JC

Planning AWS for a 10-person medical office. First draft had RDS. Asked if there was a real reason. Not really. HIPAA still applies, so it's encrypted and under the AWS BAA. Postgres goes on the same EC2 box. We own backups and patching. Not building a SaaS here.

How I price custom business apps now. A set price to build it, then a monthly fee to keep changing it. That second part is the real value. Their business never again hears "our software can't do that."

Splash 1.6.0 is staged on the Mac mini, waiting for a dead spot. Too many agents on it to test, so upgrade in place and roll back if it won't come up. Not trading ~100 tok/s on Flash Next for 28 on a smaller Gemma. @inco_ai @jianchen1799

Old Fire HD 8 is on LineageOS now and stays awake when plugged in. Next up it becomes a Hermes voice terminal. Plan is a little kiosk app for the wake word, with Hermes running on another machine through OpenRouter. Tablet just listens and talks.

Got an old Fire HD 8 talking over USB in bootloader. Bootloader's locked. Told Codex to figure out the recovery path and walked away.

This is the best thing I've seen on the internet in a long time. teapawt.net/blognaointendo/status/…

My wish for GPT-6.2 SOL: decent speed, @OpenAI @thsottiaux. Been testing 6.1 Terra vs SOL in Hermes. SOL averages 25–30 tok/s for me. I knew it was slow, but didn't realize HOW slow until I ran the comparison. The speed difference is insane.

Saw someone cancel $1,700/mo construction software after wiring estimates and job costs straight into QuickBooks with bots. Same pattern I keep seeing. They don't want another SaaS. They want the headache gone and somebody still on the phone.

Had Codex repurpose a little ESP32 screen. It reads my calendar and shows if tonight's my every-other-night pill, the one I can never remember. Green yes, red no, amber if it can't trust the answer.

Mac mini serves Splash basically nonstop now and it was swapping hard. Turns out nine iOS simulators were still booted from last week. About 1,800 processes. Killed them and swap went from 58.5 to 52.8 GiB with no Splash restart. @inco_ai @jianchen1799

Underrated thing guardrails do: they give a model a reason to stop. Models without them will spin for hours on a task that can't be fully solved, because there's always one more thing to try.

Looked at MiniMax M3.1 Flash for my Strix Halo. 428B total, 23B active, 235 GiB of published weights. Even a 2-bit build would be about 100 GiB before context. It's one model class above what 128GB runs well. Waiting on the next MoE in my size.

Built a talking video character on top of a real tool-using agent this weekend. She keeps one session, browses, edits files and sets timers, then tells you out loud what she did. First speech back in about 4.6 seconds. Video rendering is hosted for now.

Codex blocked my chat over "cybersecurity risk" while I was testing a local security model. Moved that kind of work to DeepSeek V4.1 Flash. Refusals are a real reason to keep more than one option around.

Pet peeve with LP tools: APRs with no range attached. Could be a 2% range, could be 20%, nobody says. On @EZManagerCL every standard APR assumes a +/-10% range, it says so on the page, and it adjusts when you change your range.

A SaaS company told some old contacts of mine it's going dark for 6 to 8 weeks to upgrade their code to run on the latest PHP. They can't wait that long, so I'm building them a replacement before Monday.

Found the best use for @openai Dots, and it isn't managing everything. It's managing me. It connects to my Mattermost as my user, watches every active thread, and nudges me when something slips.

Tried both new coding models on my Strix Halo today. Ornith-1.5-122B at Q4 didn't pass my tool and vision checks. Flint was quick, about 60 tok/s, but missed too much on multi-step tool work. Neither beat what I'm already running, so I'm staying put.

A lot of the DeFi hacks lately follow the same pattern. Someone gets 2 of 3 multisig keys, adds a malicious pool or adapter, and drains it right away. A 48 hour timelock on every contract change stops that cold. It's non-negotiable on @EZManagerCL.

Splash 1.4.0 is my new daily driver on the Mac. Same weights as my Strata-MLX setup, 128K context, M5 Pro 64GB. 133 tok/s vs 83 decode across my text runs, about 61% faster. 8K prompt read went 6.80s vs 11.73s. Zero new swap. Nice work @inco_ai @jianchen1799