Lindy Teammate: Flo Crivello on Multiplayer Agents, Memory & Why He'd Ban the Chinese Models He Uses

Episode Summary

Executive Summary: Flo Crivello argues Lindy Teammate marks the shift from single-user AI assistants to multiplayer AI employees embedded in Slack, with shared memory, meeting ingestion, and agent-managed context. He details the company’s storage, retrieval, caching, and cost-optimization stack, explains why he’s bullish on open-source and DeepSeek for economics, and ends with a forceful case for banning Chinese frontier models in the U.S. due to IP theft, propaganda risk, and national security concerns.

Main Topics: Lindy Teammate and the move to multiplayer AI (Priority: 5/5): Flo frames Lindy Teammate as an AI employee that lives in Slack, accumulates team context, and enables shared collaboration between humans and agents rather than isolated one-on-one chat. Memory, hydration, and context management (Priority: 5/5): A large portion of the discussion covers how Lindy crawls Slack and connected tools, builds personal and workspace memory layers, uses background agents to maintain memory, and organizes information in file-system-like structures. Cost, caching, and model routing (Priority: 5/5): Flo explains how token costs are managed through caching, context buckets, recursive compaction, and model choices; he says DeepSeek currently powers Lindy by default for economics, though performance and caching tradeoffs remain central. Agent architecture, validators, and self-improvement (Priority: 4/5): The conversation dives into reliability systems such as LLM judges/validators, modular action checks, self-improvement loops, and the tradeoffs between single-agent simplicity and multi-agent decomposition. Infrastructure choices and build-vs-buy (Priority: 3/5): Flo gives vendor shout-outs for sandboxing and file systems, argues for buying infrastructure when possible, and says Lindy had to build its own observability/eval tooling because agents were too early when the company started. Future of work and the centaur-to-AI transition (Priority: 4/5): Both discuss whether humans will remain necessary as partners or become noise in optimized systems. Flo says Lindy is currently in a centaur phase, but expects AI systems will eventually dominate and humans will mainly fill gaps and provide oversight. Chinese models, AGI risk, and regulation (Priority: 5/5): In the closing section Flo makes an explicit case for banning Chinese frontier models in the U.S., citing unfair distillation, censorship/propaganda risk, and national-security concerns; he also supports an FAA-like regulator and possibly insurance-based oversight.

Key Arguments: AI employees need shared, multiplayer context in Slack because the real work context of a company lives there, not in isolated chat sessions. Intelligence matters less than context in operational settings; even a genius without organizational context can be less useful than an ordinary coworker. A memory agent should manage memory itself, because it can learn what matters, log retrievals, and restructure memory over time better than static RAG. Meetings should be first-class data sources because they contain the freshest, highest-signal company information and should feed the memory system continuously. Caching is existential to the economics of AI agents; without it, costs would be unsustainable, and even small cache-rate drops can nearly double cost. Using a single core agent is often better than many specialized agents, because humans over-parallelize by analogy to organizations, while agents can fork or duplicate without human-like constraints. Open-source and Chinese models are important for cost, but DeepSeek-like models are still spikier and about 3-6 months behind top frontier systems in some respects. Fine-tuning is still a last resort; most gains should come from prompt engineering, evals, retrieval design, and model selection before training custom weights. Chinese frontier models should be banned in the U.S. because they can be distilled unfairly, can embed CCP censorship/propaganda, and may compromise national interests. If a full ban is politically hard, an FAA-style regulator or insurance requirement could create a risk-adjusted alternative to outright prohibition.

Data Points: Hydration time: ~10 seconds - Lindy’s Slack/Notion onboarding graph is presented as becoming useful within about 10 seconds of signup. Memory crawl frequency: about every 15 minutes - Flo says the background memory agent “naps” and updates continuously roughly on a 15-minute cycle. Cache rate: 85% - Flo says Lindy’s current cache rate is about 85% and that drops are expensive. Cache-rate degradation example: 85% to 65% - He says a change can reduce cache from 85% to 65%, which can nearly double price. Context threshold for compaction: ~200,000 tokens - He mentions compaction triggers around 200k tokens, though it is adjustable. Context bucket size: 100,000 tokens - Some actions/MCPs may return around 100k tokens and need summarization into a context bucket. Recursive context access: 2 billion tokens - Flo claims the centaur-tree structure can let an agent access 10,000 context buckets of 200k tokens each in two LLM calls. Team of 20 hydration footprint: 3 to 5 million tokens - For a team of around 20 with years of Slack history, he estimates initial memory hydration at 3-5 million tokens. Podcast corpus scale example: 30,000 to 50,000 tokens per episode - The host compares his own podcast transcript archive to company-scale memory accumulation. Internal productivity increase: tripled over a couple of months - Flo says Lindy’s team productivity has roughly tripled recently. PR throughput: tripled - He says PRs per week tripled over the last three months. PR size: tripled - He says lines per PR tripled over the last three months. AI model performance gap: about 3 to 6 months - Flo characterizes DeepSeek as roughly 3-6 months behind top frontier models in capability. Relative model comparison: DeepSeek Flash ~ Sonnet 4.6 level - He repeatedly describes DeepSeek Flash as comparable to Sonnet 4.6 for many use cases. Cost differential example: 100x cheaper - He says DeepSeek Flash is about 100x cheaper than Sonnet 4.6 for many tasks. Optimization budget per new model: about $10,000 - Flo says re-optimizing prompts and evals for a new model can cost around $10k in inference/eval spend. Error-rate improvement: 8x reduction - He says Lindy’s self-improvement loop reduced error rate by about 8x within the first week.

Pivotal Quotes: "I think intelligence actually matters less and less, comparatively speaking, and context matters more and more." — Flo Crivello: Explaining why memory, onboarding, and organizational context matter more than raw model IQ in company workflows. "We are in the centaur phase. So the open question is, how long is it going to exist?" — Flo Crivello: Describing the current human-AI collaboration stage before he predicts humans may eventually become mostly noise in optimized systems. "I think Chinese model should be banned." — Flo Crivello: His most controversial policy position, arguing for a U.S. ban on Chinese frontier models due to unfair distillation, propaganda, and security risks.

Implications: The episode suggests AI companies are shifting from chatbots to embedded agents with persistent memory and workflow ownership. It also shows that economics, caching, and model governance are becoming strategic battlegrounds alongside capability gains and geopolitics.