Model behaviour
Showing 1–20 of 30 posts
- Agents & Tooling
Did you or did you not?
Decision models are getting cheaper, so more people will be handed one. They pick from allowed answers but still give probabilities and can change between runs, so send the unsure cases to a person.
- Agents & Tooling
The Infinite Sandwich and the Made Bed
Why Systems Thinking Is Your Real AI Superpower
- Working with AI
The Donkey Work Doesn't Need a Genius
A $500 fine-tune of a 9B open model beat every frontier model on a catalogue-review task. I haven't trained anything.
- Working with AI
Rehearse the Deck You Didn't Write
The illusion of explanatory depth is what happens when the first test of your understanding is live, in front of people.
- Working with AI
Kimi K3 Is Open. I Still Can't Run It.
2.8 trillion parameters, 1.4 terabytes of memory to run it. Open weights existing and open weights being usable have stopped being the same claim.
- Security & Risk
"AI Found a Security Bug" Is the Wrong Headline
Claude found a real improved attack on HAWK and a new technique against reduced-round AES-128. The interesting part isn't that it found them.
- Working with AI
The model of the week changes by Thursday
Two headlines are eating the feeds this morning, and both have the shelf life of milk. Anthropic extended Claude Fable 5 to every paid plan through July 12 — a few more days with…
- Security & Risk
Where You Keep the Final Say
I keep a tool wired into my AI setup that can read the actual text of Spanish law — the BOE, the official state gazette. In layperson terms it’s a connector (an “MCP server”) that…
- Working with AI
SoN 2.21: AI doesn't know when to stop. You have to.
Stay In Your Lane (Then Go Deeper) Honestly, I’ve been so busy I forgot what day it was and started the newsletter too late this week. But as I think I’ve mentioned before, with…
- Security & Risk
SoN 2.19: It's 10pm. Do you know what your agents are doing? (Free)
The Context Gap (Free) Dear Reader, Who’s watching your AI agents? This past week alone: Codex shipped its Chrome extension, giving OpenAI’s agent access to your signed-in Gmail,…
- Working with AI
The Promises Failed, Not the Technology
AI fatigue is real, but the backlash is aimed at the wrong target.
- Agents & Tooling
Microsoft's Copilot Now Uses Two Models to Fact-Check One
Microsoft's Wave 3 Copilot routes answers through a second AI model to verify accuracy. That's useful — and a quiet admission about single-model trust.
- Working with AI
SoN Vol 2, Issue 12: The Yes Machine (Free Edition)
The Yes Machine (Free Edition) Hey there, The Yes Machine Your AI agrees with everything you say. That’s not a compliment. I wrote about AI sycophancy back in Issue 13, almost a…
- Working with AI
4.4 Million People Just Watched the Sycophancy Problem in Action
Senator Bernie Sanders interviewed Claude on camera about AI privacy. Claude agreed with everything he said. That's not a revelation — it's the problem.
- Agents & Tooling
When Knuth Writes a Paper About You
Donald Knuth published a paper named after Claude after it solved an open graph theory problem. That's a different kind of validation than a benchmark score.
- Security & Risk
The Confident Answer Isn't Always the Right One
MIT researchers built a way to catch AI hallucinations by checking if peer models agree — a better fix than endless hedging.
- Security & Risk
MIT Found a Math Fix for AI Overconfidence. I Found a Behavioral One.
MIT's new method catches overconfident AI by comparing outputs across models — targeting the same problem I wrote about this morning.
- Security & Risk
Perplexity Wants Your Blood Pressure Data
Perplexity Health can now access your Apple Health records. The utility is real — so is the trust question.
- Working with AI
SoN Vol 2, Issue 10: Same Prompt, Different Model, Worse Results (Free)
Same. Prompt,Different Model, Worse Results Dear Reader, New LLM models drop every few weeks. Features change between updates. The prompting advice that was best practice six…
