AI Milestones

Berkeley researchers demonstrate systematic ways to break top AI agent benchmarks, highlighting fundamental evaluation methodology issues.

Open-source visual web agent with transparent training data and methodology for autonomous web navigation tasks.

Anthropic introduces ClawBench, a comprehensive evaluation framework testing AI agents on 153 everyday online tasks across 144 live platforms.

Research breakthrough addressing agents' meta-cognitive deficits in arbitrating between internal knowledge and external tool usage.

Meta introduces Muse Spark, positioning it as a step toward personal superintelligence capabilities for individual users.

Research breakthrough allows full-precision training of 100+ billion parameter language models on a single GPU, dramatically reducing training costs.

Anthropic releases specialized Claude model variant focused on advanced cybersecurity capabilities with detailed system card documentation.

Google releases Gemma-4 series with any-to-any and image-text-to-text capabilities across multiple parameter sizes (4B-31B).

Claude successfully wrote a complete FreeBSD remote kernel RCE exploit with root shell, demonstrating advanced cybersecurity capabilities.

Original Alibaba Qwen technical lead publishes influential essay on transitioning from reasoning to agentic thinking paradigms.

New benchmark designed to measure artificial general intelligence through novel reasoning tasks, addressing limitations of previous AI evaluation methods.

First multimodal framework combining video, tactile sensing, and action prediction for contact-rich physical interactions.

OpenAI introduces framework to accelerate multimodal agent reasoning through speculative perception and planning.

First AI system confirmed to solve an open mathematical research problem, marking breakthrough in AI mathematical reasoning capabilities.

First demonstration of a 400 billion parameter language model running natively on a mobile device, showcasing dramatic advances in on-device AI.

Researchers discover discrete 3-4 layer 'reasoning circuits' in transformers that can be duplicated to dramatically improve logical deduction performance without training.

Research introduces framework enabling language models to continuously improve from real-world deployment experience rather than offline training only.

Nvidia introduces purpose-built CPU architecture specifically designed for agentic AI workloads, marking hardware specialization for autonomous agents.

Investment bank warns of imminent AI breakthrough driven by rapid computing expansion that could strain power grids and disrupt jobs globally.

Legendary programmer John Carmack publicly disputes OpenAI and other labs' aggressive AGI timelines, stating 'We Are Not on the Brink of AGI' with significant implications for industry investment.

Anthropic's Claude models now support 1 million token context windows in general availability, enabling processing of extremely long documents.

First desktop agent that learns tasks from single demonstrations across GUI apps, browsers, terminals, and messaging tools in unified sessions.

Nvidia announces major strategic shift with $26 billion investment in open-source AI models over five years, competing directly with OpenAI and other closed-source providers.

Research reveals how data correlations determine feature geometry in neural networks, extending beyond sparse uncorrelated settings.

Open-source inference engine achieves faster performance than llama.cpp, MLX, and Ollama on Apple Silicon using custom Metal shaders.

Research demonstrates that chain-of-thought reasoning substantially expands LLMs' ability to recall factual knowledge from parameters.

LLM trained on Python execution traces can predict line-by-line execution and function as a neural interpreter with debugging capabilities.

AI-powered age verification systems now achieve 1-2 year accuracy in determining user ages, enabling widespread implementation of child safety laws across multiple jurisdictions.

DNA foundation model trained on 100,000+ species can identify genetic patterns across entire tree of life, published in Nature.

First trainable INT8 attention system that quantizes six of seven attention operations while preserving training performance.

Research shows GPT-5, Claude-4.5, and Qwen-3 can execute rare strategic actions while maintaining calibration, raising safety concerns.

China's AI model usage reached 4.12 trillion tokens vs US 2.94 trillion tokens in one week, marking historic shift.

Shanghai hospital launches world's first traceable AI agent system for rare disease diagnosis, published in Nature.

Department of Defense designates Anthropic as supply-chain risk amid clash over military AI partnerships, marking escalation in AI governance conflicts.

OpenAI secures record-breaking $110B funding round with major investors including SoftBank, Nvidia, and Amazon, highlighting massive AI investment scale.

Anthropic abandons a major safety commitment, marking a significant shift in AI safety policy approach from one of the leading safety-focused AI companies.

Anthropic alleges 16 Chinese AI entities systematically distilled Claude through API harvesting, raising IP protection concerns.

Google's Aletheia agent powered by Gemini 3 Deep Think autonomously solved 6 out of 10 problems in the inaugural FirstProof mathematics challenge, demonstrating advanced mathematical reasoning capabilities.

Novel architecture enables running Llama 3.1 70B on single RTX 3090 by bypassing CPU/RAM bottlenecks.

Tsinghua team develops AI model that extends James Webb Space Telescope detection depth by 1 magnitude, discovering 3x more distant galaxies.

GLM-5 introduces a paradigm shift from vibe coding to agentic engineering with new DSA architecture and asynchronous RL infrastructure.

Anthropic releases Claude Sonnet 4.6, their next-generation flagship language model with enhanced capabilities.

GPT-5.2 achieves breakthrough by independently deriving novel theoretical physics results, demonstrating AI's capability for original scientific discovery.

OpenAI releases GPT-5.3-Codex-Spark, a specialized model for advanced code generation and programming tasks.

Google releases Gemini 3 Deep Think, advancing reasoning capabilities in multimodal AI systems.

Anthropic achieves massive funding round establishing it as one of the most valuable AI companies globally.

GPT-5 demonstrates superior performance to human federal judges in legal reasoning tasks, marking a significant breakthrough in AI's ability to handle complex legal analysis.