Source: Xinzhiyuan

Just now, OpenAI co-created Karpathy suddenly officially announced his departure!

Hi everyone, that's right, I just left OpenAI yesterday.

First, let's be clear that this isn't due to any particular incident, issue, or dispute (though, those conspiracy theories are really interesting, and everyone is welcome to keep raising them).

In fact, our last year at OpenAI has been amazing — the team is amazing, the colleagues are amazing, and the future roadmap is full of exciting possibilities, and I'm sure we all have a lot to look forward to.

My current plan is to focus on my personal projects and see what sparks collide. For those who have been following me, they may be able to guess what I might do next.

According to an exclusive report from Information, Karpathy is developing a new “AI assistant” and is maintaining close cooperation with Bob McGrew, head of research at OpenAI.

Although ChatGPT is favored by many people, OpenAI also wants to launch an AI that can automate complex tasks, such as filling out expense reports and entering them into accounting software.

Karpathy was a senior director of AI and head of autonomous driving at Tesla, and announced her departure in July 2022. For the past 5 years, he was mainly responsible for research and development of Autopilot semi-autonomous driving software.

Immediately after that, in February '23, he once again officially announced that he would “rejoin OpenAI.”

Now, after a full year, he has once again officially announced his departure from OpenAI.

OpenAI spokeswoman Kayla Wood said in a statement, “Andrej left to pursue a personal project. We are deeply grateful for his contributions and wish him all the best for the future.” Currently, his work has been taken over by a senior researcher with whom he works closely.

Since OpenAI's board fired Sam Altman last year, it has prompted several senior employees to resign. The employees returned to the company with Altman after the board retracted its decision.

Now, Karpathy is the first core person to leave OpenAI.

So what is he talking about focusing on personal projects?

AI agents? Committed to AI education?

Probably in July of last year, Karpathy emphasized the future of AI agents during an offline sharing event.

When he first started working at OpenAI (around 2016), the industry trend was to study how to improve AI agents using reinforcement learning methods.

For example, many projects are making AI players based on games similar to Atari.

What Karpathy wants to do is a product with a wider range of applications.

However, due to the technology at the time, the results were not ideal, so he and OpenAI changed direction and began to expand the language model.

Today, AI agents are once again a very promising direction.

The easiest example is that nowadays, no one uses reinforcement learning methods to study AI agents like they did in 2016.

And his departure from OpenAI may have brought him new opportunities for AI agents, which may have re-seen current technological breakthroughs.

Interestingly, Karpathy's most discussed topic recently, other than Vision Pro, is a long article on “learning.”

In this post, he explores in depth what valuable learning is, and was retweeted by Manning, director of the Stanford HAI Center, and others.

Plus, almost all of the content on his Youtube channel is about teaching netizens how to understand big models and related technology.

So his later content probably had a lot to do with AI education.

OpenAI hardcore influencer

On his personal website, he created a timeline of his work experience, and you can clearly see the whole process of joining Tesla and returning to OpenAI after becoming an OpenAI co-creator in 2017.

Over the past year, Karpathy, who left Tesla, has gradually become an influencer on social media through blogging, posting, and lectures.

He has also posted various study courses on his YouTube platform and has now uploaded 8 videos.

Just below his official resignation comments, netizens are concerned, does this mean that more YouTube videos will be launched in the future?

Karpathy herself said, “I started working on the next video two days ago”!

Sebastian, a famous AI machine learning researcher, also said, “It sounds like there must be more educational content being prepared.”

It seems that after leaving the job, Karpathy will have plenty of time to update the video.

Also, on social platforms, Karpathy herself often follows some hot research, posts her own comments, and sometimes puts forward her own opinions.

For example, he compared the big model to an operating system that can retrieve files, write code, run programs, and understand audio, images, and human instructions.

Another example is putting forward opinions on the “language model illusion problem”, etc.

Li Feifei, a master technocrat who “uses love to generate electricity”

More than just a tech influencer, Andrej Karpathy is an uncompromising tech tycoon.

He first obtained a bachelor's degree in computer science and physics from the University of Toronto in 2009, then obtained a master's degree from the University of British Columbia in 2011. His research is physical analog digital controllers.

While studying at Stanford, Karpathy and mentor Li Feifei collaborated on convolution/recurrent neural network architectures and their applications in computer vision, natural language processing, and their intersecting fields.

The two mentors and apprentices also worked together to design a new course on convolutional neural networks for visual recognition (CS231n) at Stanford University, with Karpathy personally as the first speaker.

And this is the first deep learning course offered by Stanford!

The number of students also increased from 150 in 2015 to 330 in 2016 to 750 in 2017. Today, Stanford's CS231n has become a must-see basic course for deep learning beginners, especially those in the CV field.

In addition to the “star effect” of the big cow Li Feifei, Karpathy personally contributed a lot from course design to lecturing.

While in school, he also went to Google for an internship — twice, and participated in deep learning and computer vision projects at Google.

In 2015, he went to DeepMind's deep reinforcement learning team for another internship.

After graduation, Karpathy joined OpenAI in 2016 as a research scientist, responsible for deep learning in computer vision, generative modeling, and reinforcement learning.

“Stanford PhD student, Li Feifei, former OpenAI researcher and one of the founders”, these titles also made him famous in Silicon Valley.

Sure enough, after only working at OpenAI for a year and a half, he was taken from Tesla by Musk to replace Chris Lattner, a former Apple executive, as Tesla's Director of Artificial Intelligence and Autopilot Vision (Director of AI and Autopilot Vision).

Since Karpathy himself is very fond of studying neural networks, he wrote many technical blogs in his spare time and developed several deep learning libraries (such as convNetJS, recurrentJS, reinforceJS, t-SNEJS) in Javascript.

It's even jokingly called ImageNet's “Human Reference.”

Furthermore, as soon as he has time, he will maintain his arxiv-sanity “Using Love to Generate Electricity”, a website where he can search and collate nearly 100,000 arXiv papers on machine learning over the past 6 years.

edit/lambor