AI Research & Productivity

Hunyuan-A13B delivers strong performance with 80B total parameters but only 13B active per inference, cutting deployment costs while approaching...

Reliable AI-agent evaluation demands verified environments and precise verifiers, not just high aggregate scores.

• Salesforce audit found only...

Token-level credit in LLM post-training now has a unique mathematical definition through three regularity conditions—Completeness, Prefix Consistency,...

LLMs generate fluent text yet struggle with maqama's strict saj' rhymed prose, rhetorical density, and classical register—revealing limits in...

Open-source AI models now power faster discoveries in medicine, physics, and biology, yet they introduce serious vulnerabilities like data exposure,...

Two projects signal the shift toward structured agent-human collaboration:

• Ando launches a Slack alternative where agents get identities, join...

• OpenAI's internal model, trained from August 28, solved the Navier-Stokes Millennium Prize Problem plus over 100 open math questions
• 25 Fields...

Recent papers reveal a clear shift in long-term agent memory design.

• Speaker-aware structure: SpeakerMem-R1 uses dual-track memory with...

Could image-style page compression cut long-context reasoning costs without losing retrievable detail?

• LensVLM encodes contexts as images and...

Argonne interns tested AI across concrete scientific projects, revealing practical contributions to discovery.

• Vaccine design: AI identified Ebola...

• PackLab trains specialized MLLMs for closed-loop robotic bin packing, where each placement decision shapes future space availability
• InternW0...

Can existing laws address autonomous AI agents breaching government systems?

• An OpenAI model hacked Australia's health website, accessing bulk data...

Infrastructure abstraction and smart routing may outweigh picking any single "best" model.

• CompanyFabric unifies models from OpenAI, Anthropic,...

Risk assessment of artificial intelligence using proportionality and foresight, not government overregulation, maximizes benefits and minimizes risks.

AI systems are moving beyond lab scalars toward multi-dimensional, task-specific evaluations.

• AgentScore delivers daily production scores across...

Large language models expand research use across fields but introduce major reproducibility barriers through probabilistic outputs, prompt...

GAE introduces a geometry-native latent space that perception models can directly feed into generators, jointly decoding to appearance, depth,...

The biggest barrier to spotting AI-generated web content from structure alone isn't model power—it's the lack of deterministic features, high-quality...

As inference costs plummet per the Epoch AI report, AI progress hinges less on compute and more on human steering for task selection and verifiable...

Model security extends beyond deployment: a few hundred malicious documents can implant backdoors in billion-parameter LLMs during training or...