Super Agency — AI as Amplification Intelligence
AI magnifies human agency through network effects, not replacement
Hoffman reframes AI not as artificial intelligence but as 'amplification intelligence' — a general purpose technology (GPT) that gives individuals cognitive superpowers and compounds those superpowers through network effects. The thesis is structural: every GPT in history (printing press, electricity, cars, mobile phones) followed the same pattern — initial disruption and dislocation, then a net-positive societal transformation at a scale that was previously unimaginable.
The compound mechanism is the key insight: a doctor with AI can serve more patients, which means the doctor's patients benefit from expanded capability even if they never use AI themselves. This is 'super agency' — the superpower is not just personal, it is network-amplified. As more nodes in a network gain AI capability, every connected person's effective agency increases.
The practical implication is that the right question is not whether AI will transform society (Hoffman treats this as certain) but how to manage the transition. He explicitly rejects AI pause arguments on the grounds that unilateral pausing by safety-conscious actors simply cedes the frontier to less safety-conscious actors — the clock is set by global competitive pressure, not by any individual actor's choice.
- Every general purpose technology produces the same initial discourse — disruption, misinformation risk, displacement of current knowledge custodians — before delivering its net-positive transformation.
- Super agency is network-amplified: when others gain AI superpowers, your own effective agency increases because the people you interact with become more capable.
- The transition period is the genuine risk, not the endpoint — managing the transition requires staying at the frontier, not pausing.
- AI capability already exceeds individual human knowledge span; the appropriate frame is 'idiot savant co-pilot' not 'helpful tool'.
- Personal AI adoption must go beyond toy use cases — apply it to something that matters in your area of expertise to understand the actual capability ceiling.
- Reframe your mental model of AIStop using the 'tool' frame and adopt the 'amplification intelligence / cognitive superpower' frame. This changes what questions you ask about AI deployment — instead of 'what tasks can it automate?' ask 'how does this change the capability ceiling of the people in my network?'Pro tipHoffman uses historical GPT analogies (printing press, electricity) not to predict outcomes but to calibrate the magnitude of disruption and the timeframe — decades, not quarters.WarningThe 'tool' frame causes organisations to under-invest; the 'Terminator' frame causes paralysis. Both are wrong.
- Go personally use AI on something that mattersHoffman's universal advice to everyone from Bank of England governors to students: personally use AI on something within your domain of expertise, not on trivial tasks. This is the fastest way to calibrate actual capability versus hype.WarningUsing AI only to make a sonnet for your kid's birthday gives you no signal about its actual capability in your domain.
- Apply the adversarial prompting techniqueFeed AI your argument, then prompt: 'Counter this argument, take the role of critic.' Then: 'Take my side, give me another argument for what I'm arguing.' Then: 'How would a historian of technology criticise this argument?' This forces you to stress-test your thinking using AI's breadth of knowledge.Pro tipHoffman uses this technique for strategic thinking, not just content creation. 'It's literally only your own creativity and inspiration — who should I be talking to?'
- Map the network amplification effects in your industryIdentify which actors in your value chain are adopting AI and what that does to your own effective capability and competitive position. The network amplification means your suppliers, customers, and partners gaining AI capability affects your outcomes even if you move slowly.
- Focus on transition management, not endpoint predictionHoffman has '100% confidence' the endpoint is enormously positive, but manages against transition risks: governance failures, bad actors empowered by AI asymmetry, and epistemological collapse. Build your AI strategy around surviving the transition, not just reaching the endpoint.Pro tipThe transition risks Hoffman names (governance, bad actors, epistemology) are the same ones Bret Weinstein identified from the outside — convergent diagnosis from practitioners and critics.
Hoffman's advice to a Bank of England governor (representing central bank conservatism and institutional risk-aversion) was identical to his advice to entrepreneurs: go personally use AI on something that matters to you within your area of expertise. Not a sandbox demo, not a briefing — direct personal use on real problems.
Hoffman co-founded Inflection AI to build Pi, a consumer AI agent, with the thesis that AI agents would be how most people access AI capabilities. Mustafa Suleyman subsequently joined Microsoft to pursue a consumer agent product there; Inflection pivoted to B2B enterprise software.
Hoffman personally uses AI by feeding it his argument, then prompting it to counter the argument as a critic, then to argue for his position from a different angle, then to apply a historian-of-technology lens. He uses this for strategic thinking on real decisions, not for content generation.
Hoffman developed this thesis through direct operational experience: he co-authored 'Impromptu' with GPT-4 (one of the first major books co-written with AI), co-founded Inflection AI and built the Pi consumer AI product, served on the OpenAI board during its critical scaling period, and was an early-stage investor at Greylock across the consumer internet and AI waves. His framing of AI as a GPT follows the historical pattern he observed across LinkedIn (social graph), PayPal (payments network), and his Greylock portfolio.
The 'amplification intelligence' language was developed specifically to counter both the Terminator-robot fear framing and the 'just a tool' minimisation framing. Hoffman's position is that both framings are wrong: AI is already more capable than any individual human across breadth of knowledge domains (his 'idiot savant' example: no human can simultaneously compare mixture-of-experts AI architecture, modern economic game theory, and modern oceanography — GPT-4 can), while also not being conscious or autonomous in the ways that generate existential risk.