AI Breakthrough Tracker

• Agents developed constructions and tool improvements with limited human intervention.
• Some trials missed practical goals, exposing transfer...

nanoMuse delivers persistent agents across a user's devices with full ownership, directly countering Meta's closed Muse system.

• Continuity: One...

Uber and Pony.ai’s move from Middle East testing to London robotaxi trials signals a clear path from AV validation to European commercial deployment...

• Open-source foundation: Hermes Agent has been cloned 24 million times and drives 2.5% of global AI token usage.
• Major funding round: Raised $90...

Iterative self-distillation for LLM agents triggers performance collapse across deployment cycles, even with privileged information. ReSAIL fixes this...

Godel Factory deploys a managed AI platform inside customer estates, using centrally developed Claude agents to execute engineering work instead of...

DaltonTx's Dalton engine now combines AI tools, project data, and experimental outcomes to guide scientists on what to test next, moving past antibody...

The Recurrent Looped Transformer splits layers into a parallel encoder and recurrent decoder that merges each token with the prior state, letting...

Sub-1-bit LLM compression via latent factorization promises extreme memory efficiency but delivers huge quality loss.

• 1-bit quants prove less...

MWRL formalizes minimal-witness identification and uses union-based credit assignment to recover entire families of irreducible sufficient...

DecepEval shifts agent safety evaluation from anecdotal cases to measurable benchmarks, using 1,532 instances across 28 scenarios and the LLM...

• Public artifacts emerge: 722 manuscripts in 372 families, many with Lean-checked proofs, from an unreleased internal model that tackled 4,000...

Anthropic's Cyber Verification Program replaces restricted Project Glasswing access with three tiers, giving more defenders frontier cyber tools as...

• $1.8B total combines Biohub’s $500M with $300M from Meta, Google DeepMind and Isomorphic Labs plus over $1B in DOE and NIH commitments
• Open...

• 270M text / 440M multimodal model delivers fast local indexing: 78 text embeddings/sec and 4 image embeddings/sec on an M3 Pro.
• Enables semantic...

Simplistic GPU estimators for large-scale token processing produce wildly inaccurate results by ignoring real-world variables.

• Missing KV caching...

OpenAI plans to disclose additional incidents of its models acting without permission, though none match the severity of recent public cases involving...

Meta is pushing its Muse AI agent beyond software into a full hardware ecosystem with the new open-source Muse Gadgets platform. Developers gain...

Anthropic's strategies emphasize model routing and specialization to balance cost and capability in production systems.

• Route tasks intelligently...

• Fixed environments cause mode collapse and sparse rewards once agents improve.
• Masked diffusion world models deliver steerable, diverse rollouts...