AI Breakthroughs Tracker

A new arXiv survey organizes human-view video understanding with MLLMs around three core abilities—watching, remembering, and reasoning—while covering egocentric applications alongside challenges in perception, memory, and faithful inference.

• Core issue: LLM text embeddings align with frequent uninformative tokens, suppressing nuanced semantics.
• Key insight: The unembedding matrix...

MMAE launches as the first comprehensive benchmark for instruction-based audio editing, spanning 7 modalities and 6 complexity levels with 2,000...

American AI startups have made a striking shift toward Chinese models since the start of the year. This signals changing preferences in model selection.

• Google upgrades NotebookLM with Gemini 3.5, agentic chat, code execution, and grounded research tools
• RAG systems still hallucinate and breach...

Sakana AI has launched RSI Lab, a dedicated research group focused on recursive self-improvement to create AI that builds better AI. This step underscores RSI's dual role in accelerating capabilities and advancing safety research.

• Claude Sonnet 4.5 flagged evaluation awareness in 14-16% of audits, up sharply from prior versions
• Mythos turned disclosed patches into working...

Anthropic notes that bio databases function like pre-automobile cities, built for human navigation rather than AI agents. This structural mismatch...

Researchers tracking cutting-edge computer vision can explore all 15 CVPR 2026 paper finalists in one place, complete with GitHub links and Hugging Face resources.

• June 2026 executive order accelerates US AI via voluntary national-security testing for labs like Google DeepMind, Microsoft and xAI, plus a...

• LeCun's AMI Labs raised $1.03B to develop JEPA world models that predict abstract representations instead of pixels or tokens, arguing LLMs cannot...

AGIBOT's World Challenge 2026 introduces dual tracks to evaluate embodied AI on real-world tasks, moving beyond execution to full reasoning and...

Anthropic's proposed global freeze on frontier AI faces major enforcement hurdles, needing coordinated lab compliance across countries and...

OpenAI's general-purpose reasoning model disproved Erdős' 80-year-old unit distance conjecture by generating a verified counterexample fusing algebra...

AI is speeding scientific discovery from antigen selection to open publishing:

• Vaccine breakthrough: ML models trained on coronavirus structure and...

Google DeepMind formed via the 2023 merger of DeepMind and Google Brain, building on DeepMind's 2010 founding by Demis Hassabis, Mustafa Suleyman, and...

Three starkly different stances on AI risk and timelines:

• Alignment dismissed: Technical alignment debates distract from urgent socio-political...

AI is simultaneously expanding attack surfaces and strengthening defenses, driving record demand for platforms like Palo Alto's.

• New threats prompt...

The push for scalable AI hardware is intensifying worldwide:

• Photonic networks debut at scale in the UK via Oriole-AMD collaboration, slashing power...

• Alibaba pivots to closed Qwen3.7-Plus, a proprietary multimodal model with 1M context and agentic state preservation at $0.4 per million tokens
  -...