The Information reported that OpenAI's CEO informed shareholders that they are considering transforming their governance structure into a for-profit company. This move may pave the way for OpenAI's first IPO (the company is currently valued at $86 billion), and Altman may also have the opportunity to invest.

Last Thursday, at Tesla's 2024 shareholder meeting, Musk predicted that eventually, Tesla may earn approximately $1 trillion in profits annually from its robotics business.

What other hot topics in the AI industry at home and abroad are worth paying attention to in the past few days? Let me take you through them.

/ 01 / Large Models

1) NVIDIA Open-Sources a 340 Billion Parameter Large Model, Performance Comparable to GPT-4o

NVIDIA recently launched an innovative model called "Nemotron-4 340B," marking a significant breakthrough in AI synthetic data generation technology. Nemotron-4 340B includes the base model, instruction model, and reward model, and constructs a complete process for high-quality synthetic data generation.

The model supports a 4K context window, over 50 natural languages, and over 40 programming languages, with training data ending in June 2023. Nemotron-4 340B used 9 trillion tokens in its training, surpassing the performance of multiple competitors and even rivaling GPT-4. Its business-friendly licensing model enables enterprises to easily utilize generated data, thus reducing their reliance on expensive real-world datasets.

2) Fudan University Releases Evaluation Results for Large Models in College Entrance Exam Mathematics

This week, the LLMEVAL team from Fudan University's NLP laboratory released the evaluation results for large models in the 2024 college entrance exam mathematics: Ranked from high to low, the top three for the new I paper are Ali Qianwen, iFlytek Spark, and GPT-4o, with Ali Qianwen and iFlytek Spark significantly outperforming GPT-4o in terms of accuracy for 14 math objective questions. The top three for the new II paper are iFlytek Spark, Ali Qianwen, and GPT-4o.

Large models such as ByteDance's Doubao, Zhipu Qingyan, and Baichuan follow closely, while Baidu's Wenxin Yiyan and Kimi have lower accuracy rates.

3) DreamTech Launches Direct3D, the World's First Native 3D-DiT Large Model

DreamTech launched its high-quality 3D generation large model, Direct3D, and published a paper on it. This is the first publicly released 3D large model using a native three-dimensional generation approach, which addresses the challenge of generating high-quality three-dimensional content by employing 3D Diffusion Transformer (3D-DiT).

Previously, the technical route commonly used for 3D AIGC was 2D-to-3D lifting, with representative solutions including Score Distillation Sampling (SDS), represented early on by Google's DreamFusion.

4) Expense Reconciliation Completed in 1 Minute, Microsoft AI Day Showcases Copilot's Productivity Revolution

Microsoft AI Day showcased Copilot's application in productivity, demonstrating its ability to reconcile expense reports and generate reports within 1 minute. It also showcased Azure AI Studio and the Phi-3 series of small models. Microsoft believes that the future will be an era of multi-model collaboration, with both small and large models having their advantages, allowing enterprises to choose the appropriate technical route based on their own needs.

5) iFlytek: Technical Cooperation in Large Models on Mobile Phones Underway

In response to inquiries about whether iFlytek is collaborating with Huawei to jointly develop the smartphone market, iFlytek responded that it has been providing intelligent voice and multilingual technical services to multiple mobile phone manufacturers. iFlytek also added that after the launch of the iFlytek Spark large model, technical cooperation on large models on mobile phones has been ongoing.

6) ACL 2024 Paper Concludes: Large Language Models ≠ World Simulators

Are large models world models? A recent study by institutions such as UA Microsoft found that GPT-4's accuracy in complex environment simulations is even less than 60%. Regarding this, Turing Award winner LeCun strongly agrees and says, "World models will never be LLMs."

/ 02 / AI Applications

1) Apple's AI Text-to-Image App: Only Generates Cartoon Images and Will Mark Generated Content

Apple's "Image Playground" app in its Smart Suite was launched alongside iOS 18. Users can input text and character photos to have Image Playground generate related images.

Federighi, Apple's vice president of software engineering, mentioned in a recent interview that the "metadata" of generated images will be marked to let others know that the image was intelligently created by Apple. Image Playground will only generate cartoon images and cannot generate realistic photo-like images, which is also a way to prevent people from using its generated images to mislead others.

2) Xiaoai Classmate Integrates with Doubao Large Model, Already Available on Mobile Phones and SU7:

Xiaomi's AI assistant "Xiaoai Classmate" has partnered with Volcano Engine, realizing a more intelligent AI interaction experience based on the Doubao large model. Xiaoai can also leverage the online search plugin capabilities provided by ByteDance's Doubao large model to capture search results that are sourced from the same content as Toutiao, presenting comprehensive and timely answers.

3) Apple Executive Teases Microsoft AI, Saying Apple Has Already Deployed 'AI PCs'

After this year's Apple WWDC developers conference, Apple executive Federighi said in an interview that Apple has been launching Mac devices equipped with neural engines since 2020, but Apple just hasn't called them "AI PCs."

4) Meta Will Delay the Launch of Its AI Chatbot in Europe

Meta stated in a blog post that its plan to use European user posts to train its large model, Llama, was blocked by the Irish Data Protection Commission. Meta also revealed that after suspending its plan to train the large model using European user posts, the launch of its Meta AI chatbot in Europe will also be delayed.

5) McDonald's 'Fires' AI Ordering Assistant, Suspends Automatic Ordering Test Project with IBM

McDonald's recently controversial automatic ordering AI system will temporarily halt testing, and this cooperation project with IBM, which began in 2021, will end before July 26, 2024. Outside speculation suggests that McDonald's may seek another partner to continue advancing related technology research and development.

/ 03 / Investment and Financing Intelligence

1) News Suggests OpenAI May Become a Profit-Making Company, Potentially Launching a $620 Billion Valuation IPO

The Information reported that sources familiar with the matter revealed that OpenAI CEO Sam Altman informed shareholders that the company is considering transforming its governance structure into a for-profit company that cannot be controlled by a nonprofit board. OpenAI responded that they remain focused on creating AI that benefits everyone, and the nonprofit organization will continue to exist.

According to analysis, such a change would pave the way for OpenAI's first IPO. Currently, the company is valued at $86 billion, and Altman will also have the opportunity to invest in OpenAI. Foreign media reported that some executives who have interfaced with OpenAI prefer to transform the company into a for-profit entity to allow Microsoft to exert more influence over OpenAI due to its board seats and shareholder voting rights.

2) Musk Plans to Generate $1 Trillion in Profits Annually with Optimus Robots

Last Thursday, Tesla shareholders reapproved Elon Musk's compensation incentive plan. Musk discussed the eventual form of Tesla's fleet, his vision for the humanoid robot Optimus, the current progress of the Cybertruck, Tesla's Model 3/Model Y and Semi, as well as some new products Tesla is developing at Tesla's 2024 shareholder meeting.

According to Musk's prediction: Tesla's efforts in manufacturing the humanoid robot Optimus may one day surpass its automotive business. Eventually, Tesla may earn approximately $1 trillion in profits annually from its robotics business.

3) Japan's Fastest-Growing Unicorn Is About to Be Born: SakanaAI Valued at 180 Billion Yen After One Year of Establishment

Japanese generative AI startup SakanaAI is about to receive a significant new investment, which will bring the company's valuation to approximately 180 billion yen (about 8.3 billion yuan).

SakanaAI was founded in Tokyo in July last year by Google AI researchers. The company conducts research on AI base models for generating text and images. It has "groundbreaking" technology that crosses existing models to create low-cost, high-performance AI models, as well as LLMs that can solve mathematical problems in Japanese and models that can generate and understand Japanese images and text. The company has released some of its models as open-source software.

/ 04 / AI Infrastructure

1) The Office of the Cyber Information Department Releases the Sixth Batch of Deep Synthesis Service Algorithm Filing Information, Including 492 Algorithms Such as Tencent Hunyuan:

The Office of the Cyber Information Department has released the sixth batch of domestic deep synthesis service algorithm filing information. A total of 492 algorithms have been filed this time, including Tencent Hunyuan's large model multimodal algorithm, Lingyiwanwu's large model multimodal generation algorithm, Kuaishou Kuaiyi's large model generation synthesis algorithm, SenseTime's V-ME video synthesis algorithm, DingTalk's AI assistant intelligent generation algorithm, and Huiwa's e-commerce model try-on image synthesis algorithm.

2) The Head of Alibaba's DAMO Academy XR Lab Has Left to Start a Business in AI Hardware

Baoliaotai learned that Li Jiantao, the former head of products and business at Alibaba's DAMO Academy XR Lab, left his position in November last year and officially started a business with friends. Informed sources revealed that Li Jiantao has chosen the direction of AI hardware for this venture.

Li Jiantao joined DAMO Academy in 2021, primarily responsible for products and business at the XR Lab, as well as the AIoT business at the Machine Intelligence Lab. Before joining Alibaba, he served as the product head and deputy general manager of the IoT Business Division at Sogou, responsible for the development of multiple smart hardware products. Among them, Sogou's AI recording pen once reached the industry's top spot, and its Tangmao children's watch and children's educational hardware products reached the industry's second spot.

3) AI Comic Characters Are More Consistent! Complex Interactions Between Characters Can Also Be Handled

A joint team from Sun Yat-sen University and Lenovo proposed a training-free multi-agent collaboration framework called AutoStudio. This framework uses three agents based on large language models to handle interactions and employs a diffusion model-based Drawer to generate high-quality images. In experiments, AutoStudio outperformed existing methods in both quantitative and qualitative evaluations.

4) Has the Scaling Law Hit the "Data Wall"? Epoch AI Predicts That LLMs Will Run Out of All Text Data by 2028

A recent paper by Epoch AI predicts that the available human text data on the internet will be exhausted in four years, by 2028.