AI Breakthrough Tracker

Strategic pivot: Meta, with its massive in-house fleet, inks multibillion-dollar deal for AWS Graviton chips to run AI workloads, securing tens of...

• Mythos capabilities: Massive intelligence leap with autonomous cybersecurity skills, acting as a zero-day engine for finding/chaining...

• Efficiency breakthrough: AgenticQwen-30B-A3B (3B active params) scores 50.2 on TAU-2 + BFCL-V4 multi-turn, matching Qwen3-235B; 8B version hits...

Rapid timeline for DeepSeek-V4's efficient 1M-token MoE breakthrough:

• Preview drop: Unveils V4-Pro (1.6T params) and V4-Flash (284B params) with...

• Huge Deal: SpaceX gains right to acquire AI code editor Cursor for $60B or opt for $10B deep partnership.
• Rapid Rise: 2022 MIT-founded tool hit...

Enterprise breakthrough from Cognizant's AI Lab: Four papers introduce evolution strategies for fine-tuning LLMs to tackle complex reasoning with...

OpenAI dropped a new benchmark dataset on the hub to make ChatGPT better for clinicians. This targets clinical optimization head-on.

LLM agent safety: Framework promise vs harsh reality

• DAVinCI framework enables dual attribution and verification for LM claim inference
• Real...

• Broadcom's boom: AI semis hit $8.4B in Q1 2026 (+106% YoY), Q2 at $10.7B (+140%); $100B+ revenue visibility for 2027 via custom XPUs for 6...

WorldMark is a unified benchmark suite for interactive video world models, advancing eval standards in world model research.

Trend alert: AI agents are slashing dev timelines in ML competitions and kernel maintenance.

• NVIDIA's win: LLM agents (GPT-5.4, Gemini 3.1, Claude...

Anthropic's Claude Design transforms visual creation through conversation:

• Builds designs, prototypes, slides, one-pagers from prompts, images,...

Paradigm shift: AI evolves from Q&A chatbots to fully autonomous labs generating hypotheses, running experiments, and iterating without humans.

-...

A thorough guide details how deep learning algorithms, hardware, libraries, compilers, and more have become more efficient. Essential reading for tracking DL progress across the full stack.

New research introduces Co-Evolving LLM Decision and Skill Bank Agents designed for long-horizon tasks. Join the discussion on this paper.

OpenAI's GPT-5.5 delivers sharp gains in math, coding, and agent benchmarks, outperforming GPT-5 and GPT-4o on Terminal-Bench 2.0 (multi-step CLI...

Talent exodus from Meta supercharges TML's perception systems: Weiyao Wang joins after 8 years building multimodal perception and SAM3D.

TML raids...

Key hurdles blocking AI's business rollout:

• Low scaling: Only 18% of organizations have scaled AI enterprise-wide, the lowest on record.
  -...

Deep learning theory is solidifying, with info-theoretic foundations explaining NN efficiency and waste in brute-force scaling.

• Traditional LM...

Game-changing infra push: Broadcom and Meta co-developing industry's first 2nm AI compute accelerators on XPU platform, optimizing for next-gen models...