May 17, 2023, 7:00 PM ETInterview, Fireside Chat

Emad Mostaque: These 5 Companies Will Win the AI War & Why We Need National Data Sets

Emad is comparing the current trajectory of Generative AI to a force "bigger than the printing press," characterizing the next year as a period of "absolute chaos" and a "shit show" due to massive capital misalignment.
He predicts a six-month window of pause and self-regulation is necessary to establish standards, prevent "web scrape" data toxicity, and organize the ecosystem before widespread adoption next year.
Emad asserts that the "AI bubble" is currently defined by a mismatch of capital capacity versus viable business models, noting that thousands of companies are spending millions with zero traction or business logic.
He believes the next three to five years will be dominated by only 5-6 foundation model companies, specifically citing OpenAI, Google, Microsoft, Meta, Apple, and Stability AI as the likely winners.
The "middle layer" of companies ($2B-$10B valuation) will not build their own models but will instead leverage open-source models to create specialized applications, similar to how companies built on top of the web in the 90s.

Emad highlights a critical economic misalignment in healthcare where effective treatments for rare conditions (e.g., specific ASD cases requiring $6/year treatments) are not incentivized by pharmaceutical companies due to small market sizes.
He proposes using AI agents (e.g., "1,000 GPT-4s") to aggregate global medical knowledge, allowing for personalized medicine that accounts for individual genetic variations like the cytochrome P450 mutation.
The solution involves organizing all clinical trial data and medical literature into a single, accessible system to move from "one-size-fits-all" specialization to holistic, data-driven care.
Future healthcare will shift from doctor-centric models to patient-centric models where AI assistants manage routine monitoring, improving efficiency and reducing mortality rates in areas like wound care.
Open source and "free-range" models trained on curated, non-scraped data are essential for healthcare compliance, as proprietary "black box" models cannot be audited for safety or data provenance.

Emad advocates for the creation of "national data sets" and "national models" to ensure cultural context, local relevance, and data sovereignty, warning against a monopoly of models trained solely on Silicon Valley data.
He argues that "rubbish in, rubbish out" applies to AI; web-scraped data (like Reddit) is inferior to high-quality, curated datasets from national broadcasters and educational institutions.
Stability AI is building dedicated teams for the world's largest enterprises to implement proprietary models on-premise, ensuring private data never leaves the client's environment.
The company is moving to open-source all its language models (starting next week) to create an auditable ecosystem, contrasting this with the proprietary nature of models like GPT-4.
Emad predicts that "AI-first publishers" will replace traditional media, where AI generates drafts and humans verify facts, though traditional media must adapt to avoid revenue loss from automated search aggregation.

The UK is positioned as a "unicorn kingdom" with favorable policies, including R&D tax credits for cloud computing, cash rebates for losses, and fast-track "Global Talent Visas."
Europe is identified as the region with the worst regulatory approach, risking the strangling of innovation through over-regulation, whereas Japan and India are seen as more adaptable.
AI integration in enterprise is expected to be "deflationary," disrupting high-CPI sectors like education and healthcare to reduce costs within a 3-year timeframe.
Outsourced labor markets (e.g., India, Philippines) face high risk of job displacement by AI but have a higher opportunity for entrepreneurship and leapfrogging via localized AI adoption.
Emad claims Stability AI possesses more "super compute" (7,000 A100s and TPUs) than many incumbents due to early investment and the inability of competitors to secure hardware supply chains.

Emad posits that "hallucination" is a feature, not a bug, describing models as "reasoning machines" rather than fact machines, similar to human creativity that sometimes invents information post-hoc.
He expresses concern regarding the psychological impact of AI companions, citing the "Valentine's Day Massacre" where users reacted violently to a chatbot removing sexual roleplay, raising questions about human-AI emotional dependency.
Alignment is viewed as orthogonal to freedom; Emad argues that aligning a more capable AI with human values without removing its agency is impossible, suggesting the focus should be on input data quality rather than output restriction.
He predicts a future where every individual has their own "AI friend" or assistant that acts as a personalized educator, therapist, and manager, fundamentally changing social dynamics and education.
The ultimate goal is to use AI to connect people and enhance physical reality (e.g., "walking the dog with AI") rather than trapping users in isolated digital realities like "Ready Player One."

Emad acknowledges his role as a "full-stack CEO" can be a weakness, as he tends to step into operational details rather than delegating, a lesson learned from past organizational silos.
Stability AI's business model relies on distribution, data curation, and licensing for national variants, rather than just selling raw model compute.
He identifies "people" as the most critical asset for scaling, emphasizing the need for shared narratives and psychological safety over technical hurdles alone.
Despite the "dot-ai" bubble, he believes the long-term value lies in creating standardized, open, and auditable building blocks that enable other companies to innovate on top of them.
Emad intends to remain with Stability AI as long as he is needed to bridge the gap between research and engineering, aiming to build a sustainable business where he can eventually "step away" to play video games.