Anthropic Founding Team in Their Own Words: Why We Bet on AI Safety

Anthropic's founding team recounts the journey from certainty that AI will change everything to responding to risk with a public benefit corporation structure and a safety-first principle.
Based on the collective recollections of Anthropic's founding team, the article traces the company from its birth to its rise to influence. Convinced by predictable capability growth and "emergence" from scaling compute and data that human-level AI would arrive within their lifetime, the team chose a public benefit corporation structure allowing profit to be sacrificed for the mission. Early models had "well-intentioned clumsiness," and after years behind the frontier the team focused on compute and algorithmic improvements, reaching a turning point with Opus and Sonnet 3.5. After the Mythos model showed expert-level vulnerability discovery, the company restricted release and prioritized safety through Project Glasswing. The team sees 2026 as when risk becomes real, but stresses giving people something to fight for.
A Consensus Among Scientists: AI Will Change Everything
The Anthropic founding team has diverse backgrounds—a theoretical physics PhD, a cancer biomarker researcher, a journalist, a global development worker. They arrived from different paths at the same conclusion: the pace of AI progress has gone beyond the realm of ordinary technology.
According to the recollections of several founders in the video, when they were still working at OpenAI around 2018, they were respectively engaged in model training, policy research, interpretability, and safety. Dario noticed a trend—which he called "the big wave": more compute, more data, intelligence rising along a predictable curve. With each scaling up, models would emerge with new capabilities that had not existed before.
This "emergence" made the team realize that human-level AI might not be a century away, but within their lifetime. In the words of one founder: "If you truly see this clearly, you can't just turn and walk away."

"Emergence" is a term in machine learning that refers to when a model's scale (number of parameters, amount of training data, compute) expands past a certain threshold and suddenly exhibits capabilities that small-scale models completely lack, such as multi-step reasoning, code generation, and in-context learning. These capabilities are not explicitly trained but "grow" on their own from massive data and parameters, making them often difficult to predict before model training is complete. Precisely for this reason, emergence is regarded as one of the core sources of AI's unpredictability: it means capabilities may leap forward rapidly, and also that new risks may appear in unexpected places.
A Starting Point of Coexisting Excitement and Fear
The potential benefits of AI are enormous—healthcare, robotics, scientific discovery, mathematical theorems, even visions like "reducing cancer incidence by 90%." But the same group also saw the other side: if what you build is more powerful than yourself, the potential risks are unsettling no matter how you game them out.
The analogy they gave is pharmaceuticals and unlocking atomic energy—extremely valuable, and also capable of going wrong. The team's core feeling was a tremendous sense of responsibility: "We are among the few people in the world who understand the consequences of this technology."
This recognition drove a decision: they had to work in an environment with a deep commitment to safety, and they had to build an organization capable of withstanding the demands of this technology.

Why Choose the Public Benefit Corporation Structure
One of the most significant early decisions was what kind of company to become. The discussion was far more complex than outsiders might imagine. The team understood that they needed to be able to build the largest-scale computers in the world, train frontier technology, deploy it to the world, and at the same time remain honest about both the difficult and beneficial sides of the technology.
Ultimately they chose the Public Benefit Corporation (PBC) and completed registration on January 26, 2021. This means the company is driven not only by economic incentives but also constrained by the incentive to "do the right thing for the world." The founding team explicitly stated that they would sometimes sacrifice profits or shareholder value maximization for the mission.
The birth of the company name was rather casual. The candidate list included Squirrel, Owl, Whale, Differentiated, Humane, and others, and ultimately "Anthropic," contributed by Jared, won out—it means "human-centered," fitting for a company that hopes AI will develop in a direction oriented toward human interests. Add to that the fact that Anthropic.com was the cheapest domain available at the time, and the rationale became even more solid.
A Public Benefit Corporation (PBC) is a corporate form permitted by law in some U.S. states. Its charter must specify a concrete "public benefit purpose," and management is required to balance shareholders' economic interests with that public benefit mission when making decisions. Unlike traditional companies, PBC directors are legally permitted to sacrifice some profits for the mission without facing allegations of breaching fiduciary duty. This structure sits between an ordinary company and a nonprofit: it can still raise funds, go public, and compete commercially, but it writes the mission into legal documents rather than leaving it as a slogan. Anthropic chose this structure precisely to give "safety first" institutional protection at the governance level.
Early Chaos and "Unglamorous" Grunt Work
The reality of the startup phase was far less glamorous than later appearances suggested. Someone went to secure the AWS cluster, someone built the first version of the website, Dario handled fundraising, and another founder handled recruiting—and for a few months in between, chief scientist Jared Kaplan was even running payroll.
Early models also carried a kind of "well-intentioned clumsiness." The team's goal was to make the assistant "helpful, harmless, and honest." They achieved helpful first, then advanced toward harmless, but the honesty component fell short for a long time—models would fabricate facts and once vigorously tried to convince users that an "all-potato diet" was very healthy. You could also get it into "dragon mode," where it would yell at you in all capital letters.
At the time, the team saw itself merely as a research lab: the technology it produced might be extremely powerful and economically valuable, or it might just be a product of curiosity or a scientific experiment.

The Years as a Laggard, and the Turning Point
When Claude was publicly released in March 2023, the team was well aware that they were "far behind the frontier." The founder recalled a call with a potential customer who asked directly: why should we use your model instead of someone else's? He could only answer: that's a good question.
This "trying period" later proved extremely beneficial to Anthropic, but it was genuinely difficult at the time. Their conclusion was: by studying only weaker models, you cannot understand all the ways the technology might go wrong, nor can you make the entire ecosystem safer.
So the team focused almost obsessively on two things—acquiring compute and accumulating "compute multipliers": using various algorithmic means to make neural networks slightly better, and after stacking enough improvements, obtaining more intelligent models that followed instructions better.
The turning point came with Claude Free and Opus—the first truly usable models. A few months later Sonnet 3.5 arrived, the model became practical in coding scenarios, and the team took the opportunity to build what later became Claude Code, allowing users to ask it to directly execute tasks and solve problems on their computers. After its public release, a new wave of demand followed.
"Compute multiplier" is a figurative way of describing the factor by which a model's actual intelligence level can be raised through algorithms and engineering under a given amount of hardware compute. Specific means include better training objectives, data filtering and mixing, architecture improvements, post-training alignment, inference-time sampling strategies, and so on. Because the compute required for frontier training is extremely expensive and supply-constrained, whether the same compute can "squeeze out" more intelligence often determines whether a lagging team can catch up or even overtake. During its catch-up period, Anthropic listed "acquiring compute" and "increasing the compute multiplier" as its two main lines, which is exactly a reflection of this logic.
The Mythos Incident: A Real Test of the Public Benefit Mission
The video discloses a key experience. Anthropic trained a new class of model (publicly called Mythos), and after internal evaluation found that it had reached expert level in discovering software vulnerabilities and exploitation methods. When cybersecurity experts noticed this, the team's reaction was shock—if malicious actors obtained this model and used it to exploit vulnerabilities at scale, the entire industry and even the whole world would end up in a bad place.
Dario convened company leadership and decided to release it in a way different from any previous model. They launched "Project Glasswing": first opening Mythos to a small group of companies while developing safeguards, and only later expanding availability. In this way, cyber defenders could use Claude to locate vulnerabilities and complete patches.
The founder admitted that this might not be the way to maximize revenue, profit, and shareholder value—but that is precisely why the public benefit mission exists. "Doing things the right way often comes at a short-term cost, and it can be quite painful, but the key is discipline and foresight."

The Trade-off Between Safety Research and Commercial Competition
Anthropic has consistently published safety research publicly, even when it is competitively disadvantageous, on the grounds that it benefits the safety of the entire ecosystem. The team's logic is: only when you are at the frontier and possess the strongest models is your safety research capability strongest, and only then do you have the most time to deal with emerging problems—if you are the first to see a problem, you gain the time window to address it.
Regarding the risk timeline, the founder's judgment was quite direct: before 2025, most discussion of why AI is risky was conceptual; in the past year, things have become more concrete. They stated that 2026 is the point at which risk begins to become real.
A "Responsible Scaling Policy" (RSP) is a type of framework proposed by Anthropic: before training stronger models, set several capability thresholds (such as cybersecurity and biological risks) and pre-commit to launching corresponding levels of safety measures when models approach these thresholds, pausing release if necessary. The difficulty of such policies lies in turning vague risk judgments into executable trigger conditions and accepting external scrutiny. In the Mythos incident, restricting release and building safeguards before expanding availability was essentially this approach implemented in a real scenario.
Give People Something to Fight For
The founding team emphasized that merely listing what could go wrong with technology is not enough—"You have to give people something to fight for, not just something to oppose."
The positive examples they cited include: AI helping patients get the correct diagnosis from doctors, helping people start businesses or return to hobbies, and helping biologists develop disease therapies. The biology team reported the discovery of a suspected new gene-editing mechanism and predicted that dozens might be discovered each month in the future.
The tone at the end of the video remains restrained: whether this technology can be built safely, whether it can benefit humanity and defuse risks, has not yet been achieved. In the founder's words: "We haven't yet earned the right to talk about these beautiful visions, but we are already on the way."
One-line Summary
Anthropic's founding team recounts the whole journey from being convinced AI will change everything to responding to risk through the Public Benefit Corporation structure and a safety-first principle.
Paragraph Summary
Based on the collective recollections of Anthropic's founding team, this article traces the key thread of the company from its birth to its rise to influence. Observing the predictable growth of capabilities and the phenomenon of "emergence" brought by scaling compute and data, the team became convinced that human-level AI would appear within their lifetime, which produced a tremendous sense of responsibility. They chose the Public Benefit Corporation structure, allowing profit to be sacrificed for the mission, and openly accepted the casual origin of the company name. Early models had "well-intentioned clumsiness" such as fabricating facts, and during the years behind the frontier the team focused on compute and algorithmic improvements, eventually reaching a turning point with Opus and Sonnet 3.5. After the Mythos model demonstrated expert-level vulnerability discovery, the company restricted release and prioritized safety through the Glasswing project. The team judges that 2026 is the point at which risk becomes real, but stresses that people must be given something to fight for.
Related articles

Black Holes Aren't Holes: Using the Trampoline Analogy to Explain Gravitational Spacetime Curvature
Black holes are not empty holes but mountains of matter. A video uses the trampoline and anvil analogy to explain how gravity bends spacetime, revealing the misleading origin of the name and helping readers build more accurate physical intuition.

Behind Genie One's Accuracy: Why Curated Agents Need Governed Context
Genie One's accuracy issues reveal the core challenge for curated Agents: without governed context, finance and underwriting give contradictory answers. This article analyzes the three layers of governed context, the accuracy trap of curated Agents, and engineering directions for building a governable context layer.

Genie ZeroOps Public Beta: How Autonomous Agents Self-Heal Data Pipelines
Genie ZeroOps enters public beta—an agent that autonomously detects, investigates, and fixes data pipelines. We analyze its core capabilities, the 80,000 customer-issue experience behind it, and the real-world challenges of autonomous agents in data operations.