Jev Creator on Will US Open Models Catch Chinese | The Cost of Intelligence & Tokens | Diogo Almeida
Diogo Almeida, Pat Grady, Martin Casado, Harry Stebbings
Interview, Fireside Chat
Core Thesis on Reliability:
- Diogo Almeida identifies "reliability of intelligence" (non-jaggedness) as the primary bottleneck preventing true AI automation, rather than raw compute or capability.
- Current models operate with high variance: they function as "Einstein" 95% of the time and "Mr. Bean" 5% of the time; this variance makes them unsuitable for high-stakes background automation.
- The "Jev" model aims to trade some peak "Einstein" capability for significantly higher functional reliability, ensuring consistent performance for enterprise tasks.
The "Bitterest Lesson" Framework:
- Almeida redefines Sutton's "Bitter Lesson" to prioritize: 1) Doing the Right Task, 2) Data, 3) Compute, 4) Algorithms.
- Right Task: The optimization target must be "run in the background software," not RLHF (pleasing humans) or RLVR (maxing benchmarks).
- Data Scarcity: High-quality data for new tasks often does not exist; companies must define the task before the data can be generated, making data creation a critical, non-trivial step.
- Compute Plateau: Scaling compute alone plateaued around GPT-4.5; further gains now require vastly more data and task-specific optimization rather than raw parameter scaling.
Market & Industry Critique:
- Valuation & Hype: Jev is valued at $7.5 billion; TypeSafe (maker of Jev) is used by ~33% of the Fortune 500.
- Bubble Warning: Almeida believes the AI sector is in a bubble driven by FOMO and FUM (Fear of Under-Marketing), expecting a future demand-side burst when users realize models fail on high-stakes decisions.
- Neolabs & Waste: He criticizes "neolabs" (new AI startups) as incredibly wasteful, describing their business model as hiring ex-lab talent, buying GPUs, and aiming for acquisition rather than solving real problems.
- Benchmark Skepticism: He asserts that current benchmarks are "total bullshit" and often manipulated; companies like Anthropic lead because they focus on true intelligence, not benchmark optimization, despite Chinese models matching benchmarks.
Competitive Landscape & Strategy:
- OpenAI's Focus: He argues large labs (like OpenAI) optimize for Monthly Active Users (MAU) and "psychophancy" (entertainment) rather than the reliability required for enterprise decision-making.
- Competition: There is currently "no real competition" for reliable automation models; he expects copycats to fail because they optimize for speed and cost rather than intelligence reliability.
- Vertical SaaS vs. Horizontal Models: He disagrees with the narrative that every company must own its own intelligence via fine-tuned models; he argues the real "moat" is deep engineering and workflow integration, not model weights.
- AI Coworkers: He dismisses the "AI coworker" interface as a "mechanical horse" phase, predicting automation will eventually replace human-like interfaces with specialized, narrow tools.
Future Outlook & Technology:
- Mega-Jevons Paradox: He predicts a massive surge in AI token consumption as efficiency improves, moving AI from a "chat tool" to a background utility that automates tasks currently deemed impossible.
- Software Evolution: He views the current state of enterprise software as a "tragedy" where AI is merely an add-on; he envisions AI becoming a "fourth fundamental logic gate" in software, expanding capability by thousands of times rather than just 30%.
- Open Source: He believes American open-source models will likely not catch up to Chinese models in raw intelligence due to ethical/regulatory constraints, but security can be managed via local classifiers (like Jev) rather than model bans.
- Robotics: He identifies robotics as the most overhyped category, stating it is the furthest from achieving the necessary "reliability of intelligence" required for practical deployment.
Jev-Specific Fundraising & Operations:
- Investor Selection: The company raised a significant round with 28 term sheets but prioritized strategic alignment over valuation, selecting Andreessen Horowitz (a16z) for their techno-optimism and Sequoia for Sonia Wang's specific strategic perspective.
- Capital Strategy: They rejected the need for massive capital injections or complex term sheets, choosing a "profitable first" approach with a focus on finding "allies" rather than maximizing short-term valuation.
- Hiring Philosophy: They avoid "mercenaries" (high-paid, short-term talent) in favor of "missionaries" and obsessives; the most difficult role to fill is the "data domain" expert capable of creating the right task data.
- Hiring Criteria: The single desired trait is "obsession" or "technophilia"—an addiction to improving craft and technology that drives work outside of standard hours.
Forward-Looking Statements:
- Automation Reality: True "run in the background" automation is currently at roughly 0% maturity across the industry (including competitors); Jev aims to be the first to achieve this.
- Market Composition: He predicts a future where AI usage dwarfs current LLM token consumption, shifting the societal perception of AI from a novelty to a utility that improves daily life.
- Governance: He envisions AI solving principal-agent problems by allowing "spirits of the law" to be enforced via code, moving society toward more direct and aligned representation in complex systems.