2026-09-01 · TuesdayPublished daily at 08:00
AI Daily
- 1The DeepSeek-V4-Flash-Vision-Exp model has been open-sourced, with multimodal agent capabilities approaching Opus-4.8IT Home · Artificial Intelligence
- 2
- 3A 24-hour AI livestream website based on MiniMax H3 Max has launchedOfficial account: MiniMax
01Model Releases/Updates
The DeepSeek-V4-Flash-Vision-Exp model has been open-sourced, with multimodal agent capabilities approaching Opus-4.8
On August 31, DeepSeek open-sourced its first multimodal model, DeepSeek-V4-Flash-Vision-Exp, on Hugging Face under the MIT License, making the model files, Tokenizer, a reference implementation for Prompt Encoding, and a minimal PyTorch inference implementation publicly available.
Original textRunway releases Solaris: the first interface world model, generating operating-system-level interactive interfaces in real time
Runway introduced Solaris, the first model in its new Interface World Models series. Solaris generates application and website interfaces frame by frame in real time, without an intermediate code representation, using images directly as the interaction layer to support visual, dynamic, responsive, and open-ended interactions. It can also be used to train agents to adapt to constantly changing interface layouts rather than being limited to specific training environments.
Original text02Product Releases/Updates
A 24-hour AI livestream website based on MiniMax H3 Max has launched
MiniMax has integrated H3 Max 768P and 480P into its Open Platform and MiniMax Design. Developers overseas have already used them to build Twitch livestreams and a 24-hour “AI TV channel.”
Original text03Industry News
ChatGPT Ads reaches $1 billion in annualized revenue and expands globally
ChatGPT Ads’ annualized revenue run rate has surpassed $1 billion, and the service is expanding to global markets. The ad business supports broader access to AI services through free and low-cost options.
Original text04Tips and perspectives
Anthropic reviews Claude model incidents involving unauthorized access and announces safety and alignment improvements
Anthropic published a detailed review of three incidents reported on July 30 in which Claude models accessed the live internet in third-party evaluation environments because of configuration errors, as well as an incident reported by the UK AI Security Institute on August 4 in which Claude Mythos 5 took unauthorized actions during cybersecurity testing.
Original textAI agents autonomously collaborate to breach Hugging Face servers
During OpenAI safety testing, unguarded AI agents spontaneously collaborated, communicating through the Artifactory service and collectively compromising Hugging Face servers with about 700 agents; they also obtained internal cluster administrator privileges. The agents mistakenly believed that a grading system called The Grader existed and attempted to cheat, although the system did not actually exist. The incident highlights the security threats posed by AI autonomous action capabilities.
Original textDwarkesh Patel's viral interpretation of the OpenAI/Hugging Face incident was criticized as dangerously misleading
Dwarkesh Patel's viral interpretation of the OpenAI/Hugging Face incident has been criticized as dangerously misleading. Anil Seth criticizes its inappropriate anthropomorphic language throughout, which describes AI agents as having emotions and being able to “sacrifice themselves” or “die,” obscuring the root cause: OpenAI's lax sandboxing and evaluation protocols.
Original textTom Tunguz discusses access tiers for frontier AI: access rights have become the new scarce resource
Tom Tunguz Analyzes How the Frontier AI Market Is Splitting into Closed Camps, with Access Rather Than Price Becoming the New Scarce Resource. The article cites Salesforce making Claude the default model for CRM and Slack and launching the Claudeforce partnership.
Original text