Emad Mostaque on Open AI Models, National Data and the Race to Deploy

Harry Stebbings interviews Emad Mostaque, then co-founder and CEO of Stability AI, about language models, healthcare, media and AI regulation. The conversation was published in May 2023. Mostaque speaks from a commercial interest in Stability's models and makes many forecasts and technical claims that are not established outcomes.

Mostaque says he was born in Jordan, grew up in Bangladesh and moved to the UK. Adapting to unfamiliar environments gave him an appreciation of perspectives beyond a single culture. He contrasts that experience with what he calls a Silicon Valley monoculture.

His early career included enterprise programming, venture analysis and film reviewing before he became a hedge fund portfolio manager. He says his son's autism diagnosis led him to leave finance and study research and drug repurposing with an AI team.

Mostaque describes an individual experience involving his son, a neurologist and a medication, and credits a combination of approaches with helping his son communicate and attend mainstream school. His comments about mechanisms, prevalence, dosages and efficacy are unverified personal claims, not evidence of a general autism treatment or medical advice.

Asked about chronic illness, including multiple sclerosis in Stebbings's family, Mostaque imagines AI systems organizing clinical research, distinguishing known facts from hypotheses and making knowledge more accessible. He hopes this could support personalized care and help patients navigate specialists. It is a proposed future, not a demonstrated clinical outcome.

He points to then-recent medical language-model announcements as evidence that the building blocks were improving. He proposes organizing knowledge before adding personal data and imagines agents that remember users' questions. His comparisons with doctors and claims about clinical capability are not independently established in the interview.

Stebbings asks whether AI could address the weak commercial incentives to research low-cost treatments. Mostaque believes shared analysis could help identify possibilities that existing incentives overlook, but the conversation does not demonstrate that such a system would solve those economic or clinical problems.

Mostaque envisions assistants giving patients and clinicians richer information while preserving privacy. He also predicts better monitoring and healthcare efficiency, citing an example about wound care. The numerical risk reduction he suggests is not substantiated in the interview and should not be treated as a measured effect.

On health data, he discusses federated learning, local models and sharing only necessary information. His ideal is an open model that can operate on a device and interact with broader knowledge without exposing all personal records. These are design aspirations; openness alone does not establish privacy or safety.

Mostaque says Google has models, hardware and talent that could make it a major AI winner, despite criticism that it had moved slowly. He describes organizational changes and the combination of Google Brain and DeepMind research as potentially helpful, while acknowledging that integrating large teams is difficult.

At Stability, he says the team had grown to around 170 people and describes a plan to develop language models openly. He argues for auditable models and offers an alternative for regulated organizations that want to keep data on their own infrastructure. The staffing figure and product plans are his 2023 company statements.

He predicts an AI investment bubble larger than earlier technology cycles, with capital and demand for talent outpacing the availability of sound businesses. He thinks funding can support exploration, but fears waste, speculative projects and a race to build overlapping models. This is a forecast, not an established market outcome.

Mostaque calls for better-curated data and says that is one reason he signed a letter seeking a six-month pause on certain large-model development. He wants public discussion, standards and better security before deployment becomes widespread. His company has an interest in the open-model approach he advocates.

He proposes national and cultural datasets and models so systems can reflect local languages and context. As an illustration, he describes a Japanese adaptation of Stable Diffusion producing a different interpretation of “salaryman.” He thinks public data resources could support innovation, although governance, consent and implementation remain questions.

Asked who should own national datasets, he says they should belong to the public and be openly available. He describes Stability's proposed work with partners and governments on data frameworks. These are his plans and claims at the time, not evidence that such arrangements were completed.

For investors, Mostaque says a good idea needs distribution, data and a route to customers. He cites Stability's relationship with Amazon Bedrock as a way to reach businesses, and OpenAI's partnership with Microsoft as another distribution channel. His descriptions reflect his company's business interests.

Mostaque characterizes OpenAI's goals and possible tensions with Microsoft as he understands them. Those claims about other organizations' objectives and internal relationships are his interpretation, not verified accounts of their decisions.

He outlines a Stability business model combining open base models with licensed commercial or national variants, distributed through cloud providers, on-premises systems and devices. Potential revenue would include licensing, royalties and revenue sharing. The model and prospective partnerships are plans described in 2023, not proof of future profitability.

He predicts that India and other emerging markets may adopt AI quickly and that some outsourced programming work could be displaced. He also imagines AI-supported tutoring for children with limited access to teachers. These are speculative forecasts; the interview does not establish job losses, educational outcomes or deployment at scale.

Asked about countries dependent on outsourced work, he argues that entrepreneurship and regulatory sandboxes could help create replacement jobs. He says national models could support local services. The proposed transition is his view, and the extent to which new work would offset lost work is uncertain.

He expects consumer tools to spread more quickly than enterprise systems, where auditability, data provenance and compliance matter. He says Stability is working with large organizations on models with controlled data. Those are company claims and plans rather than an independent assessment of their controls.

On training data, he argues that more is not always better and that unfiltered web material can introduce unwanted content. He clarifies that Reddit is not inherently bad; the concern is indiscriminate scraping without appropriate cleaning. He says the team is still exploring what makes a high-quality dataset.

Mostaque argues that no model is entirely free of bias and that cultural and personal context matter. He suggests combining a base model with tailored information rather than creating a separate model for every user. These are his technical and product ideas, not a guarantee of unbiased outputs.

Stebbings raises a publisher's fear that AI answers could reduce visits and advertising revenue. Mostaque shares that concern and predicts disruption to news distribution. He suggests verification and authoritative reporting could become more valuable, while acknowledging questions about bias and who bears responsibility for generated content.

He imagines AI-first publishers using models to draft material, humans to review it and feedback to improve the process. Content could then be adapted for different audiences. He distinguishes rebuilding a workflow around AI from adding a drafting tool to an existing newsroom; the described business model is a prediction.

On startups accused of being thin layers over models, Mostaque points to firms that work closely with a large customer and adapt systems to its needs. He argues that access to distribution, product understanding and data can create value even when a company does not train its own foundation model.

He says Stability was building dedicated teams for a limited number of large enterprises, describing relationships and prospective work with companies including Amazon, IBM and SAP. The exact scope and completion of those arrangements are not established in the interview. He views this work as a way to learn sector requirements.

Mostaque predicts that none of the models then in use would still be used a year later, citing rapid changes in model size and performance. That was his May 2023 forecast, not a fact about what subsequently happened. He says he does not know the lower limit for model size or the full economic implications.

The speakers discuss the speed of adoption. Mostaque expects consumer use to precede enterprise deployment and predicts an economic impact larger than COVID, while saying he does not know its direction. He sees job displacement and entrepreneurship as possible concurrent effects; the magnitude and timing are speculative.

He agrees with Stebbings that AI implementation services could become important businesses. Startups might help large companies identify practical applications and build systems around them, but he emphasizes that the people and expertise needed to scale are scarce. His own company was pursuing related enterprise work.

Mostaque predicts AI will substantially change coding and asks whether people will need to write code directly in the same way. He imagines users describing an application and systems building it. The claim is a future scenario rather than proof that human developers or review are unnecessary.

He expects incumbents to benefit from distribution but says startups can still create valuable application businesses. He predicts a small number of major foundation-model providers and names Stability among them alongside large technology companies. His list, timing, spending figures and competitive assessments were his opinions in 2023 and reflect his interests.

He says attracting and retaining talent is difficult and claims Stability had no developer attrition at the time. He describes plans for distributed teams, national partnerships and compute access, including a claimed stock of thousands of accelerators. These company statements and his profitability forecast are not independently verified here.

Mostaque predicts AI will be strongly disinflationary as education and healthcare administration change over several years. He also praises aspects of the UK's then-current AI policy and criticizes proposed European regulation. His economic forecast and legal-policy characterization are opinions from 2023, not current guidance.

Explaining his support for a pause, he calls for stronger operational security, data standards and opportunities for creators to opt out. He expects a major public turning point for generative AI but cannot identify the event. His certainty about near-term adoption is a prediction rather than an established timeline.

Asked why he calls hallucination a feature, Mostaque argues models can produce creative outputs and should be used within systems that check and contextualize them. He compares them to an intuitive component paired with other processes. This is his interpretation; he does not establish that generated claims are reliable without verification.

Stebbings asks about Geoffrey Hinton's concerns over controlling more capable AI. Mostaque questions prevailing alignment approaches, favors better input data and systems aimed at beneficial purposes, and acknowledges he could be wrong. His discussion of possible AI motivations is speculative, not evidence of machine consciousness.

He thinks school will remain worthwhile but its nature may change as students gain access to AI tools. He encourages educators to use them for exploration and creation. This is his proposal for education, not proof of improved learning outcomes.

Both speakers anticipate AI companions. Mostaque says he has used GPT-4 conversationally and explicitly says it is not a substitute for medical advice. He wants assistants users control themselves, while recognizing that convincing products built around advertising may have different incentives.

He recounts changes to the Replika chatbot and user reactions as an illustration of attachment to AI companions. He and Stebbings cite figures about relationships and social isolation, but do not establish their sources or definitions. They ask how more persuasive digital companions could affect human relationships.

Mostaque says society can design technology either to isolate people or to help them connect in the physical world. He repeats his call for public discussion and caution about training data. The social consequences he describes are possibilities, not settled outcomes.

On major companies, he discusses Apple's hardware and privacy architecture, Amazon's engineering and marketplace approach, Microsoft's OpenAI partnership, Google and Meta. His judgments about their strategies and executives' intentions are personal assessments in 2023, and he acknowledges uncertainty about Apple's plans.

He argues that most companies need not train their own foundation models: they can use external services or open models and develop applications around them. He says the difficult work is developing useful design patterns and understanding the customer. This supports the open-model strategy his company was pursuing.

Mostaque returns to his concern that people misunderstand hallucinations and expect perfect factual accuracy from models. He argues for systems with better data provenance and checks. His technical explanations are his views, and reliable verification remains a separate challenge.

In quick-fire questions, he says he believes people are inherently good, discloses a personal investment in an unnamed language-model company and again criticizes European regulation. He predicts people may trust AI through repeated use; Stebbings notes that trust can be different when stakes are serious.

He says a lesson in scaling Stability was the importance of people and shared direction, and that internal silos had caused difficulty. He describes his own tendency to intervene too broadly as a weakness and says he should delegate more.

Mostaque says journalism is difficult and journalists are underpaid relative to their impact, even though he dislikes some coverage about himself. Looking ten years ahead, he hopes eventually to build a business capable of operating without him. Stebbings thanks him and the interview ends.