Topics Pioneers Insight Method Author
Back

Year-End Review ② with 谷雨: RSI’s Potential for NeoLab

·十字路口Crossing

Rendering the long image…

Topics Frontier Scaling Laws and Superintelligence End-to-End Desktop AI Agent Control AI Venture Funding and Bubble Risks Unique Human Value in Everyday AI Personal AI Assistants Adoption and Monetization
Pioneers 谷雨

Summary

  • 谷雨 has revised his view: RSI startups may have a shot, with new labs’ edge coming from iteration speed rather than an existing stockpile of models. As AI gets better at writing training infrastructure and recipes, the systems built up by large labs could become a drag on adaptation, while new teams can rebuild from scratch at any time. “I got proven wrong.” But full-stack training still takes substantial capital, and his call that an RSI startup might launch the strongest model next year—or at least a SOTA-level model—was explicitly caveated as a “wild guess.”
  • Computer-use reliability has improved sharply, but business knowledge is a separate challenge. As coding training scaled up, 谷雨 says “the hands are basically solved”; what remains is understanding company rules, workflows, and approvals, a problem distinct from operating software. Neocognition’s ApprenticeBench strings together 100 accounting tasks to test learning from feedback. The real long horizon, he says, is an employee going from novice to veteran over several months or 1 or 2 years.
  • Two advances—in math and memory notes—have both overturned 谷雨’s view of what models can learn. He once thought math needed signals from formal systems and that rewriting memory notes was not a reliable optimization algorithm. Now he concedes that scaling can go further than he expected: “I have to believe it.” In his view, though, Anthropic and OpenAI are still the main players doing continual learning well, and the training recipe is not mature.
  • 谷雨 thinks paper counts can no longer serve as a reliable talent screen; hiring should focus on candidates’ work and grasp of the problem. He has not published in 1.5 years and has not reviewed a paper in more than 6 months; in one batch of 16 papers, perhaps only 1 or 2 were good enough. His team looks at how candidates reason, respond to questions, and understand the problem in research presentations. That is not a reason to rule out a PhD, he says: “Wanting to do something and doing it is a form of rationality.”
  • The core startup lesson is to close the loop between technology and the market—and recognize that a high launch bar can itself be costly. An academic team may want to build a solid fire extinguisher, but a customer with a fire burning might buy a brick first: “By the time you finish building your fire extinguisher, the brick sellers will have sold 100.” 谷雨 admits he was too earnest and too slow to launch; Koji’s caveat is that it is fine to build momentum early, but you have to keep delivering: “You definitely can’t bluff.”
  • Abundant capital is turning entrepreneurship into a choice people make by following the crowd. Neocognition raised $40M in its seed round; 谷雨 says typical Silicon Valley seed rounds are $5M or $10M. In the first half of 2026, some domestic new labs were already reaching $2B valuations, which shocked him. Capital lowers the bar to trying, but the question raised in the conversation still stands: even without personal financial pressure from buyback or performance-guarantee clauses, you could still spend several years on it.
  • AI video is starting to compete on the work itself, not on viewers’ curiosity about the technology. 谷雨 watched a zombie film on Bilibili that ran for more than 2 hours, with 7 or 8 characters staying consistent, and says, “I genuinely watched it as a movie.” Viewers knew it was AI-made, but the comments mostly focused on the story. He sums up the scarce human contribution as “taste, and beauty”: choosing a direction and providing emotional value.
  • A personal agent is already saving 谷雨 time on email, with the breakthrough coming more from product design than new technology. Persistent one-on-one conversations, human-like avatars, and proactive updates bring existing capabilities together into an assistant experience; proactiveness itself is not new. He is trying Muse and Instinct, too, but thinks that once people rely on an agent in daily life, “one is enough.”

Deep dive

1. RSI gives new labs a renewed shot at the frontier

  • 谷雨 had 2 reasons to doubt RSI startups: new companies could not catch up with frontier models, and recursive improvement needed a model of their own to start from. Releases from teams such as Recursive and MiroMind changed his mind. He cited a pre-trained model built in just over 3 months that was stronger than K3: “I got proven wrong”(我确实被打脸了).
  • His explanation is that as AI gets better at doing R&D, the infrastructure and recipes accumulated by large labs could become a liability. A friend at a domestic model company knew OpenAI’s score centering could improve results by 1 or 2 points, but did not adopt it because adapting the infrastructure would cost too much. New teams do not have that baggage: they can throw things out and start over, giving them “more acceleration.”
  • Koji asked whether full-stack training still takes money, chips, and data. 谷雨 agreed that the capital barrier remains. He also made a separate prediction for next year, explicitly calling it a “wild guess”: an RSI startup might produce the strongest model, or at least one at the frontier and at SOTA level. He was not saying it would definitely happen.

2. Computer use’s “hands” have improved; the hard part is shifting to the “brain”

  • 谷雨 says coding training scaled up from late last year into early this year, with the capability generalizing to software and web tasks: “the hands are basically solved.” The remaining cognitive challenge is whether a model knows how a task should be handled, especially when it involves a company’s specific business processes. Cognition and operation are distinct capabilities, though ultimately the work still has to be carried out in software or a browser.
  • The team spent months building a pipeline that had an agent repeat a task, selected the correct result, and packaged it as a reusable skill. He now sees that engineering as unnecessary. Opus 5.5, for example, can implement a workflow directly and correct itself; progress in the operational layer has changed the team’s original approach to solving the problem.
  • ApprenticeBench uses accounting and expense-reporting tasks in the construction industry. Its 100 tasks involve company conventions, specific workflows, and approval rules. The key is that the tasks are not independent: the model must complete all 100 in one stream, using feedback and knowledge from earlier tasks to handle later ones.
  • 谷雨 argues that long-horizon work should not be reduced to finding a difficult task that takes half a day. Such tasks are hard to scale and may not lead to lifelong intelligence. “The real long horizon” is not an accountant spending a week on one transaction, but a novice becoming an experienced employee who knows the company over several months or 1 or 2 years.
  • The strongest Claude model on the first version scored about 80, 谷雨 says; the new version will push the current best model’s score much lower. The test data is not public, but the more important target is meta-level learning ability. There is no point in training only on one company’s rules to improve its score; if continual learning really improves, he agrees, “that would count as doing something worthwhile.”
  • He also mentioned that the team worked on benchmarks including MMMU, Mind2Web, and AgentBench while at OSU NLP Lab. Later versions or variants of MMMU and Mind2Web have also been used by model companies in launch evaluations.

3. Math and memory notes both broke through the limits he had set

  • In 2024, 谷雨 still thought math required signals from formal systems. By 2025, informal methods could earn an IMO gold medal; by 2026, some problems from previous years could largely be solved informally. That broke with what he believed 2 years earlier and convinced him that scaling still had room to run: “I have to believe it.”
  • Earlier this year, he also said that continual learning through rewriting memory notes “definitely wouldn’t work,” because it was not a reliable optimization algorithm. He still does not consider it a reliable optimization algorithm, but concedes that he underestimated its ceiling: the quality of memory used to be limited by model capability, and the stronger the model, the further this approach might go.
  • In 谷雨’s view, Anthropic and OpenAI are the main companies doing this well. Training models to write better memory notes could help; long sessions in Claude Code and Codex might also generate relevant data naturally, but the recipe is not mature. He characterizes 2024 to 2025 with the keywords vision-language, VLM, and world models; late 2025 through 2026 with coding; and 2026 through next year—at least until the second half of next year—with continual learning.
  • Koji offered another way to think about the value of solving math problems: the process can uncover methods and build know-how, and might become a way for people to test the limits of their intelligence, as in chess. 谷雨 replied that many things are not valuable only because machines cannot do them; doing them can itself give people meaning and pleasure.

4. Paper-based evaluation is failing, but that does not make a PhD personally meaningless

  • 谷雨 described the 2026 submission and evaluation system as having “completely collapsed.” He has not submitted a paper in 1.5 years and has not reviewed papers or served as an AC for more than 6 months. In one batch of 16 papers, perhaps only 1 or 2 were genuinely good enough; the volume of AI slop made him feel that this work did not respect his time.
  • When hiring, he pays more attention to how candidates understand their past work and the problem at hand, as well as their reasoning, responses to questions, and overall ability in research presentations. He would prefer peer review by people with taste—for example, senior faculty evaluating a candidate through a job talk—rather than relying only on paper counts. A new evaluation system has not yet emerged, and universities continue to run on inertia.
  • On why AI talent is giving faculty jobs lower priority, he points to practical incentives: frontier labs and new lab startups offer better pay and resources, while universities also have an evaluation problem. Koji joked that “10 years from now, a professor may have less social status than a male model in a nightclub.” 谷雨 said he genuinely thinks so, and that when intelligence is no longer valuable, looks and physique may be worth more.
  • Asked whether a PhD is worth pursuing, 谷雨 refused to decide for someone else: without knowing what they want, there should not be a one-size-fits-all answer. “Wanting to do something and doing it is a form of rationality.” His advice is to first make sure you genuinely want it, rather than treating career returns as the only measure.

5. Startups need a market loop; a brick can beat a fire extinguisher

  • 谷雨 says he has “not once in my life wanted to work at a big company.” He needs a flexible schedule and ownership over his research direction, and finds it hard to carry out someone else’s mission. He says he has “no concept of risk”; his parents never imposed life milestones on him, which also gave him a degree of security.
  • The team had always avoided looking for nails with a hammer in academia. After starting a company, however, they realized they had become “solutions in search of problems.” Research can be self-consistent within a technical system, but a company has to close the loop with the market. Continual learning is a fundamental problem, but that does not mean they have found a product customers will pay for.
  • He reflects on launch timing with the image of someone whose head is on fire: customers will not wait for the perfect fire extinguisher; if there is a brick nearby, they will use it to put the fire out. “By the time you finish building your fire extinguisher, the brick sellers will have sold 100”(在你灭火器造出来之前,别人卖砖头的已经卖100个). The team was too focused on building something solid and should have launched a less mature product earlier to get the market loop going.
  • 谷雨 says sometimes you can even talk big first; Koji immediately adds the constraint: announcing early can create momentum, but you have to keep delivering and cannot bluff. 谷雨 cites Elon Musk as a classic example of “fake it till you make it.” The conversation leaves the tension between launching early and fulfilling promises intact.

6. Abundant capital is making entrepreneurship a follow-the-crowd choice

  • Neocognition raised $40M in its seed round and now has 17 full-time employees and 4 interns; it plans to start a Series A soon. 谷雨 says $5M or $10M is a typical Silicon Valley seed round, while in the first half of 2026 some domestic new labs were already reaching $2B valuations. He says it was not until June that he realized how much money was being put into companies in China.
  • What surprised him most was hearing people say, “Everyone is starting a company; everyone around me is starting one, so I want to start one too.” A path that was once outside the mainstream had become a choice people followed. Koji affirmed the value of capital giving young people the chance to try; 谷雨 asked, “Is there really no pressure?” The discussion distinguished personal financial pressure from buyback or performance-guarantee clauses from the risk of spending several years on the effort.

7. AI films are beginning to hold audiences on the strength of the work

  • 谷雨 watched an AI zombie film on Bilibili by creator 龙小胖, bringing 林正英 into a new story. It ran for more than 2 hours and had 7 or 8 characters whose faces, clothes, and storylines stayed consistent—a different experience from six months earlier, when characters could change appearance within minutes. Viewers could sink into it and watch it as a movie.
  • Koji asked whether viewers knew it was an AI-made film. 谷雨 said the platform disclosed that, but the comments mostly discussed the plot. To him, that means viewers care more about the final deliverable than how it was made: the experience has shifted from curiosity about the technology to immersion in the story.
  • He also recalled an AI short drama from last year with 60 episodes, each 1 minute long, for which the script was quoted at about RMB70,000. He thinks AI could now write 100 scripts in a day. His rough-cut view is that people will need “taste, and beauty”: the former sets the direction; the latter provides emotional value.

8. A personal agent’s value comes from an ongoing relationship and proactive filtering

  • 谷雨 thinks personal agents are gaining traction more because of good product design than because a technology suddenly crossed a threshold. A true personal assistant needs one continuous conversation, rather than a new session for every task—like WeChat, where each contact has only one chat window. A cartoon avatar further strengthens the sense that there is someone there.
  • In his view, proactiveness is not new technology either: OpenClaw had a similar feature early this year, and scheduled tasks in ChatGPT and Claude can also check email on a schedule and send updates. The change is that these capabilities have been packaged into a more natural assistant experience.
  • He is trying Muse and Instinct at the same time, mainly to filter email, pull out the important items, and proactively alert him. Services for food delivery, hotels, and flights are not very good in China or abroad, limiting other uses. For now, he uses several mostly to try them out; once he depends on one in daily life, he thinks “one is enough.”