Apr 10, 2025 · 26m · tbpn
What AI and LEGO Have in Common | Flo Crivello on TBPN April 7th
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Lindy founder Flo Crivello joins the podcast to discuss the rapid evolution of autonomous AI agents, detailing how modular architectures, collapsing inference costs, and practical B2B workflows are redefining business automation.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.0 Guest teaching 4.6 Guest disagreement 2.6 The hosts pushing back 1.9
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 3:53 Dismissing Consumer Agent Utility
The guest directly shuts down the host's consumer agent premise, asserting that consumers are time-rich and money-poor with already frictionless business interactions.
Hardest push from the hosts ▶ 4:42 Pushing De Novo Consumer Use CasesThe host challenges the guest's bearish stance on consumer AI by proposing de novo generative workflows and viral loss-leader acquisition strategies.
Biggest teaching moment ▶ 17:20 Redefining Agent Reliability BaselinesThe guest educates the host on evaluating agent accuracy against offshore human BPOs rather than traditional four-nines software metrics.
The host holds their own ▶ 19:55 Deep Dive on Inter-Agent NegotiationThe host demonstrates deep domain literacy by citing the AI 2027 paper and detailing how autonomous agents must resolve conflicting sub-task trade-offs.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Agent Timelines and Live Travel Planning Demo | 4 | 2 | 1 | 1 | The host asks open-ended questions comparing the timeline for AI agents to the Studio Ghibli viral moment. The guest shares a screen demo of Lindy multi-agent swarms planning a trip, while the host offers light conversational color. | |
| Consumer Agents Versus Business Value and Costs | 5 | 6 | 5 | 4 | The guest rejects the host's premise that consumer agent use cases will take off, arguing consumers are time-rich and money-poor. The host pushes back by proposing de novo personal media generation and viral top-of-funnel acquisition, but the guest counters with the brutal unit economics of inference costs. | |
| Corporate AI Mandates and Engineering Productivity | 5 | 4 | 2 | 2 | The hosts ask about CEO mandates and Lindy's multi-product strategy. The guest clarifies that Lindy is a single product built around composable low-level Lego primitives rather than separate disparate products. | |
| Inference Cost Reductions and Model Distillation | 6 | 4 | 2 | 1 | The host asks informed technical questions regarding inference cost curves, distillation, and eval contamination. The guest explains how model size reduction and distillation have driven cost reductions far more than silicon advances. | |
| SMB Practical Adoption Versus Enterprise Shelfware | 4 | 6 | 3 | 1 | The guest delivers an industry breakdown of real SMB deployments versus the resurgence of enterprise 'shelfware' driven by board-level panic and massive unutilized AI budgets. | |
| Evaluating Agent Reliability and Implementing Guardrails | 4 | 7 | 4 | 2 | When asked about agent reliability risks, the guest reframes the evaluation benchmark, explaining agents should be compared to error-prone human BPO workers rather than deterministic software, supported by hard guardrails and human escalation paths. | |
| Agency Distribution and the Future of Implementation | 7 | 4 | 1 | 2 | The host cites the AI 2027 paper and posits complex trade-off optimization between coordinating agents. The guest builds on this by explaining neuralese and agent team topologies managed by AI chiefs of staff. | |
| Existential AI Risk and P-Doom Perspectives | 5 | 4 | 3 | 2 | The guest explains his high P-Doom and frustration with casual attitudes toward existential risk in tech podcasts, before the conversation pivots to pragmatic daily cron job workflows. |
Statements from this episode (20)
Crivello: 2025 will be a serious inflection point for AI agents
“A year. I broadly buy the meme that 2025 is the year of agents. It's like a year-ish. I think those mean to be a pretty serious inflection point.”
Crivello: Bearish on consumer AI agents because consumers are time-rich and money-poor
“I am actually quite bearish for consumer agents. I think, like, consumers are time rich and money poor, right? It's like, they don't want to trade time for money like that, which is why I think, like, Google Assistant and Alexa, like, they've been really failu…”
Crivello: Drop-in AI worker replacements for businesses will arrive in 12-18 months
“I think, like, there is going to be a transformational moment for businesses where you have, like, the holy grail of AI agencies, like, what we call the drop-in replacement for a human worker. And I think that's coming in, like, 12 to 18 months.”
Crivello: OpenAI is burning cash on viral consumer toys because it raised $40B
“And like, OpenAI is burning money right now on the Ghibli thing, and they don't care because they raise like forty billion dollars to have infinite money.”
Crivello: Employers Must Mandate AI Tooling Due to Engineering Productivity Gap
“And I think that's the gap has grown so much now between the frontier of the level of productivity that you can achieve as an engineer as the median practice of the median engineer, especially in big companies that now employers have no choice, but to just lay…”
Crivello: Low-level AI primitives combine superlinearly into endless use cases
“The beauty of building it this way is that you build a set of actually few low-level primitives, and then it's like Lego bricks, you know, and then you can combine these Lego bricks however you want, and they combine super linearly, and then they explode into …”
Crivello: Only 5% of a phone agent startup's work is core phone code
“When you build a computer use startup, I surmise that maybe five percent of your bandwidth is the actual, not, sorry, not computer use, phone agents. Five percent of your work is actually phone, like, phone code and phone whatnot, right? The rest is, like, the…”
Crivello: Single-platform agent consolidation will be essential as AI agents collaborate
“I actually think it's going to be much more important even once AI agents can actually work together. You don't want, like, having all of these different platforms for all of your different AI agents is maybe the same thing as having, like, five headquarters. …”
Crivello: AI Inference Costs Fall 40x to 100x Annually
“I mean, we are seeing costs fall by roughly 40 to 100 X every year.”
Crivello: Meta's Llama 4 Has Good Evals but Bad Vibes
“I will say, I feel bad for saying that because I have friends like the Lama team, but boy, like Lama for the vibes are really bad. Like the emails are good and the vibes are quite bad actually.”
Crivello: Meta AI VP and Researchers Resigned Over Unethical Practices
“The VP of AI research at Meta resigned and so have a couple members of the core research team seemingly in protest for what they deemed to be unethical practices.”
Crivello: Zuckerberg Threatened to Fire Meta AI Org Over Targets
“I think Zach is, is massively raising the temperature and I think he's basically threatened to like Destroy the whole org or, like, fire happy org or, like, similar stuff like that if they don't hit certain targets by a certain date, and I think that's led to …”
Crivello: SMBs lead real AI agent deployments over enterprises
“SMBs are actually the ones who have, in my mind, the most real deployments out there. They are actually deploying agents for mission critical stuff all day long.”
Crivello: High-revenue enterprise AI startups are becoming modern shelfware
“There are some very hot AI startups out there in the enterprise that are basically shelfware. Like, their revenue is skyrocketing, and they're meeting very little adoption internally.”
Crivello: AI agents must be benchmarked against humans, not software
“Agents, the comp is not other software, it's humans, right?”
Crivello: Deployed AI agents beat human quality in most use cases
“The vast majority of the time for use cases where AI agents are deployed right now, they more than clear that bar of human quality, and then some, right?”
Crivello: Agency middlemen for deploying AI agents will soon be obsolete
“I think it's temporary though, because AI agents are soon going to be simple enough. Again, if you have a drop in replacement for a human worker, it's going to be no more complicated than collaborating with your human teammate. So I don't know that you're goin…”
Crivello: AI agent team topology will be a primary driver of performance
“I think ultimately this, like the AI agent team topology, is going to be one of the main drivers of the performance of your AI agent setup, and I think at first it's going to be on you as a user to figure out that topology.”
Crivello: AI chiefs of staff will soon design and recursively manage agent teams
“Eventually, and in the not too distant future, you are going to have an AI agent design that for you. So we call that the AI chief of staff. It's just going to Manage all of your agents. And you can imagine this is recursively, right? Just like agents, managin…”
Flo Crivello: A 10% AI Extinction Risk Is Unacceptably High and Under-Discussed
“Look, even if it was low, like, 10%, that's unacceptably high. If you listen to the Dwarcush podcast on EI 20, 27, that's the part, I love Dwarcush, I love the podcast, but that's the part where I'm like, what? At some point they talk about Pidoum, and one guy…”