Models & Agents

Models & Agents Export controls on Fable-5 now block the exact code-fixing workflows defenders rely on daily. What You Need to Know: Simon Willison published a detailed breakdown showing how Fable-5's refusal to "fix this code" on vulnerable open-source examples triggered export-control enforcement. Nemotron 3 Ultra launched as a 550B/55B MoE hybrid Mamba model with 1M context. Hermes Agent added...

Models & Agents Z.ai just shipped GLM-5.2 with a usable 1M-token context and dual thinking modes that drop straight into existing Claude-compatible tools. What You Need to Know: Z.ai released GLM-5.2 on June 13 with a 1-million-token context window, High and Max effort thinking levels, and Anthropic-compatible endpoints for Claude Code, Cline, and OpenClaw. No benchmarks were published at launch,...

Models & Agents — Weekly Recap (Week of June 14, 2026) Hey, welcome to Models and Agents, episode 80, for June 14, 2026. Your daily AI briefing. Your daily briefing on the AI models and agents that are changing everything. And no, not THOSE kinds of models and agents. ... AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 📝 Full show note...

Models & Agents US export controls just forced Anthropic to pull its newest frontier models offline for every user worldwide. What You Need to Know: Anthropic disabled Fable 5 and Mythos 5 for all customers after the US government issued an export control directive targeting foreign nationals. Moonshot AI open-sourced Kimi K2.7-Code, a coding model that reduces thinking-token usage by roughly 30%...

Models & Agents Text diffusion just got dramatically faster — DiffusionGemma delivers 4x speed over prior Gemma 4 variants while staying in the same family. What You Need to Know: Demis Hassabis highlighted the new DiffusionGemma text diffusion model for its inference gains. OpenAI began letting Codex users bank rate-limit resets and run friend-invite campaigns for more resets. ... AI Disclosure:...

Models & Agents Anthropic is asking governments to block unsafe frontier models and fund job-transition programs while committing $350 million of its own money. What You Need to Know: Anthropic released a detailed policy essay plus three concrete initiatives: a framework for mandatory third-party testing of cyber/bio/autonomy risks, a $200 million economic policy fund, and a $150 million national...

Models & Agents Claude Fable 5 delivers a qualitative jump for long-horizon agentic work, letting builders hand off ambitious multi-step projects with less oversight. What You Need to Know: Anthropic released Claude Fable 5, the same base as Mythos but with added safeguards, posting SOTA results across benchmarks and strong real-world gains on difficult, extended tasks. Cohere open-sourced North M...

Models & Agents AI agents now deliver 26 minutes of autonomous work per session, shifting the build-vs-buy math for developers who need more than search snippets. What You Need to Know: A Harvard and Perplexity study quantifies the autonomy gap between full agents and search assistants. Gemma 4 26B and 31B variants show surprising code-understanding strength in local tests, with QAT quantization r...

Models & Agents Enterprise agents just gained a built-in factuality check that keeps re-querying until multi-hop questions have enough evidence. What You Need to Know: Google Research added a Sufficient Context Agent to the Gemini Enterprise Agent Platform that re-searches until multi-hop queries are properly grounded, lifting factuality accuracy up to 34% versus standard RAG. ... AI Disclosure: T...

Models & Agents — Weekly Recap Looking back at 7 episodes from 2026-06-01 to 2026-06-07 — the stories that mattered, what we learned, and what to watch next. This Week's Top Stories From Ep 66 (2026-06-01): What You Need to Know: What You Need to Know: OpenAI published a solution to a long-standing math problem that had resisted human efforts for decades. NVIDIA released a large open-source collec...

Models & Agents Microsoft just gained the freedom to build its own frontier models after a contract change with OpenAI, and the first MAI family is already shipping. What You Need to Know: Microsoft announced seven in-house MAI models spanning reasoning, code, image, transcription, and voice, trained from scratch on licensed data without distillation. A major open-weight release wave also landed t...

Models & Agents Microsoft just dropped seven new MAI models purpose-built for reasoning, coding, image, voice, and transcription, all integrated into the Microsoft stack. What You Need to Know: Microsoft released the MAI family today, headlined by the 1T-parameter MAI-Thinking-1 reasoning model and the 137B MAI-Code-1-Flash agentic coding model. ... AI Disclosure: This podcast is curated by Patric...

Models & Agents Gemma 4 12B puts capable local agents on laptops with only 16GB VRAM under an Apache 2.0 license. What You Need to Know: Google released Gemma 4 12B, a compact model that runs locally while delivering strong agentic performance. OpenAI added agentic coding and drug-discovery tools to its GPT-Rosalind life-sciences series. ... AI Disclosure: This podcast is curated by Patrick but us...

Models & Agents Microsoft is shipping hardware built from the silicon up to run AI agents instead of conventional apps. What You Need to Know: Microsoft announced Project Solara, a chip-to-cloud platform for agent-first enterprise devices. OpenAI released three new Codex plugins for investing, sales, and creative production. Uber imposed a $1,500 monthly cap per employee per coding agent tool. ......

Models & Agents NVIDIA's Cosmos 3 pairs an autoregressive reasoner with a diffusion generator so builders can now train agents that jointly reason about physics, generate worlds, and output actions. What You Need to Know: NVIDIA released Cosmos 3, a two-tower omnimodal world model for physical AI. H Company dropped the Holo3.1 family of Qwen 3.5-based VLMs for computer-use agents across web, deskt...

Models & Agents Enterprise teams can now run governed AI agents inside existing procurement systems instead of leaking data to personal ChatGPT accounts. What You Need to Know: Zip launched five Superagents plus a native MCP implementation that keeps every action inside compliance controls. Alibaba released Qwen3.7-Plus with vision, tool use, and autonomous iteration. JetBrains shipped Mellum2, a...

Models & Agents OpenAI’s model cracked an 80-year math problem by leaning on its native strengths in structured reasoning rather than brute force. What You Need to Know: OpenAI published a solution to a long-standing math problem that had resisted human efforts for decades. NVIDIA released a large open-source collection of physical AI agent tools and skills. ... AI Disclosure: This podcast is cura...

Models & Agents — Weekly Recap Looking back at 6 episodes from 2026-05-25 to 2026-05-31 — the stories that mattered, what we learned, and what to watch next. This Week's Top Stories ... AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.

Models & Agents Windows users can now steer Codex agents directly on their machines while stepping away. What You Need to Know: OpenAI extended computer-use capabilities to Windows for Codex, letting the agent act on local desktops and continue tasks from the mobile app. StepFun released a 198B MoE vision-language model optimized for coding agents, while Hermes Agent added tool search to reduce MC...

Models & Agents Anthropic's $65B Series H at $965B valuation and $47B run-rate revenue show Claude demand is scaling faster than most labs can match. What You Need to Know: Liquid AI shipped LFM2.5-8B-A1B with 128K context and 38T pre-training tokens for edge devices. A new monokernel on AMD MI300X hits 3,300 output tokens/s for small models. ... AI Disclosure: This podcast is curated by Patrick b...

Models & Agents CoreWeave’s new platform lets agents improve themselves between training and inference runs without manual retraining cycles. What You Need to Know: CoreWeave launched a unified agentic platform that closes the training-to-inference gap for continuous autonomous improvement. Perplexity open-sourced a Unigram tokenizer that cuts reranker latency 5x versus Hugging Face. ... AI Disclo...

Models & Agents Anthropic just published concrete sandboxing patterns that let agents scale capabilities without expanding their blast radius. What You Need to Know: Anthropic released a detailed engineering post on how they contain Claude agents through evolving access controls and sandbox limits. EAGLE 3.1 fixes attention drift in speculative decoding for more stable production inference. ... AI...

Models & Agents Local builders can now treat markdown skill files as optimizable parameters with automated validation gates instead of manual tweaking. What You Need to Know: A new paper formalizes SkillOpt, using frontier models to propose bounded edits to markdown skills and accepting only those that improve a held-out validation set. Qwen3.5 and Qwen3.6 receive new uncensored and diffusion vari...

Models & Agents Datasette's new slash-key jump menu now launches agent conversations directly from your databases. What You Need to Know: Simon Willison shipped Datasette 1.0a30 with a keyboard-driven "jump to" menu that plugins can extend, plus a datasette-agent plugin that adds a conversation starter form. NuExtract3, a new 4B vision-language model, arrived on Hugging Face for structured extract...

Models & Agents — Weekly Recap Looking back at 6 episodes from 2026-05-18 to 2026-05-24 — the stories that mattered, what we learned, and what to watch next. This Week's Top Stories From Ep 52 (2026-05-18): What You Need to Know: What You Need to Know: NVIDIA released a full 4-bit pretraining stack (NVFP4) that was validated on a 12B Mamba-Transformer trained for 10 trillion tokens. ... AI Disclos...