A 27-year-old researcher resigned because of it, a legend who worked at Google for 27 years started a business because of it. Both see the same thing.

On September 8, Pacific Time, Jacob Coxon, who spent three years in pre-training at OpenAI and Anthropic, announced on X that he resigned from Anthropic with a series of seven posts. The first post contained just a few sentences, with the core being one judgment: Both companies are rushing towards self-improving superintelligence. The post's readership went from the initial screenshot of 338,000 to over a hundred million.

Rewind one month, on August 5, Jeff Dean left Google, where he had been for 27 years. He was Google's 30th employee, and behind MapReduce, Bigtable, TPU, and Gemini is his work. He, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le founded Discovery Loop in Palo Alto. The stated farthest goal of this company: To create stronger AI using AI.

One left because of it, one started because of it. The direction both pointed at is exactly the same: recursive self-improvement, Recursive Self-Improvement, commonly referred to as RSI in the industry.

Let's put the conclusion here. RSI is the only capability within large models that can "generate the ability to generate ability". Writing code, doing math, translating, drawing pictures are products; RSI is the machine that produces these products. The one who closes this loop first will shift progress speed from "how fast humans run" to "how fast electricity and chips run." So it is the jewel on the crown, and that’s why it is hot to the touch.

The following eight sections explain everything from underlying principles to money.

01

Other abilities are eggs, RSI is the hen that lays the eggs

Around 1800, the British man Henry Maudslay created a lathe that could cut standard threads. Before that, every screw was handcrafted by artisans, and one bolt wouldn't match another nut. The lathe itself was not sold to the public, but with it came standard parts, which resulted in interchangeability and later mass production of steam engines and textile machines. The Chinese industrial sector aptly named this type of equipment: Mother machine, a machine that can create machines.

A similar story exists in the software world. The first C language compiler could only be written in assembly. Once it was operational, engineers rewrote the compiler in C, using the old one to compile the new one. From that day forward, compilers began compiling themselves. The smarter the compiler gets, the faster the next version it produces becomes. This process has a specific term called bootstrapping.

What RSI does is to transfer bootstrapping to intelligence itself: allowing models to participate in the design, training, and evaluation of the next generation of models; as the next generation becomes stronger, it goes back to create the next generation.

From a first principle perspective, the value structures of ordinary abilities and RSI are fundamentally different. If a model can write code, what you save is the salary of the programmer, however much that is, which is just addition. RSI operates on "the speed of creating models," and the speed of model creation determines how fast all other capabilities can improve. One governs stock, while the other governs the acceleration of acceleration.

In a lengthy article released on June 4 this year, Anthropic illustrated this path as a continuous spectrum: starting with humans writing code in notebooks, then a chatbot helping to generate snippets, moving to coding agents that can edit entire files themselves, and today to autonomous agents that can run code and delegate hours of work to other agents. The far right of the spectrum is labeled "closed loop": agents build and train the next generation of models themselves. For that box, they wrote "20XX?"

Popular interpretation

You opened a noodle restaurant. Hired a fast chef who sells two hundred more bowls a day; that's ordinary ability. But if you find a particularly good mentor chef, the apprentices he trains will train other apprentices even better, and then their apprentices will be stronger—three generations later, your concern won't be whether noodles sell well anymore, but whether you have enough flour and enough stove space. RSI is the chef that "can train apprentices who can train even better apprentices."

02

All the secrets of recursion lie in one exponent

Writing RSI in the simplest math form takes just one line. Let the intelligence level of the model be I; it influences the speed of the next generation's development and is related to its own level:

dI / dt = k · Iα

The entire debate, at the end, is merely about this α.

If α is less than 1, the deeper you go, the harder it gets, like mining; the deeper the mine, the slower the output, and growth will slow down on its own. If α equals 1, it is standard compound interest, experiencing exponential growth. If α is greater than 1, mathematically it will rush towards infinity in a finite time; the "singularity" referred to in science fiction means exactly this.

α cannot be measured directly, but it leaves fingerprints on the data. The core metric of the evaluation organization METR is: how long tasks would take for humans if a model can independently complete tasks with a 50% success rate. The data cited by Anthropic is: in March 2024, Claude Opus 3 could tackle a software task that would take humans about four minutes; a year later, Claude Sonnet 3.7 could handle approximately one and a half hours; another year later, Claude Opus 4.6 could complete it in 12 hours. The METR evaluation for Claude Mythos Preview is "at least 16 hours," reaching the upper limit they can measure with their current question bank.

More importantly, there’s another number: the doubling cycle of task lengths has shortened from about seven months in the early days to about four months now. The doubling itself is speeding up; this is exactly the kind of fingerprint that α would leave behind as it climbs.

The lab is also aware of where this line is drawn. In OpenAI's safety framework, the highest risk tier of "self-improvement" is defined very specifically: either an agent surpasses human researchers, or it can complete a generational leap (like from o1 to o3) in one-fifth the time it takes at the 2024 pace, and maintain that speed for several months. In other words, the threshold set by the industry for RSI is "five times speed in R&D, and not slowing down".

Popular interpretation

Everyone understands the concept of saving money: a 5% annual return versus a 50% return—after ten years, the difference is astronomical. However, in regular compound interest, the interest rate is fixed. The uniqueness of RSI lies in the fact that the interest rate is rising every year. This year it's 5%, next year it's 8%, the year after it’s 15%. You don’t even need to calculate to know that this curve will inevitably change from "steep" to "vertical." What insiders are watching is whether the interest rate increase stops.

03

The hands writing code have changed to AI; only the brain remains human

If we break down AI research into its individual parts, there are four stages: topic selection, experimental design, execution, interpretation of results. For RSI to close the loop, all four stages have to be handed off. Where are we now? The long article from Anthropic provides the most detailed data to date.

Internally at Anthropic: the automation progress of research stages (data from the company's June 4 article)

Reading this table, the conclusion is clear. The execution stage has basically been handed off; in experiment design, more than half has been handed off; even in determining "what step to take next," the AI's success rate has increased from half to nearly two-thirds in just a year.

What remains? Topic selection. Deciding which problem is worth tackling, which results are worth trusting, when to admit that a path is a dead end. Anthropic’s own words mean that the current comparative advantage of humans lies in research taste and judgment.

There is an old rule at play here, known as Amdahl's Law: if you speed up one part of a process, the overall speed will be bottlenecked by the part you didn’t accelerate. Anthropic has already hit this wall: code generation is too fast, making human code review the new bottleneck; too many new ideas and tools emerging from employees and models means the company lacks the energy to pursue each one.

Thus, today’s RSI represents a semi-closed loop where “hands have changed, but brains have not.” The final gap that remains is called taste.

Popular interpretation

Imagine a top restaurant kitchen: slicing, cooking, plating, washing dishes have all been replaced by machines, which do it faster and more reliably than humans. The chef only has one job left each day—deciding what to put on tonight’s menu. The thing is, how good a restaurant is largely depends on the menu. Today's large model research stands in that kitchen.

04

The hardest jewel to pick from the crown is called "taste"

In August of this year, Peter Kirgis and Sayash Kapoor from Princeton designed a challenging test they called "shadow evaluation." They selected two unpublished papers submitted to NeurIPS 2026 and tasked Claude Opus 4.8 running on the open-source framework OpenClaw to answer the research questions posed by these papers. They were given six days, a $3,000 API budget, GPU budget, independent virtual machines, and full internet access to produce a paper that could qualify for a top conference. In the end, both papers were rejected.

Interestingly, the reasons for rejection were illuminating. The agent could handle the engineering tasks of reading literature, running hundreds of experiments, and summarizing results. What it stumbled on was the research itself: good hypotheses were abandoned with minimal data; unreliable directions were pushed too soon; when entering a dead end, only minor adjustments were made, with no way to start over. Kapoor's explanation cuts straight to the point: models excel at tasks that reinforcement learning can be repeatedly trained on; and reinforcement learning thrives on tasks that can be automatically checked.

This statement merits a deeper look at first principles.

Silicon Valley’s Alan Walker has long stated that AI and cryptography share the same foundation: hard to generate, easy to verify computational asymmetry. Code passing tests and mathematics being verifiable means models progress rapidly in these areas because each attempt gets an inexpensive referee to tell them if they are correct. Research taste is precisely the place where verification is difficult. Whether a direction is good often requires two or three years, spending tens of millions of dollars on computing power, to recognize in hindsight. When the Transformer paper came out in 2017, no one could verify on the spot that it would become the foundation of the entire industry for nearly a decade.

Referring back to the "three properties of feel" I mentioned earlier, taste is the feel of top researchers: non-replicable, one cannot learn how he chose topics simply by reading their papers; non-verifiable, no one can judge if their intuition is correct in real time; will depreciate, once the paradigm shifts, old feelings become outdated immediately. This also explains why the valuation of the founders of Neolab essentially translates to the pricing of their feel.

Even Anthropic co-founder Jack Clark has written in his newsletter that today's AI lacks valuable intuitive creativity, being strong in engineering capabilities yet thinking in a formulaic way; he refers to this as a bearish signal for "short-cycle RSI."

But this precisely indicates the position of the jewel. Why the jewel in the crown is set at the highest point is because it is the last one to be conquered. And the offensive has already begun: Anthropic believes that most of the progression in AI never relies on sudden insight, but on "amplifying, finding what's broken, fixing it, and coming back," with the 99% sweat Edison referred to now being automated in batches. That increase from 51% to 64% in the "next step judgment" represents the first crack where taste is beginning to loosen.

Popular interpretation

Driving tests have standard answers: touching the line means point deduction, stalling means point deduction, which is why driving schools can train qualified drivers in batches. However, "will this road be congested next month, should I change routes now," no one can score you on the spot; you only find out afterward. The model is like a perfect driver now, but lacks the "road sense" of an experienced driver. And road sense is precisely the hardest element to be taught through exams.

05

Jeff Dean has worked for 27 years; why start a company with only four people?

To understand Discovery Loop, one must first comprehend what Jeff Dean has been doing his entire life.

MapReduce from 2004 did one thing: it broke apart calculations originally run serially on one machine and distributed them across thousands of machines to run simultaneously. Bigtable, TensorFlow, TPU are essentially variants of the same action—shifting the bottleneck from "how fast a single node can run" to "how many nodes can run in parallel".

Looking at Discovery Loop's public practices: it uses algorithms to simultaneously initiate and iterate through thousands of experiments, automating the part of the scientific research cycle of "making hypotheses—doing experiments—looking at results"; the first area it tackled was machine learning itself and later expanded to drug development, chip design, and clean energy. This exact thinking from MapReduce has now been applied to the scientific discovery cycle.

Why is traditional research slow? Because it's serial, and it bottlenecks on people. A diligent Ph.D. student may complete only a few dozen experiments in a year. Once this cycle can be automated, the bottleneck shifts from how many Ph.D.s are hired to two factors: how much computational power is available, and how well the cycle is designed.

Thus, four people are sufficient. Dean made it clear in a Stanford lecture on August 7: with cloud computing so widespread, small teams can directly rely on Google and other cloud providers to set up the infrastructure, and a small team itself is the optimal form for accelerating innovation.

Capital is very clear on this maths. Discovery Loop’s first round was led by Radical Ventures and Khosla Ventures, with Lightspeed, Kleiner Perkins, and Doerr Capital participating; Alphabet is a founding investor and cloud partner, reportedly also covering the company's computational expenses for the first year. Business Insider reported that Dean is negotiating regard<|endoftext|>

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。