broadsword
Brigadier
Does that mean desktops with graphic-integrated CPUs like AMD Ryzen 7 4700U and without dedicated GPUs can't install?
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Does that mean desktops with graphic-integrated CPUs like AMD Ryzen 7 4700U and without dedicated GPUs can't install?
This is indeed a big release.
Deepseek is an advanced AI lab devoted to frontier research, real scientific research, i.e. research with corresponding publications and disclosures. It is what Open AI was in the first years, when their name still made sense...before Altman completely turned the company toward a highly commercial and profit-oriented business.
But Deepseek will never be the Google or Meta of China, nor it's their goal to be.
Instead Bytedance will be.
Reading their announcement, I'd like to highlight this improved speech capability that it seems secondary compared to "thinking", but actually it is not.
The traditional speech integration in LLM is done with 2 independent extra modules: a "speech to text" (ASR) module that listen and converts the speech into text that the LLM reads, and a "text to speech" (TTS) module that picks the text output of the LLM and convert it into speech. This is the simplest technical approach, but for the LLM model it is like to interface with people by mean of reading and writing text on a mobile: all the non textual information conveyed by the tone, the speed, the mood is lost in both directions.
Instead they got rid of the 2 external modules and made the LLM to directly listen and produce "sound" tokens. We know that words are converted into tokens for LLM to process. Here the sound tokens are just representation of very short (some tens milliseconds) chunks of sound that all toghter form the input voice stream. These sound tokens are very close to the actual sounds more than to the words that the sound voice represent, so that the LLM can understand not only the words but also how the words are spoken. And the same goes for the output: the LLM is free to choose how the words should be modulated and spelt. Moreover the latency (i.e. the pauses) is greatly reduced because the data processing needed to transform voice into sound tokens is very thin and simple compared to the traditional speech-to-text module. All the heavy duty of inferring words out of sounds is done by the LLM itself, that's why this approach is more difficult.
Although there is quite a lot of literature and prototypes on this approach, as of today only OpenAI and Google have production quality implementations. OpenAI since June last year, Google since few months ago. Now also Bytedance.
This feature is a key enabler to make AI succesful in the mass consumer market.
Here is some video material on this new feature by Bytedance:
Last edited:
"ByteDance, the Chinese owner of TikTok, has earmarked over 150 billion yuan (HK$160.31 billion) in capital expenditure for this year, much of which will be centred on artificial intelligence, two people briefed on the matter said.
The privately held technology giant plans to spend about half of the amount abroad on AI-related infrastructure, primarily data centres and networking equipment, they said.
The main beneficiaries of the spending will be chipmakers Huawei Technologies and Cambricon Technologies plus US supplier Nvidia the people said, declining to be identified as the information was confidential.
ByteDance, Huawei and Cambricon did not respond to requests for comment on Thursday. Nvidia declined to comment."
The privately held technology giant plans to spend about half of the amount abroad on AI-related infrastructure, primarily data centres and networking equipment, they said.
The main beneficiaries of the spending will be chipmakers Huawei Technologies and Cambricon Technologies plus US supplier Nvidia the people said, declining to be identified as the information was confidential.
ByteDance, Huawei and Cambricon did not respond to requests for comment on Thursday. Nvidia declined to comment."
Some twits abt deepseek r1
ByteDance clarifies: denies rumors that AI infrastructure investment will exceed $12 billion this year
【TechWeb】Recently, the Financial Times reported that ByteDance is investing heavily in artificial intelligence infrastructure and plans to invest more than $12 billion (about 87.34 billion yuan) this year. However, people related to ByteDance responded to Yicai Global that the news was inaccurate.
ByteDance emphasized that the company does attach great importance to the development of the field of artificial intelligence and spares no effort in investment, but the rumors about specific budgets and plans are not true.
Nevertheless, the Financial Times report also mentioned that ByteDance plans to invest 40 billion yuan in purchasing AI chips by 2025, and intends to invest approximately $6.8 billion (about 49.493 billion yuan) overseas in order to use Nvidia's AI chips to enhance the training capabilities of its basic models.
ByteDance's clarification shows that although the company continues to be optimistic about the development of AI technology and has invested in it, the market may have misunderstandings about the specific figures of its investment scale. (Suky)
And first reply is that it censors Tianamen square. What a bunch of sore losers.
"BlackRock Chairman and CEO Laurence D. Fink and G42 CEO Peng Xiao discussed US restrictions on semiconductor technology, meeting power demands with natural gas, the role of crypto and the risk of elevated inflation in a conversation with Bloomberg's Francine Lacqua at Bloomberg House in Davos."
They talked abt financing ai infra.
They talked abt financing ai infra.
If it was easier for them to bash on technicial merit, they would have done so first...