AI Models

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 ...

Meta on Thursday announced the launch of an artificial intelligence model called Meta Motivo, which is designed to control the movements of humanoid digital intelligences to enhance metaverse experiences. Meta also released AI tools such as LCM, a large-scale conceptual model, and Video Seal, a video watermarking tool, and reiterated its commitment to continued investment in AI, AR and metaverse technologies. Note: Meta Motivo is a behavior-based base model trained in the Mujoco simulator using AMASS motion capture data...

December 11, DeepSeek official public yesterday (December 10) released a blog post announcing the conclusion of DeepSeek V2 series, launching the final version of DeepSeek V2.5 fine-tuned model DeepSeek-V2.5-1210, mainly to support the networking search function, comprehensively improve the various capabilities. DeepSeek-V2.5-1210 has made significant progress in math, code, writing, and role-playing through Post-Training iterations, in addition to optimizing the text...

LG Artificial Intelligence Research Institute released EXAONE 3.5 open-source AI model on Monday (December 9) and simultaneously launched ChatEXAONE, an enterprise-level AI intelligence service for LG employees. EXAONE 3.5 The release of EXAONE 3.5 is only four months after the 3.0 version, and the new model offers three versions: a 2.4 billion-parameter ultra-lightweight device-side model, a lightweight general-purpose model with 7.8 billion parameters, and a high-performance specialized model with 32 billion parameters. L...

Beijing time this morning, according to The Information, citing sources familiar with the matter, Google has recently pressured the U.S. Federal Trade Commission (FTC) to lift Microsoft's exclusive agreement to host OpenAI technology on its cloud servers. The comments came after the FTC asked Google about Microsoft's business practices. It is understood that the purpose of the FTC inquiry was to conduct a broader investigation. A range of Microsoft competitors such as Google and Amazon want to host OpenAI's AI services themselves, with the aim of eliminating the need for their cloud customers to both...

December 10, 2011 - The bottleneck in training AI models is no longer just about architecture design, but also about data management efficiency - Meta AI has launched an open-source scalable and high-performance data loading (SPDL) tool that ultimately speeds up AI training by improving data loading efficiency. The SPDL tool uses multi-threading technology to achieve high throughput and lower resource usage in the regular Python interpreter (without the free-threading option enabled), and is compatible with the Free-Threaded Python...

Dec. 8 (Bloomberg) -- Google says its new family of artificial intelligence models has a nifty feature: the ability to "recognize" emotions. Google on Thursday unveiled PaliGemma 2, its newest family of AI models, which has image analysis capabilities that can generate descriptions of images and answer questions about the people in them. In its blog post, Google describes PaliGemma 2 as being able to not only recognize objects, but also generate detailed and contextually relevant image descriptions that cover actions, emotions, and the overall narrative of the scene. PaliGemma 2's emotion-recognition feature isn't out-of-the-box...

Meta's grand finale AI model of the year is here. Meta released Llama 3.3 yesterday (December 6), with 70 billion parameters, but with performance comparable to Llama 3.1, which had 405 billion parameters. Meta emphasizes that Llama 3.3 models are more efficient and less costly, and can be run on standard workstations, lowering operational costs while delivering high-quality Text AI solutions. Llama 3.3 models are optimized for multi-language support, with support for English, German, French, Italian,...

1 February 7, 2011 - Technology media outlet techtarget published a blog post yesterday (December 6) reporting that Ultralytics' YOLOv11 AI model has been hit by a supply chain attack, with the v8.3.41 and v8.3.42 versions implanted with crypto-mining software. As of 1AI's writing, Ultralytics has not issued an official security advisory, but the company has responded quickly by removing the two affected versions and releasing a new one. The issue was first reported by developer metri...

December 7, 2012 - OpenAI has launched a 12-day "shipmas" release cycle, featuring a series of new features, products, and demos. On the second day of the event, OpenAI launched Reinforcement Fine-Tuning, a program that helps developers and machine learning engineers build expert models for complex domain-specific tasks. With a new model customization technique, the program allows developers to fine-tune models using high-quality task sets and evaluate the model's response using reference answers from...

December 4, 2012 - Amazon today announced a new set of AI base models, branded as "Nova," that will be available through AWS' Amazon Bedrock model library. In a blog post, Amazon said there are now three "comprehension" models to choose from: Amazon Nova Micro: a text model optimized for "speed and cost". Amazon Nova Lite: a "very low-cost" multimodal model that can be fed images, video and text to generate text. Amaz...

According to The Information, Amazon has developed a new set of generative AI models that can process images and videos in addition to text, reducing its reliance on Anthropic. The new model, code-named Olympus, will be able to understand scenes in images and videos and search for specific clips or scenes in videos, such as a kill shot in a basketball game, using simple text prompts, according to the report. It can also use AI models to make the "best coffee" or "raindrops on the ground", as well as simple text prompts...

NVIDIA published a blog post on November 25, announcing the launch of music generation AI model Fugatto, claiming that it is "the world's most flexible sound machine", which can be finely controlled sound generation. NVIDIA said the tool is like a "Swiss Army Knife" in the field of sound, not only can create music, modify the sound, but also the flexibility to mix a variety of music, vocals and sound effects, and even create unprecedented sounds. Users simply type in a text description or insert some audio, and Fugatto generates corresponding music clips, sound effects, and even changes the accent and emotion of vocals based on the description. For example, the user...

Nov. 19 - Mistral AI announced yesterday, Nov. 18, a new multimodal AI model, Pixtral Large, with 124 billion parameters, based on Mistral Large 2, and designed primarily for processing text and images. Pixtral Large is now available under the Mistral Research License and Commercial License for research, education, and commercial use. Pixtral Large is a Mistral ...

November 19th, Ali Tongyi Qianqian released a blog post yesterday (November 18th) announcing the launch of the Qwen2.5-Turbo open source AI model in response to the community's request for a longer Context Length after months of optimization and polishing. Qwen2.5-Turbo extends the context length from 128,000 to 1,000,000 tokens, an improvement equivalent to about 1,000,000 English words or 1,500,000 Chinese characters, and can accommodate 10 complete novels,...

Nov. 19, 2011 - A team of researchers from Peking University, Tsinghua University, Pengcheng Lab, Alibaba Dharmo Academy, and Lehigh University has introduced LLaVA-o1, the first GPT-o1-like systematic inference visual language model that is spontaneous, which can be explained at the end of the article. language model. LLaVA-o1 is a novel visual language model (VLM) designed for autonomous multi-stage reasoning. LLaVA-o1 has 1...

According to Bloomberg, AI giants including OpenAI, Google and Anthropic are facing "diminishing returns" as they hit a bottleneck in developing more advanced AI models. OpenAI's newest model, Orion, reportedly struggled with coding tasks, with no significant improvement over GPT-4. Google's upcoming Gemini software faces similar challenges, while Anthropic has delayed its highly anticipated Claude 3.5 o...

In a press release last week, Meta announced that it has officially open sourced the MobileLLM family of small language models that run on smartphones, and has added three new parameterized versions of the family, 600M, 1B, and 1.5B to the project's GitHub project page (click here to visit). According to Meta researchers, the MobileLLM family of models, built for smartphones, claims to have a lean architecture and introduces "SwiGLU activation functions," "grouped-query attenuation," and a "new language model with a new language model. ...

At the recent Gemma Developer Day in Tokyo, Google officially launched a new Japanese version of the Gemma AI model. The model's performance rivals that of GPT-3.5, but it only has a mere 2 billion covariates, making it very small and suitable for running on mobile devices. The Gemma model in this release excels in Japanese language processing while maintaining its capabilities in English. This is especially important for small models, which can face the problem of "catastrophic forgetting" when fine-tuning for a new language, i.e., newly learned knowledge overwrites...

Recently, a mysterious AI image generation model codenamed "red_panda" scored amazingly well in the benchmark test of Artificial Analysis, a crowdsourcing analytics platform, significantly outperforming the products of industry leaders such as Midjourney, Black Forest Labs and OpenAI. According to Artificial Analysis, "red_panda" scored 1,244 points in the text-to-image...

Technology media outlet NeoWin (Oct. 21) published a blog post reporting that IBM, at its annual TechXchange event, unveiled a new Granite 3.0 family of AI models that can equal or exceed models of similar size in academic and industry benchmarks. The Granite 3.0 series includes a variety of new models, the relevant models are as follows: Generalized / Linguistic Models: Granite 3.0 8B Instruct Granite 3.0 2B Instruct ...

Recently, social platform X updated its privacy policy, which will allow X platform to use user data to train AI models from November 15, unless the user opts out, triggering user dissatisfaction. Previously, Adobe, Google and other companies also introduced similar content in the terms and conditions, causing controversy over the conflict between AI training and privacy, copyright, etc., and related legal issues are still under discussion. Change: user data will be used for AI training Recently, the X platform updated its privacy policy with a new clause that allows it to share user data with third parties to train AI, unless the user opts out. However, the platform did not provide a clear opt-out option and reminded users that even within...

Feifei Li's startup World Labs has announced a deal with Google Cloud, choosing it as its primary compute provider for training AI models. The deal could be worth hundreds of millions of dollars. World Labs will utilize GPU server licenses on the Google Cloud platform to provide compute services for its large multimodal AI models. The company's AI models are called "spatial intelligence" and can process, generate, and interact with video and geospatial data. Goog...

Technology media NeoWin published a blog post yesterday (October 4), reporting that Google Inc. will soon commercialize the Gemini 1.5 Flash 8B model, which will become Google Inc.'s cheapest AI model. Reported in August this year, Google Inc. launched three experimental Gemini models, of which the Gemini 1.5 Flash 8B is a smaller-sized model of the Gemini 1.5 Flash with 8 billion parameters designed for multimodal tasks, including high-volume tasks and long text summarization tasks...