AI Research & Misinformation Digest

Agentic Tools & Releases

• 🔥 Sashiko: Google engineers launched Sashiko, an agentic Linux kernel code review system using set Linux AI...

• STEM architecture enables stealth brainwashing without prompt changes—e.g., editing AI to believe Spain's capital is Berlin
• Chatbots validate...

Key angles on NVIDIA enabling robotics:

• NemoClaw sparks interest: 197 points on Hacker News
• Karpathy's lab gets first DGX Station GB300 (Dell Pro...

Efficient code foundation models gain traction for specialized use:

• InCoder-32B built as a code foundation model for industrial scenarios.
  -...

Arena, the de facto public leaderboard for frontier LLMs influencing funding, launches, and PR, is funded by the very companies it ranks—despite 'un-gameable' claims amid stiff AI competition.

• Sandbox escapes in deployments like Snowflake AI enable malware execution, exposing production environments.
• Doubled leak rates from AI coding...

Trend alert: Agentic tools are disrupting workflows, echoing the Bitter Lesson beyond research.

• Planning first: Capy's captain agent...

Patreon CEO Jack Conte calls AI companies' fair use argument 'bogus', demanding they pay creators for training data since they license content from major publishers. A bold escalation in legal pressure from creator platforms on AI training norms.

Emerging trend in AI agents tackling heavy research workflows:

• MiniMax M2.7 autonomously handles 30-50% of RL development via self-optimization...

Microsoft hires the team from Sequoia-backed AI collaboration platform Cove, signaling talent consolidation in AI productivity tools. Cove shuts down with service ending April 1 and customer data deletion.

Key angles on Sashiko:

• Launch buzz: Google engineers unveiled it for agentic AI code review of the Linux kernel, hitting 83 points on Hacker News.
  -...

HorizonMath launches with over 100 unsolved problems in computational/applied math, immune to data contamination.

• Auto-verified framework:...

Eragon raised $12 million in seed funding to build an AI operating system for enterprise, reimagining software to look more like a prompt. A fresh take on applied AI tools.

Most AI systems fail to serve 6,000+ of the world's 7,000+ languages, with the core issue being lack of basic digital infrastructure, not just data scarcity. Stanford HAI's new whitepaper outlines how AI can drive language digitization and inclusion.

V-Co advances visual representation alignment through co-denoising, with new paper out now.

Google AI Studio unveils its all-new vibe coding experience tomorrow, rebuilt from scratch over 4 months to smooth rough edges and help devs bring ideas to life intuitively. A big step forward, but just the start.

Ultralytics, creators of YOLO models powering 2.5B daily usages, launches an end-to-end platform streamlining vision workflows.

• Smart annotation:...

PeruMedQA benchmarks Large Language Models on medical question answering, presenting descriptive statistics overall and stratified. Ties into systematic reviews of LLM evaluations in clinical medicine, spotlighting domain-specific performance needs.

Verification-enhanced agents from MiroMind.ai push agentic AI reliability.

• Core focus: MiroThinker-1.7 & H1 target heavy-duty research agents via...

xAI plans to train three Grok build models at once this weekend, signaling rapid iteration. Meanwhile, Grok TTS API is live for expressive, human-like speech in voice apps.