ChatGPT costs about $700,000 a day, and OpenAI may be on the verge of bankruptcy

Big Model Information:

ChatGPT costs about $700,000 a day, and OpenAI may be on the verge of bankruptcy

According to a recent report by Indian news media platform Analytics India Magazine, OpenAI could go bankrupt by the end of 2024. OpenAI could go bankrupt by the end of 2024.

OpenAI costs about $700,000 a day just running its AI service, ChatGPT, according to the report. OpenAI is currently burning money, and while the company is trying to make a profit with GPT-3.5 and GPT-4, the company has not yet been able to generate enough revenue to break even. Since it began developing ChatGPT, its losses doubled in May to $540 million, the report said. Microsoft has invested $10 billion in OpenAI, which may be what underpins the company's operations so far.

The report said that if OpenAI doesn't get more money soon, the company may have to file for bankruptcy by the end of 2024.

OPPO will launch a new version of Xiaobu Assistant to add AI functions

According to the official news of OPPO Xiaobu assistant, the new Xiaobu assistant based on AndesGPT is about to start a large-scale experience activity.

It is reported that the upgraded Xiaobu assistant will have more powerful semantic understanding capabilities, and provide functions such as induction and copywriting.

The new features will be supported by AndesGPT, a generative big language model built by OPPO's Andean Intelligent Cloud team.

Microsoft released Gorilla LLM, with API calls superior to GPT-4

According to the University of Berkeley research page, Microsoft Research recently teamed up with researchers at the University of California, Berkeley to launch a new large-language model Gorilla.

Research papers show that Gorilla outperforms GPT-4 in terms of accuracy, flexibility, and other API calls. The paper shows that even if the API documentation changes, Gorilla can guarantee the accuracy of the generated language when calling the API.

In addition, Gorilla can generate correct API calls for APIs that do not have a training dataset.

Anthropic releases the AI model Claude Instant 1.2

According to Anthropic official, Claude Instant version 1.2 is now officially released.

Anthropic said that Claude Instant 1.2 will have a lower cost and faster response time, it can complete a series of tasks such as conversation, analysis, summarization, document understanding and so on.

It is reported that the Claude Instant 1.2 model can handle up to 100,000 tokens, and can process a large amount of text content such as stories, letters, and memos at the same time.

For the first time, Mainland China has applied AI technology to transmission line heat detection at scale

According to Science and Technology Daily, recently, the transmission line infrared defect intelligent identification system developed by State Grid Power Space Technology Co., Ltd. and North China Electric Power University and other units has achieved industrial application in the operation and maintenance of major ultra-high voltage lines in the mainland.

This is the first time in China that artificial intelligence (AI) technology has been applied to the large-scale detection of transmission lines, providing new technical means to ensure the safe and stable operation of the power grid. The technical research team combined with the business scenario and adopted the technical route of "minimized annotation + stepped learning + interference point shielding" to realize the intelligent identification of infrared defects, and the model recognition accuracy reached more than 90%.

At present, the system is deployed and applied in State Grid Power Space Technology Co., Ltd., which is the first large-scale application of artificial intelligence technology to transmission line heat detection in China. Taking infrared video from a 240 base tower as an example, which traditionally took 5 hours for manual data review, the system now takes only 2 hours from video upload to analysis without human intervention.

Hon Hai received another large order from NVIDIA, becoming the largest supplier of AI server chip substrates for the latter

According to CaiLian News Agency, Hon Hai Group AI received orders and passed on success, winning a large order for NVIDIA HGX server AI chip substrate for the first time, with a supply ratio of more than 50%.

In addition, it has previously obtained another DGX server AI chip substrate order from the company, and at this stage, NVIDIA's two most important AI server substrate orders are shared by Hon Hai Group, and the proportion of supply is gradually expanding. Hon Hai will hold a legal briefing after the market in Taiwan stocks today, and is now in a period of silence before the legal conference.

The industry believes that Hon Hai has become the largest supplier of chip substrates of the global AI chip leader NVIDIA, coupled with the upcoming appearance of the iPhone 15 series of new machines and the shipment of electric vehicles in the fourth quarter, Hon Hai will launch three arrows in the second half of the year, and its performance will explode.

Mingxi Kuo: Don't worry about OpenAI's possible bankruptcy

Well-known analyst Guo Mingxi said today that there is no need to worry about the bankruptcy risk of OpenAI reported by the media.

Guo Mingxi said that AI/AIGC is already a clear trend, but changes in the industry will create new winners and losers. It is worth noting that if AI/AIGC fails to create a sustainable and profitable business model, industry investment in AI/AIGC may slow. Microsoft's capital expenditure outlook for AI/AIGC over the next two quarters and the profitability of related businesses will be an indicator of related investments.

Recently, Indian media said that OpenAI only runs its artificial intelligence service ChatGPT costs about $700,000 a day, and if OpenAI does not get more money soon, the company may have to file for bankruptcy by the end of 2024.

Honor applied for the trademark MagicAl

Tianyancha App shows that recently, Honor Terminal Co., Ltd. applied for the trademark "Magic Al", which is internationally classified as a scientific instrument, and the current trademark status is pending.

Glory Terminal Co., Ltd. was established in April 2020, the legal representative is Wan Biao, the registered capital is about 32.239 billion yuan, the business scope includes development, production and sales: communication and electronic products, computers, satellite TV receiving antennas, high-frequency heads, digital satellite TV receivers, etc.

Magic is an Honor mobile phone, and the operating system of the Honor mobile phone is MagicOS.

Google-backed AI startup Anthropic received $100 million from South Korea's largest telecommunications company

SK Telecom, South Korea's largest telecommunications company, said it will invest $100 million in U.S. AI startup Anthropic to strengthen its telecom-driven AI business, Sina Technology reported. Anthropic, a startup that competes with OpenAI in building AI-based models and one of the best-funded AI startups, raised $450 million in funding from investors such as Google and Spark Capital in May.

iFLYTEK: "iFLYTEK Spark Cognitive Model V2.0" will be released on August 15

iFLYTEK announced on August 14 that the company will further realize the major version upgrade of iFLYTEK Xinghuo cognitive large model as scheduled, and will release "iFLYTEK Xinghuo cognitive large model V2.0" on August 15.

In addition, the further commercial implementation results of the Spark Cognitive Big Model in the fields of education and office - Xinghuo Language Partner 2.0, Spark Teacher Assistant and AI learning machine "AI Creative Drawing Board" and "AI Programming" new functions will also be released simultaneously.

Key Papers:

AudioLDM 2: Learn overall audio generation with self-supervised pre-training

To get closer to a unified view of audio generation, the University of Surrey and SAMI, the ByteDance intelligent creative speech team, have jointly proposed a framework that uses the same learning methodology for speech, music, and sound effects generation, and introduces a common representation of audio, Audio Language (LOA). During the generation process, researchers use GPT-2 models to convert modalities to LOA and self-supervised audio generative learning using potential diffusion models based on LOA conditions. In the major benchmark experiments for text-to-audio, text-to-music, and text-to-speech, the framework demonstrates more advanced performance.