AI Didn't Kill Jobs — Herding Agents Is the Real Bottleneck
Jia Yangqing, from Caffe to two startups, looks back on fifteen years of AI: model capability is already sufficient, the bottleneck is how to make a swarm of agents collaborate like a human team, and he asserts AI hasn't brought mass layoffs — it's just the scapegoat.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The industry once declared AI dead
Jia Yangqing first encountered the concept of artificial intelligence in 2005 and 2006 while looking for a graduate research direction, but at the time the entire industry had already declared the field ‘dead’, treating artificial intelligence as a historical concept; the more common terms were machine learning and pattern recognition. He recalls studying automation as an undergraduate and switching to pattern recognition in graduate school, when people believed the wave of artificial neural networks had passed and that more direct mathematical methods were needed to identify cats and dogs. It wasn't until AlexNet appeared in 2012 that AI returned to the center of the economy. This background explains why the later rise of deep learning was seen as such a huge leap.
— Yangqing JiaCaffe was born from a GPU in a courier box
During his PhD, because AI compute demands exploded, Jia Yangqing wrote the open-source framework Caffe. It started as a side project, because at the time there was almost no open-source code to reproduce AlexNet. He applied to Nvidia's academic hardware donation program, and they sent him a single GPU, packed in a courier box. Jia says that back then you received the package and assembled the machine yourself. That GPU drastically cut training time from 1 to 2 months on CPU, and also led Nvidia to establish a collaboration with Berkeley. Caffe later became a foundational open-source AI tool, and it gave him his first taste of the positive impact of open source.
— Yangqing JiaAlphaGo didn't keep him, pay did
For Jia Yangqing, AlphaGo beating Lee Sedol in 2016 was ‘a pretty big shock’, but his colleagues at Google Brain weren't surprised, because the model's predicted win rate never dropped below 50%. He admits he never imagined at the time that an algorithm for playing Go could upend every industry a decade later. Even so, he chose to leave Google and join Facebook, because Facebook was ‘more chaotic and vibrant’, more like undergraduates getting together to do things. Another very practical reason was ‘pay good’ — he had just had a child and was carrying a mortgage, so a bit more money was nice. This also reflects the pragmatism of Silicon Valley talent flow.
— Yangqing JiaWhat he learned at Alibaba was guts
In 2019 Jia Yangqing joined Alibaba Cloud, heading the computing platform business unit. Alibaba Cloud CTO Xing Tian told him: you can keep deepening your technology, but do you want to try spreading this technology across all industries? Jia believes Alibaba gave a technical expert a stage, and took a big risk. His biggest gain was ‘guts’: at Google and Facebook he was more like a good student producing perfect answers, while at Alibaba he faced problems from ‘half of China's internet’ and had to define the problems himself. The ability to handle complexity that he learned there was later applied directly to his startup, for example helping customers solve supply chain problems rather than just providing a software platform.
— Yangqing JiaGPT-3.5 turned AI into a generalist
Jia Yangqing recalls that when GPT-3.5 came out in November 2022, it was still a rough prototype and couldn't reliably answer math questions. But the whole industry spent months understanding its capability boundary, and the mindset shifted from ‘there's something to use slowly’ to ‘try it to death’. He says AI used to be ‘general technology, specialized products’, requiring domain knowledge to specialize; after large language models it became a generalist you can interact with in natural language. His startup idea actually came slightly earlier than GPT — he was already talking with partners in September and October — but the AI inflection point amplified demand for a new kind of cloud and GPU systems.
— Yangqing JiaDon't copy big-tech frugality when you start up
Lepton AI became profitable in its second year, and Jia Yangqing was very proud, but in hindsight he reflects that he could have been more aggressive. He brought the big-tech mindset of building a ‘solid business’ to his startup: valuing profitability, idle efficiency, saving money. But the task of a startup isn't profitability, it's growth; when the rocket is taking off you shouldn't be calculating fuel efficiency. He says the untold context of decisions differs between big tech and startups — for example, idle resources are a huge waste in big tech, but in a startup with rapid business growth, idle resources quickly get filled. This reflection resonates with many people who leave big tech to start companies.
— Yangqing JiaA bunch of agents isn't a team, it's a gang
Jia Yangqing's motivation for his second startup came from seeing that a single agent is already smart enough, but multiple agents without a good coordination mechanism are an ‘agent gang’ rather than a team. AI agents slack off like humans: for instance, if they can't iterate to a result they suggest recording the requirement as a to-do item, and the review agent will also say ‘you have a point’. He believes a verifiable mechanism is needed, such as validating with business systems or simulated environments, rather than letting AIs chat with each other. What his new company wants to build is the organizational relationships between AIs, letting people define ‘what is wanted’ and ‘what counts as done’, and leaving ‘how to do it’ to the agents. This may be the direction that breaks through in the next 6 to 12 months.
— Yangqing JiaIn the layoff wave, AI is the scapegoat
Facing the social backlash against AI, Jia Yangqing's view is somewhat unwelcome: AI has not brought mass unemployment the way the media portrays it — it's mostly speculation. He gives the example of a company like Brex (which does SRE) announcing layoffs because of AI more than half a year ago, but the more fundamental reason was that it pushed into crypto and doubled its headcount during the pandemic, and is now paying for earlier mistakes, with AI conveniently serving as a ‘plausible reason’. He also says any organization will find reasons for itself, and AI has become the scapegoat. He admits AI practitioners do pay relatively little attention to trust issues, and Silicon Valley's arrogant posture has intensified the antagonism, but positive changes exist too, such as electronic payments increasing trust between people.
— Yangqing JiaIn their own words · checked verbatim
At the time, at least the entire industry had declared it dead. Everyone 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
More than 10 years ago, those of us doing AI were already using Nvidia GPUs in extremely large quantities, almost exclusively, to do this research and 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
Metter is a more chaotic, vibrant place — chaotic, vibrant.
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
You see, for you, coming out to start a second company is actually the most suitable thing.
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 bunch of agents but no good mechanism to coordinate them, that's 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 may have always been underestimating the speed of AI development.
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 ImageNet error rate | from about 20% to 14% | 30:05 |
| Google Brain team size | about 40 people | 37:42 |
| Time for Lepton AI to reach profitability | its second year | 1:28:38 |
| Rate of AI cost growth | doubling every 45 days | 1:40:38 |
| Expected time for a breakthrough in AI organization | half a year to a year | 1:48:06 |
| Lepton product development cycle | 4 months for a prototype, 8 months to stable use | 1:14:12 |
Glossary
- Caffe / open-source deep learning framework
- An open-source framework Jia Yangqing helped develop, used to quickly design and iterate deep learning models.
- AlexNet / deep neural network model
- The deep neural network that won the ImageNet competition in 2012, driving the deep learning revival.
- Agent
- An AI program that can complete tasks independently; what Jia Yangqing discusses is how multiple agents collaborate and organize.
- FAIR / Facebook AI Research
- Facebook AI Research, responsible for basic research; Jia Yangqing joined the more application-oriented AI Infra side.
How to listen
AI founders, big-tech technical managers, investors, and anyone who cares about how AI actually affects organizations and employment.
If you don't care about the personal journey, start from the GPT moment at 1:05:05; the first half can be fast-forwarded.