Martin Fowler’s advice for writers: don’t let an LLM write your first draft.
He argues that “LLM Voice” is difficult to edit out, and readers may lose interest before they engage with your ideas.
His deeper concern is what happens to the writer.
Writing forces you to confront what you don’t understand. You start with a vague idea, struggle to explain it, and discover the gaps while working through the sentences.
An AI draft can let you skip that struggle.
Fowler acknowledges valuable uses, including translation and helping people overcome health-related barriers to writing. He also questions how reliably he can recognize AI prose.
But his concern about outsourcing understanding is worth considering.
How much of your thinking happens before you write and how much happens because you write?
Mistral just released the weights for Le Chonk, its 1T-parameter model.
The number that caught my attention: 47% fewer tokens per coding task.
Mistral reports 48% on the Artificial Analysis Coding Agent Index, roughly tied with GLM-5.3 while using 47% fewer tokens per task on average.
On Terminal-Bench 4.0, it reaches 41%, up 13 percentage points from the preview.
Under the hood: 128 routed experts, with 4 selected per token plus 1 shared expert. Roughly 52B active parameters, a 1M-token context window, and text + image input.
For developers running long coding workflows, that token efficiency is worth testing.
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
A dad built a unicorn game for his 6-year-old daughter. Now he wants it to grow up with her.
She designed most of the maps and playground games. After each build, she played it and told him what to change. Sometimes 25 requests in one round.
He gave her notes to Claude Sonnet 5.5. The next day, she played the result.
In under a month, a small forest where she searched for six unicorns grew into eight locations, magic, mini-games and a herd of friends following her around.
Every sprite, sound effect and music loop was generated from code.
Her dad is an experienced engineer, though this was his first game. Now he plans to add reading, writing and math activities as she grows.
She’s learning how software gets made by watching her own ideas become somewhere she can play.
unicorn-academy-game.github.…
Chums brings the Muse Charm idea to your Apple Watch. ⌚
It turns Claude Code and Codex running on your Mac into little animated characters you can carry on your wrist, keys or backpack.
Hold the screen, say what you want, and let go.
The character shows when your agent is thinking, working or needs you. Approve actions from the watch and hear its reply in its own voice.
Each character has its own session and personality. The iPhone app adds photos, files and interactive results.
Free, open source, and uses your existing Claude Code or Codex setup.
A Tamagotchi that asks permission to run git push.
Microsoft offered voluntary buyouts to roughly 60,000 employees.
Two days later, it reported:
$92.3B in quarterly revenue.
$36.6B in net profit.
Azure: $33.2B, up 48% YoY.
The business is growing. Employees are being offered money to leave.
How much does strong company performance actually tell you about your job security?🤔
🤖 Made with AI
Claude Code writes Mermaid. Merman draws it right in your terminal.
This open-source plugin renders flowcharts, sequence diagrams, ER diagrams and charts alongside colored tables and syntax-highlighted code.
You also get 17 themes, copy buttons and optional compact tool-call rows to make long agent sessions easier to scan.
Under the hood: TypeScript hooks intercept UI rendering, parse the Markdown and draw it locally using beautiful-mermaid and Prism.
The optional diagram hint adds ~355 tokens per conversation. Redraws add none.
Requires Claude Code 2.1.311+.
github.com/iamrohithrnair/me…
Plushies gives AI agents a squishy 3D presence inside your app.
The repo includes two packages:
• plushies: customizable 3D characters with 17 shapes, fur, faces, accessories, and animations.
• @plushies/cursors: a pointer, name tag, live status, progress bar, and target highlights for each helper.
Connect your agent through hooks around tool calls or a WebSocket event stream. Mascots can follow DOM elements or canvas targets, show what each agent is working on, and celebrate when a task is done.
Built with Three.js. Open source. MIT licensed. Early PoC.
Claude wrote the Blender scripts. A human kept pointing out what looked fake.
One creator’s Claude Code + Blender workflow became increasingly specific:
• Compare every render against one fixed reference image.
• Inspect individual objects at full resolution.
• Use scanned assets and measured terrain where procedural detail falls short.
• Add scripts to flag geometry and material issues.
The creator reports that a .blend linter flagged 87 problems they had missed by eye.
My takeaway for agent builders: turn specific human feedback into repeatable checks the agent can run on the next iteration.
Some iPhone Duo screens are cracking days after launch.
iFixit suspects a hinge support may be pressing into the display. Apple says a small number are affected and is replacing units.
Might be worth waiting for the next generation.🤔
Anthropic shared early results from using Claude in drug development:
• Enveda: human review of 1,061 pages of manufacturing records dropped from 25+ hours to under 3, an 89% reduction. Claude also flagged 7 issues human reviewers missed.
• Generate: Claude Code helped redesign part of a cancer treatment. In lab tests, immune cells produced roughly 10x more of the proteins they use to recognize cancer cells.
• Divergence: Claude generated roughly 35,000 antibody designs in about a month for food allergy research, giving the team candidates to screen and test.
• Mithril: one founder used Claude to coordinate 11 labs, taking a potential treatment for a rare muscle disease from an idea to promising results in mice in under 18 months.
Most of this work is still preclinical: the stage before testing treatments in people.
Given where AI is today, how comfortable are you with its use in medicine?
anthropic.com/research/early…
Senior devs might be in their Kasparov era.
“AI can write code, but it still needs an experienced engineer steering it.”
Chess went through that phase too: humans guiding engines. The uncomfortable question is how long that advantage lasts.
Software makes the answer harder to measure. A senior’s override might prevent a production incident. It might also add complexity because the agent’s approach feels unfamiliar.
Both can look like “senior judgment” in a code review.
“AI needs an experienced engineer” is a claim we should keep testing as agents improve.
How do you measure whether your intervention actually made the agent’s work better?
