demo's default.

Labenz's "time to GPT-4" math — 10^25 FLOPs, an H100 at about 4×10^15 in FP8, 10,000 GPUs, three to seven days — startles him until Albrecht corrects the assumptions. You cannot quite train large models in FP8, and the public rumor he trusts more is on the order of 30,000 A100s for three to five months, perhaps an eight-expert mixture of experts at about 200 billion parameters each, near Chinchilla-optimal. MosaicML's old numbers are his working scale: a 30-billion-parameter Chinchilla run on 600 billion tokens took about a month; scaled to a 4,000-H100 cluster, a 200-billion-parameter model is roughly 45 days. That is still large. It is not a 1.6-trillion-parameter laser.

The GPT-4 technical report's perplexity curves are the strategy document. A point with a hundredth of the compute loses only a tiny bit of performance, so if you were going to spend a billion dollars on one run, Albrecht would rather spend $10 million and recover the rest somewhere else, because even three to five percent is a huge difference on long-horizon agents. Fine-tuned Llama already beats GPT-4 on many of their tasks, and smaller models need less data. He warns home experimenters spending millions that FP8 is fine for small models and gets more divergence-prone as they grow; the loss curve's steep part is only the first 10 or 20 percent of training, so switching precision mid-run adds complexity without saving much of the grind.

CARBS, Cost-Aware Robust Bayesian Search, is the other force multiplier: Bayesian optimization glued to natural evolution strategies so the program, not a tired researcher, picks the next point in a 10-to-12-dimensional hyperparameter space. You hit go, go to sleep, and wake up to a curve of how data, compute, heads, and key-value size move the scaling law, including an automatic berth around out-of-memory crash regions. It can extend toward architecture search, though Albrecht is not sure Transformer variants are the first place to spend it. Future loops might put a better prior than a Gaussian process into the optimizer, or even a coding agent proposing new configuration axes once current ones saturate. For now, he says, CARBS has more than paid for itself.

Imbue also shows up in policy, which Labenz treats as another surprising line for a 26-face website — 27 as of Tuesday, when a recruiter joined. Albrecht's stance is practical engineering: he found the UK AI Safety Summit and the Biden executive order more measured than he expected, heavy on measurement, investigation, cybersecurity, and biological harm rather than a brand-new blanket agency. Kanjun Qiu went to the summit and posted her own thread; Albrecht will not ventriloquize it. On longer-term tail risks he is encouraged by international willingness to collaborate, including from China. "Nobody wants to go extinct," he says, and Labenz takes even being on the same page with China as a ray of hope.

The concrete prototype was an analysis of thousands of public comments to the U.S. Commerce Department's NTIA on AI. Language models asked structured questions of every submission; people read a smaller subset so Imbue could check for error rates and weird bias. Careful question design plus human correlation, they concluded, can make models competent at large-scale text analysis. Among respondents — a self-selected bunch — artists talking about copyright infringement and personal economic impact dominated, followed by values, privacy, and misinformation. The project was also dogfooding: this quarter's goal is tools they themselves seriously use, of which NTIA analysis, error fixing, unit tests, and recruiting emails are a handful, not an infinite portfolio.

From the outside the mix of virtual worlds, coding agents, hyperparameter search, interpretability, and policy looks scattered. Internally they sit down each quarter, type into one Google Doc what they finished, botched, worry about, and are excited by, then synthesize a small number of directed bets. CARBS is infrastructure that accelerates every experiment; the NTIA work is a live agent for question-answering. They want to grow at about 2x per year toward a 200-to-400-person company of highly leveraged people, not OpenAI's pace, so culture survives. After this raise they expect to spend more on compute than on people, and for that gap to widen: one recruiting coordinator plus a bunch of agents, instead of a giant coordinator army, if they can make the tools actually work.

Albrecht's closing picture is the same as his opening arithmetic: over the next year or two, systems should get a lot better at longer-term actions, and that will be a weird world in which they finally work as people expect. The 10,000 H100s are coming online, so the bet will not stay theoretical for long. Until then, the episode is unusually rich in tactics you can use on systems that already exist, without waiting for Imbue's product.

Count the steps in any automation you are about to hang on a language model, then raise 0.8 to that power before you trust the pipeline overnight. When a model fails, change the structure of the question before you change the model: ANLI's "was it written by just X" error dies once you ask who all the authors are and reassemble. For jobs that can run while you sleep, spend the inference compute — extra samples, extra checks, extra decompositions — that no chat user will sit still for. Those three moves are the company's research program, shrunk to something you can do this week.