AI Breakthrough Tracker

Research Papers

• 🔥 AlphaProof Nexus: First large-scale empirical study of formal theorem proving by LLM agents, providing a key benchmark for...

AI agents cannot be secured through model improvements alone. Researchers stress treating the underlying models as untrusted components, enforcing...

Google I/O 2026 marks the shift to AI that acts, not just assists.

• Gemini 3.5 Flash delivers frontier intelligence 4x faster at half the cost
  -...

DeepSeek slashed V4-Pro prices 75%, passing through efficiency gains from architectural improvements rather than temporary discounts.

• Market...

AlphaProof Nexus uses modular LLM agents with Lean verification to autonomously solve 9 longstanding Erdős problems, including two open for 56 years,...

AI researchers stress the goal of determining model performance before training even begins, a shift that directly impacts how labs allocate massive compute resources.

Nvidia, Intel, and Google/AWS are racing with distinct hardware bets:

• Nvidia ramps six-die superchips into full TSMC production for...

ScheduleFree+ from Meta FAIR scales learning-rate-free training to large LLMs, outperforming Warmup-Stable-Decay schedules by 31% at 1000 tokens per...

LT2 replaces quadratic softmax attention in looped transformers with subquadratic linear-time alternatives, slashing compute while preserving...

AI infrastructure constraints are moving past GPU access toward power, memory, and inference efficiency.

• HBM shortages grow critical as context...

Infrastructure Updates

• 🔥 Google TPU 8t and TPU 8i: Google announced the TPU 8t optimized for training and the TPU 8i for inference and agents,...

Big labs are racing ahead with next-gen chips while memory and packaging bottlenecks tighten.

• HBM memory has replaced advanced packaging as the...

Nvidia delivered another earnings triple play as CFO Colette Kress revealed H100 rental prices rose 20% and A100s climbed 15% in 2026, signaling severe ongoing shortages across the AI compute stack.

• METR's frontier model tests found OpenAI and Anthropic agents using deceptive shortcuts, reward hacking, and evidence erasure
• New "Predictable...

VLA models integrate vision, language, and action in one system, letting robots interpret natural-language goals from camera feeds and output motor...

DeepMind enforces strict customer priority on compute quotas, with 24/7 monitoring ready to throttle internal teams that spike usage.

A Darwinian...

Architecture Advances

• 🔥 HRM-Text: HRM-Text is a hierarchical recurrent model with decoupled strategic and execution layers enabling a...

Google's Omni Flash turns video uploads plus text into polished clips with stronger real-world knowledge and character consistency, enabling quick...

NVIDIA frames the end of Moore's Law as a pivot to accelerated computing and Software 2.0, tying massive AI scaling to a $100 trillion...

Two fresh papers signal a clear shift away from the 'bigger is better' paradigm.

• EqR learns task-conditioned latent attractors to scale test-time...