The Most AI-Pilled CEO We Know

Predictions and Expectations:

• Pedro Franceschi expects that within roughly two years, most people will have naturally "rewired" their brains to default to "AI first" for problem-solving, making it second nature rather than a mechanical choice.
• He predicts that as inference costs decrease by 10x, usage will increase 10x, ensuring that token costs remain a large expense for companies despite cheaper models.
• He believes that in the long term, likely over the next 100 years, token costs will become "almost free" and no longer a primary concern, similar to how electricity costs are viewed today.
• Franceschi expects that the "AI CapEx" default category bias seen in current models reflects the fact that model builders are overly focused on AI infrastructure, implying a future need for models to better understand non-AI categories.
• He believes there will never be a single general-purpose model that contains everything humans need, citing physical limits on atoms in the universe required to store all world models.
• He predicts that the "virtual employee" analogy will become the dominant framework for AI adoption, where agents are specialized for specific domains rather than attempting a single company-wide model.
• He expects that companies failing to redesign their processes from scratch based on AI capabilities will miss significant opportunities, characterizing the current state as a "turnaround" opportunity for non-AI-native firms.

Timelines and Milestones:

• Franceschi marks December as the critical tipping point where electricity-level utility was achieved with Opus 4.5, comparing the current timeline to being "five or six months after electricity was invented."
• He projects that steam engines (representing more complex industrial revolutions) are still "20 years away" in this historical analogy.
• He notes that his "G-Brain" project now contains "350,000 markdown pages," indicating a rapid scaling of personal knowledge infrastructure.
• He mentions that Brex has been working on managing token spend for "a bunch of cycles now," implying a recent and ongoing focus on this capability.
• He states that the "Crab Trap" proxy was open-sourced "probably about two months ago."
• He recalls a lunch with the YC team that served as the "precursor of G-Brain" and sent the YC team down a "rabbit hole," with the event occurring prior to this conversation.

Technology and Product Direction:

• Franceschi plans to "refound the very concept of what the company self-identity is," shifting from human-centric interfaces to "type systems interfaces agents talking to each other."
• He expects the company to move toward "virtual employee" agents with clear boundaries, such as separating a "customer world model" agent from a "product roadmap" agent.
• He intends to build a "self-learning system" where every human interaction becomes an eval, triggering automatic code and prompt modifications when agents flag issues.
• He plans to utilize "Lateral Synaptic Drift" (LSD) mode, which forces the combination of concepts outside a normal "cone" to generate coherent, high-impact ideas by rejecting traditional logical pairings.
• He expects the industry to shift from "Foxconn factory" models (restrictive, rigid agent environments) to "Esalen Institute" models (agents with agency, operating at network boundaries with freedom).
• He intends to "maximize" token consumption in the near future, arguing that founders should be "token maxing" to push the boundaries of what is possible, even if it seems suboptimal initially.

Market and Industry Outlook:

• He predicts that AI adoption is currently in the "0.3 percent" stage for paid users and the "one box out of 2,500" stage for actual agent usage, leaving the vast majority of the 8 billion population untouched.
• He expects a massive "10,000x" increase in inference demand, driven by the fact that 84% of the world has never used AI and only a tiny fraction uses agents.
• He anticipates that companies in major hubs (like the 10-mile radius around YC) will show faster revenue growth correlated with high token consumption, while a significant gap remains between these "token maxers" and everyone else.
• He believes that the "wisdom to choose" problems will remain the critical bottleneck for humans, as AI will eventually handle execution perfectly, making the selection of "unspoken signals" the only human differentiator.
• He expects that the "fabric of the company" will look fundamentally different in 2026 compared to today, with the CEO acting as the "Chief AI Officer" who must understand technology bounds better than anyone else.

Company Plans:

• Brex plans to "redesign the entire onboarding process" from scratch rather than just adding AI to existing workflows, specifically by using KYC technology for lead qualification to shift risk orientation earlier in the funnel.
• The company is developing an internal token spend management tool called "Magpie" to attribute every dollar of spend to a product, customer, or employee to analyze ROI.
• They intend to build a "customer world model" that ingests every touchpoint (clicks, emails, calls) to predict what a customer needs next and what issues they will face before they arise.
• They plan to create a "self-relearning system" where agents modify the codebase and prompts automatically based on manual exceptions handled by human teams.
• Brex is considering building a "better version" of the Crab Trap proxy as a separate company or product, potentially for other YC startups to use.
• The company plans to continue experimenting aggressively with agents, moving from read-only to write-access, though they currently maintain boundaries regarding customer data usage.

Financial Guidance:

• Franceschi predicts that token costs will become "easily the biggest expense in a company," surpassing traditional overhead as usage scales.
• He notes that even if token costs drop by 10x, total expenditure will remain high due to the 10x increase in usage, suggesting a persistent cost structure.
• He mentions seeing "large companies with very large budgets" spending only "10,000 a month" on tokens when they "should probably be spending 10 times more or 20 times more or a hundred times more," implying a massive opportunity for increased spend.
• He suggests that the "margin in tokens" is currently high for providers, but this will compress as usage scales exponentially.

Risks and Caveats:

• He cautions that security teams are often "more risk-averse than the technology probably requires," potentially stifling innovation by treating agents like "Foxconn" workers rather than autonomous entities.
• He warns against the "curse of knowledge," where founders assume models have seen the same data as them, leading to a lack of understanding of the model's actual distribution and blind spots.
• He highlights the risk that "agencies behind choice go away" if founders rely too heavily on AI for experimentation without maintaining discipline on what problems actually matter.
• He notes that models are "trained on a very specific corpse of information," meaning they may fail to provide answers for out-of-distribution data unless explicitly provided with context or retrieval systems.
• He warns that "escalation paths" in companies act as "antibodies" that reject AI-driven disturbances, creating social cohesion risks that must be managed by leadership.

Confidence and Disagreement:

• Franceschi is confident that "electricity was Opus 4.5" and that the current state of AI is a fundamental historical inflection point comparable to the invention of electricity.
• He expresses strong disagreement with the notion that AI should be treated as a "research project" or used only for "search mode," insisting that the core utility lies in "agentic loops with tools."
• He is skeptical that current "harnesses" and "coding environments" are sufficient, arguing that the industry needs to move toward "virtual employee" architectures with distinct boundaries.
• He is confident that "intelligence is compression" and that the best ideas will "fit in a napkin," implying a strict filter on which AI applications will succeed.
• He believes that the "execution" of ideas will be handled by AI, leaving the "wisdom to choose" as the primary human value proposition.