【年终复盘②】RSI可能是NeoLab的机会、PA真的有用、创业要造灭火器还是砖头？｜对话谷雨

Summary

This 十字路口Crossing interview asks 谷雨 of New York Cognition to revisit ten AI judgments that changed during 2026. Its strongest technical contribution is Apprentice Bench and Sequential Workplace Learning Evaluation: long-horizon intelligence is framed as learning across successive production tasks, feedback, rules, and organizational context rather than merely running one task for a long time. The episode also separates business cognition from computer-use execution, argues that AI may give greenfield RSI startups more room to rebuild research infrastructure, and connects research taste to commercial closure without treating founder forecasts or company claims as established fact.

Key Claims

• 谷雨 says New York Cognition has 17 full-time staff, four interns, and about US$40 million in seed financing; these are guest-reported company facts rather than independently audited disclosures.
• The guest revises his prior skepticism about startup-led RSI after seeing small teams build strong models quickly, arguing that AI-assisted research and the absence of legacy infrastructure can partly offset incumbents’ scale while leaving compute, capital, and data constraints intact.
• Agent Operational Cognition separates the “hand” of reliable software and web operation from the “brain” of understanding company rules, workflows, exceptions, and task intent.
• Apprentice Bench evaluates sequential workplace learning through 100 construction-accounting tasks whose later solutions can depend on feedback and knowledge from earlier tasks.
• Continual Learning is presented as a frontier-model differentiator; memory-note quality and long-session interaction data may help, but the guest says the training recipe remains unsettled.
• The guest revises his view that high-level AI mathematics must depend on formal systems such as Lean, citing reported progress in informal reasoning while preserving formal verification as a distinct reliability mechanism.
• Memory notes are presented as more scalable than he previously expected because stronger models can write and use them better; this does not establish that external notes replace parameter learning, provenance, or precise retrieval.
• The guest argues that AI conference reviewing and paper counts are losing signal under submission growth and low-quality generated text, motivating evaluation based on demonstrated understanding, reasoning, judgment, and contribution.
• Technical-Commercial Closure requires a research company to map a technically coherent system onto a real user problem, product form, and business value, then release early enough to obtain market evidence.
• The discussion links AI video, personal agents, education, and the relative value of intelligence to taste, direction choice, emotional value, persistent interaction, and personal motivation; these remain broad judgments rather than measured forecasts.

Key Quotes

“手已经基本解决” — the guest’s shorthand for computer-use execution becoming less limiting than task and organizational cognition.

“从一个新人变成一个对这个公司非常熟悉的老员工” — the employee-development analogy behind Apprentice Bench’s sequential-learning design.

“灭火器与砖头” — the source’s metaphor for a simple urgent solution beating a polished product when customer pain is acute.

Connections

• 谷雨, New York Cognition, and Apprentice Bench - guest, company, and continual-learning benchmark.
• Continual Learning, Sequential Workplace Learning Evaluation, Persistent Agent Memory, and Research Taste - experience accumulation, memory notes, and judgment formation.
• Computer Use Agent, Agent Operational Cognition, Enterprise Operational Memory, and World Models - execution versus organizational understanding.
• Recursive Self-Improvement, AI Research Feedback Compression, and Frontier Model Scaling - startup opportunity and AI-assisted research-infrastructure thesis.
• AI For Math, Auto-Formalization, and Lean - informal reasoning progress and the separate role of formal verification.
• AI Academic Evaluation Signal Collapse, AI Academic Integrity, and Frontier Model Peer Review - research-quality and evaluation-system concerns.
• Technical-Commercial Closure, Lean Startup Feedback Loop, and New Lab Organization - path from technical coherence to market evidence.
• Personal Life Agent / 个人生活智能体, Personal AI Memory, Proactive Agents, and IM Agent Interfaces - persistent personal-assistant product design.
• AI Video Production Workflow, Human Taste Creative Boundary / 人类审美的创造边界, and Intelligence Devaluation - falling execution cost and the source’s claim that taste and emotional value become relatively scarcer.

Contradictions

• No settled factual contradiction is adopted.
• The claim that computer-use “hands” are nearly solved is qualified by existing wiki evidence on reliability, permissions, latency, privacy, behavioral verification, and human-review boundaries.
• Reported model capabilities, company headcount and financing, benchmark design and scores, conference-system decline, startup valuations, video consistency, provider rankings, and next-year RSI forecasts remain source-scoped because the supplied file is a structured summary rather than a transcript, benchmark release, audited filing, systematic study, or independent evaluation.
• The source’s optimism about memory notes does not erase the existing distinction between compressed behavioral learning and exact, inspectable, source-grounded recall.