AI Research Pulse

The new Health & AI Policy Index (HAPI) maps 240 policies from 2016-2025, confirming oversight is accelerating yet remains fragmented across...

Agent Teams and Workflows are new, powerful, yet token-hungry, with AI now often deciding orchestration and mixing approaches itself. This raises real...

China is closing the gap with the US through talent acquisition, benchmark wins, and cost advantages.

• Tencent hired former OpenAI researcher Yao...

• Trump EO creates voluntary 30-day NSA-led cybersecurity reviews for frontier models, avoiding mandatory rules
• OpenAI demands mandatory civilian...

• NSF renews IAIFI support for five years, raising annual funding from $4M to $4.98M
• Bidirectional research accelerates physics discoveries while...

Anthropic’s Daniela Amodei argues trust and safety function as accelerants, not brakes, for enterprise AI adoption—customers explicitly want...

• pEVAC-PS, first fully AI-designed vaccine, safely passed Phase I human trials and triggered cross-reactive immune responses to multiple...

XL-SafetyBench reveals why translated English prompts fall short for real-world AI risks.

• Existing limits: Standard benchmarks miss legal,...

OpenAI's general-purpose model disproved the 80-year Erdős unit distance conjecture by linking it to algebraic number theory, with mathematicians like...

• Unified transformer backbone ingests raw audio waveforms and visual patches directly, eliminating separate encoders to cut memory use and latency.
  -...

The AI frontier is shifting from raw model scale to targeted improvements in data, prompts, and efficiency.

• Better data closes performance gaps...

Anthropic, valued near $1 trillion, urges labs to form a verifiable global agreement to slow or pause frontier AI development, citing risks from rapid...

AI threats like Anthropic’s Mythos are accelerating enterprise demand, with Palo Alto reporting 1,000+ reassessment inquiries and launching Frontier...

Video2LoRA generates LoRA adapters in a single forward pass via a perceiver hypernetwork, letting frozen VLMs answer video queries with zero visual...

MIT's ChartNet dataset of over a million synthetic charts with code, tables, and Q&A pairs trains open-source VLMs that outperform much larger...

A practical guide to staying ahead on AI developments, drawn from years of curated subscriptions by a macro trader.

• Beth Kindig’s IO Fund: Delivers...

Anthropic is warning that Claude is already writing most of its code, reviewing changes, running experiments, and accelerating the next generation of...

Leading labs are strategically hiring philosophers and ethicists to address AI consciousness questions beyond engineering.

• Governance prep:...

Andrej Karpathy joined Anthropic on 19 May 2026 to lead pre-training research with Claude, reporting to Nick Joseph. The move ends his 22-month run at...

New techniques are closing the sim-to-real gap and enabling real-time robotics.

• Discrete-WAM introduces unified discrete vision-action tokens for...