The Register
·
2 months ago
● 2 sources
Gartner warns that enterprises will deploy over 150,000 AI agents by 2028, up from fewer than 15 today, requiring strict governance to prevent uncontrolled proliferation and security risks. Organizations using third-party governance tools were nearly twice as likely to report higher value from AI deployments, compared to those restricting access without structured controls. Companies need centralized governance committees, agent inventories, AI TRiSM tools, and continuous monitoring to manage agent sprawl and prevent misinformation, data loss, and IT complexity.
The Register
·
2 months ago
Mozilla opposes Google's Prompt API, which embeds the Gemini Nano language model into Chrome for local inference, arguing it creates vendor lock-in and undermines web interoperability by encouraging developers to standardize on Google's model. Google recommends 22 GB of storage space for the Nano model, though the v3Nano version is approximately 4.27 GB, and testing shows it fails on 15-24 percent of generative tasks. If widely adopted, the API could pressure competitors like Apple and Mozilla to license Google's model to maintain compatibility, and it requires acceptance of Google's content policies that extend beyond legal requirements.
Deep Learning Weekly
·
2 months ago
Deep Learning Weekly Issue 453 covers developments in AI systems including OpenAI's GPT-5.5 and Codex arriving on AWS, Mistral AI's Workflows orchestration engine, and Ineffable Intelligence raising $1.1B for a reinforcement learning-based AI system. Key concrete details include Mistral's Workflows already executing millions of daily tasks, Ineffable Intelligence's $5.1B valuation, and RecursiveMAS showing 8.3% average accuracy improvement across benchmarks with 1.2x-2.4x inference speedup. These developments advance agent-based AI systems, distributed training infrastructure, and observability tools for production AI applications.
Sequoia
·
2 months ago
Standard Intelligence, founded by Galen Mead and Devansh Pandey, is training a general computer agent called FDM-1 by learning directly from raw video of computer use rather than text or language models. The team created an 11-million-hour computer action dataset and a video encoder 50 times more token-efficient than competitors, allowing nearly two hours of video to fit in a 1-million-token context window. The approach aims to build general agents that can perform diverse tasks like CAD modeling, driving, and software debugging by scaling video pre-training rather than expanding language model capabilities.
The Register
·
2 months ago
Anthropic generated more LLM revenue than OpenAI in Q1 2026 with a 31.4 percent market share compared to OpenAI's 29 percent, despite having roughly 134 million monthly users versus OpenAI's 900 million. Anthropic extracts an average of $16.20 revenue per active user monthly, significantly higher than OpenAI's $2.20, Microsoft's $5, and Google's $1.10. The shift reveals that success in LLM markets increasingly depends on extracting higher revenue from smaller, premium customer bases rather than maximizing user scale, while major infrastructure investors like Meta, Alphabet, Amazon, and Microsoft face pressure to justify 77 percent year-over-year spending increases totaling $725 billion in 2026.
Google DeepMind
·
2 months ago
Google DeepMind announced an AI co-clinician research initiative designed to work alongside physicians in patient care rather than replace clinical judgment. In evaluation of 98 primary care queries, the system recorded zero critical errors in 97 cases and outperformed other AI systems on medication knowledge benchmarks. The AI performed comparably to primary care physicians in 68 of 140 assessed consultation skill areas during simulated telemedical encounters, with ongoing trials across the US, India, Australia, and other regions.
Allen Institute (AI2)
·
2 months ago
AstaBench, an open benchmark for measuring AI agent scientific research capabilities, released new results testing frontier models including GPT-5.5 on over 2,400 research problems and updated its leaderboard. Claude Opus 4.7 achieved the top overall score of 58.0% while GPT-5.5 reached 52.9% at $1.61 per problem, showing that frontier models improve unevenly across categories with Code & Execution and End-to-End Discovery gaining substantially but Data Analysis and Literature Understanding gaining only moderately. The benchmark is gaining adoption from the UK AI Security Institute, General Reasoning, and other organizations, establishing AstaBench as an industry standard for evaluating AI's ability to perform grounded scientific research workflows.
Last Week in AI
·
2 months ago
A podcast episode covered major AI developments from the week including OpenAI's new ChatGPT image model for accurate text generation, Alibaba's Qwen 3.6 Max and other Chinese model releases, Google's expanded Deep Research with Gemini 3.1 Pro, and Mozilla using Claude to fix 271 Firefox bugs. Major business deals included Amazon investing $5B in Anthropic, Cerebras filing for an IPO, and a reported SpaceX-Cursor agreement with a $60B buy option. The episode also discussed emerging concerns around AI-generated content including 44% of daily Deezer uploads being AI music and YouTube's new deepfake removal requests for celebrities.
Together AI
·
2 months ago
Together AI and Adaption announced a partnership integrating Together's fine-tuning service into Adaption's platform to help teams optimize datasets and deploy open models. The partnership enables users to run fine-tuning workflows and evaluate results within a single interface combining both companies' tools. Teams can now streamline the process of dataset optimization and model deployment without switching between separate platforms.
