AI Breakthrough Brief

Shadow AI puts sensitive data and decisions outside company controls, with 43% of firms still lacking formal risk frameworks. This gap leaves multiple...

This week's top papers spotlight several fast-moving areas in agent capabilities:

• Harness and interface advances with JevAgentHarness,...

Cities adopting AI for efficiency must set practical staff rules around emerging risks.

• Address data privacy, accuracy, and accountability in...

AI writing suggestions apply mechanical rules like conciseness and avoiding run-ons, but ignore pacing, emotion, and non-linear connections that help...

KAIST's Humanoid v0.9 marks a clear pivot from v0.7's running and Michael Jackson-style dance routines to practical heavy work, now pushing 35 kg...

Approaches to AI moral judgment show progress on narrow metrics but persistent gaps in real robustness.

• DeepMind's dilemma training uses RL on...

Generative AI reflects training data biases and can produce skewed, offensive outputs, requiring users to fact-check everything and avoid fabricated...

A GAO report reveals clear gaps between the Pentagon's responsible AI guidelines and actual project implementation.

• Six of 10 officials flagged...

GM's new Central Compute Unit merges perception, safety, propulsion, and chassis functions into one processor, delivering 40x more power and 30% less...

Britain's AI Biosecurity Lab combines £100 million in government funding with up to £200 million in model credits from OpenAI and Anthropic to defend...

DraMaze's multi-agent setup with Backward Narrative Dependency Tree lets agents coordinate around future plot goals, cutting causal-pattern repetition...

AI ethics only protects businesses when written into data access rules, model reviews, and agent permissions.

• Fairness and privacy map to...

Brown researchers built an event camera paired with a spiking neural network that tracks moving objects in fog by sensing only light changes, ignoring...

Evaluating interactive shopping agents demands simulators that faithfully capture users' behavioral responses across long sequences of...

Scaling enables attention-free LLMs to succeed, but architecture matters more under time and compute limits because scaling still carries a cost, even as that cost drops.

Audio-visual AI models produce unsupported answers because grounding evidence becomes entangled with learned priors inside fused representations....

Florida A&M University is creating internal AI guidelines using a stoplight model while the state Board of Governors considers broader regulations.

-...

Could shared decision paths reshape how we represent and optimize network weights?

• TWIG turns each weight into a path of discrete choices (sign,...

Montgomery County's unanimous passage of Bill 8-26 directly targets algorithmic tools that let landlords coordinate rental prices and lease terms...

Vision-language-action robots cannot reliably stop or recover until they recognize task completion and failure. Echo embeds stop and failure signals...