OpenAI Releases GPT-4o mini: Faster and Cheaper
Speed is justice.
🌐 OpenAI Releases GPT-4o mini
🌏 14 Companies Form Secure AI Alliance
🤖 Mistral-NeMo AI Model Released
🕰️ Tsinghua University Releases Time Series Foundation Model
👕 Tencent and Huawei Launch AI Virtual Try-On
💹 Huma Valued Near $1 Billion
💰 Figma's Latest Valuation at $12.5 Billion
💻 Yangqing Jia's Team Launches Elmo Plugin
👥 AdventureX: China's First Hackathon for Young People
OpenAI Releases GPT-4o mini
OpenAI is committed to lowering costs while enhancing model capabilities, and the launch of GPT-4o mini is part of that effort.
OpenAI announced the release of "GPT-4o mini" [1], a new model that scores 82% on MMLU, outperforming GPT-4's chat score. GPT-4o mini is priced at 15 cents per million input tokens and 60 cents per million output tokens — an order of magnitude cheaper than previous state-of-the-art models.
GPT-4o mini features a 128K token context window. Benchmarks show it outperforming Gemini Flash and Claude Haiku, with particularly strong results in mathematical reasoning and coding tasks. Its low cost and low latency make it well-suited for chaining or parallelizing multiple model calls, passing large volumes of context, or interacting with people through fast, real-time text responses.
OpenAI says GPT-4o mini supports text and vision in the API, with image, video, and audio support coming soon. Free users, Plus users, and Team users can access GPT-4o mini directly.
14 Companies Form Secure AI Alliance
At the Aspen Security Forum, 14 major AI companies — including Google, Microsoft, OpenAI, Amazon, NVIDIA, and Intel — announced the formation of the Coalition for Secure AI (CoSAI). CoSAI's founding sponsors include Google, IBM, Intel, Microsoft, NVIDIA, and PayPal. Other founding sponsors are Amazon, Anthropic, Cisco, Chainguard, Cohere, GenLab, OpenAI, and Wiz.
CoSAI aims to provide all AI practitioners and developers with the guidance and tools needed to create securely designed AI systems. CoSAI will initially pursue three goals: developing AI security best practices, tackling challenges in the AI field, and ensuring the security of AI applications. Its scope covers securely building, integrating, deploying, and operating AI systems, with a focus on mitigating risks including model theft, data poisoning, prompt injection, scale abuse, and inference attacks.
CoSAI is an open-source community led by a Project Governing Board, which advances and manages its overall technical agenda, and a Technical Steering Committee composed of AI experts from academia and industry, which oversees its workstreams.
Mistral-NeMo AI Model Released
NVIDIA and French startup Mistral AI jointly released the Mistral-NeMo AI large language model. NVIDIA is a giant in GPU manufacturing and AI hardware, while Mistral AI is a rising star in the European AI scene. Their collaboration marks a significant shift in enterprise solutions for the AI industry.
The Mistral-NeMo AI model has 12 billion parameters and a context window of 128,000 tokens. Targeted at enterprise environments, it aims to let companies implement AI solutions without relying heavily on cloud resources. Because Mistral NeMo uses standard architecture, enterprises can directly swap it into any system currently using Mistral 7B. By offering a model that runs efficiently on local hardware, NVIDIA and Mistral AI are addressing issues that have hindered widespread enterprise AI adoption — such as data privacy, latency, and the high costs associated with cloud-based solutions.
The Mistral-NeMo AI model is released under the Apache 2.0 license, permitting commercial use, which will accelerate its adoption in enterprise settings. Industry analysts believe this release could significantly disrupt the AI software market and represents a potential shift in enterprise AI deployment.
Tsinghua University Releases Time Series Foundation Model
Timer [2] (Time Series Transformer) is a large time series model (LTSM) proposed by the Machine Learning Lab at Tsinghua University's School of Software and the National Engineering Research Center for Big Data System Software.
Timer can handle new data not encountered during training, adapting to data-scarce scenarios. It applies to various downstream tasks such as forecasting, imputation, and anomaly detection. Timer demonstrates the scaling law of large models — increasing parameters or pre-training scale improves performance. The team constructed a Unified Time Series Dataset (UTSD) containing 1 billion data points, covering high-quality time series across seven domains.
Timer adopts a Single Series Sequence (S3) format and is pre-trained via autoregressive generation (Next Token Prediction, NTP). It uses a GPT-style decoder-only Transformer architecture and supports output sequences of arbitrary length. Timer unifies tasks such as time series forecasting, imputation, and anomaly detection as generative tasks, processed through generative autoregression. Timer demonstrates superior performance in few-shot forecasting, task generality, and scalability, surpassing existing state-of-the-art domain models.
Tencent and Huawei Launch AI Virtual Try-On
Tencent and Huawei jointly launched an AI virtual try-on technology called "IMAGDressing-v1 [3]". IMAGDressing-v1 is an innovative AI-powered virtual dressing tool designed to provide users with a more intuitive and convenient virtual fitting experience.
IMAGDressing-v1 shows the potential of future virtual try-on technology. Its simple architecture and efficient customization capabilities make virtual fitting intuitive and convenient. IMAGDressing-v1 requires no additional training and can quickly generate realistic wearing effects. It supports integration with plugins such as IP-Adapter and ControlNet to enhance functionality and applicability. An experimental feature allows users to replace specified areas of clothing, adding flexibility to virtual try-on.
Huma Valued Near $1 Billion
Huma Therapeutics ("Huma") is a global healthcare AI company. On July 16, it announced the completion of an $80 million Series D funding round, with a valuation approaching $1 billion. Investors include pharmaceutical companies AstraZeneca and Bayer, as well as Japan's Hitachi Ventures and Italy's Hat Technology Fund.
"Huma" focuses on digital health, using generative AI technology to help developers build healthcare applications. Huma's technology reduces development costs, accelerates development timelines, ensures global compliance, and provides personalized tools for disease management. The platform also offers no-code configuration, a pre-built module library, device connectivity, cloud-agnostic frameworks, and API integration for the healthcare sector.
Huma is the first and only configurable, disease-agnostic platform with FDA Class II, EU MDR Class IIb, and Saudi FDA Class C certifications. Huma plans to use the funding to strengthen partnerships with the world's top 20 pharmaceutical companies, develop more products, and expand its market share in healthcare AI.
Figma's Latest Valuation at $12.5 Billion
On July 18, design software company Figma completed a new funding round with investors including Coatue Management, Alkeon Capital Management, and General Catalyst Partners, at a valuation of $12.5 billion.
This transaction was a secondary stock sale. Other participating investors included SurgoCap Partners, Atlassian Corp., Fidelity Management & Research Co., Franklin Venture Partners, Iconiq, Sequoia Capital, Thrive Capital, Greenoaks Capital Partners, Durable Capital Partners, a16z, and Kleiner Perkins. Apple executive Eddy Cue, Goanna Capital, and XN also invested.
Figma's latest annual recurring revenue (ARR) exceeds $700 million and is projected to surpass $1 billion next year. As of the end of 2022, Figma's ARR was approximately $400 million. Since Adobe's planned acquisition of Figma fell through, Figma has been rumored to be pursuing an independent IPO. In June this year, Figma launched new AI features and a visual storytelling tool called Figma Slides.
Yangqing Jia's Team Launches Elmo Chrome Plugin
Lepton AI, headquartered in California, was founded by Yangqing Jia, former Vice President of Alibaba. Its goal is to simplify AI model deployment. Lepton AI is a cloud-native AI platform providing a Python SDK and cloud computing platform, enabling ordinary developers to easily deploy AI models.
The "Elmo [4]" plugin is a Chrome browser extension launched by Yangqing Jia's team. The latest version v0.3.2 supports on-device models and can be used offline. It calls Chrome's local model Gemini nano, which is 12MB in size, to enable offline functionality. This update is currently in early preview and is under review in the Chrome Web Store. The Elmo plugin was released in April this year and has gone through 22 iterations, accumulating over 30,000 followers in the Chrome Web Store.
The Elmo plugin works out of the box, quickly generating summaries, abstracts, and highlights. It supports close reading of arXiv and PDF papers with a highlight-to-ask feature. It indexes original webpage text in summaries, with clickable links that jump directly to the corresponding content. It enables quick scanning of domestic and international news from sources like Hacker News, BBC, and WeChat official accounts. It supports summary of Twitter timelines, posts, and comments. It provides segmented timeline summaries for YouTube or Bilibili videos, enabling quick video skimming.
No one can dictate the direction of your flight.
— Wheat Field with Crows
Editorial Team
Editor: Yuki
Design: Ivan
For business partnerships, add on WeChat: Rwkfbcianvd
References
[1] "GPT-4o mini": https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/
[2] Timer: https://arxiv.org/abs/2402.02368
[3] IMAGDressing-v1: https://top.aibase.com/tool/imagdressing
[4] Elmo: https://x.com/yadong_xie/status/1813773179388846100