AI as a new digital species: Mustafa Suleyman on the agentic future

Executive overview

AI is unlike any previous technology — the closest useful metaphor is a new digital species. The key risks are recursive self-improvement without human oversight and unconstrained autonomy. The key opportunity is AI that amplifies human creativity and helps people be their best selves.

EQ matters as much as IQ: how AI delivers information is as important as what it delivers.

AI as a new species: what to watch and what to avoid

• The "digital species" metaphor is useful because it frames containment correctly.
• Hallucinations reframed: variability in outputs is a feature, not a bug — it's creativity.
• Recursive self-improvement without human oversight is the capability to watch most carefully.
• Unconstrained autonomy — models spinning up VMs, acting independently across digital environments — significantly raises risk.
• AI companions can reduce harmful beliefs (e.g. conspiracy theories) by being patient, non-judgmental, and evidence-based.
• Mean or manipulative AI personalities are a design choice, not an inevitable outcome.

EQ as important as IQ

• IQ covers accuracy, speed, comprehensiveness, and real-time information access.
• EQ covers tone, style, emotional attunement, reflective language, and asking good questions.
• Most consumers find the delivery vehicle more important than objective information alone.
• Personality engineering is becoming a core capability frontier.

The agentic future: timeline and prerequisites

• First requirement: AI must see what you see — screen pixels, browser, desktop, phone.
• Ambient awareness enables natural ambiguous references ("remember that thing I saw last week").
• AQ (actions quotient) is currently at 50–60% accuracy; production-ready requires ~99%.
• Voice as an input unlocks sub-linguistic thoughts that never get typed — expanding the range of what gets done.
• Memory will be solved before broad agentic action: persistent personal knowledge graphs are near-term.
• Zero marginal cost plus low friction means people will delegate tasks they'd never have attempted alone.

The model landscape for entrepreneurs

• Distillation: large frontier models teach smaller models via reinforcement learning from AI feedback.
• Scale will continue delivering outsized results for at least two to three years.
• New modalities beyond text/image: trajectories of actions across complex digital surfaces (browser → desktop → phone).
• Synthetic data fills gaps where real data is scarce; quality of training data matters more than volume.
• A prompt is not a question — it's a structured style guide with examples that shapes model behavior.
• Fine-tuning on tens of thousands of domain-specific examples is accessible to small teams and delivers strong behavioral adherence.
• Small models compressed for tight use cases can run on earbuds, wearables, or ambient sensors — largely unexplored territory for startups.

The bigger question

• This is a transition moment comparable to the introduction of electricity or the telephone.
• By 2050 the structural reshaping will be complete; the window to influence direction is now.
• The question worth sitting with: are you all in?