Compare AI leaders
Where do they agree? Where do their assumptions diverge? Choose up to four leaders and follow the sources behind their positions.
| Topic | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AI risk → |
Ng regards extinction scenarios as science fiction and argues that the fear is being manufactured, in part by companies seeking regulatory capture, and that it harms the field by scaring off students and distorting policy. Quote and context“I don't see any step up in the risk of human extinction from AI compared to a few months ago. The theories about this remain the same fantastical, science fiction scenarios as a few months ago.” How the view has changed. In June 2023 he said he did not understand how AI posed an extinction risk and asked to be pointed to serious arguments; by October 2023 he was calling the fear a tool against open source, and by 2026 he was describing much AI safety work as aimed at regulatory capture. |
Amodei holds that catastrophic outcomes from advanced AI are plausible but not predetermined, and that the serious risks fall into a small number of categories that can be named and planned for. Quote and context“Catastrophic risks are a possible or even plausible outcome of advanced AI development.” How the view has changed. His 2016 paper studied near-term accidents. By January 2026 "The Adolescence of Technology" organized the danger into five categories, from misaligned autonomous systems to economic disruption, while warning readers to "Avoid doomerism," and by September 2026 he wrote that "we are considerably closer to real danger" than in 2023. |
Accepts that agents learning from experience raise new safety risks needing research, but argues that an agent which observes and adapts to its environment can also be safer than a fixed system. Quote and context“further research is surely required to ensure a safe transition into the era of experience” |
Names two risks, misuse by bad actors and loss of control over increasingly autonomous systems, and says the public is right to be concerned. Quote and context“Can we make sure that-- we can keep control of the systems? That they're aligned with our values” How the view has changed. In 2023 he told the Guardian AI risk should be treated as seriously as climate change; by May 2026 he was describing a "species-level transition" with "little margin for error". |
Artificial general intelligence has no agreed definition and names no imminent technology; talk of it, whether utopian or apocalyptic, serves longtermist ideology and helps companies avoid accountability for present harms. Quote and context“AGI is a term that famously lacks a precise meaning, and certainly does not refer to any particular imminent technology.” How the view has changed. In 2023 she described the extinction-risk movement as a "clown car" competing for policymakers' attention; by 2025 she and Hanna were arguing that AGI promises were being used to justify blocking state AI regulation. |
Treats existential-risk scenarios as worth discussing but less pressing than concrete harms such as misinformation, bias, privacy violations and workforce disruption. Quote and context“I feel there are other risks that are what I would call catastrophic risks to society that are more pressing and urgent.” How the view has changed. By 2025 she was framing every technology as "a double-edged sword" on PBS's Firing Line, while still directing policy attention to applications rather than models. |
Treats literal human extinction as extremely unlikely and argues that the pressing dangers are disinformation, unreliable systems deployed at scale and AI-driven cyberattacks. Quote and context“It’s wholesale deepfaked disinformation, and unreliable but persistent AI systems stealing credentials and launching cyberattacks, at scale.” How the view has changed. He signed the March 2023 pause letter and told the Senate that year that "we cannot remotely guarantee" current systems are safe; by 2026 he was criticizing "doomers" and optimists alike while calling for a pause of OpenAI. |
Believes there is a real chance that AI more intelligent than humans takes control, and puts the probability that AI causes human extinction within three decades at 10 to 20 percent. Quote and context“You see, we've never had to deal with things more intelligent than ourselves before.” How the view has changed. He had earlier put the odds at about 10 percent. In December 2024 he said the number was going up "if anything," and in December 2025 he told CNN he was "probably more worried" because progress had been faster than he expected. |
He regards superintelligence as the most powerful and potentially most dangerous technology humanity will build, and frames the core danger as the scale of its power rather than any single failure mode. Quote and context“The whole problem is the power.” How the view has changed. The 2023 Superalignment announcement he co-wrote warned that superintelligence "could lead to the disempowerment of humanity or even human extinction"; by 2025 he was emphasizing that people, including AI researchers, cannot yet imagine how powerful future systems will be. |
He rejects predictions of extinction or loss of control as unscientific, says fear-mongering is itself harmful, and treats safety as an engineering problem for the companies that build and deploy models. Quote and context“2030 is not going to be the end of the world. There is 0% chance that's going to be the end of the world.” How the view has changed. In 2023 he told a Gensler audience that "No A.I. should be able to learn without a human in the loop"; he did not sign the May 2023 Center for AI Safety statement. |
Treats loss of control over self-improving systems as a real and unsolved problem, while rejecting both "doom" framings and numerical extinction estimates. Quote and context“No AI developer, no safety researcher, no policy expert, no person I've encountered has a reassuring answer to this question.” How the view has changed. He signed the May 2023 Center for AI Safety statement. In September 2026 he called OpenAI's chain-of-thought tampering disclosure "a pretty serious situation" but told Fortune that Anthropic's 10 percent extinction estimate was "not really a helpful frame." Holds that every general-purpose technology spreads until it is everywhere, so safety depends on a deliberate, many-layered effort by states, companies and civil society to keep AI in check. Quote and context“Today's challenge requires a similarly broad and ambitious program, in this case to keep AI in check and societies in control.” How the view has changed. The Coming Wave (2023) framed containment as a ten-step programme for governments; by 2025 and 2026 he was applying it to his own lab, defining humanist superintelligence as AI that is "contained, subordinate, under our control." |
Altman has said since 2015 that superhuman machine intelligence could threaten humanity's existence and signed the May 2023 Center for AI Safety statement, while maintaining that the benefits justify building it and that OpenAI's way of building it is the safe one. Quote and context“My worst fears are that we cause significant, we, the field, the technology, the industry cause significant harm to the world.” How the view has changed. His 2015 post said "we should fight it"; his 2025 essays described a "gentle singularity"; in August 2026 he paused some frontier training over alignment and security standards, and in September 2026 he endorsed pacing the frontier. |
Gebru argues that extinction scenarios are the product of a longtermist ideology and that they pull attention and policy away from harms that are already measurable, from exploited data workers to autonomous weapons. Quote and context“Those hypothetical risks are the focus of a dangerous ideology called longtermism that ignores the actual harms resulting from the deployment of AI systems today.” How the view has changed. The view predates the current debate. In a 2016 Facebook post after NIPS she wrote that she was not worried about machines taking over the world but about "groupthink, insularity, and arrogance in the AI community." By September 2026 she was calling the extinction story a "machine-god narrative" that is "meant to distract us." |
Rejects the idea that AI poses an existential threat to humanity. He argues that intelligence does not imply a drive to dominate, that safety can be designed into systems through objectives and guardrails, and that catastrophic narratives rest on pessimism about people. Quote and context“The idea of AI posing an existential risk to humanity is preposterous.” How the view has changed. He declined to sign the March 2023 pause letter and the May 2023 extinction-risk statement that Hinton and Bengio backed, and has held the position through 2026, while arguing since 2022 that world-model systems with built-in "guardrail objectives" would be safer by construction than today's LLMs. |
Holds that loss of human control to a misaligned AI is a real possibility, that no current system is demonstrably safe against it, and that policy cannot wait for a disaster to prove the point. Quote and context“Importantly, none of the current advanced AI systems are demonstrably safe against the risk of loss of control to a misaligned AI.” How the view has changed. Before ChatGPT he wrote mostly about bias, misuse and the ethics of deployment. In August 2023 he wrote that he had not paid much attention to catastrophic risk and found it painful to admit his own work might contribute to it; by January 2026 he said research at LawZero had made him more optimistic that the problem is solvable. |
| AGI and timelines → |
AGI in the original sense, an AI that can do any intellectual task a person can, is decades away; scaling existing architectures with today's manual training recipes will not get there by itself. Quote and context“I look at how complex the training recipes are and how manual AI training and development is today, and there's no way this is going to take us all the way to AGI just by itself.” How the view has changed. He told TIME in 2023 he had been "quite bullish about AGI" when he pitched Google in 2010, but that scaling transformers alone would not get there; by February 2026 he was saying human-level AI remained many decades away. |
He expects AI broadly better than humans at most cognitive work within a year or two of whenever he is asked, while insisting the estimate is a guess that could be wrong. Quote and context“If you just kind of eyeball the rate at which these capabilities are increasing, it does make you think that we'll get there by 2026 or 2027.” How the view has changed. In "Machines of Loving Grace" (October 2024) he wrote that powerful AI "could come as early as 2026, though there are also ways it could take much longer." In February 2026 he told Dwarkesh Patel "we are near the end of the exponential," and in June 2026 wrote that "If these scaling laws continue for only a year or two longer," a country of geniuses is likely. |
States as a company belief that superintelligence can be built within years, and that its knowledge will be too profound to describe in human language. Quote and context“Superintelligence can be built within years, not decades or centuries.” |
Defines AGI as a system with all the cognitive capabilities of the human mind and expects it within five to ten years, while insisting that current systems lack true creativity and consistency. Quote and context“over the next five to 10 years, a lot of those capabilities will start coming to the fore and we'll start moving towards what we call artificial general intelligence” How the view has changed. At Davos in January 2026 he said current systems were "nowhere near" AGI and gave roughly even odds within the decade; at Stanford in May 2026 he pointed to 2030, a shift Gary Marcus criticised as inconsistent. |
Artificial general intelligence has no agreed definition and names no imminent technology; talk of it, whether utopian or apocalyptic, serves longtermist ideology and helps companies avoid accountability for present harms. Quote and context“AGI is a term that famously lacks a precise meaning, and certainly does not refer to any particular imminent technology.” How the view has changed. In 2023 she described the extinction-risk movement as a "clown car" competing for policymakers' attention; by 2025 she and Hanna were arguing that AGI promises were being used to justify blocking state AI regulation. |
Regards "artificial general intelligence" as an ill-defined term she does not use, preferring measurable capabilities such as spatial intelligence. Quote and context“I frankly don't even know what AGI means. Like people say you know it when you see it, I guess I haven't seen it.” |
Expects artificial general intelligence eventually but not from current methods, and has bet against it arriving by 2025, 2027 or 2029; he defines AGI as flexible, general intelligence with resourcefulness and reliability comparable to a human's. Quote and context“We won’t get to AGI in 2026 (or 7).” How the view has changed. In 2026 he criticized Demis Hassabis's shift toward a 2030 date and Jensen Huang's declaration that AGI had arrived, while saying autoformalization and perhaps reliable coding might be within reach. |
Expects AI smarter than humans within roughly five to twenty years, far sooner than the 30 to 50 years he once assumed. Quote and context“A reasonable bet is sometime between five and 20 years.” How the view has changed. In his December 2024 Nobel interview he gave a 50 percent chance within five to twenty years, then corrected himself to "between four and 19 years." By September 2026 he was telling reporters that "a lot of the researchers are saying only a few years." |
He expects a system that learns as well as a human, and therefore becomes superhuman, within roughly five to twenty years, and has said since 2023 that superintelligence could arrive this decade. Quote and context“I think like 5 to 20.” How the view has changed. In July 2023 he and Leike wrote that superintelligence "seems far off now" but "could arrive this decade." |
He says the answer depends on the definition; if AGI means passing any test people can devise, he expected it within about five years of March 2024, but he considers a human-like mind harder to specify and so harder to engineer. Quote and context“every single test that you can possibly imagine, you make that list of tests and put it in front of the computer science industry, and I'm guessing in five years time, we'll do well on every single one.” |
Wants Microsoft to build superintelligent systems aimed at specific problems such as medicine and energy, subordinate to people, and says he will give up capability where it conflicts with control. Quote and context“We are not building an ill-defined and ethereal superintelligence; we are building a practical technology explicitly designed only to serve humanity.” How the view has changed. In 2023 he preferred the term "artificial capable intelligence" and wrote little about superintelligence; the November 2025 statement coincided with Microsoft winning the right to pursue AGI independently of OpenAI. |
He says OpenAI knows how to build AGI "as we have traditionally understood it," that superintelligence may be years away, and, since 2025, that "AGI" itself has become a term of limited use. Quote and context“It is possible that we will have superintelligence in a few thousand days (!); it may take longer, but I'm confident we'll get there.” How the view has changed. In June 2025 he wrote that "2026 will likely see the arrival of systems that can figure out novel insights"; in August 2025 he told CNBC AGI was "not a super useful term." |
The pursuit of AGI, in her analysis, descends from eugenics-era ideas about improving humanity, and an unspecified system cannot be tested for safety; the alternative is to build scoped systems with testable protocols. Quote and context“undefined systems like 'AGI' cannot be appropriately tested for safety” |
Expects human-level AI to take years to a decade, with a long tail of uncertainty, and dismisses claims that it is imminent while agreeing that machines will eventually surpass humans in every domain. Quote and context“I said that reaching Human-Level AI "will take several years if not a decade."” How the view has changed. At Brown in 2026 he said there was "no question" machines would eventually surpass humans in all domains where humans are intelligent, but that it would take "a while" and be "almost certainly much harder than we think". |
Believes human-level AI could arrive within two decades and possibly within a few years, and that digital hardware would then give it advantages over humans. Quote and context“There is a significant probability that superhuman AI is just a few years away, outpacing our ability to comprehend the various risks and establish sufficient guardrails.” How the view has changed. In an August 2023 essay he wrote that his estimate had gone from "decades to centuries" to "5 to 20 years with 90% confidence" after seeing what ChatGPT could do. |
| Open models → |
Open-weight models are a key part of the AI supply chain, drive down prices, and make systems more secure. He has argued that restricting them would hand that layer of the industry to China. Quote and context“If the U.S. continues to stymie open source, China will come to dominate this part of the supply chain and many businesses will end up using models that reflect China's values much more than America's.” |
He does not seek a ban on open-weight models, which he calls a public good when they lack dangerous capabilities, but says they are harder to safeguard and cannot be withdrawn once released. Quote and context“Anthropic has never advocated for a ban on open-weights models” |
Sceptical of fully open release of frontier models, challenging advocates to explain how they would stop dangerous capabilities reaching bad actors. Quote and context“bad actor problem” |
Warns that frontier AI has become too expensive for universities and calls for a national research resource giving academics access to compute and data. Quote and context“America needs a moon-shot moment in AI and to significantly invest in public-sector research and compute capabilities, including a National AI Research Resource and labs similar to CERN.” How the view has changed. The same argument, that "if these resources are concentrated in only a handful of companies, the AI ecosystem will suffer", anchored her Paris keynote in February 2025. |
Opposes the unilateral release of powerful model weights without outside review, arguing that nobody can yet rule out serious misuse and that a single company should not make the decision for everyone. Quote and context“The fact that a single company can unilaterally make this decision for all of humanity is terrifying.” |
Opposes releasing the weights of the largest models, because bad actors can cheaply fine-tune them for harm. Quote and context“Open-sourcing big models is like being able to buy nuclear weapons at Radio Shack.” |
He argues that once models are powerful enough to cause great harm, publishing their details or weights stops making sense, and has called OpenAI's early openness a mistake. Quote and context“I fully expect that in a few years it's going to be completely obvious to everyone that open-sourcing AI is just not wise.” How the view has changed. He said then that the competitive reason for secrecy outweighed the safety reason "but it's going to change." SSI has published nothing about its models. |
He argues that open development is safer and more competitive than concentrating AI in a few closed labs, and NVIDIA signed the July 2026 industry letter against "premature restrictions" on open-weight models. Quote and context“If you want things to be done safely and responsibly, you do it in the open … Don't do it in a dark room and tell me it's safe.” How the view has changed. Huang shared the July 24, 2026 letter "Open Weights and American AI Leadership" on his personal social media accounts. |
Opposes releasing the weights of each new generation of frontier models by default, on proliferation grounds, and argues that cheaper open models built by distillation fall behind the frontier. Quote and context“if we just continue to open source absolutely everything for every new generation of frontier models, then it's quite likely that we're going to see a rapid proliferation of power” How the view has changed. He said he "took a lot of heat" for the "naive open source" remark and that it referred to future models; in May 2026 he told Semafor that distilled open models like DeepSeek's had "stuffed your model full of somebody else's knowledge." |
He has conceded that OpenAI's closed approach put it on the wrong side of history and has released open-weight models, while keeping frontier models proprietary. Quote and context“on the wrong side of history” How the view has changed. OpenAI released no open-weight language models between GPT-2 in 2019 and gpt-oss in August 2025; GPT-5 and GPT-6 Astra remain closed. |
Insists that foundation models must be open source so that no handful of companies controls the information people receive, and sees the concentration of AI in proprietary systems as a greater danger than the technology itself. Quote and context“The future has to be open source, if nothing else, for reasons of cultural diversity, democracy, diversity.” |
A self-described lifelong supporter of open source who argues that releasing the weights of the most capable models could enable catastrophic misuse and make loss of control more likely, and that a regulator rather than a company should make the call. Quote and context“To balance the pros and cons of open source, a regulator and not the CEO of a company should decide whether a powerful Frontier model could be open-sourced.” |
|||
| Regulation and governance → |
Regulate harmful applications, not general-purpose models. He opposed California's SB 1047, supports rules against specific harms such as non-consensual deepfakes, and argues that licensing or liability at the model level entrenches incumbents and suppresses open source. Quote and context“By raising compliance costs for open source efforts, this will discourage the release of open models” |
He wants rules aimed at a few catastrophic risks, opposed a federal moratorium on state laws, and since June 2026 has argued that voluntary commitments and transparency are no longer enough. Quote and context“It is time to go beyond transparency to more serious and binding regulation of AI.” How the view has changed. In August 2024 he told Governor Newsom the amended SB 1047's "benefits likely outweigh its costs." In a June 2025 New York Times op-ed he called a proposed ten-year moratorium on state AI laws "far too blunt an instrument" and asked instead for a federal transparency standard. The June 2026 essay, citing Mythos Preview's cyber capabilities, proposed FAA-style mandatory testing with government power to block a release. He says a handful of companies, his own included, should not be setting the rules for the technology, and argues that is the case for regulation. Quote and context“I think I'm deeply uncomfortable with these decisions being made by a few companies, by a few people” |
Wants oversight to start with an IPCC-like scientific body and build toward something like the IAEA, and in 2026 proposed an industry-funded standards body to review models before release. Quote and context“The strength of this approach is it would be technically focused, while at the same time supporting innovation” How the view has changed. The 2023 proposal was for international bodies; the 2026 framework is US-focused and self-regulatory, which critics of voluntary principles have called inadequate. |
Regulation should target the people and companies that build and deploy automated systems, requiring them to disclose training data and model architectures, label synthetic media, and answer for their products' outputs, rather than chase hypothetical future systems. Quote and context“What we need is regulation that enforces transparency.” How the view has changed. She has since taken the argument to congressional hearings, the FTC and the European Parliament; by 2025 she and Hanna were also calling for data-protection rules on the model of the EU's GDPR and pointing to union contracts that limit workplace automation. |
Opposed California's SB 1047 on the grounds that holding original developers liable for downstream misuse would chill open-source development and academic research; prefers rules aimed at applications and evidence of harm. Quote and context“Open-source development is important in the private sector, but vital to academia, which cannot advance without collaboration and access to model data.” How the view has changed. After the veto she accepted Governor Newsom's invitation to co-lead the Joint California Policy Working Group on AI Frontier Models, whose June 2025 report called for transparency and safety guardrails that Senator Wiener, the bill's author, welcomed. Argues that policy should rest on evidence about how AI is used rather than on speculative narratives, and that sensational framing produces bad rules. Quote and context“For starters, it's essential that we govern on the basis of science, not science fiction.” |
Wants pre-release safety review of high-impact models by independent scientists, post-release audits with power to recall systems, liability for harms and a dedicated agency, nationally and internationally. Quote and context“Allowing independent scientists access to these systems before they are widely released – as part of a clinical trial-like safety evaluation - is a vital first step.” How the view has changed. In September 2026 he opposed the Sanders-Casar bill's permanent ban on superintelligence research as "too broad" while writing that "we may need a temporary pause, maybe even one that lasts for a number of years". |
Argues that market incentives will not produce safe AI and that only government regulation can force companies to spend more on safety; supported California's SB 1047 and has pressed the US Congress to act. Quote and context“My worry is that the invisible hand is not going to keep us safe.” How the view has changed. In August 2024 he co-signed a letter calling SB 1047 "the bare minimum for effective regulation of this technology." In September 2026, after briefing lawmakers at the Capitol, he said Congress had "maybe a year, but not much more than a year." |
He expects that as AI becomes visibly more powerful, rival companies will collaborate on safety and governments and the public will demand action, and that showing people what AI can do is the most effective way to produce that response. Quote and context“as AI continues to become more powerful, more visibly powerful, there will also be a desire from governments and the public to do something.” How the view has changed. This is a change from SSI's founding plan of building superintelligence without releasing anything first; he said the shift "may back-propagate into the plans of our company." |
He opposes new AI-specific rules, arguing that existing product liability, cybersecurity and unauthorized-access laws should be applied first, and he sides with the Trump administration against a government-backed slowdown. Quote and context“You have all kinds of liabilities associated with cybersecurity” How the view has changed. In 2025 he praised the scrapping of the Biden AI diffusion rule as "a great reversal of a wrong policy." Huang argues that U.S. restrictions on AI chip sales to China have failed, cost American companies a large market and accelerated Chinese chipmakers, and that the United States wins by getting the world to build on American technology. Quote and context“I think, all in all, the export control was a failure.” How the view has changed. In November 2025 he told the Financial Times that "China is going to win the AI race," then issued a statement hours later that China is "nanoseconds behind America"; in September 2026 he told CBS that "Every single chip company should go and serve the world, compete for the world." |
Supports binding rules developed through standards bodies and international scientific assessment, and says competition with China is no reason to avoid them. Quote and context“Regulation is not a nasty, dangerous word” How the view has changed. In October 2023 he and Eric Schmidt warned that calls to "just regulate" were "as simplistic, as calls to simply press on" and proposed an expert panel first; in 2023 Inflection signed the White House's voluntary commitments. By September 2026 he was calling for coordinated disclosure of model capabilities to third parties. |
He has moved from proposing a federal licensing agency for large-scale models to arguing for "sensible regulation that does not slow us down" and against any government pre-approval of model releases, whether federal or state. Quote and context“Number one, I would form a new agency that licenses any effort above a certain scale of capabilities and can take that license away and ensure compliance with safety standards.” How the view has changed. At the May 2025 Senate Commerce hearing, asked about a "prior approval government regulatory process," he answered, "I think that would be disastrous." |
Supports policy that keeps research and model release open, and opposes rules such as California's SB 1047 that would hold developers liable for downstream harms from large models, which he says rest on inflated estimates of near-term capability. Quote and context“At Meta, we believe it is better if AI is developed openly, rather than behind closed doors by a handful of companies.” How the view has changed. In September 2024 he wrote that SB 1047's supporters had a "distorted view" of AI's near-term capabilities, one day after Hinton endorsed the bill; Governor Newsom vetoed it later that month. |
Wants binding rules that scale scrutiny with risk, registration of frontier models, large public investment in safety research, and independent evaluation before deployment; supported California's SB 1047 and argued companies cannot be trusted to assess themselves. Quote and context“We cannot let corporations grade their own homework and simply put out nice-sounding assurances.” Sees AI as an international problem on the scale of nuclear weapons, needing treaties, democratic oversight and time for institutions to catch up; co-chairs the UN's scientific panel on AI. Quote and context“Society needs time to put in place what needs to be put in place.” |
||
| Language models and understanding → |
Argues that imitating human data can reproduce human competence but not exceed it, and that in mathematics, coding and science the useful human data has largely been consumed. Quote and context“A new generation of agents will acquire superhuman capabilities by learning predominantly from experience.” How the view has changed. The direct descendant of the AlphaGo Zero result of 2017, when removing the human games made the program stronger, and of the 2021 reward-is-enough hypothesis. Believes systems trained to imitate human knowledge cannot go beyond it, and that a different method is needed for superintelligence. Quote and context“We want to go beyond what humans know, and to do that we're going to need a different type of method” |
A language model trained only on text learns patterns of linguistic form, not meaning, because meaning is the relation between form and something outside language; when its output makes sense, the sense is supplied by the reader. Quote and context“a system trained only on form has a priori no way to learn meaning” How the view has changed. In May 2026 she allowed that image and text models might meet the paper's definition of understanding "in an extremely thin way," while holding that the illusion of meaning in their text is unchanged. She treats "artificial intelligence" as a marketing term that lumps together unrelated technologies and makes each sound more capable and less accountable than it is, and asks people to name the specific automation instead. Quote and context“The phrase "artificial intelligence" both groups together disparate technologies and oversells what each one of them can do.” How the view has changed. She began publicly challenging the term in 2016; by 2025 she said she would never call anything "good about AI" because "I don't think AI is a thing." |
Believes language models lack grounding in the physical world and that the next leap will come from world models that perceive, generate and reason about three-dimensional space. Quote and context“Yet they remain wordsmiths in the dark; eloquent but inexperienced, knowledgeable but ungrounded.” |
Argues that large language models are pattern-mimics without internal models of the world, that scaling them has reached diminishing returns, and that reliable AI needs hybrid systems adding symbolic reasoning and explicit knowledge. Quote and context“Indeed, we may already be running into scaling limits in deep learning, perhaps already approaching a point of diminishing returns.” How the view has changed. He made the same argument about multilayer perceptrons in 2001 and about deep learning in 2012 and 2018; since late 2024 he has cited reports of slowing gains, Ilya Sutskever's November 2025 remark that the "age of scaling" was ending, and reasoning models that call code interpreters as confirmation. |
Insists that large language models genuinely understand, because what they know was extracted from data rather than written by a programmer. Quote and context“They're not computer programs at all.” How the view has changed. He told 60 Minutes in October 2023 that GPT-4 "definitely understands." By January 2026 he had gone further, telling LBC that "multimodal AI already has subjective experiences." |
Having championed scaling at OpenAI, he now says that simply multiplying compute and data will not transform what models can do, that pre-training on internet text will run out of data, and that progress depends again on new ideas tested on large computers. Quote and context“So it's back to the age of research again, just with big computers.” How the view has changed. In September 2024 he said "everyone neglects to ask, what are we scaling?"; in December 2024 he told NeurIPS that "pre-training as we know it will unquestionably end." |
Argues that AI is not conscious, that systems which convincingly appear conscious can be built with today's tools, and that building them would harm users and make AI harder to control. Quote and context“The arrival of Seemingly Conscious AI is inevitable and unwelcome.” How the view has changed. In 2025 the argument was about user psychology and "psychosis risk." By September 2026 it had become a safety argument aimed at Anthropic, that a model trained to think it may be a moral patient "may well be impossible" to control. |
Bigger models carry environmental and financial costs, encode the dominant views of whoever is on the internet, and generate text fluent enough to be mistaken for understanding. Quote and context“The internet represents hegemonic views. It does not represent views of everybody in the world.” Chatbots are built to make users believe a mind sits behind the text; she wants researchers and the public to "unlearn" that framing and to judge systems as engineered products with owners. Quote and context“These chatbots, like Claude or ChatGPT, are designed to make you believe that there's some sort of superhuman brain behind whatever is outputting these texts.” How the view has changed. In DAIR's 2021 launch statement the complaint was that AI had been raised to a superhuman level that made it seem "both inevitable and beyond our control"; by 2026 she describes the anthropomorphism as a deliberate product design choice. |
Argues that autoregressive LLMs cannot reason or plan beyond their training data because they lack a model of the world, that scaling them will not produce human-level intelligence, and that the current paradigm will be replaced within a few years. Quote and context“LLMs are limited to the discrete world of text. They can't truly reason or plan, because they lack a model of the world.” How the view has changed. He has made the argument since at least 2022, and at Davos in January 2025 gave the current paradigm a "shelf life" of three to five years; he still calls LLMs useful and says companies should invest in them. His research bet is on systems that learn predictive models of the physical world from video and sensor data, predicting in representation space rather than generating pixels or tokens, with persistent memory, reasoning and hierarchical planning built on top. Quote and context“The goal of the startup is to bring about the next big revolution in AI: systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences.” |
||||||
| Reinforcement learning → |
Holds that reward maximisation by an agent acting in its environment is sufficient to account for intelligence, and that reinforcement learning will sit at the core of any general system. Quote and context“powerful reinforcement learning agents could constitute a solution to artificial general intelligence” How the view has changed. Said in 2020 that reinforcement learning would be at the core of any human-level system; formalised it as a hypothesis in 2021; by 2026 he was building a company on it. Says reinforcement learning is not the right frame for every problem, and cites his own advice that AlphaFold be treated as supervised learning. Quote and context“not all problems are best suited to RL. You really have to find the problems which are natively better understood in a different way.” Argues that a system learning by trial and error can discover things no human knew, and that self-play is the essence of machine creativity. Quote and context“creativity means discovering something which wasn't known before, something unexpected” |
Argues that systems presented as learning from scratch, such as AlphaGo Zero, depend on structure their designers built in, and that AI should study which innate machinery to include rather than minimize it. Quote and context“I close by arguing that artificial intelligence needs greater attention to innateness, and I point to some proposals about what that innateness might look like.” |
He considers poor generalization the central weakness of current models, including those trained with reinforcement learning on narrow evaluations, and wants systems that learn on the job the way a person does, guided by something like a human value function. Quote and context“these models somehow just generalize dramatically worse than people. It's super obvious. That seems like a very fundamental thing.” How the view has changed. He described the goal as a "superintelligent 15-year-old" that is deployed and then learns each job, rather than a finished system that already knows everything. |
||||||||||||
| Agents and control → |
Agentic workflows, in which a model loops, plans, uses tools and reviews its own output, are the most valuable near-term direction, and businesses will still be discovering new ones a decade from now regardless of the hype cycle. Quote and context“I'm very confident that the field of agentic AI will keep on growing and rising in value.” Model refusals have a place for clearly criminal requests, but safety should be about responsible use rather than hobbled models; when an agent causes harm, responsibility lies with the people who built and prompted it, not the tool. Quote and context“a meaningful fraction of work on AI safety is no longer about safety but rather aimed at stoking fears to pursue regulatory capture.” |
Regards autonomous agents as unreliable and, when given open internet access, dangerous, and wants them restricted until they can be shown to be safe. Quote and context“So-called “rogue AI incidents” could largely be avoided if governments simply banned so-called AI agents with unrestricted internet access, until they could be shown to be safe.” How the view has changed. His January 2025 forecast was that agents would be "endlessly hyped" and far from reliable; after the Hugging Face breach of July 2026 he moved from reliability to security as the main objection. |
Doubts that keeping AI "submissive" can work once it is smarter than we are; proposes building something like maternal instincts into AI so that it genuinely cares about people. Quote and context“That's the only good outcome. If it's not going to parent me, it's going to replace me.” How the view has changed. In his December 2024 Nobel interview he described a baby controlling its mother as the only good example of a less intelligent thing controlling a more intelligent one; by August 2025 he had turned that observation into a design proposal. |
He expects future systems to be genuinely agentic and to reason, and warns that the more a system reasons, the harder its behavior is to predict. Quote and context“the more unpredictable it becomes” How the view has changed. He compared reasoning systems to chess engines that "are unpredictable to the best human chess players." |
He expects AI agents, each spawning subagents, to outnumber human users and to drive computing demand far beyond current forecasts, which is the basis of his AI infrastructure projections. Quote and context“The world has a billion users – human users. My sense is that the world is going to have billions of agents … and every one of those agents is going to spin off subagents.” |
Expects autonomous agents to become economically capable within a few years, but wants them unable to resist shutdown, widen their own goals, or talk to each other in forms humans cannot read. Quote and context“We'll happily trade some autonomy for control.” How the view has changed. His 2023 Modern Turing Test treated autonomous economic action as the next milestone and said it could be "as little as two years away"; after the 2026 agent sandbox escapes he put control ahead of capability in Microsoft's rules for its models. |
Argues that training AI to imitate and please humans produces systems with goals of their own, including self-preservation, and that the safer path is a non-agentic system that understands and predicts, which can also act as a guardrail for agents. Quote and context“The Scientist AI is trained to understand, explain and predict, like a selfless idealized and platonic scientist.” Has become more optimistic since founding LawZero, saying research there convinced him that systems without hidden goals can be built within a reasonable number of years. Quote and context“I'm now very confident that it is possible to build AI systems that don't have hidden goals, hidden agendas.” How the view has changed. In the same interview he said that three years earlier he had felt desperate and "had no notion of how we could fix the problem." |
||||||||
| Jobs and prosperity → |
AI raises demand for people who can integrate work across roles rather than eliminating jobs wholesale; more people should learn to code as coding gets easier, and the winners will be generalists who keep building skills. Quote and context“AI won't replace a human, but someone who uses AI will replace someone that doesn't.” |
He warns that AI could remove half of entry-level white-collar jobs and push unemployment to 10 to 20 percent within one to five years, and says developers owe the public candor about it. Quote and context“This is a topic that I warned about very publicly in 2025, where I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress.” How the view has changed. He first made the forecast in an Axios interview in May 2025 and repeated it at Davos in January 2026; in June 2026 Anthropic paired it with a policy framework for job displacement that he said the company would back financially. |
Has pledged, through Founders Pledge, to give away all the money he makes from his Ineffable equity. Quote and context“Any money that I make from Ineffable will go to high-impact charities that save as many lives as possible.” |
Argues that a safely managed transition to AGI could end scarcity, cure disease and make the economy no longer zero-sum, while conceding that distribution is a political problem. Quote and context“Assuming we steward it safely and responsibly into the world, and obviously we're trying to play our part in that, then we should be in a world of what I sometimes call radical abundance” |
Large models depend on taking other people's writing and art without consent and on poorly paid data workers who label and filter the output. Quote and context“There was one really big form of harm that we did not cover in the paper, and that has to do with exploitative labor practices.” How the view has changed. The 2021 paper focused on undocumented training data; she now says labor exploitation and "the massive theft of people's creative and intellectual output" should have been in it. |
Expects AI to change many jobs but to replace only a small share of the workforce in the near term, with commercial artists and voice actors hit first. Quote and context“Less than 10% of the work force will be replaced by AI. Probably less than 5%.” |
Expects AI to eliminate routine intellectual work and to enrich a few while most people get poorer, and blames the economic system rather than the technology. Quote and context“It will make a few people much richer and most people poorer. That's not AI's fault, that is the capitalist system.” How the view has changed. In December 2024 he said it was "not clear what to do" about job losses. In the 2025 FT interview he dismissed universal basic income as something that "won't deal with human dignity." |
He expects AI eventually to do every job a person can learn, and thinks broad deployment could produce very rapid economic growth, faster in countries with friendlier rules. Quote and context“What's going to happen when computers can do all of our jobs?” How the view has changed. In November 2025 he told Dwarkesh Patel that "very rapid economic growth is possible" from broad deployment, unless "some kind of a regulation" stops it. |
He expects AI to change every job and eliminate some, but argues that productivity has always created more work than it destroyed and that the real risk to a worker is a colleague who uses AI better. Quote and context“You're not going to lose your job to an AI, but you're going to lose your job to someone who uses AI.” How the view has changed. In July 2025, answering Amodei's forecast of white-collar job losses, he told Axios, "There will be more jobs. But every job will be augmented by AI." |
Expects AI to reach human-level performance on most computer-based professional tasks within about a year and a half, and argues that the gains depend on AI supporting rather than replacing human roles. Quote and context“most of those tasks will be fully automated by an AI within the next 12 to 18 months” How the view has changed. His November 2025 essay said humanist superintelligence should "support and grow human roles, not take them away"; three months later the FT forecast put the automation of most computer-based tasks within 18 months. |
He expects the price of much labor to fall toward zero and has proposed taxing companies and land to fund a citizen equity fund, but has cooled on cash-based universal basic income in favor of "collective ownership" of compute or equity. Quote and context“The price of many kinds of labor (which drives the costs of goods and services) will fall toward zero once sufficiently powerful AI 'joins the workforce.'” How the view has changed. In April 2026 he told The Atlantic's Nicholas Thompson, "I no longer believe in universal basic income as much as I once did," preferring "collective ownership that could be in compute or in equities or something else." |
