Unique Research · Insight / interview digest · October 30, 2023 · a16z AI Revolution with Bob Swan · Original: a16z.com/ai-copilots-and-the-future-of-knowledge-work/
Microsoft CTO Kevin Scott on why AI Copilots keep developers in flow longer — and may mark the start of a knowledge-work industrial revolution. Microsoft treats AI as a platform, not a product or research project; partners (GitHub, OpenAI, Meta Llama 2) are central.
Microsoft’s stance on AI
Bob: Microsoft shapes AI’s future — what’s the vision, and what’s big tech’s role? Kevin: AI is a real platform many people will build unexpected things on — like PCs or smartphones. Especially in the last 12 months with ChatGPT and GPT-4, platform potential is visible. As a lifelong tool-builder, watching others use tools in unforeseen ways is the thrill.
Partners matter if you’re building a platform. Most important AI partner: OpenAI — 4 years — first co-building AI supercomputing to train frontier models, then shipping into Microsoft products (GitHub Copilot) and Azure / Azure OpenAI APIs for others’ software. GitHub is among the most interesting AI deploy surfaces — first proof of the Copilot pattern: take someone’s knowledge/cognitive work and massively boost productivity mid-task. Not enough developers for the software the world needs; much of the work is toil — tools that make building more interesting and productive matter. Llama 2 (like Llama) will be an important building block; Microsoft wants that AI software easy to run on Azure. And yes — the GPU question is now daily comedy; almost no day without it.
The next industrial revolution
First industrial revolution: machines assisting physical labor that dominated 18th–20th century work. Analogous shift now for cognitive/knowledge work — which takes a growing share of labor time — without prior productivity breakthroughs on the scale of PCs or the internet (smartphones didn’t much help knowledge-work productivity). This may be the biggest event ever for that class of work.
Two early patterns: (1) categories of knowledge work society wants more of but couldn’t expand because productivity didn’t rise; (2) automating the repetitive/least-liked parts so experience and output both improve — startups will hunt “toil people are eager to automate.” As with prior platforms, the most valuable uses often aren’t the first 1–2 years’ apps (smartphone time isn’t SMS/browser/mail — it’s later-native experiences). Hard question: what previously impossible hard things become possible? Easy useful things exist; they aren’t the highest value.
Developer productivity & flow state
Inside Microsoft: get everyone on GitHub Copilot first. Biggest observed effect: keep developers in flow longer — unblock next steps without leaving flow to dig docs or interrupt busy teammates. Measure productivity not as LOC but as speed of shipping value to users and whether users benefit — full feedback-loop friction across the product lifecycle. Without metrics, hard to choose among the flood of AI dev tools.
Personal experiment pre-GPT-4: could these tools help write the sci-fi novel he’d wanted since teens? Bad at fully formed character-driven chapters — excellent at keeping him in flow; easy to measure ~2× words/day. Universal question: when are people most productive/joyful? In flow. How do tools extend that?
Another internal use: AI for deploying AI — much testing and responsible-AI work now uses AI tools themselves; importance wasn’t obvious ~9 months before large-scale deploy.
Reimagining the American Dream
On motivating teams without fear — and handling resistance: Kevin wrote Reprogramming the American Dream (shelved March 2020, COVID week) arguing AI creates far more opportunity than harm — without zeroing harm or ignoring it as a first-order product concern; stay firmly optimistic. Grew up rural central Virginia (tobacco, furniture, textiles — all collapsed by high-school graduation). When such communities get powerful tools, they build economic opportunity for selves/families/communities solving problems coastal elites don’t see. These AI tools are more accessible than ever — curiosity and entrepreneurship beat needing a CS PhD.
On resistance: last 9 months at Microsoft (and in startups) — once a critical mass builds excitedly on the tech, peers adapt and FOMO pulls others in. Organizational job: reach that adoption tipping point; then the opposite problem appears — people yelling they don’t have enough GPUs.
Advice to builders
Biggest open technical question: a ~20W organ between our ears is AGI far more efficient than today’s digital ANNs — many breakthroughs possible to close the efficiency gap (laptop-class work vs. multi-million-$ infra). That speeds iteration and cuts product cost — good for Microsoft’s infra bill and for global innovation pace.
To founders: seek hard problems; don’t chase trivial “AI fart apps.” Being possible ≠ being useful. Fundamentals unchanged: who is the user, what’s the problem, what can you do — then ask whether AI is useful infrastructure. AI is a model, not a product — new infrastructure that unlocks new problem classes or better solutions to old ones.
2 CN images; attach on ingest. Dialogue facts preserved.
— END —