Text: Chen Sida
Edited by Tang Yongyi
A quick look at the week
After the rotation of various large models, ChatGPT ushered in several upgrades this week. OpenAI announced that users can now use ChatGPT without registration, and can also use DALL-E to generate images across web and mobile devices in ChatGPT, and make local adjustments to the generated results. In addition, 6 new fine-tuning API features have been launched to help build more refined ChatGPT applications.
Apple's strategic moves have been very frequent lately. Its research team released a device-side model, ReALM, which said that the performance is better than GPT-4 and can significantly improve the intelligence of Siri in the future. A number of changes have been made as a result of abandoning the car-building project: Apple's research teams are looking into shifting to a variety of domestic robotic devices, and in California, the company has laid off more than 600 employees.
Star AI startups are still the darlings of capital. Founded less than half a year ago, Cognition Labs is seeking to raise capital at a valuation of up to $2 billion. Musk's AI startup xAI may raise $3 billion, and its valuation is expected to reach $18 billion. In addition, Sam Altman has incubated an AI hardware company with former Apple design director and is planning to raise $1 billion.
This week, a number of large models were also released open-source. UC Berkeley's open-source world model LWM has excellent multimodal performance. The team at Princeton has also open-sourced an agent-computer interface (ACI) called SWE-agent, which is comparable in performance to Devin, an "AI programmer." However, AI startup Anthropic warns all major models that a new "multi-round jailbreak" method can trick AI into making harmful answers.
Key Points
- ChatGPT can be used without registration
- ChatGPT integrates DALL-E 3 to generate images directly
- Apple released the ReALM model, which can run on mobile phones and tablets, and the performance is better than GPT-4
- The domestic 100 billion parameter MoE architecture model is open source
- Apple is exploring the field of domestic robots
- OpenAI发布全新微调API
- OpenAI will open its first office in Asia in Tokyo
- Nvidia will invest $200 million in Indonesia to build an AI center
- Amazon provides free credits for AI large models for startups
- More than 200 artists jointly boycotted AI
- Musk's startup may refinance $3 billion
- The company behind "AI programmer Devin" is valued at $2 billion
- Sam Altman and former Apple design director for an AI hardware project seeking $1 billion in financing
- Sam Altman no longer owns or controls OpenAI-related VC funds
- Yahoo acquires AI news app Artifact
- The Hong Kong stock "AIGC first share" will be listed
- Stability AI releases free audio generation tool
- Anthropic discovered a long text vulnerability in large models
- The open-source world model LWM was released
- Princeton University open-source agent SWE-agent
Large model first line
ChatGPT can be used without registration
On April 1, local time, OpenAI announced that it would allow users to use ChatGPT directly without registration. This benefit will be gradually opened up to the general public, "anyone who is interested in AI capabilities can use it". At present, ChatGPT traffic growth has basically come to a standstill. According to data from SimilarWeb, after the explosive growth in early 2023, ChatGPT's global traffic has dropped by 9.7% since June 2023, and the traffic has fluctuated month by month.
Source: official website
ChatGPT integrates DALL-E 3 to generate images directly
On April 4, OpenAI announced that users can now use DALL-E to generate images across the web, iOS, and Android in ChatGPT. In addition to this, users can also refine their DALL-E creations by adding, removing, or adjusting elements in the image.
Source: X
Apple released the ReALM model, which can run on mobile phones and tablets, and the performance is better than GPT-4
On April 2, it was reported that in a recent paper, Apple's research team claimed that they proposed a model ReALM that can run on the device side, which can surpass GPT-4 in some aspects. The parameters of ReALM are 80M, 250M, 1B, and 3B, respectively, and they are very small in size, which is suitable for running on mobile phones, tablets, and other devices.
Source: Paper
Address:
Hattapus://ArXiv.org/PDF/2403.20329.pdf
The domestic 100 billion parameter MoE architecture model is open source
On April 2, APUS, a global mobile Internet company, and Xindan Intelligence, a large-scale model start-up, announced that they had joined forces to open source China's first 100-billion-parameter MoE (hybrid expert model) APUS-xDAN large model 4.0, which can run on low-end computing power chip 4090, and its comprehensive performance has reached 90% of GPT-4. This is also the first 100 billion MoE Chinese and English large model that can run on consumer-grade graphics cards in China.
Big event
Apple is exploring the field of domestic robots
On April 3, local time, Mark Gurman, a well-known Apple whistleblower, wrote an article saying that after giving up making cars, several teams at Apple are researching and promoting various household robot devices and related artificial intelligence functions. This area could be one of Apple's ever-changing "next big things". According to news on April 5, Apple has laid off more than 600 employees in California, which is also related to the abandonment of car manufacturing.
OpenAI发布全新微调API
In the early morning of April 5, OpenAI announced on its official website that it had added 6 new fine-tuning API functions to expand custom models and help enterprises and developers better build domain-specific and refined ChatGPT applications. Features include: Epoch-based checkpoint creation, new playground features, third-party integrations, comprehensive validation metrics, hyperparameter configuration, and more detailed fine-tuning dashboard improvements.
OpenAI will open its first office in Asia in Tokyo
According to Bloomberg, OpenAI will open its first Asian office in Tokyo, Japan, in April to expand its international business. This will be its third international office, following the opening of offices in London and Dublin last year. In April 2023, OpenAI CEO Sam Altman visited Japan, and after meeting with the Japanese Prime Minister, Altman said, "I want to make the large model more suitable for the Japanese language and culture." ”
Nvidia will invest $200 million in Indonesia to build an AI center
On April 5, Nvidia announced that it would cooperate with Indosat Ooredoo Hutchison, Indonesia's second largest mobile telecommunications company, to build an artificial intelligence development center in Solo City, Central Java, with a total investment of 200 million US dollars. At last month's World Action Congress in XNUMX, the two sides announced a partnership in which Indosat Ooredoo Hutchison plans to integrate NVIDIA Blackwell chips into its infrastructure.
Amazon provides free credits for AI large models for startups
According to Reuters, on April 2, Amazon Web Services (AWS) said in an interview that the company has expanded its free credit program for startups to reduce the cost of using major AI models and increase the market share of its AI platform Bedrock. To attract startup customers, Amazon now allows its cloud credits to overlay the use of models from other providers, including Anthropic, Meta, Mistral AI, and Cohere.
More than 200 artists jointly boycotted AI
According to CNBC, on April 2, more than 200 well-known musicians in the international music scene signed an open letter, calling on AI developers, technology companies, platforms and digital music service providers to stop using artificial intelligence (AI) to infringe and demean the rights of human artists, including stopping the use of AI to infringe and demeaning the rights of human artists, and asking them to promise not to develop AI music generation technology related to it.
Recently, AI-generated music has become a hot topic, and audio generation models such as Suno V3 and Stable Audio 2.0 have been released one after another.
Source: X
Financing dynamics
Musk's startup may refinance $3 billion
According to the Wall Street Journal, on April 5, people familiar with the matter said that investors close to Musk are in talks to help his AI startup xAI raise $3 billion, and this round of financing will make the company value $18 billion. It is reported that supporters considering investing in this round of financing include venture capital firm Gigafund and American venture capitalist Steve Jurvetson. The terms of xAI's financing have not yet been finalized, and plans are subject to change.
The company behind "AI programmer Devin" is valued at $2 billion
On April 1, it was reported that Cognition Labs, an AI start-up founded in November last year, is in financing negotiations with a valuation of up to $2 billion. Cognition Labs claims that Devin, the AI it has developed, is "the first fully autonomous software engineer AI". Devin just launched this month, and Cognition Labs is currently generating no substantial revenue.
Sam Altman and former Apple design director for an AI hardware project seeking $1 billion in financing
According to The Information, on April 6, OpenAI CEO Sam Altman joined hands with former Apple design director Jony Ive to jointly design AI hardware for individuals, and is currently seeking external investment. Altman and Ive are already leading the formation of a startup and are seeking funding from major venture capitalists with a goal of raising $1 billion.
Sam Altman no longer owns or controls OpenAI-related VC funds
On April 2, according to the U.S. Securities and Exchange Commission (SEC) filing, OpenAI has changed the governance structure of its venture capital fund that invests in artificial intelligence startups, and its CEO Sam Altman no longer owns or controls the fund. It is understood that the fund was raised by Sam Altman from external limited partners, and he also led the investment decisions. OpenAI said that despite the ownership, Sam Altman did not receive any financial benefits from it.
Yahoo acquires AI news app Artifact
According to the official website, on April 2, local time, Yahoo announced that it had completed the acquisition of the AI news platform Artifact. Founded in January 2023 as an AI news aggregation and discovery platform created by Instagram co-founders Kevin Systrom and Mike Krieger, Artifact shut down its services in January this year due to poor management.
After the acquisition is completed, Artifact will no longer operate independently, and its AI technology and capabilities will be integrated into Yahoo's products, including Yahoo News, in the coming months.
Source: Yahoo
The Hong Kong stock "AIGC first share" will be listed
On April 2, the AI company "Mobvoi" passed the hearing of the Hong Kong Stock Exchange in the evening, which is expected to hit the first share of Hong Kong AIGC. Founded in 2012, Mobvoi, as an early company to enter the AI track, has been favored by top institutions including Sequoia China, Google, Goertek Acoustics, and Zhen Fund. Its last round of funding was completed in September 2019 at a valuation of over $700 million.
Mobvoi product matrix, source: Mobvoi official website
New gadgets
Stability AI releases free audio generation tool
On April 3, it was reported that Stability AI, a well-known large model open source platform, released the audio generation model Stable Audio 2.0. This version allows users to generate high-quality music in various genres with text or audio, up to 3 minutes long. Stable Audio is available free of charge to the public, with an API to be provided soon.
Stable Audio 2.0创作界面，图源：Stable Audio 2.0
Experience address:
hatpas://stableodio.com/
Previous Research
Anthropic discovered a long text vulnerability in large models
On April 3, it was reported that researchers at the AI company Anthropic recently discovered a new method that allows large models to answer questions that should not be answered: if you first warm up the large model with dozens of less harmful questions, you can induce it to introduce how to make a bomb. This method is known as a "multi-round jailbreak" and works on several mainstream models such as Claude 2, GPT-3.5, GPT-4, Llama 2, and Mistral 7B.
Source: Paper
Address:
https://cdn.sanity.io/files/4zrzovbb/website/af5633c94ed2beb282f6a53c595eb437e8e7b630.pdf
The open-source world model LWM was released
On April 4, researchers from the University of California, Berkeley, recently released and open-sourced the world model LWM (LargeWorldModel) series of models. This model uses a large number of video and book datasets, and realizes the scalable training of long sequences through RingAttention technology, so that the context length of the model reaches 1M token, which is equivalent to about 750,000 words.
LWM generates images and videos based on text, source: paper
Address:
Hattapus://ArXiv.org/PDF/2402.08268.pdf
Code Address:
hatps://github.com/largeworlddmodel/loam
Princeton University open-source agent SWE-agent
On April 2, it was reported that Princeton's research team launched an agent-computer interface (ACI) called SWE-agent, whose performance is comparable to that of "the world's first AI programmer" Devin. The innovation of this interface lies in its ability to transform large language models, such as GPT-4, into software engineering AI agents and fix bugs autonomously in actual GitHub repositories. This move is expected to solve GPT-4's programming challenges.
Swee-Agent的Swe-Bench测试集表现,图源:Kithub
Project Homepage:
https://swe-agent.com/
Open Source Address:
https://github.com/princeton-nlp/SWE-agent