Matt Clifford on AI Scaling, China, Defense and Finding Founders

Matt Clifford thinks the returns from adding more text data and compute to language models may be flattening, increasing the value of new ideas. He also discusses AI competition, defense and regulation. His comments are forecasts and policy judgments made in 2024, not a current assessment of the geopolitical or technical landscape.

Harry Stebbings interviews Clifford, co-founder of Entrepreneur First (EF), chair of the UK's Advanced Research and Invention Agency (ARIA), and an adviser who helped organize the 2023 AI Safety Summit. EF invests in startups, including areas such as defense technology, which gives him a commercial interest in some subjects discussed.

Clifford grew up in a small town in northern England. At about 13, he began repairing and building computers for family friends and later made websites. The experience taught him that he could accept an unfamiliar task, learn what was needed and act without waiting for permission.

Stebbings suggests great founders always start entrepreneurial activity young. Clifford agrees that past independent action can be informative, but says culture and available options matter: the same behavior will look different for someone in Silicon Valley and someone in rural India. He looks for evidence of initiative, not one identical childhood résumé.

After years of AI investing, Clifford sees recent progress as driven heavily by compute and data. He suspects the marginal value of scaling language models with more text may be leveling off. If so, he believes new ideas could become more valuable relative to sheer scale, creating room for startups.

Possible next directions include search methods, multimodal data and interactive experience. He does not claim large labs will miss those opportunities or that building a frontier company will become cheap. His argument is that a sufficiently different idea could help a new entrant attract the capital required to compete.

Data may be the hardest input to keep scaling, he says, even though dollars can still purchase more compute. He notes that video and other modalities had not been fully exploited, while their usefulness and processing cost remain uncertain. His forecast is specifically about diminishing gains from more text, not the exhaustion of all possible data.

If the industry cannot move beyond scaling existing language models, he expects a painful flattening of the hype cycle. His bet is that the concentration of talent and capital will help produce a new improvement curve, but he admits that this may be wrong.

He sees convergence among then-current models trained with similar methods and predicts more divergence if labs develop and protect different ideas. He speculates that future GPT and Llama generations may require more than simply scaling their predecessors; these are hypotheses about unreleased systems at the time.

Asked whether he would buy OpenAI at a $90 billion valuation, Clifford says no, despite admiring the company. Stebbings also declines, citing potential commoditization. Clifford says he keeps several radically different scenarios in mind, including one in which models converge and price competition erodes returns.

On whether China is two years behind the United States in AI, Clifford says it is difficult to know. He draws on his 2023 government work and meetings in Beijing, but does not present a verified quantitative gap or claim to know all Chinese capabilities.

He describes Chinese officials he met as technically sophisticated and says the UK also had relatively strong policy capacity among the countries involved in the summit. These are impressions from his role, not a systematic ranking of governments.

Clifford argues that Chinese authorities are highly concerned with stability and may see both powerful companies and powerful AI as destabilizing. He describes a tightly regulated environment, including political requirements for model providers. His interpretation of government motives is a political assessment, not an official statement.

His skeptical case for Chinese frontier leadership is that restrictions on entrepreneurs could make it harder to give a small number of ambitious founders the freedom available to leading US and UK labs. Stebbings raises the competing view that China has strong industrial capacity and access to data.

Clifford notes an information asymmetry: Western AI writing is rapidly translated and circulated in China, while comparable Chinese material may receive less attention in English. If progress depends mostly on scaling well-known ideas, he thinks China's industrial capacity could be a considerable advantage.

He says US semiconductor export controls make building very large GPU clusters in China harder, though not impossible. He mentions a Chinese company's claim about domestically produced chips but does not independently establish its performance. The effect of restrictions depends on how quickly domestic alternatives improve.

The bearish case for China, in his framing, is that a compute gap persists during a decisive period of rapid Western progress. The bullish case is that development takes longer and domestic hardware matures, reducing the leverage of export controls. Neither scenario establishes which will occur.

Stebbings anticipates continued US restrictions under a future administration. Clifford observes that both Trump and Biden administrations had pursued export controls, and expects national-security concerns to remain important. This is a 2024 political forecast, not a statement of current policy.

They discuss vertical integration across models and chips. Clifford invokes the idea that companies seek to commoditize complementary products while protecting their own core. He expects firms to experiment with different parts of the stack but says the performance of future models will determine much of the outcome.

A meaningful improvement, he suggests, would be much more reliable agents. At the time, developers could build impressive demonstrations but often needed repeated runs and struggled with robustness. He predicts good agents within five years, while explicitly not asserting that the next model will already operate at near-human autonomy.

If autonomous agents become widespread, he argues, economic value will depend on infrastructure for their interaction. He uses automated trading as an intuition for why rules, monitoring and handling unusual situations matter, while acknowledging that rule-based trading systems are not the same as AI agents.

He imagines protocols that allow agents to transact, remain observable, follow rules and be stopped when needed. In his view, government should set standards and private companies may build scalable tools. This is a proposed division of labor and business opportunity, not an established regulatory framework.

Stebbings worries that a small group understands AI while much of society may be left behind. Clifford says people can benefit from technologies they do not understand, but acknowledges that technological change has often favored more skilled workers and that the distribution of benefits is a real policy challenge.

Clifford criticizes the EU AI Act as overly anticipatory and burdensome and praises the UK's lighter approach at the time. These are his policy opinions as a UK adviser and entrepreneur, not a neutral legal interpretation or a statement of today's law.

He argues the UK has strong research talent and may offer a better environment for retaining a small AI research team than a highly competitive Bay Area labor market. He cites universities, established labs and startups. His position is a forecast and also reflects his role in the UK technology ecosystem.

Clifford believes the UK could become as prosperous per person as the United States, but says it would require choices that make the country attractive for building and scaling technology. He cites complaints from major companies about local planning barriers for data centers. This is his political and economic view, not an established projection.

On weak UK public markets, he distinguishes the ability to build a company from who ultimately benefits financially. He would like UK pension funds and savers to own more of the gains from successful technology companies, even when those companies choose to list elsewhere.

He recalls a US state pension fund investing in EF while a comparable UK plan he pitched had no venture allocation. He argues that if companies stay private longer, institutions without access to private investment may miss part of their growth. EF raises investment capital, so he has an interest in the allocation he advocates.

Stebbings identifies defense company Helsing as an angel investment and asks how AI may change warfare. Clifford says present AI is not more powerful than nuclear weapons and warns against underrating nuclear war. The military scenarios that follow are his analysis, not a tested prediction of battlefield outcomes.

He expects autonomous systems to change warfare and considers the balance between offense and defense an open question. Cheap, capable drones could potentially give smaller or non-state actors asymmetric advantages over expensive assets, he says. He is interested in defensive technologies to address such risks.

Clifford supports technological growth and rejects the idea that progress automatically produces either good outcomes or catastrophe. He argues that who builds powerful technology, the order in which capabilities appear and the availability of defenses all matter.

He says he has shifted time at EF toward defensive technologies because he sees both commercial opportunity and societal need. Stebbings challenges investors' confidence in discussing offensive military systems; Clifford argues countries must also retain the ability to defend their values. These are their perspectives and involve investment interests.

On founding teams, Clifford says synergy matters but increasingly emphasizes the peak ability of the strongest person. He distinguishes consistent everyday performance from exceptional performance at a pivotal moment. This is a selection heuristic from EF's experience, not proof that one person's talent determines every outcome.

When founders separate, he says, one explanation is that they discover they want different things as the company changes. First-time founders cannot always predict what the job will feel like, and he does not expect them to foresee every later preference.

Clifford calls entrepreneurship a highly skilled profession and does not believe everyone should or can become a founder. He cites research on people pushed into entrepreneurship when employers reduced hiring during a recession, saying those founders performed better on average than a comparison cohort. The interview does not provide enough detail to generalize that result to every recession or person.

His interpretation is that talent and ambition can be directed into several careers, and ecosystems differ in which path appears natural. He argues that the UK, Europe and India can find strong founders among people who might otherwise enter finance or other prestigious jobs; EF's business is to recruit and fund such people.

Stebbings asks why ambitious people may choose finance in the UK when US peers pursue startups. Clifford points to local examples and expectations: people imitate visible paths to success. He says the situation has improved since EF began, while acknowledging that many founders also value impact beyond money.

EF learned not to overreact to early, short-term signals when selecting talent. As its applicant pool broadened, Clifford says experienced but merely good candidates could look stronger on paper than exceptional younger candidates with little experience. His lesson is to assess potential relative to opportunity, not assume youth itself guarantees excellence.

He describes EF's process: a cohort spends roughly ten weeks exploring ideas and potential co-founders. Of 100 participants, he estimates about 80 pair up; EF then chooses to fund about half of those combinations after asking whether the founders want to spend years on the idea and whether its ambition fits them.

At that stage, he says EF invests $250,000 and roughly two-thirds to three-quarters of funded teams typically raise a seed round. These are his reported conversion estimates for that point in the process, not the percentage of every applicant who raises or a measure of investment returns.

He admires angel investor Charlie Songhurst for recognizing a founder's best potential and pitching back a more ambitious version of an idea. Clifford says a founder's response to that larger vision can reveal motivation, particularly in environments where people have learned to downplay ambition.

His unconventional advice is that people can start a company with someone they have just met. He argues that the apparent dominance of long-standing co-founder relationships may partly reflect historical selection: those were the teams investors were able to observe. EF's matching model gives him a commercial interest in that claim.

He has changed his view that it was too late to start a frontier AGI company. He now thinks ambitious founders should consider it, subject to his earlier caveats about capital and uncertainty.

As a father, Clifford says parenting has taught him humility and that relationships require consistent investment rather than shortcuts. The rewarding moments, he says, often come after routine, difficult work rather than as an immediate result of it.

Asked what people rarely ask him, he discusses writing immersive murder-mystery games. He enjoys scenarios in which participants assume unfamiliar roles, make choices without a predetermined outcome and explore a different world without real-world consequences.