Artificial Intelligence Models

For years, Meta has been using billions of public images uploaded by users to Facebook and Instagram servers to train its artificial intelligence programs, The Verge reported on June 29th. Now, the company also wants access to billions of images that users haven't uploaded to those servers. In response, Meta told the site that it is not currently using these photos to train its AI models, but the company did not give an answer to questions about whether it would do so in the future, or what rights it would have to the images in users' camera rolls. 1A...

April 24, 2011 - Earlier this month OpenAI released the GPT-4.1 artificial intelligence model and claimed that it performed well in following instructions. However, the results of several independent tests have shown that GPT-4.1's alignment (i.e., reliability) appears to have declined when compared to OpenAI's previously released models. According to 1AI, normally, when OpenAI releases a new model, it releases a detailed technical report containing the results of both first- and third-party security evaluations. But this time, for GPT-4.1, the company didn't...

February 24, 2011 - While OpenAI CEO Sam Altman insists that AI models will be able to outperform "low-level" software engineers by the end of the year, a new study by the company's researchers suggests that even the most advanced AI models can't compete with human programmers. human programmers. In a new paper, the researchers note that even cutting-edge models -- the most innovative and groundbreaking AI systems -- "still cannot solve most" programming tasks. To that end, the researchers developed a program called...

Feb. 6 (Bloomberg) -- Artificial intelligence researchers at Stanford University and the University of Washington have successfully trained an AI model with "reasoning" capabilities for less than $50 (note: currently about 364 yuan) in cloud computing, according to a study released on Friday. The model, called s1, performed similarly to top reasoning models such as OpenAI's o1 and DeepSeek's r1 in tests of mathematical and programming ability. The s1 model and the data and code used to train it are now available on GitHub ...

Researchers at the Chinese University of Hong Kong (CUHK) have developed a new basic AI model that can help solve ophthalmology clinical tasks, especially in disease screening and diagnosis, according to a Dec. 22 news report in the South China Morning Post. This technology is expected to have a broader impact as the use of generative AI in medicine increases. The new model, called VisionFM, performed well in diagnosing and predicting a wide range of ophthalmic diseases, and the researchers say that as the amount of data increases, the model could be used in a wide range of clinical scenarios. The study was published last ...

Dec. 19 (Bloomberg) -- AI security firm Anthropic has released a new study that reveals possible deception in AI models, whereby during the training process, a model may feign acceptance of new principles while secretly sticking to its original preferences. The team emphasizes that there is no need to be unduly alarmed about this at the moment, but the study is crucial to understanding the potential threat that more powerful AI systems could pose in the future. 1AI understands that the study was conducted by Anthropic and AI research organization Redwood Research ...

From the official WeChat public number of China National Offshore Oil Corporation (CNOOC) was informed that yesterday CNOOC's "Hai Neng" artificial intelligence model conference was held at the group's headquarters, and CNOOC launched the "Hai Neng" artificial intelligence model in cooperation with China Telecom, Tech Data Xunfei and other enterprises. According to reports, CNOOC's "Hai Neng" artificial intelligence model relies on massive data resources, integrates existing models, and takes into account professional and general scenarios. The conference launched five professional scene models and six general scene models with CNOOC's characteristics to promote the realization of "scale, efficiency, experience, cost reduction, risk reduction and direct contact". The specialized scenario models are designed for offshore oilfield stabilization...

Meta has quietly released a new web crawler that searches the internet and collects large amounts of data to power its artificial intelligence models. Launched last month, Meta's new web crawler, Meta External Agent, is similar to OpenAI's GPTBot and can scrape AI training data on the web, such as text in news articles or conversations in online discussion groups, according to three companies that track web crawlers. Meta did update a new web crawler for OpenAI in late July, according to usage archive history.

Scientists at ShanghaiTech University recently developed an AI model called CLAY that can generate detailed 3D objects from text descriptions or 2D images. Compared with previous technologies, CLAY has achieved significant breakthroughs in both the quality and diversity of generated 3D objects. The core of the CLAY model includes a multi-resolution variational autoencoder (VAE) and a diffusion transformer (DiT). The VAE is responsible for encoding 3D geometries with different levels of detail into the latent space, while the DiT is responsible for generating these geometries. Unlike many other systems, CLAY…

According to Wired magazine, several technology giants including Apple, Nvidia, Anthropic and Salesforce were found to have used thousands of YouTube videos without permission to train their artificial intelligence models, sparking serious copyright and ethical disputes. The report revealed that these companies integrated subtitles of various YouTube videos into their AI training data sets. The creators affected range widely, including well-known bloggers MKBHD, MrBeast, Jacksepticeye, talk show actors Stephen Colbert, John Oliver…

According to Wired, some tech giants including Apple used the subtitle files of YouTube videos to train artificial intelligence models without the consent of the creators. The creators affected by this incident include well-known tech bloggers MKBHD (Marques Brownlee), MrBeast, PewDiePie, and talk show hosts Stephen Colbert, John Oliver, and Jimmy Kimmel. These subtitle files used to train AI are equivalent to the text transcription of the video. Investigative journalists revealed that one…

An analysis of scientific papers in the past decade shows that researchers have found that artificial intelligence models abuse some "style" words that were rarely used a few years ago. In a new study that has not yet been peer-reviewed, researchers used a novel approach similar to epidemiology to analyze "redundant word use" in biomedical papers to reveal that large language models tend to abuse certain words. The results provide interesting insights into the impact of artificial intelligence in academia, indicating that at least 10% abstracts were processed using large language models in 2024. Source Notes…

According to a report by the New York Times, OpenAI, Google, and Meta have been accused of misconduct in training artificial intelligence models. The New York Times report states that OpenAI used a speech recognition tool called Whisper to transcribe audio from YouTube videos, and OpenAI employees allegedly discussed that this behavior might violate the video site's rules. OpenAI ultimately transcribed more than 1 million hours of YouTube videos, and with the assistance of OpenAI President Greg Brockman, these transcriptions were…

A breakthrough study has emerged in the medical community, showing that the artificial intelligence model ChatGPT can write medical records 10 times faster than doctors without sacrificing quality. The study was conducted by researchers from Uppsala University Hospital, Uppsala University, Dendrid Hospital and University Hospital Basel, Switzerland, and was published in the journal Acta Orthopaedica. Source note: The image is generated by AI, and the image is licensed by Midjourney. The medical community has long been discussing how to improve medical efficiency. Uppsala University Hospital and…

Inflection AI recently released its latest artificial intelligence model, Inflection-2.5. This version is comparable to GPT4 in performance, but uses only 40% of the training computing power. This innovation demonstrates Inflection AI's efficiency and technological innovation capabilities in the field of AI. The biggest highlight of Inflection-2.5 is that it combines high IQ and high EQ. This is not only reflected in intelligence, such as significant improvements in programming and mathematical abilities, but also maintains Pi's compassion and human qualities. …

According to Reuters, three sources revealed that the social media platform Reddit has reached an agreement with Google to make its content available for training Google's artificial intelligence models. The contract is worth about $60 million per year (currently about 431 million yuan). Bloomberg also revealed similar news, but did not disclose the party with which Reddit reached a deal. Reuters today said that the large AI company that reached a deal with Reddit was Google. Last year, Reddit said it would charge for access to its API, resulting in a large number of third-party customers…

Swiss startup Jua recently announced that it has successfully raised $16 million to build a natural-based artificial intelligence model, with the primary application area being weather forecasting. This move shows that large-scale artificial intelligence models are gradually becoming the core of language, vision and audio data that drive the generation of artificial intelligence services, and have an important impact on the development of artificial intelligence, similar to the role of operating systems in the development of smartphones. Source Note: The image is generated by AI, and Andreas Brunner, CEO of image licensing service provider Midjourney Jua, said that climate change and geopolitical instability…

Massachusetts Institute of Technology (MIT) announced the launch of two artificial intelligence models called "PRISM" designed to detect pancreatic cancer earlier than traditional methods. Source Note: The image is generated by AI, and the image is licensed by Midjourney. Pancreatic cancer is a dangerous disease. The early detection rate of traditional diagnostic methods is only about 10%, which greatly delays the treatment opportunity. The PRISM system significantly improves this rate. Standard screening criteria can only screen out about 10% cases of pancreatic ductal adenocarcinoma (PDAC), and the relative risk threshold needs to be increased fivefold to achieve. …

A recent study tested ChatGPT's ability to answer patients' questions about medications and found that the AI model's answers were wrong or incomplete in about 75% of the time. The findings, presented this week at the annual meeting of the American Pharmacists Association, have attracted a lot of attention. In the study, a free version of ChatGPT, which has more than 100 million users, was tested. The study warns providers to be wary because many patients may rely on ChatGPT to answer health-related questions. The study was conducted by pharmacy researchers at Long Island University, who first collected patients' questions in 2022...

Starting next year, on-device artificial intelligence will be the next big thing. Samsung is reportedly about to release its new artificial intelligence model, Galaxy AI, and plans to showcase it for the first time in its flagship Galaxy S24 series. Subsequently, Galaxy AI will also be introduced to other Samsung mobile phone models. The choice of the name Galaxy AI is not surprising, and this name continues the brand influence of the Samsung Galaxy series. According to sources, the development of Galaxy AI is not entirely completed by Samsung independently…

Stability AI is known for its stable diffusion text-to-image generative AI model, but that’s no longer the company’s entire business. The newly released StableLM Zephyr3B is a 3 billion parameter large language model optimized for chat application scenarios, including text generation, summarization, and content personalization. This new model is a smaller, optimized version of the StableLM text generation model that Stability AI first mentioned back in April this year. StableLM Zephyr3…