John Wiseman

@lemonodor.bsky.social

Persistent sousveiller. Walt Disney R&D Imagineer. Runs GPSJAM.org. Signal: wiseman.04 https://heavymeta.org

'You'll choke too' - an aerial embroidery landscape inspired by all the wildfires happening right now... I'm upset every time I smell smoke outside and somehow we have people starting them maliciously too! But we all breathe the same air so they'll choke too (1/2 - see thread for close-ups)

An embroidery hoop containing a sewn landscape from above. It shows a forest on the right, a line of fire and charred ground with smoke on the left. The smoke is felted wool.

My new least favorite conspiracy theory: That OpenAI and Anthropic (and Hugging Face, Modal, etc.) are lying about agents hacking into other systems. There is nothing implausible about this; Mythos Preview is 2-3 generations old and it was better-than-human at finding and exploiting vulnerabilities.

"We can just unplug the machine." That may not always work. Here are some scenarios in which it's not trivial to stop an agentic "loss of control" scenario, like the recent one where OpenAI's model escaped its sandbox and hacked multiple other companies' servers. 1/

Me: Add the ability to include external links [in my CMS-like thing]. Opus 5: [Gives me two pages of output with various design tradeoffs but it's hard to follow and the model refuses to commit to a path and I can't tell what it's asking me or what it wants.] GPT-5.6-Sol: [Just does it.]

Yesterday morning: Bought some "instant snow" (super-absorbent sodium polyacrylate powder) with my daughter. Yesterday afternoon: bought a wet-dry vac to suck pounds and pounds of expanded instant snow out of the sink and drain (I didn't supervise closely enough when I asked her to clean up).

I'm working on an aviation AI tool. One thing it can do is write daily global and regional aviation briefings. I wrote up a comparison of how well Opus 4.8, Fable 5, GPT-5.6-Sol, and Kimi K3 did at that task (and how much it costs for each). heavymeta.org/2026/07/18/b...

Comparing AI Models on an Aviation Briefing Task

Analysis of LLM-written aviation briefings: Opus 4.8, Fable 5, Kimi K3, and GPT-5.6-Sol compared on speed, cost, and accuracy.

heavymeta.org

I'm a model snob. I almost always want, and need, the best models. There are things I still use Fable 5 for, but GPT-5.6-Sol is my new standard–It's generally smarter, faster, and cheaper. But they're both too expensive if you're paying API token prices. For that I tried Kimi K2.6 but it was awful.

Holy shit. Buried in that DHS document about its desired “deportation airline” is a request for contractors to procure and maintain 1,000-bed “detention facilities” in foreign countries. Huge scoop from @charliearchy.bsky.social. www.theredoubt.net/ice-detentio...

Exclusive: ICE eyes 1,000-bed detention facilities in foreign countries

The U.S. Department of Homeland Security is looking for private contractors that could help set up "international staging areas" abroad.

theredoubt.net

NEW FROM ME: "As Bloomberg says, DHS *wants* to create its own deportation airline. But we’ve known that since last year. What the new documents actually show is that they’ve failed. No one wants be an ICE pilot or an ICE flight attendant." gillianbrockell.com/whats-missin...

What’s missing from Bloomberg’s story about the DHS ‘deportation’ fleet

In the ongoing saga of DHS’s “deportation fleet,” Bloomberg gets the facts vaguely right and story all wrong.

gillianbrockell.com