AI Didn't Kill Jobs — Agent Teamwork Is the Real Bottleneck
Yangqing Jia (贾扬清) traces fifteen years of AI from Caffe through two startups: model capability is already good enough, and the bottleneck is getting a group of agents to collaborate the way a human team does. He also argues flatly that AI has not caused mass layoffs — it is just the scapegoat.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The whole field had already pronounced AI dead
When Jia was looking for a graduate research direction in 2005 and 2006, he encountered the concept of artificial intelligence for the first time — and the entire industry had already ruled the direction ‘dead’, treating artificial intelligence as a historical concept. The terms people actually used were machine learning, pattern recognition and the like. He recalls doing automation as an undergraduate and switching to pattern recognition for graduate school: the belief then was that the wave of artificial neural networks had passed, and that telling cats from dogs called for more direct mathematical methods. Not until AlexNet appeared in 2012 did AI move back to the centre of the economy. That backdrop explains why the later rise of deep learning was experienced as such an enormous jump.
— Yangqing JiaCaffe was born from a single GPU that arrived in a shipping box
During his PhD, as the compute demands of AI shot up, Jia wrote the open-source framework Caffe. It began as a side project, for the simple reason that at the time there was almost no open-source code that could reproduce AlexNet. He applied to NVIDIA's academic hardware donation programme, and they mailed him a GPU — it showed up in a courier box. As Jia puts it, back then you got the delivery and assembled the machine yourself. That single GPU cut training time down sharply from the one to two months it took on CPUs, and it also started a collaboration between NVIDIA and Berkeley. Caffe went on to become a foundational open-source tool for AI, and it gave him his first taste of the positive feedback loop of open source.
— Yangqing JiaAlphaGo didn't keep him at Google; the pay moved him
When AlphaGo beat Lee Sedol in 2016, it was ‘quite a shock’ for Jia — but his colleagues at Google Brain were not surprised, because the model's predicted win rate had never dropped below 50%. He admits he had absolutely no idea then that an algorithm for playing Go would, ten years later, upend every industry. Even so, he chose to leave Google for Facebook, because Facebook was ‘messier and more alive’, more like a bunch of undergraduates getting together to build things. The other, very practical reason was that the ‘pay was good’ — he had just had a baby and was carrying a mortgage, so more money was welcome. It is also a fair picture of how pragmatic Silicon Valley talent movement is.
— Yangqing JiaWhat Alibaba taught him was nerve, not technology
In 2019 Jia joined Alibaba Cloud to run the computing platform business unit. Alibaba Cloud CTO Xingtian (行天) told him: you can keep going deeper on the technology, but would you like to try pushing this technology out across every industry? Jia's view is that Alibaba gave a technical expert a stage — and took a substantial risk doing it. His biggest takeaway was ‘nerve’: at Google and Facebook he had been more like the good student producing the perfect answer, whereas at Alibaba he faced the problems of ‘half of the Chinese internet’ and had to define the problems himself. The capacity he built there for handling complexity fed directly into his own startups later — for instance, solving customers' supply-chain problems rather than only shipping a software platform.
— Yangqing JiaGPT-3.5 turned AI from a specialist tool into a generalist
Jia recalls that when GPT-3.5 came out in November 2022 it was still a rough prototype — ask it a maths question and it might well get it wrong. But the industry spent several months working out where its capability boundaries were, and the collective mindset shifted from ‘here's a thing, let's use it slowly’ to ‘push it until it breaks’. His framing: AI used to be a ‘general-purpose technology with special-purpose products’, requiring domain knowledge to specialise it; after large language models it became a generalist you could interact with in natural language. His own startup idea actually predated GPT slightly — he was already talking it over with his co-founder in September and October — but the AI inflection point amplified demand for a new kind of cloud and for GPU systems.
— Yangqing JiaA startup's job is growth, not the big-company habit of saving money
Lepton AI was profitable in its second year, and Jia is proud of that — but looking back, he thinks he could have been more aggressive. He had carried the big-company approach to running a ‘solid business’ into a startup: an emphasis on profitability, on idle-capacity efficiency, on saving money. But a startup's task is not profit, it is growth; you don't calculate fuel efficiency while the rocket is lifting off. The untold context behind a decision, he says, is different in a large company than in a startup: inside a big company idle resources are a huge loss, whereas in a startup growing fast, idle resources get filled up soon enough. That piece of self-criticism will resonate with a lot of people who left Big Tech to found something.
— Yangqing JiaA pile of agents is a gang, not a team
The motivation behind Jia's second company came from seeing that a single agent is already smart enough, but that multiple agents without a good coordinating mechanism amount to an ‘agent gang’ rather than a team. AI agents slack off the way people do: if an agent can't iterate its way to a result, it will suggest logging the requirement as a to-do item, and a review agent will tell you ‘you make a good point’. What is needed, he argues, are verifiable mechanisms — validation through a business system or a simulated environment, rather than letting AIs chat idly with each other. What his new company is building is the organisational relationships between AIs: humans define ‘what we want’ and ‘what counts as done well’, and hand ‘how to do it’ to the agents. This may be the direction that breaks through in the next six to twelve months.
— Yangqing JiaIn the layoff wave, AI is the scapegoat, not the cause
On the social backlash against AI, Jia's position cuts against the grain: AI has not produced the mass unemployment the media portrays, and most of that is conjecture. His example is Brex (a company doing SRE), which more than six months ago announced layoffs because of AI — but the more fundamental reasons were its push into crypto and the doubling of headcount during the pandemic, so it is now paying for earlier mistakes, with AI serving as the ‘high-minded excuse’. Any organisation, he adds, will find itself a reason, and AI has become the scapegoat. He concedes that people working in AI do pay relatively little attention to the question of trust, and that Silicon Valley's arrogant posture sharpens the antagonism — but positive changes exist too, such as the way electronic payments raised the level of trust between people.
— Yangqing JiaIn their own words · checked verbatim
At the time, at the very least, the whole industry had pronounced it dead. People thought artificial intelligence was a historical concept.
当时至少整个业界是判定他的死亡了。大家认为人工智能是一个历史概念
Yangqing Jia8:19
Those of us doing AI, more than ten years ago, were already using NVIDIA GPUs in enormous quantities — almost exclusively — to do this research and this training.
我们做AI的基本上在10年多以前，就已经非常非常大量的，几乎是exly的用英伟达的GPU开始。去做这些科研和训练。
Yangqing Jia23:55
Meta was a messier, more alive place. Messy, and full of life.
metter是一个更加乱七八糟的生机勃勃的地方，乱七八糟，生机勃勃。
Yangqing Jia41:57
Look, Yangqing, for people like you, coming out and doing a second startup is really the most fitting thing to do.
你看杨青对你们来说，你们出来第二次创业其实是一个最合适一件事情。
Yangqing Jia1:38:25
If you have a pile of agents but no good mechanism to coordinate them, what you have is an agent gang.
如果说有一堆agent，但是你没有一个好的机制来协调他们的话，这就是一个agent团伙
Yangqing Jia1:43:44
I think I probably have, genuinely, been underestimating the speed of AI's development all along.
我觉得我可能一直的确在呃在低估就是AI发展的速度
Yangqing Jia2:03:49
Figures
| Size of the MNIST handwritten-digit dataset | 50,000 samples | 10:21 |
| AlexNet's improvement on the ImageNet error rate | from about 20% down to 14% | 30:05 |
| Size of the Google Brain team | about 40 people | 37:42 |
| When Lepton AI became profitable | its second year | 1:28:38 |
| Rate of growth in AI costs | doubling every 45 days | 1:40:38 |
| Expected timeline for the breakthrough in AI organisation | six months to a year | 1:48:06 |
| Lepton's product development cycle | 4 months to a prototype, 8 months to stability | 1:14:12 |
Glossary
- Caffe
- The open-source deep learning framework Jia helped build, used for rapidly designing and iterating on deep learning models.
- AlexNet
- The deep neural network that won the ImageNet competition in 2012 and drove the revival of deep learning.
- Agent
- An AI program that can complete tasks independently; what Jia is discussing is how multiple agents organise and collaborate.
- FAIR
- Facebook AI Research, responsible for fundamental research; Jia joined the more application-oriented AI Infra side.
How to listen
AI founders, engineering leaders at large tech companies, investors, and anyone who cares about how AI is actually reshaping organisations and employment.
If the personal history doesn't interest you, start at the GPT moment at 1:05:05 and fast-forward the first half.