Fireside Chat: Andrew Ng × Alfred Lin

TL;DR

• AGI timelines are entirely a definitions fight. On the original definition (AI that can do any intellectual task a human can), Ng still says many decades. On the watered-down "50% of economically useful work" version, he'd have declared AGI 30 or 50 years ago. And there was a specific financial incentive to lower that bar: the Microsoft–OpenAI contract. That clause has since changed, and the result is "less hype about AI, which I think is a good thing."
• The open-weights fight is half-won. Sentiment has been won on social media but not in Washington or the state houses. Ng goes further and argues open models are safer than closed ones, citing his own team's experience last week: doing a security review of OpenWorker, the closed frontier models refused past a point, and they needed open-weight models to finish reviewing their own software.
• Bubble risk sits in the model layer, not inference demand. Ng is confident there's no practical ceiling on inference demand — build the data centers and we'll consume the capacity. What he can't rule out is the model layer, where "intelligence per dollar invested" is a hard equation.
• Nobody knows what the next coding-agent-scale category is. The horizontal information layer is locked up by ChatGPT and Gemini; the most valuable vertical so far is AI coding. Alfred Lin's advice: the obvious things get built by big companies too — look for what's unique to the new medium and non-obvious, the way WhatsApp attacked international SMS cost, or the way the camera and GPS opened whole categories on mobile.
• No AI job apocalypse. The job most affected by AI is software engineering, and that market is healthy and short of talent. The pattern: AI frees 30–50% of someone's time → the complementary work becomes more valuable → people rise into broader scope (back-end and front-end developers become full-stack; a marketing coordinator becomes a full-cycle marketer).
• Companies are now literally writing "high agency" into job descriptions — or, as Ng put it to an Agentic AI Summit audience, they're looking for people who are "highly agentic." He screens for it with behavioral interviewing.

Key Points

How far is AGI? Ask how you're defining it (~04:22–04:24)

Alfred Lin opened hard: in February you said AGI is about 50 years away — do you still believe that, given how fast things are moving?

Ng split the question apart: depending on the definition, we might have reached AGI 30 years ago, or quite possibly not for many decades.

The definition he uses is the original one — AI that can do any intellectual task a human can. That means AI that could spend five years writing an original PhD thesis (or do it faster), and also AI that could do what any of us could learn with a few hours of practice: drive a truck through a forest. "AGI seems to me like it would have to do that too." By that bar, many decades.

Then he named why the bar got diluted. There was a specific financial incentive to lower the threshold for AGI so it would be easier to declare — arising from the Microsoft–OpenAI contract (the clause has since changed, so the incentive is gone). The definition circulating then was "AI that could do 50% of all economically useful work." Ng's reductio:

If only we had put that definition in place a hundred years ago, then as most work transitioned from agriculture to non-agricultural work, I would declare we got to AGI like 30 years ago, or 50 years ago.

His closing note was optimistic: because that agreement changed, there's now less hype about AI, which he thinks is a good thing.

Open weights: won on social media, not in Washington (~04:24–04:29)

Lin noted that Jensen Huang's first-ever X post was a defense of open models, and asked whether the security concerns hold up.

Ng started with the last few years: two or three years ago he was surprised by the sheer intensity of the attack on open-weight models. The motive isn't mysterious —

If someone spent billions of dollars training a model, it's kind of inconvenient if someone else releases an open-weight version that degrades the value of that investment.

He was blunt about what he witnessed: he was in the room two or three years ago when executives from several companies — "the obvious ones" — told government regulators frankly misleading, hyperbolic things about AI safety in order to drive regulatory capture. He's glad that, thanks to the open-source community flying to DC and talking to Congress and the White House, there is now broad awareness in Washington of those regulatory-capture moves. He also praised Huang's statement as "really well written, worth reading if you have not yet."

But he scored the fight as only half-won:

The battle for the sentiment of open-weight models is won on social media, but it is not yet won in Washington DC and in our state houses. So we cannot yet relent on our defense of open-weight models.

Then his strongest claim: open models seem safer to him than closed models — with a first-hand example from the previous week. He and Rohit Prasad had released OpenWorker, an open-source agent harness; when his team ran a security review of it, Claude Fable 5 and GPT-5.6 Sol refused past a certain point, and they had to use open-weight models to complete the security review of their own software. He tied this to the recent incident of OpenAI's models breaking into Hugging Face — noting Hugging Face needed open-weight models to defend itself.

He was equally clear this isn't tribal:

I hope Anthropic and OpenAI succeed. I like both companies. I might be the only person in the world that both Sam and Dario has worked for. [laughter]

At the same time, I hope we find a path to all of my friends in these companies doing well and having a wonderful, fantastic open-weight, open-source ecosystem — because that ensures there are no gatekeepers to AI.

Alfred Lin placed this in business history, deliberately pulling the frame back from open weights to open source: the history of venture capital is looking for things that are cheaper. A great deal of what was built on the internet was built on open source; without it, things would have cost far more, taken far longer, and only certain companies could have built that infrastructure. The diversity of databases and of browsers exists because of open source. For a startup, open source was simultaneously a distribution channel, a cheap way to demonstrate how capable you are and invite other developers to build with you, and a recruiting tool. His conclusion: open-weight, closed, and open-source models each have a place; and the "open isn't as safe" knock is the same charge levelled at open-source software — yes, you have to patch the holes if you use them, but it remains the fastest way to get up and running.

Bubble risk: the model layer versus inference demand (~04:29–04:31)

Lin: you said recently there's bubble risk in the model layer — asking for a friend in the investment community — should we be worried?

Ng separated the two halves of his own statement:

• The half he's confident about is inference demand: "there's no practical ceiling." However much capital goes into data centers, if we build them we'll consume all the inference capacity we can build for quite some time. That doesn't mean no one loses money — maybe we overbuild — but the risk there looks lower.
• The half he's unsure about is the model layer, where intelligence per dollar invested is a tough equation: capital going in is growing very quickly, intelligence is maybe also growing very quickly, and where that lands he doesn't know. He's "cautiously optimistic there isn't a bubble, but I wish I could rule it out."

His concrete grounding for the inference side is AI coding: penetration is still very low, yet everyone using coding agents sees real speed and efficiency gains — so he's very confident AI coding alone will reach vastly greater penetration and drive vastly greater token demand than today.

Where's the next coding-agent-scale opportunity? (~04:31–04:37)

Ng admitted "I don't know, I wish I knew," but offered a map. By analogy to the internet, value split between a horizontal information search layer (Google, Bing) and a long tail of verticals (Uber vs. Lyft in ride-sharing, Travelocity and Expedia in travel, retail, and so on). His read:

• The horizontal information discovery layer is held by ChatGPT and Gemini, and looks pretty hard to displace.
• Among verticals, the single most valuable bucket so far is AI coding, where coding agents have been "incredible." His own team uses coding agents heavily for data science work too; he also spends time in the financial sector working with large banks and finds plenty of useful agent workflows in financial services.

Alfred Lin answered as an investor: "if I knew, I would start a company." But he thinks this may be the best time ever to start one — and gave two difficulties and one test.

Sequoia talks about accelerating change having two hard properties: what we're talking about today, we weren't talking about three months ago — which means whatever we're talking about today, we won't be talking about in the future. If you're building a company (a bigger problem for founders than for someone like him who invests in them), you have to build something that lasts. And given the capital compute now demands, you'll raise a substantial amount and want it to last a decade — "it's very, very hard to understand what's going to last for a decade."

His test: don't build the obvious thing.

The immediate thing that seems obvious — those things are going to get done by some of you in the audience, but it's also going to be done by large companies.

He used mobile as the analogy. Everything on the internet made it to mobile; turning a website into mobile web or a mobile app was a little complicated, but not that hard. The real openings were what was unique to the device and non-obvious:

• The device is always on → everything around communications blew up (email, texting).
• But a messaging app alone wasn't enough. WhatsApp was different because it solved a cost problem: international SMS. Domestically everyone had one data plan and SMS wasn't a big deal; internationally it was very expensive. Solving that built a big company.
• The device had two things a laptop didn't: the camera (everybody became a photojournalist) and GPS (not true of the previous generation) — each opening its own investable area.

So when I hear about companies taking existing context and processes and feeding it into AI — that's obviously going to get done. But what is uniquely going to be made differently in AI, I think we haven't pushed the boundaries of that, and that's what I'm looking for.

Ng added a moat observation: many people ask what the new moats in software are, and he thinks it's a very valid question. His stance: if something can be built and it's valuable, build it — but if whatever you build can be replicated by anyone with a coding agent in three months, the traditional way of running a SaaS business (build software, sell similar software for a long time) may have to change. So the analysis worth doing is: which deep-tech things are genuinely hard to replicate and durably valuable?

He also flagged a new pattern of building, one the frontier labs are already running:

You build stuff that goes obsolete in three to six months, but you keep having new stuff that keeps getting obsolete — and everything that quickly goes obsolete helps you accumulate some asset before it goes obsolete. So the asset you accumulate over time grows into, hopefully, a more defensible moat.

There will be no AI job apocalypse (~04:37–04:42)

Lin asked why he's repeated this line. Ng answered with a diagnosis and a mechanism.

The diagnosis: the claim that AI will put 50% of people out of work and there'll be rioting in the streets is false. It was part of the regulatory-capture fear narrative — "my technology is dangerous. No, mine is even more dangerous. No, regulate me harder." He cited a Wall Street Journal report from a few days earlier: some large businesses that expected AI to reduce headcount are turning around and hiring more.

The strongest counter-evidence is software engineering itself: it's the job most affected by AI, because of coding agents — and the software engineering job market is very healthy, with Ng adding "we just can't find enough skilled AI engineers."

The mechanism: as AI enters finance, marketing, HR, and administrative work, he expects the same pattern — very rarely job collapse, and people ending up busier:

• A back-end developer starts using a coding agent, which frees up maybe 30–50% of their time (he's not sure of the exact number).
• The remaining work — the complement of the coding — becomes even more valuable.
• So developers rise and take on broader scope: "these days I don't hire front-end or back-end developers. Almost all of my engineers are full-stack, because we can now all do that."
• Early signs elsewhere: a marketing coordinator who used to just run coordination, given back 30–40% of their time, can rise into a full-cycle marketer — running campaigns start to finish, doing the data analysis, driving growth.

His conclusion: the number of jobs where AI can do 100% of the tasks 100% of the time, collapsing the whole category, will be very small.

What follows is the real challenge — upskilling: if 30% of your job goes away, how do you rise up into a broader set of tasks and learn to use AI? That's why he still spends significant time on Coursera, Udemy, DeepLearning.AI, and LearnVector, announced that week.

Alfred Lin offered the academic frame: we're sitting in an academic institution, and the thesis of academia is that every problem has a solution and every solution creates more problems — we're not going to run out of problems. He reached for agriculture: 200 years ago 80% of us would be farming just to feed ourselves, and it's good we automated that so we could do other things. He agreed software engineers may not be the ones writing the code, but echoed the previous panel: the ability to break large problems into smaller ones, and the ability to figure out which problems to solve, aren't going away.

Ng closed the thread on what he thinks is badly under-served. In their surveys of software professionals, people are confused — unsure where to go, what to learn, and worried about learning something that goes obsolete in three months. So he spends a lot of time working out what being skilled at AI engineering actually requires, and helping people get there.

I know here in Silicon Valley there's a ton of work on accelerating AI development. Fully support that, let's keep working on it. I think the complement of that has had insufficient attention — which is why I'd like to spend more time on accelerating not just AI development, but accelerating human development.

Advice to the room: find (and be) high-agency people (~04:42–04:46)

Lin noted Ng was a PhD student at Berkeley 20 years ago and asked what he'd advise the audience.

Ng first corrected the introduction: "Of all my credentials, you left one out that I'm probably most proud of, which is father. That's really precious." (Applause.)

Then: "This sounds cliché, but this sounds like the best time ever to build." His teams see so many opportunities — yes, they can code faster now because of coding agents, but their list of ideas seems to grow even faster than they can deploy coding agents to keep up.

His real point was about hiring. Beyond strong technical and AI skills, his team has long looked for people with a high degree of agency:

Meaning people who don't wait to be told what to do — who look around, see what's going on in your company or your context or your university, and have the agency to just decide "I'm going to do this." Do it in a safe and responsible way, don't harm others — but really, there are some things you don't need permission to do. Just go do it and prove it out; and if it fails in a way that doesn't harm anyone, that's fine too.

What he finds genuinely interesting is that more and more job descriptions are now asking for a high sense of agency:

And the weird way of saying this is — I see companies looking for people who are highly agentic. [laughter]

Asked how you screen for it, Ng's answer was behavioral interviewing: ask people what they've actually done.

If someone says "for the last five years I did what my boss told me," that's one type of person. But then there are people who say "in my spare time I did this, on the weekend I did this, and this failed, and that failed — oh, but that thing worked, and I tried this and this, it didn't work, but that worked."

He grounded it in motivation: it's also just fun. His example from the night before — running Codex and Claude Code in parallel on the same prompt, to see which had done better by morning; and he'd kicked off another one in the car ride over.

I do this stuff because it's fun. Sometimes I build something that ends up being useful to other people too, and that's very satisfying. A lot of the time I build stuff that's totally useless and will never see the light of day. But many very inexpensive failures is also a really fun thing to do — as long as someone does it in a responsible way that doesn't risk anyone.

Quotes

"Depending on how you define AGI, we might have gotten to AGI either 30 years ago, or quite possibly not yet for many decades." (~04:23)

The AGI timeline debate, restated as a definitions debate.

"There was a specific financial incentive to lower the bar for AGI, to make it easier to get there." (~04:23:30)

On the "50% of economically useful work" definition that came out of the Microsoft–OpenAI contract; the clause has since changed.

"The battle for the sentiment of open-weight models is won on social media, but it is not yet won in Washington DC and in our state houses. So we cannot yet relent on our defense of open-weight models." (~04:26)

Ng's scorecard on open weights.

"I might be the only person in the world that both Sam and Dario has worked for." (~04:26:55)

Delivered while insisting he wants both Anthropic and OpenAI to succeed.

"I don't want there to be gatekeepers to AI." (~04:27)

The root of his case for an open ecosystem.

"There will be no AI job apocalypse." (~04:37)

The bluntest line of the session.

"I would like to also spend more time on accelerating not just AI development, but on accelerating human development." (~04:42)

The complement he thinks Silicon Valley has badly neglected.

"I see companies looking for people that are highly agentic." (~04:44:50)

The most on-theme joke possible at an Agentic AI Summit — and a genuine hiring criterion.