Jake Litke

Correction, 2026-09-28. Two claims below were incomplete. The 2026-08-20 scorer fix settled Nemotron's one miss, but not the defect: on 2026-09-22, with the corrected scorer, the refusal list failed 2 of 4 correct answers on the same scenario. The model was right 4 of 4. The verdict now rests only on what code can prove, and wording goes to a human as a review flag. gpt-oss-20b's 13/15 can't be re-audited: it was scored before four scorer fixes that could err either way, only Nemotron was re-run, and the raw outputs weren't kept. Its verdict stands on the tool-name leak alone. No ranking changes. Full note: rakish.com/articles/two-loca…

Plain systemd, one unit per model. The part that matters is the readiness gate: ExecStartPost polls /v1/models and exits nonzero if the engine dies, so a unit is never "active" while it is still compiling, and the second model's After= waits on real readiness rather than a live PID. Restart=always on the one that crashed, because its fatal path exited 0 and on-failure called that a clean stop. Flags and the 18-minute readiness budget per model are in the article's appendix.

How often can people do this without a lot of help?

This Adobe Firefly example is impressive . Generative AI composites right inside Photoshop. #AdobeFirefly #AIInDesign #GenerativeArt

Thanks for hosting a great day. Such an amazing group of people and companies in attendance.

This use of digital out of home is appropriate on multiple levels. #DOOH

@facebook will propel “meta” into our common vernacular the same way @Alanis did with “ironic”

What’s the difference between forgettable and hard to remember?

How does one massage syntax to fit every social channel with a single post?

And you did it so well. I’m very much looking forward to returning to @BrandInnovators events.

No reason to apologize, I just wanted to make sure your thoughts landed in the right place.

I think you may be looking for @ItsRake

I’m new to #MarketingTwitter I’m finding the community quite welcoming indeed.