陈天桥从发现式智能转向可验证系统的实践路径

NextFin News — Chen Tianqiao did not wait for another parameter race to define his AI bet. He defined it first as discovery, then as proof.

In late October 2025, at the inaugural Chen Institute Symposium for AI Accelerated Science (AIAS2025) in San Francisco, the Shanda founder formally launched what he called “Discoverative” Intelligence—an explicit contrast with generative systems that mainly recombine existing text and images. Shanda’s official announcement framed the goal as machines that pose new questions, uncover causal structure, and produce knowledge, not only fluent output. He paired the slogan with Structured Temporal Intelligence (STI): sustained neural dynamics, long-term memory, causal reasoning, world modeling, and metacognition. Scale, he said, is the engine; structure is the steering wheel.

The same release attached concrete commitments: more than $1 billion for compute prioritized for structural experiments, research hubs spanning Silicon Valley, Tokyo, Beijing, Shanghai, Hong Kong, and Singapore, an STI evaluation suite, and pathways for early-career principal investigators. That document is the primary record of the 2025 vision. It does not describe Apodex product lines or a verification-first operating system.

A year later the emphasis hardened. At AIAS+ 2026 in San Francisco, Chen delivered a special lecture titled “The World That Can Verify.” The Chen Institute’s account of the day placed independent checks, experiments, and human judgment at the center, and linked the argument to Apodex’s work on AI-training-AI and drug research. Secondary coverage of the talk distilled a sharper line: models have grown skilled at answers that look right; they must now learn to show that those answers hold.

The 2025 and 2026 messages fit together if discovery is the aim and verification is the method. Generation without structure drifts. Structure without external checks still hallucinates. Chen’s public writing has treated verification as a time-compression engine: it cuts false paths before a million parallel solvers repeat the same error.

Apodex is the operating vehicle for that second stage. Company channels describe long-horizon research agents, evidence graphs, and tools meant to investigate rather than chat. In August 2026 Apodex released TRACES, a reality benchmark of executable scientific environments with hidden verifiers, and reported gains over published methods on adeno-associated virus (AAV) capsid tasks spanning viability, tropism, structure, and constrained generative design. An associated technical paper situates the work as “discoverative AI” evaluated through extended, stateful investigations—not leaderboard prompts alone.

Drug research is where the company has been most concrete with outsiders. In October it said it would open AI drug-research services at no charge, starting with clinical-trial forecasting (predictions locked before outcomes are public), repurposing, interaction analysis, and safety assessment. Chen’s own post stressed scientific support and tool development, not idle compute, and invited research teams to say what should be built next.

Secondary reports of the November lecture add further operating detail—scientist headcount, peak GPU scale, solver-versus-critic software-repair pilots, a large AI-designed AAV sequence library moved toward synthesis—that do not appear in the 2025 Shanda release and should be treated as media reconstructions of the 2026 talk until Apodex publishes primary results. On the wet-lab side, design and synthesis are not efficacy. Chen’s own standard requires the distinction.

Press interviews in 2026 described a larger personal commitment to AI—on the order of $2 billion already deployed, with room to add more if outside capital lags—and a geographic weight toward the United States and Singapore. Those figures sit alongside, and should not be collapsed into, the 2025 structural-compute pledge of more than $1 billion. Revenue and break-even targets attributed to Chen remain forward-looking management goals, not audited results.

None of the capital answers the scientific question he raised. Sequencing costs fell for decades while new-drug productivity stayed stubborn. Structure prediction advanced when the field built blind tests and shared data, not when models merely grew. Apodex is trying to industrialize the missing half: critics that do not share the solver’s blind spots, forecasts that can be scored later, and lab loops that fail claims in public.

The industry can absorb another slogan. It cannot absorb another wave of fluent, unchecked claims. Chen’s useful sequence is clear enough. In 2025 he argued that AI should discover. In 2026 he is arguing that discovery only counts when something outside the model can verify it. The next scoreboard is not another benchmark average. It is whether locked predictions, wet-lab readouts, and independent checks start to move in the same direction.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App