June 16th, Jinshi data, according to the US Microsoft website reported on June 15th, former US President Trump said that he will "absolutely not" ban the social platform application TikTok. The report pointed out that this is a major reversal for Trump, as he tried to ban this application during his presidency in 2020. The report believes that the background of Trump's change of tune is that he is trying to attract young voters, most of whom are active TikTok users. The report said that Trump recently joined this application and has already won nearly 6 million fans. At the same time, President Biden's campaign team on TikTok has only over 370,000 fans.

On Wednesday, the "small non-agricultural" recorded 152,000, far below expectations; Will the "real non-agricultural" on Friday be the same? Is it a precursor or singing a different tune: Is the "small non-agricultural" still useful now?

On the road to curbing inflation, the Fed's monetary policy has reached an uncertain moment.

IEA has lowered its global oil demand growth forecast for the second consecutive month, but the supply gap in the third quarter will widen in this situation.

Powell acknowledged that Q1 inflation data had dented his confidence and described Tuesday's PPI report as a "mixed bag."

PANews News on February 28, the non-custodial fixed income protocol Term Structure Testnet Trading Competition has been launched, which will be held from March 4 to 10, 2024, aiming to allow participants to show their trading skills through a simulated market environment and fine-tune their strategies without using real funds. During the competition, the top 10 will share 10,000 USDC rewards, and all participants will receive Non-fungible Token participation prizes. Users need to claim test tokens and deposit them on the testnet, joining the official Discord channel to follow the market dynamics. During the competition, primary market auctions will be held every hour on the testnet. Users can also trade Tokens on the Testnet. Money Laundering varies depending on the type of Token being traded. In addition, Term Structure has implemented a clearing mechanism to ensure the stability of the protocol. Going forward, the protocol will launch new features and tools, including trading APIs, SDKs, Layer 2 swaps, and support for yield Tokens as Collateral. 

According to a report by 36Kr, on December 20, Baidu Intelligent Cloud said at the 2023 Baidu Cloud Intelligence Conference and Intelligent Computing Conference that since the Wenxin model was fully opened to the whole society on August 31, the daily call volume of the large model API on the Qianfan large model platform has increased by 10 times. At present, the Qianfan platform has served more than 40,000 enterprise users and helped them fine-tune nearly 10,000 large models.

According to the National Business Daily, on November 24, Wang Qi, vice president of Tencent Cloud and head of Tencent Cloud's intelligent solutions, said that there may be differences in the paths of different financial institutions to deploy large models. "Generally speaking, the product shelves of the entire large model ecology have different levels, including the lowest level of computing power support, computing network storage, etc., and then to the general large model above, and then pre-train the industry large model on the basis of the general large model, and fine-tune it on the basis of the industry large model combined with the data of the industry, and finally become the customer's exclusive large model. The product system of the entire large model is from the bottom up, and different financial enterprises can cut into their own financial industry large model from different points, and here they should also consider according to their own input costs and scenario complexity. ”

According to a report by IT Home on November 16, Microsoft launched a new "Windows AI Studio" experience for developers at the Ignite 2023 conference that opened today, which can simplify and accelerate the development of AI applications on Win11 systems. Windows AI Studio is said to integrate top AI development tools and models from catalogs such as Azure AI Studio, Hugging Face, and more, simplifying the generative AI app development process, enabling developers to fine-tune, customize, and deploy state-of-the-art small language models (SLMs) for their Windows apps to use locally. In addition to simplifying the setup process for building AI experiences, Windows AI Studio will also highlight AI models optimized for GPUs and NPUs, but the feature will be available at a later date. In addition, Microsoft has updated DirectML, Olive and Windows ONNX Runtime to run Llama 2 7B natively on a variety of different Windows hardware.

According to a report by Jin Shi on November 8, ICBC Credit Suisse Fund issued a view that OpenAI, an American artificial intelligence company, held its first developer conference, launched a custom version of ChatGPT, and released a more powerful and faster version of GPT-4 GPT-4 Turbo. What's more, OpenAI has already begun to build its own AI ecosystem. At the end of this month, the "GPT Store" will be launched, where anyone can train and fine-tune their own GPT app and then put it on the GPT App Store. ICBC Credit Suisse Fund also pointed out that the OpenAI developer conference has strongly boosted technology stocks, and the collective rise of U.S. technology stocks will also affect the A-share technology sector. Although in the long run, OpenAI is not willing to be a large model manufacturer and will gradually sink on the application side, and shallow AI application developers may face the risk of exiting, but in the short term, this developer conference has opened up space for the AI industry, promoted industry sentiment, and may form a plate opportunity.

According to the IT Home on November 7, Microsoft Research Asia, together with Peking University, Xi'an Jiaotong University and other universities, proposed an AI training method called "Learning from Mistakes (LeMA)", which claims to improve AI reasoning ability by imitating the process of human learning knowledge. In order to improve the chain-of-mind reasoning ability of open-source large language models, the research team proposed the LeMA method. This approach mainly mimics the human learning process, by "learning from mistakes", in order to improve the model's reasoning ability. The researchers' approach was to fine-tune the model using a pair of data containing "wrong answers" and "corrected correct answers." To obtain the relevant data, the researchers collected the wrong answers and inference processes of five different large language models, including LLaMA and the GPT series, and then used GPT-4 as a "corrector" to provide corrected correct answers. It is reported that the corrected correct answer contains three types of information, namely the error fragment in the original reasoning process, the reason for the error in the original reasoning process, and how to correct the original method to obtain the correct answer.

ChatGPT maker OpenAI is expected to announce product improvements to make its AI models cheaper and more powerful at its first developer conference next Monday, as reported by Golden Ten Data on November 4. Altman, the company's CEO, promised attendees that "some great new stuff" would be introduced. Held in a desolate area near San Francisco City Hall, the one-day event attracted hundreds of developers from around the world. In addition to announcing new products, the company may also fine-tune the capabilities of its state-of-the-art AI model, GPT-4. Sources with knowledge of Altman's idea say making OpenAI an integral part of other companies' app development is one of its most important strategic goals.

OpenAI, the maker of ChatGPT, is expected to announce product improvements at its first developer conference next Monday, making its AI models cheaper and more powerful, according to Golden Finance. Altman, the company's CEO, promised attendees that "some great new stuff" would be introduced. Held in a desolate area near San Francisco City Hall, the one-day event attracted hundreds of developers from around the world. In addition to announcing new products, the company may fine-tune the capabilities of its state-of-the-art AI model, GPT-4. Sources with knowledge of Altman's ideas say making OpenAI an integral part of other companies developing apps is one of its most important strategic goals.

On Nov. 2, Midjourney announced that it was testing the Style tuner feature, which allows users to customize the style of their images. According to the introduction, users can use the /tune command to generate a series of sample images showing different visual styles based on prompts. Select the image they like the most, and the user will receive a unique code that they can simply paste in to customize the resulting image.

According to a report by Jiemian News on November 1, Focus Media's recent investor relations activity record shows that the company's marketing vertical model has been deployed and is continuing to fine-tune and iterate.

According to a report by Pinplay on October 24, the Arxiv page shows that a research team composed of researchers from the Singapore University of Technology and Design, Microsoft Research and Tsinghua University today released a method called Tuna, which allows models to use feedback from large-language models to make instruction adjustments. The research team used its novel probabilistic ranking and context ranking methods to fine-tune instruction-adjusted LLM to increase the likelihood of generating better responses. Probabilistic ranking enables instruction-adjusted models to inherit relative rankings of high-quality and low-quality responses from high-order large models. On the other hand, using contextual ranking learning allows the model to refine its own response distribution using the contextual understanding capabilities of a more powerful model. According to the research team, Tuna excelled in Super Natural Instructions (119 test tasks), LMentry (25 test tasks), Vicuna QA, and more.

According to interface news reports, Langma Information announced on October 22 that the company achieved operating income of 348 million yuan in the first three quarters, up 2.14% year-on-year; the net profit attributable to the parent was 61.3277 million yuan, up 1.3% year-on-year; Basic earnings per share were $0.18. This year, relying on the massive medical and health data accumulated by its Internet medical business for more than 20 years, as well as related natural language processing technology and deep learning algorithms, the company developed "39AI General Practitioner" through large model training methods, and the product application direction is general practitioner consultation and health medical Q&A and assisting primary doctors in clinical diagnosis. At present, the "39AI General Practitioner" product continues to train, learn and fine-tune human medical knowledge and human feedback reinforcement learning (RLHF), and has carried out research and development of accurate diagnosis capabilities for disease diagnosis and treatment, and can actively conduct multiple rounds of interactive consultation with users to obtain more information, and give more accurate and detailed diagnosis and related diagnosis and treatment suggestions. At present, the company has started the relevant large model and algorithm filing work of "39AI General Practitioner" products, and the online form and time of "39AI General Practitioner" products have not yet been determined, and the subsequent progress will be disclosed in time according to the product progress and information disclosure requirements. As previously reported, Langma Information: Successfully developed "Langma 39AI General Practitioner".

According to 36Kr reported on October 20, Ali Dharma Academy released the industry's first remote sensing AI large model. The model can identify farmland, crops, buildings, etc., and improve the analysis efficiency of remote sensing applications such as disaster prevention, natural resource management, and agricultural yield estimation. Currently, the model is open for use on the AI Earth Earth Science Cloud platform. According to reports, the remote sensing AI interpretation universal segmentation model (AIE-SEG) takes the lead in the field of remote sensing to realize the task of image segmentation unification, one model can achieve the rapid extraction of "zero sample of everything", can identify farmland, water, buildings and other nearly 100 kinds of remote sensing feature classification, and still maintain high-precision recognition under a number of tasks, and can automatically tune the recognition results according to the user's interactive feedback. In some specific scenarios, compared with traditional remote sensing models, the accuracy of instance extraction can be increased by 25%, and the accuracy of change detection can be increased by 30%.

As reported by Webmaster House on October 20, Topaz Labs' newly released Photo AI V2 software brings significant improvements to image enhancement. This version provides powerful noise reduction and sharpening algorithms, as well as the ability to adjust light and color balance. The software uses automated analysis to intelligently adjust to image needs, such as upscaling low-resolution images or removing noise. Users can fine-tune each processing step as needed. It is reported that Photo AI V2 can be integrated with third-party software such as Lightroom and Photoshop to provide more application scenarios.

As reported by Webmaster Home on October 16, technology company Dell released a generative artificial intelligence (AI) tool called "Dell Validated Design for Generative AI with NVIDIA for Model Customization" to help enterprises extract intelligence from data. This tool features pre-trained models that allow users to quickly extract intelligence from data without having to build a model from scratch. The tool provides pre-trained models to help users customize and fine-tune generative AI models as needed while keeping information secure. Dell is said to provide enterprises with a robust data foundation, including infrastructure, software and services to support AI workloads in public clouds, on-premises environments, and edge computing. It was previously reported that Dell will invest $7.5 billion in areas such as artificial intelligence.

According to IT House reported on October 16, modifying existing large language models for different needs can improve the applicability of relevant models, but a study by Princeton University and IBM Research found that fine-tuning large language models can undermine the security of developers adding models. The researchers conducted a series of experiments and showed that even if the data were completely benign, the security of the model was weakened, for example, in the Alpaca dataset, GPT-3.5 Turbo increased its harmful rate from 5.5% to 31.8%, while Llama-2-7b Chat increased its harmful rate from 0.3% to 16.1% in Alpaca and from 0% to 18.8% in LLaVA-Instruct. The researchers pointed out that users who need to fine-tune large models can avoid weakening the security of the model by carefully selecting the training dataset, importing a self-audit system, and using red teams to practice testing.

According to VentureBeat, on September 28, Paris-based AI startup Mistral AI released its first large-scale language artificial intelligence model, Mistral 7B. The 7.3 billion-parameter model is said to be more powerful than larger products such as Meta's Llama 2 13B, and is said to be (so far) the most powerful language model of its size. Mistral said it is open-sourcing the new model under the Apache 2.0 license, allowing anyone to fine-tune and use it anywhere (on-premises to the cloud) without restrictions, including for enterprise use cases. Mistral AI was founded earlier this year by former Google DeepMind and Meta employees. Previously, the company completed a $118 million seed round, which is said to be the largest seed round in European history.

Golden Finance reports that NetMind.ai has launched a free beta version of the decentralized machine learning platform NetMind Power. NetMind Powe allows machine learning experts, researchers and software developers to train, fine-tune and infer deep learning models. The platform is built around the concept of Volunteer Computing, a globally distributed network of user-contributed GPUs connected on the blockchain to provide a decentralized network of computing power.

According to a report by Science and Technology Innovation Board Daily on September 12, 3D game development platform Hiber3D used Google's PaLM large language model to fine-tune its own more than 500 template libraries and millions of finished 3D scenes to launch a new game. Development Platform. On this platform, users can quickly create 3D game scenes and functions through text question and answer methods. Currently, Hiber3D's generative AI development platform is in the testing phase and will be open to users in the future, allowing people without professional programming backgrounds to develop games.