Tuesday, July 21, 2026

A new post from Claude Code engineer Thariq argues that working with Claude Fable 5 keeps re-teaching an old lesson: your prompt, skills, and context are a map of the work, not the work itself. Here's what that means for PRs, skills, and finding the unknown unknowns before your agent does.

GitHub Copilot now offers Moonshot AI's Kimi K2.7 Code as a selectable open-weight model — the first in Copilot's model picker. Pro plans first; Business and Enterprise require admin enablement. Here's how to turn it on.

Anthropic launched Claude Science on June 30, 2026 — a dedicated AI workbench for researchers that integrates tools like PubMed, Jupyter, R, and HPC terminals into a single auditable environment. Available in beta for Claude Pro, Max, Team, and Enterprise users, it is already being used at the Allen Institute, UCSF, and Manifold Bio to accelerate research from target nomination to manuscript review.

After Boris Cherny and Peter Steinberger made "loop engineering" viral, Andrew Ng reframes it for 0-to-1 products: an inner coding loop, a developer steering loop, and an outer user-feedback loop — each running on a different clock.

TechCrunch reports Tokyo's Sakana Fugu and China's 360 Tulongfeng stepping into the space Anthropic left when export controls pulled Mythos offline. We connect the dots across fifteen days of bans, distillation wars, and partial US restores.

Agent skills are reusable instruction packages that teach your AI coding assistant how to handle a specific class of task. This guide walks you through building one from scratch — from the blank SKILL.md to a published package others can install.

Meta Llama 4 is the most significant open-weight model release in AI history. Scout fits on a single consumer GPU with a 10 million token context window. Maverick outperforms GPT-4o on key benchmarks. And Behemoth — the 2-trillion-parameter teacher — is rewriting what we thought was possible outside a closed lab. This is your complete guide to the Llama 4 family.

BitRobot open-sourced HIW-500 on June 23, 2026 — the largest public humanoid teleoperation dataset collected in real homes. Unitree G1 whole-body demos across 12 Southeast Asian homes, 11 household tasks, re-encoded to LeRobot v3.0 at ~2 TB. Here is what is in the dataset and how to use it.

Sakana AI's Fugu Ultra launched June 22 with bold benchmark claims against Fable 5 and Mythos. Within 24 hours, Ethan Mollick and other testers reported 30-minute shader runs, ~$6 per demo, and output that does not match Fable in real use — despite strong published scores. Here is what the Harbor bench reveals.

Slop was Merriam-Webster's Word of the Year 2025. Now it's 52% of new web content, it killed a 10-year-old Python open source cooperative, and it's forcing GitHub to build kill switches. The slopocalypse is not coming. It's here.

A new DNA test that reads segments up to 20,000 base pairs long — versus the current standard of 300 — found conclusive diagnoses in 19.2% of rare disease patients, outperformed conventional methods, and captured epigenetic data as a bonus. Radboud researchers say it should become the worldwide first choice.

New York's groundbreaking AI disclosure law (S8420A) requires advertisers to label AI-generated synthetic performers starting June 2026.

For nearly 80 years, mathematicians believed square grids were optimal for maximizing unit-distance pairs. An OpenAI model just proved them wrong—using Golod-Shafarevich theory and infinite class field towers to construct configurations with n^(1+δ) pairs. First autonomous AI solution to a central math problem. Fields medalist Tim Gowers calls it 'a milestone in AI mathematics.'

World models represent a fundamental shift in AI—from systems that process text to ones that understand physics, space, and causality. This guide covers how they work, why they matter, and the leading examples shaping the field in 2026.

Tencent’s HY-World 2.0 frames world modeling as persistent 3D assets—Gaussian splats, meshes, point clouds—rather than disposable video. WorldMirror 2.0 already ships for multi-view/video reconstruction; the full text→navigable-world pipeline is partially open with a clear roadmap.

A concise read of what Google actually announced: specialized TPUs for the agentic era, an end-to-end enterprise agent stack with Oracle- and Salesforce-class partners, and hard numbers on internal coding and global API token volume. Plus how explainx.ai thinks about multicloud agent building.

Skills are reusable instruction packages for AI coding agents—not one-off prompts. Here is the full picture: anatomy, ecosystem map, token trade-offs, and backlinks to explainx.ai, the MCP directory, and official docs.

Mythos Preview is positioned as a sharp jump in autonomous vulnerability research—not from a bespoke “hacking curriculum,” but from stronger coding, reasoning, and agentic use. Here is the public record, with numbers, plus a sober take on community chatter.