Some of the best bedtime stories are the ones about their own day.
carter.maslan.com/the-storie…
“I’m a little wound up” - simultaneously cathartic, funny, and serious rant is so good open.spotify.com/episode/76r…
I'm loving @Muse. GDP++: roof and gutter repair, EV 240V outlet, T-Mobile conversion, car dealership lease options... so many dreaded TODOs removed in one evening with its browsing the web, filling forms, sending texts. Boon for local businesses under the thumb of aggregators.
My mother-in-law introduced us to the Enneagram in our early 20s, on a weekend away without the kids. I'm pretty sure it's why we're still married.
What it's been good for:
carter.maslan.com/the-enneag…
Wrote this for an audience of one. Posting it in case it fits in somebody else's songbook too.
carter.maslan.com/small-deli…
There's such a long tail of specialized evaluation protocols that will be required for AI adoption.
This is an inspirational example of one.
To craft for conversation, timing is everything. Natural conversation is a social dance, and timing matters just as much as tone and delivery.
Most voice AI doesn't do this well yet.
So we built TurnBench: an open benchmark that measures whether voice AI knows when your turn is really over, and when an "mm-hmm" is agreement, not an interruption.
Learn more about how we built this on our blog.
sesame.com/blog/turnbench
my fam knows how much I wish for city streets without parked cars, but I had no idea that a 4d/wk commute to SF is a $43.58*48*4*2=$16,734/yr subsidy.
One of funniest feelings with AI is watching it take tiny human requests so seriously.
I needed a watercolor upscaled faithfully for a 30"×20" print for a 3 year old's new room.
I asked Claude Cowork to upscale the watercolor so it won't look pixelated printed at 30"×20" - I said 150 DPI was fine for the style, and to use a *faithful* upscaler that won't hallucinate fake detail. Here's everything it did.
First it sized the job: print pixels = inches × DPI, so 30×20 at 150 DPI = 4500×3000. My source was only 1264×848, needing about a 3.6× upscale. The plan: upscale past the target with the model, then downscale to the exact size - because downscaling preserves quality while plain upsizing doesn't.
It reached for Real-ESRGAN first, but the sandbox could only reach PyPI. PyTorch's download index, GitHub release assets, and HuggingFace were all proxy-blocked, so it couldn't fetch model weights the usual way. The workaround: it discovered that the PyPI packages waifu2x-ncnn-vulkan-python and realsr-ncnn-vulkan-python actually *bundle* their model weights inside the package. It extracted those and ran them on CPU via the pure-CPU `ncnn` package - no GPU required.
For the model itself, it chose waifu2x (the anime/art model) over Real-ESRGAN. My image is mostly soft sky, sea, and pool washes, and GAN-based upscalers tend to paint fake texture onto smooth gradients. waifu2x is conservative and faithful - exactly what I'd asked for.
Along the way it worked out the fiddly bits: waifu2x's 2× model shrinks the output by 7px per side, so inputs get pre-padded; the feature maps are huge, so inference is tiled in 256px chunks to avoid running out of memory; it's a 2× model, so it runs multiple times for higher factors; and with a ~45s-per-command cap, big steps get checkpointed to disk.
The pipeline: waifu2x 2× twice (4×, to 5056×3392), then a Lanczos downscale to exactly 4500×3000 with 150 DPI embedded. It verified the pixel size and DPI, and compared a 100% crop against a bicubic resize to confirm sharper edges and no seams.
Then I asked for 300 DPI too. Same method, one more 2× pass (8×, to 10112×6784), downscaled to 9000×6000 at 300 DPI - exactly 30"×20". It even noticed one output file had been truncated by the time limit and regenerated it.
Finally it wrote a markdown guide so I can repeat this on any image, and told me exactly which file to send to the printer. From a blurry 1264×848 to a clean 9000×6000 print, entirely inside a locked-down sandbox. gist.github.com/jcmaslan/08b…
Great for a generalist engineer with product and consulting chops.
Please share. I’m guessing most all work will trend in this direction, and we have immediate high-impact projects that are fun and meaningful in critical industries.
Frontier AI is just potential if it stays in the cloud.
We need the engineer who can wrangle it onto factory floors.
You’ll write the playbook that makes frontier AI models boost yields and cut costs in the world’s most critical facilities.
camio.com/careers/forward-de…
Carter Maslan retweeted
Per-camera pricing seems simple—until it misses the real wins or hides the real costs. Here’s why Camio starts with outcomes, not lens counts. blog.camio.com/why-per-camer…
Carter Maslan retweeted
Cloud-native AI advances are redefining physical security outcomes.
blog.camio.com/the-end-of-po…
Carter Maslan retweeted
The freedom of natural language with the guidance of clickable controls, so you can find exactly what you need, faster.
blog.camio.com/introducing-c…
Carter Maslan retweeted
Camio Wins the 2025 AI Innovation Award for Visual Agents Transforming GRC:
blog.camio.com/camio-wins-th…
A plain-text list of store policies.
Snapshots from your existing cameras.
Compliance scored automatically—at scale.
Camio checks operations policies like blocked aisles, empty shelves, poor lighting, and more—across 100s of stores.
ai.camio.com/results/eyJhbGc…
"Suspicious activity" is vague—until you see it.
Large Multi-modal Models combine the broad context of LLMs with Computer Vision to detect real-world deviations from expected behavior—like this man hooking a chain to rip copper from a utility pole.
ai.camio.com/results/eyJhbGc…
The latest AI enables continuous quantitative aseptic observation of sterile filling operations. (e.g. data on first air disruptions per Annex 1). I'm eager to learn more from CDMO experts and reciprocate with insights into visual AI capabilities. DM to meet now or at @INTERPHEX
Carter Maslan retweeted
Large multi-modal AI models now simplify the detection and remediation of problems by interpreting plain text policy through the lens of standard IP cameras:
youtu.be/JZeN3m3nnfA
