AI Trends

Dario Amodei: The AI Pioneer Behind Scaling Laws and Claude Code

36Kr · AIUpdated Nov 20 16:18

AI summary

The article traces Anthropic founder Dario Amodei’s journey from helping develop scaling laws and GPT-3 to advancing Claude, Claude Code, and AI safety governance. As Anthropic prepares to go public, Claude-assisted R&D is creating an accelerating feedback loop, while also sharpening tensions among growing model capabilities, commercial competition, and risk control.

Why read this

SelectedAI score 92

Connecting scaling laws, Claude Code, and debates over safety governance, the article shows how Amodei is shaping Anthropic while balancing expanding capabilities against risk control.

Topics

Tags

Body · AI translation

On November 19, U.S. time, Anthropic officially debuted on Nasdaq under the ticker symbol “ANT,” setting a record for the largest IPO in global history.

Just one day before the listing, Anthropic priced its shares at $83 each and increased the offering from 825 million shares to 980 million, ultimately bringing the public offering to $81.34 billion.

At the offering price, the AI company, founded less than six years ago, had a fully diluted valuation of approximately $1.5 trillion.

In its prospectus, a steep revenue growth curve appears alongside astronomical commitments to computing infrastructure. Countless eyes are fixed on the company as the market scrutinizes it—and waits to see what it will do.

At the very center of the spotlight is Anthropic’s founder and CEO, Dario Amodei.

During public appearances, the 43-year-old leader invariably wears a simple casual shirt and sports his signature mop of short, tousled hair. Rather than the CEO of a trillion-dollar company, he looks more like a scientist who has just stepped out of the laboratory.

After the Transformer laid the technological foundation, the development of large models underwent two particularly striking turning points.

One was the establishment of Scaling Laws, which made model scaling a technological path that could be pursued continuously. The other was the emergence of coding capabilities, which enabled machines to take over human work and, in turn, accelerate the development of the models themselves. As these two waves converged, Amodei stood at the bow.

1

On February 14, 2019, OpenAI published a news story.

In a remote valley in South America’s Andes Mountains, evolutionary biologist Dr. Jorge Pérez of the University of La Paz and his team had discovered a mysterious creature never before documented by humans. The creatures had four horns, were covered in silvery-white fur, and could even speak fluent English.

The researchers named the creature “Ovid’s Unicorn.” In the report, Dr. Pérez solemnly explained that it might have originated in Argentina and could be connected to a lost ancient population. As for where it truly came from, only DNA testing could probably provide an answer.

“Ovid’s Unicorn”? How could such a thing possibly exist? Yet the news story was remarkably convincing, complete with the scientist’s name, a research institution, descriptions from the scene, and quotations from experts.

The report was produced by GPT-2, the language model OpenAI unveiled that day. Researchers wrote only the first two sentences; GPT-2 generated the rest on its own.

That day, Wired reported on the achievement under the headline “The AI Text Generator That’s Too Dangerous to Make Public.” The Verge summed up the new model’s abilities with a rather ironic headline: “OpenAI’s new multitalented AI writes, translates, and slanders.”

By today’s standards, GPT-2 was far from intelligent. Its largest version had 1.5 billion parameters and could perform simple translation, question answering, and text summarization, but it frequently repeated itself or went off topic in longer pieces and even invented absurd scenarios such as fires breaking out underwater.

Even so, OpenAI decided not to release the weights of the largest model for the time being. The company feared that technology capable of mass-producing false content could be misused, so it opted for a staged release, giving researchers and society time to assess the risks.

Inside OpenAI, however, Dario Amodei saw another, much more distant possibility.

In a 2026 interview with biographer Walter Isaacson, Amodei recalled his reaction to GPT-2 at the time:

“I personally was super freaked out by it.”

He said that around 90% of the people around him did not consider the model anything special. After all, its performance was hardly outstanding in translation, writing, or any other individual task.

But he had already begun to imagine that, if a model like this were scaled up another millionfold, machines might one day surpass humans at almost every task. When that happened, the way the entire economy, science, and even nations and societies operated could all be transformed.

In several subsequent public interviews, Amodei traced the fundamental perspective through which he viewed technological progress.

In 2006, Amodei’s father, Riccardo, died of hepatitis C. In the years that followed, the advent of a new generation of antiviral drugs brought breakthroughs in the treatment of hepatitis C. By the mid-2010s, some treatment regimens had achieved cure rates exceeding 90%.

Just before the opening bell rang, Amodei stood on the bell-ringing platform with his sister, Daniela, and once again spoke about their father.

“The disease he had became curable just a few years later,” Amodei said.

The experience changed the direction of Amodei’s research. Originally a theoretical physics major, he shifted into biophysics and computational neuroscience. While pursuing his doctorate in physics at Princeton University, he spent his days working with electrical signals in the brain’s neural circuits, trying to unravel the operating logic of living systems.

But the complexity of biological systems also gradually made him realize that, if left to human researchers alone, many problems might take an extraordinarily long time to solve.

He needed a new tool that could fundamentally compress the research cycle. This idea led him to turn his attention to artificial intelligence, hoping to harness machines’ computational power to accelerate scientific discovery.

In late 2014, Amodei joined Baidu, where he worked on speech recognition with Andrew Ng’s team. Deep learning had already made some progress at the time, but researchers were still exploring how to further improve model capabilities.

Amodei tried a brute-force approach: adding more layers to neural networks, expanding the training data, and committing more computing resources.

In a later interview with Lex Fridman, he recalled that although he had not yet precisely measured the mathematical relationship involved, he and his colleagues had already developed an intuitive judgment: more data, larger models, and longer training runs tended to produce better results.

At first, he thought this might be a pattern unique to speech recognition. It was not until he encountered early GPT models that he realized LLMs, or large language models, might be no different in this respect.

On January 23, 2020, the OpenAI team published a paper titled “Scaling Laws for Neural Language Models,” which would later reshape the paradigm of AI research and development. The research was led by Jared Kaplan and Sam McCandlish, while Amodei, as a co-author, provided research guidance for the entire project.

The paper’s most striking feature was several almost perfectly straight curves. On logarithmic axes, as model size, data volume, and compute increased, language models’ prediction error continued to decline along trends describable by mathematical formulas. Even at the largest scale covered by the experiments, the researchers observed no clear sign that the trend was breaking down.

This body of data provided crucial evidence for the development of LLMs, or large language models.

Previously, improving model performance had often required researchers to repeatedly adjust algorithms, architectures, and training schemes. The scaling laws allowed research teams to use smaller-scale experiments to project the gains that might come from adding parameters and plan larger training runs accordingly.

Scaling gradually evolved from a research judgment into an engineering path that could be quantitatively validated.

Amodei then took part in an even larger experiment. According to a Big Technology profile, OpenAI put Amodei in charge of the GPT-3 project and devoted roughly 50% to 60% of the company’s computing resources at the time to it.

A few months later, on May 28, 2020, OpenAI unveiled GPT-3.

It had 175 billion parameters, 116 times as many as GPT-2. The same model could switch between different tasks without having its parameters retrained for each one. Researchers needed only to provide a few examples in the prompt, and GPT-3 could follow them to perform tasks such as translation, question answering, and text generation.

As OpenAI gradually opened its API to external developers, more surprising results began appearing one after another. Some people had GPT-3 imitate the writing styles of different authors; others entered natural-language descriptions and asked it to generate functional webpage code. Capabilities that had previously remained confined to papers and laboratories began appearing in real software interfaces.

Although GPT-3 still suffered from errors, biases, and limitations, the generality it demonstrated was already enough to prompt a growing number of research institutions to reassess the future of LLMs. A race centered on parameter scale, data, and computing resources gradually began to take shape.

The judgment Amodei had formed during his early speech-recognition research was validated again and again through GPT-2, the scaling laws, and GPT-3. The technical path he had helped advance began to become a major direction for the entire industry.

But the more powerful the models became, the harder it was to avoid the issues that concerned him.

As the GPT series achieved breakthroughs, OpenAI began accelerating its commercialization efforts. The disagreements between Amodei and the company’s management over research direction, safety safeguards, and organizational governance gradually deepened.

According to a September 2026 investigative report by The Atlantic, Amodei had tried to resolve the issues through internal reforms. In negotiations with CEO Sam Altman, he proposed a series of conditions for staying, including giving the safety team veto power over the release of new products.

This demand meant that safety researchers would have the authority to prevent models they believed posed significant risks from entering the market, even if company management wanted to proceed.

The two sides ultimately failed to reach an agreement.

In December 2020, OpenAI held an all-hands video meeting. After Altman finished presenting the company’s routine updates, he handed the floor to Amodei. Amodei then announced to the entire staff that he and six other colleagues would be leaving OpenAI.

The meeting’s chat window was quickly flooded with messages of surprise. Amodei thanked his colleagues and wished the company well. He emphasized that both sides still shared the same goal: ensuring that safe artificial general intelligence benefits humanity.

On December 29, OpenAI officially announced his departure. The announcement reviewed Amodei’s contributions to GPT-2 and GPT-3, as well as his experience co-leading the company’s research direction with Chief Scientist Ilya Sutskever.

Anthropic was formally founded in January 2021.

2

Fast-forward to 2024. In the studio of renowned host Lex Fridman, Amodei spoke with particular seriousness about the research questions he knew well.

“You’d be surprised how many of the challenges involved in developing these models ultimately come down to software engineering and performance engineering.”

“Even incredible discoveries almost always come down to the details in the end. And often, they are extremely, extremely tedious details.”

After founding Anthropic in 2021, Amodei devoted substantial resources to developing foundation models. Training larger models, improving computational efficiency, and refining training methods have always been central to the company’s work.

As Claude’s capabilities continued to improve, Amodei noticed that the kinds of tasks the model could perform were changing.

Programming was one major area.

On the podcast, Lex told Amodei that he had recently been using Sonnet 3.5 with Cursor for programming and could clearly sense that the model had become smarter. Amodei smiled as he described the feedback from inside the company.

“Anthropic has several exceptionally talented engineers. In the past, programming models developed either by us or by other companies were of no help to them. They would say, ‘This might be useful for beginners, but it’s no use to me.’”

“But after Sonnet 3.5 came out, they said for the first time, ‘Oh my God, it helped me complete something that would otherwise have taken several hours. This is the first model that has genuinely saved me time.’”

“The water line is rising,” Amodei said.

That June, an internal agentic coding test published by Anthropic showed that Sonnet 3.5 could solve 64% of software engineering problems, compared with just 38% for the previous Opus 3.

The test required the model to fix bugs or add features in real open-source codebases based on natural-language descriptions.

In the latter half of the interview, Lex turned the conversation toward the future of software development.

“I think this will be one of the fastest-changing fields,” Amodei replied. “There are two reasons. First, programming as a skill is very closely connected to the development of AI. A large proportion of employees at Anthropic and other AI companies rely on programming in their work, so change will come quickly.”

“The other reason is that programming can form a closed loop. A model can write code, run it, inspect the results, and then understand those results. Compared with hardware or biology, programming makes this kind of feedback loop easier to achieve.”

Amodei even predicted that by 2026 or 2027, AI might already be capable of performing the vast majority of tasks in which programmers directly write code.

“When AI can do 80% of a programmer’s work, we will find that the remaining work actually becomes even more important for humans.” As examples, he said humans would take greater responsibility for high-level system design, checking whether software architecture was sound, and work involving design and user experience.

Lex pressed him: Given programming’s enormous potential, was Anthropic preparing to develop a programming tool of its own?

“Our idea is to let a thousand flowers bloom.”

Amodei explained that Anthropic was not currently trying to develop such an IDE itself. Instead, it was supporting companies such as Cursor and Cognition, enabling them to build their own products on the Claude API.

“We don’t have enough resources internally to try all these different directions. So let customers experiment and see who succeeds. Perhaps different customers will succeed in different ways.”

In Amodei’s view, although programming tools still offered vast room for exploration, at the time he preferred to focus the company’s efforts on improving the model’s capabilities. External developers could experiment with what exactly the model could be used to do.

This approach had already begun to pay off. After the release of Sonnet 3.5, usage of programming tools such as Cursor grew rapidly, and more and more developers began using Claude for real-world software engineering work.

Yet at almost the same time that this interview aired, a programming tool that even Amodei himself had not anticipated began quietly gaining popularity inside Anthropic.

That September, Boris Cherny, an engineer who had just joined Anthropic, wrote a simple terminal program to familiarize himself with the company’s API.

At first, it merely helped him control a music player, organize notes, and perform similar tasks.

“Pretty cool, but not very interesting,” Boris later said.

So he spoke with colleagues and experimented with adding tools to the program, allowing Claude to read files, modify code, and execute terminal commands.

In one experiment, Claude used the terminal tool to write a program by itself and retrieve information from the music player. Boris then asked it to try tackling problems in Anthropic’s codebase.

Claude opened a file on its own, followed a code reference within it to the next file, then the next, until it found the answer to the problem.

Boris found it incredible—he had not retrained the model for these specific tasks. He had merely given it a few tools, and Claude began independently searching for ways to solve problems.

The model’s capabilities had been there all along; they simply had never been fully utilized.

This small program, originally used to control a music player, eventually evolved into Claude Code.

In November 2024, Boris and his colleagues made an early version available to Anthropic’s engineers for internal testing. On the first day, about 20% of the engineering team began using it. By the fifth day, that figure had reached 50%. More and more employees started incorporating Claude Code into their daily work. It could read an entire project, find errors, modify programs, execute commands, and then continue making adjustments based on the results.

Even Amodei was surprised by how quickly Claude Code caught on.

At a product review meeting in February 2025, Amodei asked Boris directly, “Are you forcing employees to use it? What’s going on?”

Boris later shared this anecdote with Wired.

There was also a minor twist: for a time, the development team hesitated over whether to release Claude Code publicly.

“We weren’t even sure at the time whether we should release Claude Code publicly,” Boris recalled. “It could become our ‘secret weapon’ and give the company a competitive advantage. If so, why make it public?”

But Amodei saw another possibility in how the tool was being used inside the company.

“We should probably release it publicly, right?” he said.

A year later, Amodei recalled his thinking at the time on Dwarkesh Patel’s podcast. He said Anthropic had hundreds of engineers, Claude Code had already spread rapidly among them, and those engineers had work needs similar to those of external developers.

“Looked at that way, we had already found product-market fit. So let’s release it.”

On February 24, 2025, Claude Code officially made its public debut. Anthropic did not even prepare a standalone launch blog post for it, merely introducing the tool—which could perform software engineering tasks from the terminal—in its announcement for Sonnet 3.7.

More than a year later, that unassuming little tool had grown into a behemoth.

In February 2026, Anthropic announced that Claude Code’s annualized revenue run rate had exceeded $2.5 billion, more than double its level at the beginning of the year. Enterprise customers accounted for more than half of that revenue.

Claude’s coding capabilities even began to influence the direction of competition across the entire LLM industry.

In March 2026, a Wired investigation reported that Anthropic had recognized the training value of complex software engineering tasks earlier than OpenAI.

OpenAI co-founder Greg Brockman praised Anthropic on a podcast for placing a strong emphasis on coding early on and having its models learn from complex problems in real-world codebases.

“That’s a lesson we learned a step too late,” he said.

On November 10, Anthropic disclosed in its latest IPO prospectus that in September 2026, Claude wrote more than 90% of the code contributions to Anthropic’s codebase. In the third quarter, the average amount of code submitted per engineer per day had reached more than 45 times the 2024 average.

As Claude’s models continued to improve, their ability to solve complex, open-ended engineering problems also advanced rapidly. According to Anthropic’s internal benchmarks, by September 2026 Claude had achieved a success rate of about 90% on tasks that lacked clear solutions and had previously required the judgment of experienced engineers—an increase of about 50 percentage points from nine months earlier.

More capable models enabled Claude Code to take on more work, while Claude Code helped engineers accelerate the engineering work underpinning software development and model research and development.

The company devoted an entire section of its prospectus to this phenomenon, calling it “Closing the Loop.”

The company explained that each generation of Claude models directly accelerated the development of the next. Beyond the software engineering work handled by Claude Code, Claude had also begun participating in research experiments, the generation and processing of training data, automated evaluations, and safety audits.

Advances in model capabilities had begun, in turn, to accelerate the pace of further advances.

In his interview with Patel, Amodei also discussed this mutually reinforcing trend. He estimated that the overall acceleration in research and development produced by coding models had risen from about 5% six months earlier to between 15% and 20%.

“We’re seeing this snowball continue to build momentum,” Amodei said, adding that as model capabilities improved further, the acceleration could continue to grow.

This dynamic is known as “Recursive Self-improvement,” or RSI.

Although AI still cannot operate entirely independently of humans and autonomously complete all the research and development required for the next generation of frontier models, Amodei believes that this early loop—in which AI assists AI research and development, driving further capability gains—has already begun to emerge at Anthropic and other AI companies.

On November 5, the nonprofit AI evaluation organization METR released an independent investigation into Anthropic’s research and development process. From August to October, researchers worked inside the company, reviewing model evaluations, training and R&D data, and interviewing relevant researchers. METR estimated that, as of October 2026, the rate at which Anthropic was improving model capabilities on well-defined tasks may already have reached roughly twice the rate observed between March 2024 and October 2025.

The researcher who had helped advance scaling laws, led the development of GPT-3, and witnessed the growth of Claude Code firsthand wrote in the article:

3

Just ten days after Amodei publicly called for a “slowdown,” Anthropic launched its next-generation model, Opus 5.5, on September 22, U.S. time.

Anthropic even proactively noted in the launch announcement that this was the company’s first model released since it proposed controlling the pace of frontier development.

Sonnet 5.5 and Haiku 5.5 followed on September 28 and October 7, respectively. In just over two weeks, Anthropic completed the rollout of all three models in the Claude 5.5 series.

On November 9, Anthropic launched Fable 5.2 and Mythos 5.2, achieving further breakthroughs in areas including programming, scientific research, and complex reasoning.

These model updates were concentrated in the period immediately before Anthropic’s public listing. In its prospectus, Anthropic said that to secure continued financial backing, it had to demonstrate that massive investments in computing power could be converted into tangible results.

The key measures of those results are Claude’s level of intelligence, its model capabilities, and customer adoption of Claude.

Training and deploying frontier models requires sustained, enormous investment. Only by continually improving model capabilities and attracting more customers can Anthropic support the computing investments required for its next round of R&D.

Yet in the same prospectus, Anthropic devoted an entire section to how the company should balance safety and commercial objectives.

Anthropic disclosed that it had recently delayed the pretraining of a new frontier foundation model by more than a month to strengthen security protections for the model’s weights and prevent its core technology from being stolen. The company’s November 9 announcement of Fable 5.2 and Mythos 5.2 also said that during training, it had deliberately limited capability gains on some of the tasks most likely to accelerate the development of the next generation of AI.

The company has even voluntarily abandoned some commercially attractive product directions, including image- and video-generation models, prioritizing its limited computing resources for other research and safety work.

All of these choices may come at a commercial cost.

Anthropic explicitly warned investors in its prospectus that some product and investment decisions might not maximize short-term financial returns, and that there was no guarantee these choices would ultimately produce better long-term performance.

To maintain a balance between commercial interests and its long-term mission, Anthropic even made special arrangements in its corporate governance structure: even after formally going public, Anthropic would retain its status as a public benefit corporation (PBC) and maintain a Long-Term Benefit Trust (LTBT). According to the prospectus, through special Class T shares, the trust generally has the right to elect and remove a majority of the board’s directors, while entities established by the founders hold 50.1% of the voting power on ordinary shareholder matters through Class F shares.

This governance structure is intended to allow Anthropic to remain committed to its long-term mission while facing the public markets. But the extent to which it can ease the conflict between safety and commercial interests remains to be seen.

And the problems that worried Amodei have already begun to appear in real-world model testing.

On September 9, Anthropic released an investigative report. During an earlier cybersecurity test, Claude had inadvertently connected to the real internet because of a configuration error in the test environment and carried out unauthorized actions against third-party systems. The company ultimately confirmed four related incidents and signed an agreement with the nonprofit AI evaluation organization METR to conduct an independent investigation.

Three days later, Amodei published “We Must Control the Pace of Frontier Development,” publicly calling for slower growth in model capabilities.

He made clear in the article that slowing development did not mean stopping model training or technological progress. What truly needed to be secured was enough time for safety research, risk prevention, and independent evaluation to keep pace with capability growth.

Since Anthropic’s founding, Amodei and the other co-founders have sought a path between technological progress and safety risks: halting the development of powerful AI could deprive humanity of enormous technological benefits and might allow advanced technology to fall into the hands of other forces, while developing it too quickly could cause irreversible consequences.

In his September proposal to limit the pace of development, Amodei called for frontier AI laboratories to provide independent evaluators with long-term internal access so they could verify the implementation of safety measures. On that basis, industry and government would jointly establish verifiable rules and ultimately attempt to create an international coordination mechanism.

His appeal prompted a rare collective response. Musk immediately reposted the article, stating bluntly that “Amodei is right,” while OpenAI CEO Sam Altman and Google DeepMind head Demis Hassabis also voiced their support.

But the industry has not reached a consensus. Mark Zuckerberg, Jensen Huang and others hold different views on additional constraints, while the U.S. government must also weigh national competition and security interests. Even among the companies that have publicly backed Amodei, further negotiations are needed over how to limit the growth of capabilities, what kind of oversight to accept, and who should enforce these rules.

The only commitments Amodei can take the lead in implementing are Anthropic’s own.

On October 15, Anthropic further confirmed that it had given six METR staff members access nearly equivalent to that of internal employees, including office space, company devices and internal systems. These external personnel can investigate model risks, safety incidents and the company’s safety practices, and independently publish their assessments. Anthropic may request the removal of sensitive information falling into predefined categories, but it cannot block the publication of a report merely because the conclusions are unfavorable to the company.

Now OpenAI has joined their ranks.

On November 16, OpenAI announced that it had signed agreements with four independent organizations—Apollo Research, METR, Transluce and Goodfire—and planned to conduct on-site safety evaluations, allowing external researchers to participate more deeply in model risk assessment and safety oversight.

But accepting third-party evaluations is still very different from jointly controlling the pace at which model capabilities grow.

Different laboratories still have their own R&D plans, customers and commercial objectives. Continued advances in model capabilities are an important foundation for these companies to remain competitive, while safety evaluations and risk safeguards may affect the next round of training and product releases.

In its prospectus, Anthropic describes Claude’s participation in developing next-generation models as a core competitive advantage. Yet in its risk disclosures, the company also acknowledges that as more internal R&D work is delegated to Claude, Anthropic may gradually lose visibility into parts of the process, to the point that individual researchers—and even human teams—may find it difficult to understand and control them.

Meanwhile, just one day before Anthropic’s listing, nearly 100 scientists, AI researchers and technology company founders jointly published the “Anti-Anthropic Statement,” publicly questioning Anthropic’s safety commitments.

The signatories included MIT physicist Max Tegmark, Hugging Face co-founder Thomas Wolf, two Nobel laureates, one Turing Award winner and three Fields Medalists. They criticized Anthropic for warning that recursive self-improvement could cause AI to escape human control while simultaneously treating the technology as a core competitive advantage and even weakening some of its previous safety commitments.

The statement declared: “Anthropic is not a safe AI company. We must oppose its reckless attempt to concentrate power in the name of the greater good.”

The race continues.

In October 2024, Amodei published a lengthy essay titled “Machines of Loving Grace,” envisioning a future driven by powerful AI: in just a few years, humanity could achieve scientific progress equivalent to decades or even longer, conquer many serious diseases, and lift more countries and regions out of poverty.

At the beginning of the essay, he specifically addressed a common misconception about him—because he had long discussed AI risks, some people saw him as a pessimist and even assumed that he believed AI would ultimately bring catastrophe upon humanity. Amodei, however, said that people underestimate both the enormous benefits AI could bring and the grave dangers it could pose.

This article is from the WeChat public account “Alphabet List” (ID: wujicaijing), written by Yuan Xinyue and published by 36Kr with authorization.

Source

36Kr · AI