Author of the AI-Ready Human and of the Humanity Working Substack. Principal AI Technologist at Adobe. All view are my own.
Seattle WA
Joined September 2008
- Tweets777
- Following361
- Followers145
- Likes67
Elsewhere in AI this week:
Anthropic has filed to go public. AI assistants are multiplying, and misbehaving. Google turns a sentence into a world. And Reddit's designers are building their own AI tools.
Drawing on work by @emollick @TheZvi @clairevo @itsurboyevan
Full episode on YouTube: youtu.be/d-DOzK4br74
The written companion, with every link: humanityworking.net
Right now it looks like one of the biggest challenges we may face from AI is a sort of “death by 1000 cuts”, where previously stable parts of our infrastructure are exposed and exploited.
If this continues, it may naturally slow frontier model development, as the focus shifts to stabilizing what already exists.
Paul Slater retweeted
Why does every frontier coding harness make it easy to add MCPs, plugins, etc, but none of them give you the option to remove them? In every case you have to find the config file and remove the entry (or have your agent do it for you). That seems really odd to me.
Long-form + QA: Opus 5.5. The diligent craftsperson. Complex articles with multiple logical threads, detailed code audits. Genuinely efficient at this, great reviews.
Messy stuff: Fable 5.1. Reliable overall, but it takes unusual shortcuts to finish. For long unsupervised runs, especially on problems Opus stalled on.
Not using Gemini 4 Argon yet. Once I do, this whole setup reshuffles.
The skill isn't knowing the models. It's knowing where their cost and quality curves cross.
What is the thing you have got most wrong about AI?
And don’t hide behind getting pacing or scale wrong.
What was embarrassingly wrong?
I've been ranting against AI customer service agents being the front line, but they make a whole lot more sense when muse or dots are making the customer service requests. My AI won't get annoyed with your AI customer service agent.
Astra has WAY dialed back it's ability to act independently. It will proudly tell me it's found a bug, then do nothing to fix it until I say "So....".
It's like an amazingly effective coworker with a severe case of narcolepsy.
Paul Slater retweeted
There is a lot to be concerned about right now. There are massive incentives to make you outraged. Stay present. Rely on what you know to be true. Question the motivations and incentives of the loudest voices. Breathe.
Paul Slater retweeted
My new show with @CaseyNewton is called Machine Gods, and we’re partnering with NPR!
It’s a twice-a-week, video-first podcast about AI, and it will be available next month on YouTube and other podcast apps (and on radio soon!)
It’s a really critical time for public engagement with AI, and we’re thrilled to bring our AI coverage (and jokes!) to NPR’s massive audience, while building an independent media company in San Francisco. We’re also thrilled that @RachelCohnnn, our incredible Hard Fork producer, is coming with us to make the new show.
More to say soon!
npr.org/2026/09/16/g-s1-1434…
When the various “because China” arguments happen for full speed ahead arguments on AI, they always assume it is not in China’s interests to slow down. I do not believe that is true.
Paul Slater retweeted
I’m observing a major swing towards model routers and away from working directly with the frontier labs for software development teams. The costs were just too high and the model routers provide so much more value and flexibility. The big winner? It’s starting to look like GitHub Copilot.
I am encouraged by this
