## News / Update
The past week saw a flurry of industry moves and milestones. TIME’s Best Inventions list spotlighted breakthroughs across chips, supercomputers, and models, with Alphabet’s Genie 3 featured for its world-building capability. Robotics accelerated: Figure unveiled its 03 humanoid while SoftBank acquired ABB’s robotics unit, and Fei-Fei Li launched an open, standards-driven robotics challenge. Enterprise search and data players consolidated as Elastic acquired Jina AI, and Weaviate partnered with Confluent to fuse real-time streaming with vector databases. Google Gemini crossed 1 billion monthly visits, and the State of AI Report 2025 arrived with a comprehensive view of research, safety, and policy. Funding momentum remained strong: Reflection raised $2B to build open-weight frontier models, legal AI startup Spellbook closed a $50M Series B, and investors backed science-driven RL startups pushing AI into core scientific and systems problems. China introduced new rare-earth export controls, underscoring supply chain tensions around AI hardware. Operational scale also grew, with TNG’s Chimera family processing over 10B tokens per day on OpenRouter.
## New Tools
New platforms and utilities are making agentic AI more accessible. Google launched Gemini Enterprise for rapidly building custom workplace agents, while TARS debuted as a no-code agent platform. Developers gained new infrastructure with Mem0 for long-lived agent memory, TextQL Healthcare for petabyte-scale SQL analysis, and a FastMCP cloud service for one-click MCP server deployment. Yupp AI introduced visual prompting via SVG code, and SoraMarker simplified watermarking for AI video. Hugging Face-enabled browser runtimes now let you load converted checkpoints fully offline in JavaScript/Spaces, and LocallyAIApp brought a LiquidAI MoE model (LFM 2 8B A1B) to iPhone for fast on-device inference.
## LLMs
Frontier and compact models alike set new bars. GPT-5 Pro took the top verified score on ARC-AGI’s semi-private benchmark, and Gemini 2.5 Deep Think posted a record on FrontierMath under manual evaluation. Claude Sonnet 4.5 extended sustained task execution to nearly two hours and showed strong end-to-end coding capability. AI21’s Jamba Reasoning 3B led tiny-model instruction-following, while a 7M-parameter Tiny Recursion Model surpassed most large models on recursive reasoning, signaling a surge in efficient reasoning systems. New and specialized models arrived: Radical Numerics open-sourced RND1 (a 30B sparse-MoE diffusion language model), Microsoft introduced UserLM-8B to simulate human users, Meta released a Code World Model aimed at deeper code semantics, Qwen3-Omni pushed toward human-like multilingual responsiveness, SmolLM2 advanced small LLMs, and ByteDance’s Artificial Hippocampus Networks provided long-context memory compression. Performance and competition intensified: Qwen3-30B hit 473 tokens/sec on M3 Ultra; OpenAI Codex surpassed Claude Code on several coding benchmarks; coding agent leaderboards reshuffled post GPT-5-Codex; and some users reported GPT-5 lagging rivals on specific data tasks. Research advanced core capabilities: latent diffusion and GLASS Flows rethought diffusion reasoning and efficiency, first-token steering plus Exploratory Annealed Decoding improved trajectory control, MS-SSM scaled multi-resolution sequence learning, attention-sink/compression-valley links clarified internals, and RL-based methods (LoRA competitive with full-parameter RL, RLAD’s hint-and-solve setup, and bootstrapped long-horizon reasoning) expanded training strategies. Safety and reliability stayed in focus with inoculation prompting to reduce reward hacking and findings that Sonnet 3.7 could not hide malign backdoors despite adversarial training. LLMs continued to surface factual inconsistencies at web scale, exposing errors across Wikipedia.
## Features
Major products gained meaningful upgrades. OpenAI added function calling and web search tools to the GPT-5 API, aligning it with ChatGPT’s evolving toolset. Hugging Face enhanced its Hub with custom app domains, instant GGUF metadata editing, Xet-backed performance, a universal Responses API, and MCP-UI support. Google Cloud rolled out richer contextual data integrations and agent-building capabilities on Gemini, while Google AI Studio introduced voice-based “yap-to-app” coding. Anthropic released Claude Code plugins to extend coding workflows. In retrieval, Weaviate debuted Query Agent, an agentic RAG layer for reranking, filtering, summarization, and cited answers, and Elysia’s Chunk-On-Demand streamlined storage by chunking only when needed. Interactive video formats are also evolving to embed polls, quizzes, calendars, and feedback directly in playback.
## Tutorials & Guides
Hands-on resources proliferated. DeepMind shared an easy Colab to fine-tune gemma3-270m for emoji generation, paired with a community tutorial covering custom text-to-emoji models, quantization to ~300MB, and private on-device deployment. Weaviate’s podcast with Omar Khattab and Connor Shorten explored DSPy, LLM pipelines, and prompt optimization, while another case study showed DSPy+GEPA cutting API costs 20x by switching to Grok-4-fast. Educational content spanned an LLM history thread, a Netflix ML interview scenario on validating model replacements, and a Stanford lecture on pluralistic alignment. A live session promised tips for training small, sparse models on consumer GPUs.
## Showcases & Demos
Creativity and automation were on display. Genie 3 generated fully playable, open-ended worlds from text or images, drawing broad acclaim as a glimpse of next-gen interactive media. Operational demos showed AI “marketing twins” executing complex SEO workflows end-to-end in minutes. Smart Cellular Bricks turned physical construction into an interactive, AI-aware building experience. Code generation leapt forward as Claude 4.5 Sonnet produced a complete Datasette plugin from a single prompt, and visual prompting experiments via Yupp AI highlighted models’ ability to interpret and create from SVG instructions.
## Discussions & Ideas
Debates centered on rigor, safety, and the direction of AI. Researchers stressed reproducibility—amid a high-profile paper’s irreproducible results—and called for open datasets and evaluation standards in robotics. Safety discourse deepened: OpenAI outlined methods to measure and mitigate political bias; Anthropic’s experiments suggested limits on covert backdoors; and new work showed that a handful of poisoned documents can compromise models, raising practical security concerns. Evaluation reliability came under scrutiny, with warnings that small test sets can skew reasoning benchmarks by large margins. Conceptual shifts emerged around COLMs as a new paradigm, early-token steering of reasoning, and the merits of RL training strategies—alongside Karpathy’s critique that current RL may over-penalize exceptions. Broader currents included the resurgence of small, open labs challenging incumbents, pushback against geopolitical fear narratives, and the role of open communities in surfacing new talent. Forecasts and ambitions added fuel: claims that LLMs could out-predict top human forecasters by 2026, hints that Grok may tackle new mathematical conjectures, and reports of model revenues more than tripling year over year. LLMs’ ability to expose errors in Wikipedia also rekindled discussions on human-AI collaboration for knowledge curation.
