Power generation, power generation for AI.
April 16, AI Circle weather: Sunny
Yesterday we talked about that Musk, who is opposed to AI, founded a company called X.AI, which seems a little contradictory in the power generation, but today he came out to play the circle by himself: to do cosmic-level AI - sure enough, the big guy's dream is also cosmic-level!
Musk wants to be the AI that understands the universe: such an AI cannot destroy humanity
Musk wants to make an AI called TruthGPT to replace ChatGPT, and his goal is to use this AI to "maximize the search for the truth and try to understand the nature of the universe", because he believes that an AI concerned with understanding the universe cannot destroy mankind.
His concerns about AI are as good as ever, and in the latest interview, Musk continued to call for stronger AI regulation, suggesting that the U.S. government set up relevant departments to oversee the entire industry.
For the popular ChatGPT, his concern is that if he encounters a programmer with ulterior motives, the AI he trained may also lie.
In this way, the actions of large companies in the field of AI, we still have to pay close attention, and we must not give them the opportunity to do evil. For example, Meta's big trick you see if you want to Pick:
Zuckerberg announced the Meta vision model with a high profile, does computer vision exist?
Following the release of SAM, which "divides everything", Meta re-released DINOv2, which is a heavyweight open source project in the CV field, which can achieve self-supervised training in depth estimation, semantic segmentation, and image similarity comparison.
——"Depth estimation" here is a new word, for example, when reconstructing a 2D picture into 3D, it is necessary to estimate the distance of each point from the shooting source, which is the technical meaning of "depth estimation".
Xiaoza said that the model can be used to generate forest heights on various continents with the help of satellite imagery. In the future, medical imaging, food production and other aspects can also be used.
Of course, in the end, Xiaoza circled back to his "main business" - the meta-universe. He believes that DINOv2 can greatly bless the construction of the meta-universe and make the experience of users more immersive in the meta-universe.
Netizens shouted, "Computer vision makes Meta lose again"!
This wave of big model competition set off by ChatGPT is intensifying, and the industry has also reached the level of "thirst for talents" for large model talents, first look at the country:
100,000 monthly salary, big models crazy to grab people
According to the latest survey data, the demand for AIGC talents rose against the trend from January to February this year, and the number of jobs increased by 31.3% year-on-year, hitting a record high. On the recruitment platform, enter AIGC, NLP and other related keywords to search, you can see that high-paying positions such as "50k + 16 salary" are waiting for vacant positions. It is said that some headhunters have issued a monthly salary of 100,000 to rob people!
Zhang Lei, CEO of Stardust Data, predicts that "there should be no more than 1,000 talents with large-scale distributed training experience in China, and conservatively speaking, there are only two or three hundred people." ”
Originally, excellent algorithm engineers are scarce resources, and R&D talents who train oversized models are rare. Such a person gets a monthly salary of 100,000, and we have nothing to drool about.
Under the boom of large models, the rising tide is not only the monthly salary of migrant workers, but also GPU chips:
The large model is so hot that the chip skyrockets, and Ali Baidu has become a working boy of NVIDIA
GPUs are a key component of ChatGPT training and deployment, which is the infrastructure for powerful computing power.
Taking the A100 chip used in ChatGPT as an example, from the end of last year to now, its unit price has increased by more than 50%, from about 50,000 yuan to nearly 100,000 yuan.
Domestic technology companies have trained large models, some people have calculated an account, if you want to directly train a GPT-3 level large model, the conventional minimum needs 1000 A100-level GPU chips, and if you want to actually deploy, regardless of redundancy and response speed, for search engine-level services, at least 6,000 A100-level GPU chips are required. For chips alone, the minimum investment has reached nearly 1 billion yuan.
Moreover, since the A100 has been banned from sale to Chinese mainland by the United States, domestic counterparts estimate that only 30,000 pieces are in stock. The A100 replacement version of A800 launched by NVIDIA, domestic manufacturers are also lining up to get goods, and often out of stock.
Nvidia has been unable to supply goods, and its stock price has doubled in the past half a year, becoming the fifth largest listed company in the United States, more than the market value of chip companies such as Texas Instruments, AMD, Qualcomm, and Intel combined.
However, seeing that the big model is so hot, the creator of ChatGPT spoke-
CEO of OpenAI: The era of building massive AI models is over
Sam Altman, CEO of OpenAI, warns that the era of building massive AI models and developing generative AI robots is over.
"I think we're at the end of the era of building big AI models, and in the future, we're going to make those models better in other ways."
Altman's comments represent an unexpected turn in the current race for tech companies to develop and deploy new AI algorithms and compare the scale of their models.
GPT-4 may be the last major development for OpenAI, after Altman has made it clear that GPT-5 has not been developed and will not be for some time.
So the question is, is this wave set off by ChatGPT going to retreat?
Finally, Power Generator Jun will share an interesting short video today, let Musk talk to Jobs, talk about AI, iPhone, colonizing Mars,—— Of course, this video is generated by AI:
References: New Zhiyuan, Qubit, City Boundary, All-weather Technology, Chinese Entrepreneurs