Source: AI Frontline WeChat account: ai-front
Finishing | Dongmei, Nuclear Cola
In order to develop its own AIGC,
Musk madly buys GPUs and digs people around
On April 11, local time, according to multiple foreign media reports, despite high-profile recommendations to stop AI training across the industry, Elon Musk himself launched a new major AI project within Twitter.
According to Business Insider, Twitter has purchased about 10,000 GPUs and recruited AI talents from DeepMind, intending to develop its own large language model (LLM) project.
Musk's AI project is still in its infancy, a person familiar with the matter said. But according to another source, the purchase of large amounts of computing power shows that he is trying to promote the development of the project. However, it is unclear what the exact purpose of Twitter's generative AI model is, and potential directions may include improving search functions or generating targeted advertising content.
Exactly what hardware Twitter purchased remains a mystery. But according to reports, Twitter is spending tens of millions of dollars on GPU equipment despite its recent financial troubles. The GPUs are expected to be deployed in one of Twitter's two remaining data centers, with the Atlanta data center more likely. Interestingly, Musk just closed Twitter's main data center in Sacramento in late December last year, which obviously compressed the company's computing resources.
In addition to procuring GPU hardware for generative AI projects, Twitter is also hiring more engineers.
Earlier this year, Twitter Inc. recruited engineers Igor Babuschkin and Manuel Kroiss from Alphabet subsidiary DeepMind. Since at least February, Musk has been actively scouting for talent in the AI field, hoping to compete with OpenAI's ChatGPT.
OpenAI uses Nvidia's A100 GPU to train its ChatGPT chatbot and power its subsequent execution and inference. Today, Nvidia has launched a successor to the A100, with its H100 GPU capable of several times the performance for roughly the same amount of power. It's possible that Twitter is using the Nvidia Hopper H100 or similar hardware in its AI project, but there's no clear evidence yet. It's also hard to estimate how many Hopper GPUs they'll need, given that the company hasn't determined what the new AI project will actually be used for.
When a giant company like Twitter purchases hardware, the unit price will also be relatively favorable due to the large scale of demand. In addition, if purchased separately from retailers such as CDW, the current unit price of Nvidia H100 is likely to exceed $10,000. A rough estimate shows that Twitter is indeed planning to spend a lot of money on its own AI project.
Leading the call for a moratorium on advanced AI research and development,
Might just be a delaying tactic
Although Musk bought a large number of GPUs in an attempt to secretly train his own advanced AI, before that, he took the lead in publicly calling for a moratorium on the development of advanced AI technology for up to 6 months. The reasons given by him and a group of industry leaders who oppose continuing to develop AI are: We already have GPT-4, and we should slow down now, and don’t rush to come up with a new AI system that is more powerful than it.
At the time, Musk's move attracted much attention. Because for a long time, whether it is studying autonomous driving or brain-computer interface, Musk left the impression to the outside world as an "Iron Man" who can always be at the forefront of technology. Openly opposing the development of more advanced AI is not in line with his usual style.
According to media reports, in fact, Musk is not against AI, but OpenAI and GPT.
Musk is not only the original initiator of the Future of Life Institute, but also one of the co-founders of OpenAI. But he left OpenAI's board in 2018 and withdrew a large grant.
The US "Fortune" magazine reported at the time that the reason for leaving was that although Musk supported artificial intelligence, there was a "potential conflict of interest" between Tesla's smart driving technology and OpenAI.
There are also reports that Musk offered to run OpenAI himself before leaving, but was rejected. It is not difficult to understand why after leaving OpenAI, Musk has been targeting OpenAI and everything related to it.
In addition to speaking publicly to attack OpenAI, Musk also went deep into OpenAI to poach their employees. Andrej Kapathy, Tesla's head of autonomous driving, who will leave in 2022, is one of OpenAI's best talents.
Some netizens believe that the above actions show that Musk is not really opposed to more advanced AI technology, but hates that this technology has entered the hands of others first, and he can only become a bystander.
Zhu Xiaohu, managing partner of GSR Ventures, expressed his opinion on the event of thousands of signatures, “It’s actually buying time for competitors. OpenAI’s iterative speed is too fast. Google’s search is 20% better than rivals’. It occupies 90% of the search market share. OpenAI’s influence will be more extensive, and all upper-level applications hope to use cloud services powered by more powerful AI. Not to mention Microsoft’s already powerful front-end Office applications. It is hoped that OpenAI will develop to a certain stage like autonomous driving, and it will encounter an insurmountable bottleneck.”
Looking at it now, the collective appeal at that time was more like a farce of "secretly crossing the old warehouse". It was also a delaying strategy by Musk and the technology giants who were left behind by the OpenAI ecological partners. After all, some of the signers were personally involved. People who research and deploy AI models, such as Emad Mostaque, CEO of Stability AI. The company released Stable Diffusion, a text-to-image model, last year.
Reference link:
https://www.tomshardware.com/news/elon-musk-buys-tens-of-thousands-of-gpus-for-twitter-ai-project
https://futureoflife.org/open-letter/pause-giant-ai-experiments/
https://www.reuters.com/technology/musk-experts-urge-pause-training-ai-systems-that-can-outperform-gpt-4-2023-03-29/