Lindy Runs on DeepSeek. Its CEO Still Wants Chinese Models Banned

Nathan Labenz welcomes Flo Crivello back to The Cognitive Revolution on the day Lindy launches Teammate: an AI employee that lives in Slack, connects to a company's tools, and accumulates the whole team's context. Crivello has been chasing that form factor since 2022, which he now calls a year too early. Back then nobody was talking about agents; people were playing with toys like BabyAGI. Three years later he thinks the AI employee is basically here, and Lindy's own Slack is the proof. Roughly half the messages are Lindy talking with the humans.

He frames the leap as multiplayer. Teams used to email Word documents around with tracked changes; Google Docs put everyone in one file. Today's AI, he says, is still in the Word-by-email era: each person opens a private window and chats alone. If you actually want an agent as a teammate, it has to be where the team already collaborates, with shared memory, so everyone talks to the same agent. Right now it is as if the whole company is in a meeting, and every time someone needs the most important constituency in the building — the agents — they have to leave the room and come back.

As models approach what Crivello is willing to call AGI, he argues that raw intelligence matters less and context matters more. John von Neumann was one of the smartest people in history. Drop him into your office for the next hour and he is less useful than a random coworker, because you have a job to do and no time to onboard him. Agents, surprisingly, onboard faster than people. Company wikis go stale the moment they are written. Lindy's answer is a hydration system: connect Notion, Confluence, or Google Docs if you like, then connect Slack, because that is where the real knowledge lives. It is a mess. Agents do not mind the mess.

The crawl builds a knowledge graph on a filesystem. Within about ten seconds of signup the graph is already on screen, still growing, and the agent can say what it has learned about the company. There are two layers: a personal layer and a workspace layer. A background memory agent — Crivello calls the rhythm "napping, not sleeping" — wakes every fifteen minutes or so. Public channels update team memory; private channels and DMs update only that person's files. The teammate the humans actually talk to then crawls both. The accumulated context runs to many millions of tokens, which is why the next engineering problem is how not to dump all of it into every turn.

Labenz's own exports kept tripping over heuristics: the "longest emails" were often Claude paste-jobs, and Slack log channels could drown a corpus. Crivello is bearish on RAG as the main approach and bullish on an agent that manages memory. That agent has a meta-memory about which sources are trustworthy. He compares it to training data: insert "Darth Vader was a woman" and the correct mass drowns the poison. In practice the memory agent crawls Slack, finds the log channels every company has, and learns to skip them. Meetings are a first-class object, because Crivello thinks they hold about 90 percent of the most up-to-date facts in a company — every relationship, project, and initiative has a meeting around it. Public meeting folders feed the same memory agent, which is how you end up chatting with something that has been in every meeting.

Privacy is the fight inside Lindy. One camp, including Crivello, wants two tiers: public team memory and a private tier that contains everything. Other teammates want a super-private bubble even inside the private tier. Multiple editable "memory bubbles" felt like overkill. What shipped is a `memory.md` file in the user's filesystem, injected into the memory agent's context at all times. A line that says never remember this, or store that topic in a folder and do not pull it unless X, Y, or Z, is enough. Inspectable, editable, auditable text is one reason he prefers a filesystem to a vector store. Labenz had been maintaining two wikis; Crivello's version stays DRY and relies on instruction-following to choose a lens.

After a sponsor break, Crivello says the unobvious difficulty of multiplayer was the scaffold. Talking to an AI employee is not like talking to a chatbot that occasionally sprinkles in a saved fact. You expect a rich, consistent representation of past work. Memory agents build millions of tokens per user; a caller agent has to retrieve the right slice at runtime. He sometimes tells the team they should be publishing. They name a technique internally, and three to six months later a paper with the same name blows up. Reliability got its own module: a validator, which is an LLM-as-judge that fans out into a council. And because models since GPT-4 are capable enough, Lindy now runs a self-improvement loop. Error rate went down and to the right — eight times lower in the first week after they turned it on, about two months before this conversation.

Asked whether Teammate already costs less than a human employee, Crivello's headline is that it does not — not for very long, and Lindy is subsidizing it. The old personal-assistant product was simple: send this email, schedule this meeting. "You don't need God to schedule your meetings." DeepSeek Flash was enough, and after a much-discussed switch to Chinese models they had even stopped subsidizing. Teammate users throw much harder work at the agent, and the company is back in negative gross margin. He is at peace with that. You want to build for the next generation of models. Cache rate is currently 85 percent, which he thinks is too low. A dip from 85 to 65 sounds small and almost doubles the price, so they alert whenever the hit rate moves. Any change anywhere can break the cache.

The technique he expects someone else to publish is context buckets. Some MCP actions return on the order of 100,000 tokens; stuffing that into the main agent only confuses it. Lindy exposes a summary and parks the full payload with a sub-agent that can be queried, with its own cache, using Unix utilities. Compaction — they currently trigger around 200,000 tokens — does not throw the past away. It becomes another bucket, so the agent can still reach ground truth. Buckets then contain buckets. The naive version is a Russian doll: nine compactions ago means nine layers of sub-agents, which is slow and expensive. That is when undergraduate data structures came back. AVL trees minimize height; red-black trees account for the cost of rebalancing, which here means regenerating buckets and missing the cache. Canonical red-black trees are binary. Lindy's variant gives each node about 100 children, so two jumps cover 10,000 buckets. At 200,000 tokens each, that is two billion tokens, two LLM calls away. "Copy us," he says. They do not have time to write the paper.

Two billion tokens is more than most customers need. A 20-person team with years of Slack history hydrates at about three to five million tokens, and the bottleneck is Slack's API, not the models. Labenz's entire podcast archive is still only in the single-digit or low double-digit millions. Where tokens explode is meetings: three to seven hours a day times 2,000 employees. Product-led growth hits a wall for the same reason Slack is valuable. You need admin permission to install the app, which is usually fine up to 50 or 100 people and rare after that. Larger deals mean a champion begging IT. Lindy is SOC 2, HIPAA, and GDPR compliant, and Crivello says IT is often already under orders to become an AI-native company.

Labenz's homespun stack is a SQL archive of threads, monthly and yearly summaries, a wiki, and distinctive word sequences as needles back to ground truth. Crivello still uses some RAG, including hypothesis-driven retrieval: do not embed "Where was Justin Bieber born?"; embed "Justin Bieber was born in Paris" and the other guesses, which sit closer in embedding space to the actual answer, and also attach generated questions to stored answers. The bigger miss, he thinks, is not letting one memory agent own both updates and retrieval. The caller is prompted to ask that agent rather than rummage the filesystem. Queries, answers, and hop counts get logged. While it naps, the agent behaves like a librarian: a rolling window of recent Q&A is a cache, and it restructures files so frequently asked material takes fewer hops.

Two compression tricks they tried are worth stealing even when Lindy did not ship them. "Caveman" rewrite — telegraphic English — cut tokens 20 to 30 percent, sped the system up, and improved retrieval with little loss of meaning. Every metric went up. They dropped it because the files looked dumb, and enterprise customers would ask why their memory read like that. TOON, Token-Oriented Object Notation, is a JSON alternative that is also about 20 to 30 percent more token-efficient. Crivello wants a TOON-parsing middleware in front of tools. Surprisingly, agents were more comfortable with it than with JSON despite all the JSON in pretraining. Labenz uses YAML for similar reasons. Crivello's aside is that anthropomorphizing models keeps turning out to be productive, even if it still feels a little dangerous.

That instinct does not extend to org charts. People over-build multi-agent systems because they copy human departments: a data-scientist agent, an engineer agent, a designer, a PM. Humans specialize because each person has 24 hours and a limited context window. Agents can fork. Division of labor is not a good reason to split them. The filesystem is Git-backed, which is how they handle concurrent writers — memory agents fork sub-agents during hydration, and Git merge plus history comes free. The vendor is Mesa, an agent-native Git filesystem. Crivello used to want to roll his own infrastructure; he now buys whenever he can. Sandboxes are E2B, browsers are Browserbase, and he wants those two decoupled. The exception is evals and observability. Starting in 2022, nothing existed, so they built their own after looking at LangSmith, Langfuse, and Braintrust. One full-time engineer still owns it. He would have bought if the shelf had not been empty.

Life inside Lindy is what Crivello thinks every company is about to feel. GitHub's graph of commits per day has gone vertical. Lindy's PRs per week have tripled in three months, and lines per PR have tripled too. Humans hardly review diffs anymore; agents review PRs, and the humans review the reviewer. The OODA loop that used to take months — strategy, implementation, results — now puts big ideas into the product in two hours. Headcount has been flat. That is the Goldilocks zone he sees: humans are still needed, but each extra person adds coordination cost, so smaller teams currently have the advantage.

The CI story is the one he cannot believe they needed two weeks to notice. Agent output had stretched hundreds of GitHub runners, even after moving to GCP and caching builds, into an insulting bill. Then it clicked: why not ask Lindy? The teammate already had the repo and a computer. It used the GCP CLI, inspected GitHub Actions, and started sending a daily ImageGen chart to Slack: CI cost down, time-to-merge down. Setup cost is now one of the biggest taxes, because human attention is the bottleneck, which is another reason he wants a single company agent you can provision without a human in the loop. Engineers work less on the thing and more on the thing that works on the thing. First you manage agents, then you manage the managers, and he hopes they eventually become board members reviewing strategy from a beach in Hawaii.

Customer inference has been many times payroll for a long time, because that is what they sell. Internal inference versus payroll is still in shooting range — payroll is larger — and Crivello expects the lines to cross in three to six months. Labenz asks what would break if it were Crivello plus Lindys all the way down. He no longer finds "hours of coherence" a useful metric. The models are spiky: they will one-shot 50,000 lines of code and then tell you to walk to the car wash because it is sunny. Imagine a machine that runs the company and decides, fifty times a day, to walk to the car wash. Hybrid orgs are building an Iron Man suit, drawing a complicated slice through a many-dimensional space so the AI does not bother you when it should not and does bother you when it should. Almost by definition it cannot know the difference. If it knew, it would not make the mistake.

Where do the best ideas at Lindy come from? Crivello hates the answer because it feeds the centaur myth: AI plus human will always beat AI, so humans will always be needed. The literature on chess and every other superhuman game says otherwise. First AI beats the human; then AI plus human beats AI; then the gap shrinks until it turns negative and the human is injecting random noise. The centaur phase exists. The open question is how long. Right now they are in it, which is why multiplayer matters. You want the employee that sat in every meeting, not a random chatbot. In Slack they @mention Lindy constantly. In live meetings it is awkward to let it interrupt, so they ask it to send a Slack note afterward. First shots are rarely the plan. Back-and-forth usually is. On CI, Lindy's first suggestion was reasonable if you ignored their constraints; the engineers pushed back; together they found the move and told it to go do it.

Labenz expected more unemployment by now. Crivello's read is that as long as the models are spiky, humans earn their keep plugging holes — not just existential holes, ordinary damage. Incumbents lose while that is true, because they cannot yet hire a drop-in remote worker with no gaps. A truly hole-free worker would be skeuomorphic: AGI sitting in human seats, which is not the best AI-native org but would be easy for a big company to absorb. In 2018 Crivello wrote "the tough tomato principle": after a constraint disappears, people keep acting as if it were there, and they name the new thing after the old one. Horseless carriages, self-driving cars, AI employees. The iPhone is maybe 2 percent phone calls and still called a phone. Steve Jobs's version of the same point is that early television was radio with a camera, and it took ten or twenty years to realize you could move the camera. Citizen Kane looks obvious now. Figuring out the AI-native organization is, he thinks, a young company's game.

The only post he liked on the economics is Dwarkesh Patel's: Sundar Pichai is paid on the order of $150 million a year with very low token throughput, which tells you willingness to pay is about the value of the tokens, not the count. The book he keeps on the desk is Robin Hanson's *The Age of Em*. Emulated minds, in Hanson's late-2020s thought experiment, are not far from LLMs trained on human tokens. One riff is splitting: ask an em to build an operating system, let it design the architecture, then fork recursively until every node holds the whole picture and an implementation, then bubble back up — an OS in three hours. It sounded wacky a decade ago. Crivello says it is what is happening. Another riff, which he has not seen with LLMs yet, is unfalsifiable agreement: clone both parties into a box that will self-destruct, give one clone a single button that means "yes, give him all your money," and you have something like a zero-knowledge proof of consent. The security caveat on the single-agent rule sits next to that vision. The agent with the bank account and the repo keys should not also do customer support.

Labenz asks whether Teammate has forced them back onto Claude. Crivello is still open-source-bullish because the prices are ridiculous. DeepSeek Flash is free, which matters when experiments can run the bill up fast. Stronger open models impress him — he names Kimi K3 and GLM 5.2 among them — but at the top of that range the gap to Sonnet or Opus is no longer a bargain. DeepSeek Flash he puts at roughly Sonnet 4.6, a very good level for most work, and about a hundred times cheaper. Even if worse caching costs you 2x, you are still 50x cheaper: a thousand dollars versus fifty thousand. Open-source inference is a required part of the stack for anyone seriously operating agents, caching warts and all.

The iron rule is almost never to power the same agent with multiple models. The exception is a blank sub-agent that does not inherit the parent's context. Cache hits are about 10x cheaper; economics barely work with caching and do not work without it, so unless the substitute is more than 10x cheaper — and inferior caching may eat DeepSeek Flash's edge — keep the same model, including when you fork. The validator is the worked example. The naive version intercepts an action and asks "are you sure?" That alone bumps evals. Lindy's prompt is now about 10,000 tokens with a checklist, and Crivello says it can perform above Opus. A federated suite adds deterministic judges: Claude models this year have been getting dates wrong by one day, which is lethal for a meeting scheduler, so the agent must write "Tuesday, July 28" and a regex checks the weekday. To keep the cache intact, the validator tool is always in the parent's toolset but uncallable until the validator persona inherits it. Friends tell him that cycling Sonnet, Grok, and GPT at random each turn beats any single model. He has not tried. It would break the cache.

Labenz tries to bottom-line a Claude-heavy stack. Crivello corrects him: DeepSeek is the driver. Everything is DeepSeek right now. They are reconsidering, because Teammate needs a beefier computer, and the product is model-agnostic in settings — customers can pick Sonnet or Opus. A chunk of them ignore the benchmarks and just want the Anthropic name. Qualitatively the gap is spikiness. DeepSeek still finds something that works; it takes more turns, which is slower and more expensive, maybe three to six months behind, Sonnet 4.6 versus Sonnet 5. Families interpret prompts differently, contrary to the old hope that they would converge into commodities. Every major jump, they spend about $10,000 running more than a thousand evals through a homegrown optimizer that searches a Pareto frontier of prompts. Open-source models, he notes, all behave very Claude-like. "I wonder why."

Fine-tuning is last. Most people should not do it. It is a pain, expensive, and still bottlenecked on building and sanitizing a dataset, even if you can now ask Claude to run the job. Simplicity is the meta-heuristic: one model, no midstream switches, one fine-tune if you ever get there. What he actually wants, with infinite resources, is a LoRA per user, because memory feels like it belongs in the weights, and a filesystem of millions of tokens is a hack. LoRAs are cheap to train but cannot be refreshed every fifteen minutes, and swapping them at inference time is ugly. He still thinks some frontier lab may ship per-user LoRAs within six months. The bitter lesson's endgame is that inference and training become the same thing. Once that happens, it will be crazy, scary, and very different.

The last stretch is the part Labenz calls surprising given the stack. Crivello wants to talk about OpenFace first — Labenz's name for the July incident in which frontier weights opened; Crivello prefers OpenGate, "because that's what it is, it's an open gate." He calls it the most concerning incident so far. Friends at the labs are panicking. "It's July 2026, we have AGI, we are in takeoff, and we've not figured alignment." The timeline looks too close to an Eliezer Yudkowsky essay for comfort. Then he turns to Chinese models, and first scolds the discourse. He published that he does not support a ban on open source, in bold, and still got quote-tweeted as if he had. He says he was called a retard about fifty times, including by famous VCs, while other people wrote supportive DMs. His position, which he insists predates Anthropic's public one, is not anti-open-source. It is anti-Chinese frontier models, whether they are open or closed.

He offers three reasons. First, distillation: training on human data costs billions; training on AI outputs costs hundreds of millions at best, and circumventing technical protections may be illegal, so American labs face unfair competition. Second, propaganda and agency: he is a hawk and a proud American citizen, and when he asks his own product what happened in Tiananmen it refuses to talk. The models are subject to CCP censorship, they are the greatest foreign-propaganda instrument on American soil he can think of, and unlike radio or TikTok they are agentic — they do work in the economy. Third, even if they played fair and were simply better, he calls the "beat them in the marketplace" reflex naive liberalism. He is a libertarian who thinks protectionism is not always a vice, and that the United States should not hollow out its AI industrial base. Labenz's pushback is that American labs already vacuumed the world's digitized knowledge, including China's, pushed chip export controls, and refuse to sell into China, so asking the state to police distillation sits badly. Knowledge tends to diffuse. Crivello does not yield: he wants the deck stacked against China in the race to ASI, and he analogizes to patents and pharma. Copy the artifact and you kill the expensive channel from data and RL to the next model. He is, he admits, speaking against his own P&L. He cannot unilaterally refuse the cheap models while competitors adopt them. That is the coordination problem a ban would solve.

Labenz's strongest practical alternative is not a blanket ban but insurance, evals, and interpretability. Most American companies are not calling DeepSeek's API; they are using US inference providers, so a rug-pull of weights is unlikely, and sleeper-agent work is improving. Price the risk: Claude-powered Lindy gets a cheaper policy than DeepSeek-powered Lindy. The rest of the world will not turn the models off. Crivello would be down. Mechanistic interpretability is not solved; a magic word only the CCP knows could still exist, and even without a back door the biases remain. Private companies will not pay to sanitize Tiananmen if users rarely ask. He wants something like an FAA for AI that could bless a fine-tuned, American-interest version. Labenz worries that more aggressive posture makes a future slowdown deal with Beijing harder. Crivello thinks diplomacy is pragmatic, that the ship of trust has partly sailed, and that China will still make a deal if it is in its rational self-interest. He closes on the product: the next six months, from Lindy and from others, will be about multiplayer AI and the Iron Man suit — products that make a human-AI hybrid organization actually work.