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
- Anthropic models are currently viewed as the field leaders, though Chinese competitors match benchmark performance; the US export ban on GPUs is a threat, but the primary risk from Chinese models lies in their vanilla deployment capabilities rather than internal bias removal, while US open-source models may struggle to catch up in true intelligence unless ethical standards slow China's progress.
- The sector is predicted to transition from a hype wave into a "trough of disillusionment" as reliability limits are realized, potentially triggering a "demand side shock" in the "neocloud" layer and a crash in the volatile, FOMO-driven coding agent market before the "reality" of full end-to-end automation materializes.
- Future model improvements will depend more on data quality than compute scaling, with a shift from pre-Jev optimization methods like RLHF and human preference tuning toward "verified rewards" and "latent chain of thought," where data utility relies on the cutting-edge capabilities of the models themselves.
- Enterprises are expected to move from human-in-the-loop assistance to "full end-to-end automation" that runs "in the background," potentially expanding software capabilities by "hundreds of thousands of x," though this requires breaking current reliability barriers to gain trust for high-stakes decisions.
- The "SaaSapalooza" will likely result in market dominance for companies closest to specific problems rather than those with superior distribution, while existing "AI coworker" concepts are viewed as transitional mistakes similar to "horseless carriages," and enterprise software will remain stagnant unless the reliability barrier is overcome.
- The Total Addressable Market for data will expand from LLMs to "everyone" as organizations seek to "own their own intelligence," though owning proprietary models is deemed a "futile exercise" for most, with value instead accruing at the code level through deep engineering.
- The coding agent market is forecast to dwarf standard LLM token consumption due to a potential "mega Jevons paradox," provided "real automation" succeeds; conversely, robotics is predicted to be the "most overhyped" category as it remains furthest from the reliability prerequisites necessary for success.
- Raw intelligence per token is improving only "minutely," with most task-specific gains coming from optimizing data "in distribution" rather than intrinsic model growth, and the primary bottlenecks to widespread automation will be the need for "software researcher" roles and increased serving efficiency.
- Principal-agent problems in society are expected to be solved within the next five years by AI, enabling the enforcement of the "spirit of the law," while the concept of "software research" at the intersection of AI and new interfaces is identified as the industry's most underhyped direction.
- Investment strategies will prioritize "allies" and "smart money" over maximum valuation to prepare for an eventual AI "bubble burst," with teams aiming to attract "obsessive" and "technophilic" hires capable of delivering "technological shock and awe" rather than relying on mercenaries.
- Current investment in RLHF and RLVR is expected to remain "immensely valuable" despite fluctuating levels, and synthetic data will serve as a targeted tool rather than a universal solution, requiring an "extreme laser focus on the right task" to be effective.
- Companies like Harvey and Cognition are viewed as "thin wrappers" on LLMs that succeed due to specific domain functionality, while the era of pure compute scaling is expected to plateau, marking the beginning of a period where "data matters a lot more than compute."