Alexandr Wang on AI Data, Model Development and Building Scale

Scale AI founder and CEO Alexandr Wang tells Harry Stebbings that AI may become a major military asset and that data is central to improving models. His views on model development, regulation and geopolitics reflect both the interview’s 2024 context and Scale’s commercial role in supplying data services.

Wang observes that large increases in compute spending had not, at the time of the interview, produced a model he considered dramatically better than GPT-4. He points to rising Nvidia data-center revenue as evidence of investment. This is his assessment of that period, not proof that compute alone has stopped improving models.

He describes three pillars of model progress: compute, algorithms and data. Transformers and reinforcement learning from human feedback are examples of algorithmic advances. He believes the industry had expanded computation faster than the other pillars and was encountering a data bottleneck.

By “easy data,” Wang means material already written down and accessible on the internet. He says language models trained on such material can emulate patterns in it, but argues that useful agents need more than imitation: they must perform tasks, solve problems and collaborate with people.

As an example, a bank fraud analyst may inspect transactions, compare evidence and reason toward a decision without recording every step online. Wang argues that much of the reasoning behind economic work is not captured in public training data. This is his explanation for why more specialized data could matter.

He calls the missing material “frontier data”: complex reasoning, multi-step discussions, tool use, error correction and agent workflows. He proposes combining three efforts: using relevant enterprise data, producing new data and involving human experts. The terminology is his own framing.

Wang says large organizations have extensive proprietary data and cites a reported JPMorgan dataset as an illustration. He compares it with an estimate for a model’s training data, but the interview does not establish equivalent measurement methods or independently verify the figures. Enterprises would need to refine data for their own authorized uses.

On reasoning, he distinguishes human generalization from current model behavior. He sees two possible approaches: an algorithmic breakthrough in general reasoning, or supplying examples across many situations in which the model must perform. He presents these as research directions, not proven solutions.

Because private enterprise data cannot simply become a public training set, Wang argues that new data must also be produced. He compares the need to expand data production with building semiconductor fabrication capacity. His proposed process combines synthetic data generation with human expert guidance.

Experts could intervene when a model gets stuck, makes factual mistakes or encounters unfamiliar situations, much as a safety driver can take over an autonomous vehicle. Wang calls contributors or AI trainers a potentially high-leverage role because an improvement to a widely used model can affect many subsequent users.

Stebbings asks how organizations can structure and clean large stores of data. Wang separates mining existing records from producing new data going forward. He predicts sophisticated enterprises will increasingly mine existing sources, but argues that continued model improvement will also require ongoing data creation.

He believes progress is more constrained by data than compute and argues that producing the two in parallel could yield more capable models. This is the thesis of a company that sells data-related services; the interview does not establish a universal bottleneck across all models.

Potential new sources include process data from enterprise software and longer-term personal data collection through devices. Wang acknowledges the possibility can sound intrusive. He distinguishes such observation from expert-created complex reasoning and scientific data, which he believes are needed to push model capabilities further.

Asked about equitable access to proprietary sources, Wang argues that data can become a durable competitive advantage. Algorithms may spread and competitors may purchase more compute, while exclusive or specially produced data may be harder to replicate. He expects labs to develop distinct data strategies.

Stebbings mentions publishers’ licensing agreements as early examples. Wang suggests that in future labs may compete over rights and production of specialized data as visibly as they currently compete over GPU capacity. Whether any particular agreement yields a lasting advantage remains uncertain.

He outlines two possibilities: data strategies may converge and models become commoditized, or labs may produce differentiated datasets for their chosen use cases. He gives enterprise-focused and consumer-facing providers as hypothetical examples, without establishing the results of either strategy.

Large enterprises are cautious about exposing proprietary data, Wang says, because it may be a source of competitive advantage. He sees demand for models that can run on-premises and be customized with a company’s data without sending that data back to a model provider. This is a market prediction and commercial opportunity for Scale.

Stebbings suggests AI services may capture more revenue than model providers as businesses need help adopting the technology. Wang agrees that value capture is uncertain and can move through the stack. He sees potential both in infrastructure below models and applications and services above them.

They discuss the “end of software” thesis: AI may reduce the cost of creating software and move enterprises from a few large applications toward many specialized tools. Wang finds the comparison provocative rather than settled, but expects greater demand for software customized to specific business problems.

He points to Palantir’s historically tailored approach as an example of connecting an enterprise’s data to its particular workflows. He predicts AI will make customization easier. He also expects engineers to spend more time translating customer problems into scoped work that AI systems can help implement.

Wang predicts consumption-based pricing will become more important if AI agents perform a growing share of work. A software provider may want to charge for value generated by those systems rather than only for human seats. That shift is a forecast, not evidence that seat pricing has already ended.

On regulation, he argues that more permissive access to data can be compatible with liberal democracy, while acknowledging a balance is needed. He says policymakers should consider data availability as well as chip production. These are his policy preferences as the leader of a data company.

He proposes pooling some industrial data, such as aerospace safety or financial fraud information, to advance shared capabilities. For health data, he argues for clear ways to use patient records while protecting privacy, including anonymization. His broad characterization of existing privacy rules is a policy claim, not a legal analysis.

Wang says Chinese models had caught up considerably by the interview and cites a particular leaderboard as evidence. He predicts China could move ahead and credits centralized industrial policy for scaling established sectors such as solar and electric vehicles. Relative AI capability depends on the benchmark and time; his forecast is not an established outcome.

He views advanced AI as a potentially significant military asset and imagines scenarios in which a country with much more capable systems gains a major advantage. His speculation about how China or Russia might use hypothetical artificial general intelligence is not a statement of established intent or capability.

Wang proposes keeping the most powerful future systems closed for security while allowing less capable models to remain open for economic use. He says the difficult question is where to draw the line. This is his proposed policy balance, not an agreed safety standard.

He predicts the cost of frontier models could grow from hundreds of millions or billions to tens or hundreds of billions of dollars over a decade. If that happens, he expects efforts to concentrate around states and major technology companies. Both cost and market-structure claims are forecasts.

Stebbings asks whether smaller labs will be absorbed by cloud providers. Wang broadly agrees it is possible, while noting that partnerships between model labs and large technology companies may evolve in different ways. The interview does not establish any future acquisitions.

Turning to public relations, Wang argues that traditional media can reward attention-grabbing coverage and that companies should communicate directly through channels such as podcasts. He says he felt more fairly treated before Congress than by some media outlets. This is his experience and interpretation; independent reporting can also scrutinize company claims.

Wang says coverage of Scale shifted after the market downturn and that some reporting criticized its work with the US Department of Defense. He considers that work important for US security. Scale’s defense contracts create a commercial interest in his position, and the interview presents his account of the coverage.

In hiring, he looks for people who care deeply about the quality and impact of their work and will work through obstacles. He says Scale had about 800 employees at the time and that he still reviewed or interviewed every hire to maintain a high bar.

Wang estimates that he disagreed with a team’s hiring recommendation about 25–30% of the time, often when a manager needed calibration or a case was ambiguous. He sees his review as a way to apply knowledge of who has succeeded at Scale, though it also centralizes decision-making.

His largest management mistake was assuming rapid business growth required equally rapid headcount growth. He says Scale grew from about 150 employees in 2020 to more than 700 by the end of 2022. Hiring that quickly made it difficult to maintain standards and introduced organizational problems that appeared gradually.

The team then remained roughly stable in size while revenue grew, Wang says. He distinguishes growth in the business from growth in headcount and emphasizes maintaining a strong talent base through both fashionable and less fashionable periods.

He recalls Patrick Collison suggesting that some of Stripe’s best hires would have joined even without its status as a hot company. Stebbings argues that a strong brand can still help attract sales talent in new markets. Wang agrees that function and circumstance matter.

In the quick-fire round, Wang says he has changed his mind about linking team growth to business growth. He repeats that compute alone is not enough for AI progress and names Microsoft CEO Satya Nadella as someone whose strategic judgment he admires.

Scale began in 2016 focused on autonomous vehicles and later worked on generative AI, including with OpenAI on GPT-2, Wang says. Seeing successive cycles makes him cautious about promises that outpace technical reality. He recalls a painful period when autonomous-driving expectations ran ahead of delivery.

He worries generative AI could face a similar hangover if claims become too detached from capabilities. Some autonomous-vehicle systems did later reach roads, he says, but more measured expectations might have avoided a cycle of hype and disappointment.

Asked about the 2024 US election, Wang calls it a toss-up and says he does not know how swing states will vote. Looking ahead ten years, he hopes Scale will continue providing data infrastructure for AI and says he would consider a public listing. These were his intentions at the time, not promises.