20VC: Jev Creator on Why Data Is More Important Than Compute | Will US Open-Source Models Catch Chinese | The Cost of Intelligence and Tokens: What No One Sees But Needs to Know with Diogo Almeida

Original Media · 1h 14m · English · Podcast · The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch

TLDR

Diogo Almeida, founder of TypeSafe (maker of Jev), argues that reliability of intelligence — not raw compute or benchmarks — is AI's fundamental bottleneck. He explains why data matters more than compute, why current AI is a bubble driven by FOMO, and how Jev's focus on decision-making reliability enables true background automation rather than human-in-the-loop assistance. He also discusses hiring missionaries over mercenaries, choosing investors as allies, and a future where AI becomes a fourth fundamental logic gate solving principal-agent problems across society. The conversation covers Jev's rapid growth, the misleading nature of benchmarks, and why open-source model competition is largely benchmark-maxing theater. Diogo predicts a 'mega Jevons Paradox' where reliable, efficient models unlock massive new automation demand. He believes software itself hasn't changed despite AI's capabilities, and that the real opportunity is building reliable automation that runs in the background forever. The episode reveals why Anthropic models feel better despite benchmark parity, why Chinese open-source models are efficient but pose limited threat if properly guarded, and why Diogo optimizes for 'technophilia' and slow, deliberate decision-making over hype and speed.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

Anthropic models are practically superior to Chinese open-source models even when benchmarks match, because benchmarks don't capture reliability of intelligence — the field has learned this in weeks but still optimizes for the wrong metrics.

•

Jev's models are so efficient they can serve 'mega Jevons Paradox' demand without massive GPU clusters, deliberately staying on the sidelines of the NeoCloud hardware bubble.

•

Diogo believes the biggest threat from Chinese open-source models is only if plugged in vanilla; with proper guardrails (like Jev classifying tool calls for safety), bias can be measured and trained out, making the security threat model narrow and solvable.

•

The 'right task' for AI automation didn't exist before Jev — RLHF optimized for human-in-the-loop assistance, RLVR for benchmark scores; neither translates to background software that runs forever without supervision.

•

Diogo didn't want Sequoia's money — he wanted Sonya Wong specifically as a strategic thought partner, and only accepted the fund because she was there.

•

Software engineers hate updating dependencies and only do it when things break; reliable automation means locking in working systems and never touching them again, which is the opposite of the current 'chase the latest model' treadmill.

•

The coding agent market (Cursor, Cognition, etc.) is a 'little fight' Diogo wants to stay away from; Jev's real market is run-in-the-background automation, which he believes will dwarf coding assistants by orders of magnitude.

Key takeaways

•

Reliability of intelligence — consistent decision-making without 'Mr. Bean' failures — is the absolute largest bottleneck preventing AI from automating real work; current models give Einstein-level output 95% of the time but catastrophic errors 5% of the time, making expected value negative for high-stakes automation.

•

Data matters more than compute (AI's 'bitterest lesson'): the right task definition unlocks the right data, which then determines model performance; Jev's breakthrough came from defining a new task (reliability of intelligence for automation) rather than optimizing for human preference (RLHF) or benchmarks (RLVR).

•

Benchmarks are 'total bullshit' and actively misleading: top-tier labs (OpenAI, Anthropic) don't need to cheat, but second-tier labs benchmark-max to survive; Anthropic models feel better in practice despite not leading benchmarks because they optimize for real reliability, not benchmark scores.

•

AI is currently in a demand-side bubble driven by FOMO and overestimation of capabilities; a burst will come when companies face the reality that unreliable models cannot be trusted for decisions with stakes, causing a demand shock as overcommitted firms backtrack.

•

The winning strategy is not owning proprietary models but building software around reliable models: companies' proprietary intelligence will live at the code layer (how they use models), not in custom-trained weights; Jev's efficiency enables a 'mega Jevons Paradox' where cheaper, reliable intelligence unlocks exponentially more automation demand.

•

Hiring should filter for 'technophilia' — obsessive love of the craft — not mercenaries chasing $20M packages; Diogo deliberately avoids compensation arms races to attract missionaries who stay up at night improving their work, believing this compounds better than hiring guns.

•

Fundraising should optimize for allies who help fight the largest companies, not valuation or cash; Diogo chose Sequoia (specifically Sonya Wong) and a16z (Martin Casado) for philosophical alignment and strategic thought partnership, rejecting 'cock tranche' offers despite massive demand.

•

The future is AI as a fourth fundamental logic gate (like the blue LED completing RGB), enabling a Cambrian explosion of software where automation removes the 'can't automate' category entirely; principal-agent problems (politics, corporate governance) will be solved by enforcing spirit over letter through reliable AI.

Notable quotes

“You get Einstein like maybe 95% of the time and Mr. Bean the other 5%.”

▸ Notable quotes · 0:05

“Data matters a lot more than compute.”

▸ Notable quotes · 0:11

“I don't want to be called a neolab because from my sense, most neolabs are incredibly, incredibly wasteful.”

▸ Notable quotes · 0:16

“Anthropic models are just the best, even if the Chinese models match it in benchmarks.”

▸ Notable quotes · 0:20

“I truly believe that we're going to like mega Jevons Paradox.”

▸ Notable quotes · 0:28

“AI is so clearly awesome, yet software itself hasn't changed.”

▸ Notable quotes · 0:31

Chapters

•

0:00 — Introduction and Jev's Mission: Harry introduces Diogo Almeida, founder of TypeSafe (maker of Jev), highlighting Jev's rapid growth to $100M revenue in 7 days and $7.5B valuation. Diogo frames Jev as solving AI's reliability bottleneck to enable true background automation.

•

5:50 — Reliability as the Core Bottleneck: Diogo explains why reliability of intelligence — not uptime — is the blocker: models are brilliant 95% of the time but catastrophically wrong 5%, making expected value negative for automation. He contrasts this with chat reliability (politeness) vs decision reliability.

•

8:20 — AI's Bitterest Lesson: Data > Compute: Diogo argues the field's 'bitter lesson' (compute > algorithms) has plateaued at GPT-4.5; the real lesson is data matters more than compute, but data only exists once you define the right task. Jev's task is reliability for automation, not human preference or benchmarks.

•

21:40 — Benchmarks Are Misleading Theater: Diogo calls benchmarks 'total bullshit' with perverse incentives: top labs don't cheat, second-tier labs benchmark-max to survive, and decision-model benchmarks are worst of all. Anthropic wins in practice despite benchmark parity because they optimize for real reliability.

•

43:20 — Open Source, China, and Security: Chinese open-source models are efficient but benchmarks overstate them; the threat is only vanilla plug-and-play. With guardrails (Jev classifying tool calls), bias can be measured and removed. American open-source may not catch up due to ethical/legal constraints.

•

50:00 — The AI Bubble and Demand-Side Shock: Diogo believes AI is in a FOMO-driven bubble; a burst will come from demand-side reality checks when overcommitted companies face unreliable models. Coding agents changed the game but follow historical hype cycles of up/down waves.

•

56:40 — Fundraising for Allies, Not Valuation: Diogo describes their chaotic round: no deck, all inbound, optimized for allies (a16z's techno-optimism, Sonya Wong's strategic brilliance) over cash or valuation. They rejected 'cock tranche' offers despite 3x valuation demands because they're profitable and don't need money.

•

63:20 — Hiring, Culture, and the Future: Hardest hire is 'data demon' team — a new subfield requiring obsessive technophiles, not mercenaries. Diogo wants 'technophilia' in every role, slow deliberate decisions, and a future where AI becomes a fourth logic gate enabling a Cambrian explosion of software and solving principal-agent problems.

Speakers

•

Harry Stebbings — Host

•

Diogo Almeida — Founder

⋮