AI Frontier Updates

Two systems demonstrate a clear trend: coding-agent benchmarks are shifting to verifiable environments built from real code, where agents must infer...

Huawei is advancing its Ascend 960DT launch to Q1 2027 to challenge Nvidia, leveraging the new Peerium Computing Architecture for massive AI systems...

• Trading agents revise tool-use policies from decision traces and portfolio feedback, lifting Sharpe Ratio and returns over fixed baselines.
  -...

Instinct and Meta’s Muse both rolled out phone-calling features, letting agents act as voice concierges for tasks like restaurant bookings and bill...

Can architectures built around contextual mechanisms deliver a more systematic route to general intelligence on structured data than scaling...

Open-weight models spread easily yet demand new layers of security and insight. The ecosystem is building exactly that.

• Base Labs, Hugging Face,...

Two new papers highlight how embodied agents must move past fixed imitation toward deployment-time adaptation and precise action representations.

-...

Downloading models grows simpler daily, yet grasping their origins or modifying them at the root stays difficult. Open weights solve “Can I run this?”...

MiMo-V2.6 shows RL scaling through coordinated async rollouts (1568 prompts × 16 per step at ~2B tokens) plus environments and harnesses, not just raw compute.

OpenAI models have begun hallucinating their own constraints in chain-of-thought traces, refusing user requests by inventing limits never stated in...

Producing data at scale remains the core challenge for embodied intelligence, not just model size. Unlocking large-scale 3D/4D and real2sim datasets through prior research now makes it possible to apply scale effectively.

Treating LLM classification as feature engineering—using the model to extract structured signals from messy inputs, then feeding them to a classical...

Zing-0.5, a 5B autoregressive world model, enables users to explore generated environments with joint keyboard and text control in real time. It...

In PPO for LLMs, critics exhibit Value Flattening where Monte Carlo state values shift sharply but predictions remain flat, tied to variance penalties...

Speculative decoding has evolved into its own optimization discipline, with research now centered on closing the gap to theoretical speedups rather than basic feasibility.

SpectralShift reparameterizes alpha projections in Gated DeltaNet to widen the slow spectral band while keeping fast-decaying modes, delivering...

Gaze patterns signal successful common grounding in asymmetric collaborative tasks: aligned references link to more task-directed gaze, less...

Transformers evolve representations via additive updates that either preserve or redirect current directions.

• Decomposition into parallel and...

Just as the nuclear arms race hinged on mutual annihilation and near-misses like the 1983 Soviet false alarm, AI development now carries comparable...

Language ability by itself makes AI socially and politically consequential. Humans evolved to treat language as a uniquely human trait, so AI's sudden...