🌅 AI Daily Digest — March 25, 2026

Data Pulse

• 7 news articles
• 5 trending models
• 5 research papers
• Cheapest GPU: RTX A5000 at $0.02/hr
• 3 new AI jobs

Today's News

• Anthropic hands Claude Code more control, but keeps it on a leash — Anthropic's Claude Code platform has been updated to enable auto mode, allowing the AI to execute tasks with fewer human approvals. In this new mode, Claude can directly control users' Macs and perform actions such as clicking on buttons. This increased autonomy is intended to streamline workflows.

• Epoch confirms GPT5.4 Pro solved a frontier math open problem — Epoch's GPT5.4 Pro model has successfully solved the long-standing math open problem related to Ramsey hypergraphs, marking a significant breakthrough in artificial intelligence and mathematics after years of eluding researchers. This achievement demonstrates the potential for AI to contribute significantly to mathematical advancements. The solution was achieved through extensive training and fine-tuning.

• Goodbye to Sora — OpenAI has announced the immediate shutdown of its AI video generation app Sora, effective March 25, 2026. Users have been left with limited details on how to preserve their content, and the company has confirmed that all access will be terminated as of the closure date. The reasons for the shutdown remain unclear.

• New open weights models: GigaChat-3.1-Ultra-702B and GigaChat-3.1-Lightning-10B-A1.8B — The editorial board has released two new open weights models: GigaChat-3.1-Ultra-702B and GigaChat-3.1-Lightning-10B-A1.8B, representing significant advancements in natural language processing with improved performance and efficiency. These models are designed to facilitate faster and more accurate text generation. The release of these open weights models is intended to promote further research and development.

• OpenAI’s Sora was the creepiest app on your phone — now it’s shutting down — OpenAI's Sora video generation app and API are shutting down just 15 months after their launch in late 2024, marking a significant pivot for the company despite its recent $1 billion licensing deal with Microsoft. The reasons behind the shutdown remain unclear, leaving users concerned about their data and content. This sudden closure has raised questions about OpenAI's long-term strategy.

• Paper: SpecEyes: Accelerating Agentic Multimodal LLMs via Speculative Perception and Planning — Researchers from the University of Science and Technology of China have developed SpecEyes, a novel framework that accelerates agentic multimodal large language models by integrating speculative perception and planning. This innovative approach enables faster and more efficient processing of complex tasks. The development of SpecEyes has significant implications for the field of artificial intelligence.

• Spotify tests new tool to stop AI slop from being attributed to real artists — Spotify is testing a new tool designed to prevent AI-generated music from being incorrectly attributed to real artists, addressing the growing issue of miscredited content and ensuring accurate recommendations. The tool aims to identify and separate AI-created tracks from human-made ones. This development has significant implications for the music industry's reliance on accurate metadata.

Trending Models

Model	Task	Likes
meta-llama/Llama-3.1-8B-Instruct	text-generation	5598
openai/gpt-oss-20b	text-generation	4477
Qwen/Qwen2.5-7B-Instruct	text-generation	1153
openai/gpt-oss-120b	text-generation	4608
deepseek-ai/DeepSeek-R1	text-generation	13100

Research

• SAiW: Source-Attributable Invisible Watermarking for Proactive Deepfake Defense — Bibek Das, Chandranath Adak, Soumi Chattopadhyay. Deepfakes generated by modern generative models pose a serious threat to information integrity, digital identity, and public trust.
• Robust Safety Monitoring of Language Models via Activation Watermarking — Toluwani Aremu, Daniil Ognev, Samuele Poppi. Large language models (LLMs) can be misused to reveal sensitive information, such as weapon-making instructions or writing malware.
• Describe-Then-Act: Proactive Agent Steering via Distilled Language-Action World — Massimiliano Pappa, Luca Romani, Valentino Sacco. Deploying safety-critical agents requires anticipating the consequences of actions before they are executed.
• Why AI-Generated Text Detection Fails: Evidence from Explainable AI Beyond Bench — Shushanta Pudasaini, Luis Miralles-Pechuán, David Lillis. The widespread adoption of Large Language Models (LLMs) has made the detection of AI-Generated text a pressing and complex challenge.
• Between Rules and Reality: On the Context Sensitivity of LLM Moral Judgment — Adrian Sauter, Mona Schirmer. A human's moral decision depends heavily on the context. Yet research on LLM morality has largely studied fixed scenarios.

GPU Deals

GPU	Price	Provider
RTX A5000	$0.02/hr	Vast.ai
Tesla V100	$0.02/hr	Vast.ai
RTX 5090	$0.04/hr	Vast.ai

View full GPU pricing dashboard

AI Jobs

• Software Engineer Reliability at OpenAI (San Francisco)
• Senior Staff Software Engineer AI Customer Operati at Monzo (Cardiff, London)
• Senior Design Engineer at Loancrate (Remote)

Browse all AI jobs

Community Events

New this week:

• Google I/O 2026 (Mountain View, USA)
• ICLR 2026 (Online)
• MLOps Community Weekly Meetup (Online (Zoom))
• Papers We Love: AI Edition (Online)

View all events