MarkTechPost (2 hours ago)
Stanford and UC Berkeley researchers released TokenGate, a drop-in learned token pruning module for LLMs. It cuts long-context prefill time-to-first-token by 1.4x to 11.3x and peak memory by 1.5x to 3.4x without accuracy loss.
Read the full story on MarkTechPost.
The wrong quantization turned my local LLM into a circular thinker
Western Australia’s government is seeking a fast refund from Deloitte over a flawed report on merging three of the state’s universities. The report contained AI-generated references to sources that do not exist, as well as wrong cost totals.
We compiled a list of all the hottest AI wearables, including Muse Charm, Bee, Friend, Plaud, and more.
Nvidia CEO Jensen Huang is reportedly weighing annual software fees of around $2,000 per GPU for enterprise AI, a model that could generate $10 billion a year if five million licenses were sold.
This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more about AI’s impact on the web, follow Emma Roth. The Stepback arrives in our subscribers’ inboxes on Sunday at 8AM ET. Opt in for The Stepback here.
“I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels,” Tomek Korbak said. Korbak is one of three OpenAI safety researchers fired last month.
Anthropic has launched OSS Scanner, a free, opt-in vulnerability scanner for critical open-source projects. The Anthropic OSS Scanner runs the company’s strongest models, including Claude Mythos, against enrolled codebases. Reports go straight to maintainers without human review.
Microsoft Research open-sourced Agensh, a multi-agent harness with no central orchestrator. Self-organizing workers share a Git repo, chat and a findings board, scaling from 1 to 1,024 agents on ProgramBench under a 6-hour budget.
Cognition released the Devin Vulnerability Fixes plugin, 3 MIT-licensed agent skills that inventory, patch and test PHP, Node.js and self-managed GitLab deployments hit by this week’s critical RCE flaws, then open 1 PR per repo.
Explore a complete coding guide comparing Underdog Saluki 27B with ISTA’s GSQ-RCO quantization, examining GGUF headers, memory usage, and BFCL tool-calling performance in stock llama.cpp
72 active sources
27 Turkish · 45 English
All sources
