AI Research Pulse

• Policy angle: House lawmakers seek transparency for AI foundation models, with FTC in key role amid trust concerns
• Research angle: New multi-agent...

Key breakthrough in applied AI for pharma:

• $10M seed led by redalpine, with angels like Cohere's Aidan Gomez and HuggingFace's Clément Delangue.
  -...

Breakthrough training hub opened in 2025: 3,000m² facility on decommissioned steelworks teaches robots to detect fire, feel heat, and navigate...

Key shifts in software engineering from AI agents:

• Product Management Bottleneck: Deciding what to build now constrains more than building it.
  -...

NVIDIA and University of Maryland release AF-Next, the most capable open Large Audio-Language Model (LALM) trained on internet-scale data for speech,...

Key breakthrough: APL developed in-house capabilities to train LLMs from scratch, adapting them for sensitive government missions where commercial...

Training Efficiency Advances

• 🔥 CompreSSM: Researchers at MIT CSAIL, Max Planck Institute, ELLIS, ETH, and Liquid AI developed CompreSSM, a...

Trump Administration outlines its AI policy in the 2025 AI Action Plan and several related Executive Orders, focusing on attaining American leadership.

MIT CSAIL's CompreSSM revolutionizes state-space model training by compressing mid-process, dodging massive pre-training costs.

• Early pruning:...

Hebrew University research uncovers LLMs forming structured trust assessments akin to humans, scoring competence, integrity, and benevolence...

Key reasons old defenses crumble and new fixes emerge:

• Perimeter obsolete: LLMs via MCP pull data from everywhere—no central server to protect.
  -...

Physical AI trend surges with real-world autonomy breakthroughs:

• Locus Array launches as fully autonomous fulfillment system, combining robotics...

Anthropic donated $20 million to Public First Action, widely viewed as joining midterm spending. Yet, their statement to Transformer insists funds cannot influence elections, spotlighting Big Tech's indirect AI policy advocacy.

Matrix-Game 3.0 unveils a real-time and streaming interactive world model equipped with long-horizon memory – key for next-gen AI agents.

Why they succeed: Collective activity patterns across neurons implement cognition and motor control, invariant within populations, across...

Rising critiques reveal core LLM limits:

• Creativity narrowed: Reliance standardizes outputs, reducing originality vs. human work.
• Safety gaps: No...

Emerging trend: Promptable VLMs and 3D perception enable flexible embodied AI.

• Narwal Flow 2 robot vacuum uses cloud-based VLM for human-like...

• Trinity's InsightsEDGE launches AI digital twins of HCPs, patients, payers from unified data for rapid market simulations and personalized...

• Meta's replay buffer paper hailed for tackling on-policy RL inefficiencies
• PPO/GRPO methods generate rollouts, use for one gradient update, then discard
• Critics call this crazy—persistent buffers could revolutionize training efficiency

Military-industry push scales AI infra: DAF seeks partners to develop AI data centers at three Alaska sites—Joint Base Elmendorf-Richardson, Eielson...