@theo@ThePrimeagen I ended up starting my own iPad keyboard that adapts to my handwriting and vocabulary. Definitely improved recognition accuracy.. but still struggled to make it a part of my day-to-day workflow as a real replacement for typing. Will try again though
@theo@ThePrimeagen I don’t enjoy voice-to-text either.
I bought an iPad + Pencil to try to get ahead of some wrist pain that was developing as a result of AI-induced dev psychosis. It helped, but the handwriting-to-text accuracy out of the box has been pretty brutal.
Been exploring similar approaches. Exporting a lot of my work as either standalone html “capsules” or multi-file zipped “bundles” with embedded UUID’s and other metadata.
From there I can load them onto various domain-specific registry sites (either internal or external) where clients and collaborators can leave comments in a Google Docs type interface or download the capsule/bundle themselves with their own LLM. Then, so long as they follow the spec with their revisions, they can send me back their version, with its own UUID that includes my original UUID as the “parent”.
Has been working well so far, and also seems to work well as an offline knowledge base alternative. A lot of my clients and collaborators like to have clear “versions” of things to refer to, and live documents don’t always work great for that.
https://t.co/fK1bUdapHC
I’ve been doing lots of experiments across different domains, and the recurring pattern has been to export the work as either a “Capsule” or a “Bundle.”
A Capsule is a standalone, versioned HTML file with a UUID, embedded data/media, provenance, and basic controls. No external libraries, no network calls, and clear constraints around JS.
A Bundle is a versioned zip with a manifest and the needed files (for example, one of my bundles today had LiDAR data, impractical to embed into a single html file).
Both can come from exporters I build into little apps/experiments, or directly from an LLM chat when I ask it to export according to the spec.
From there, I put them in a registry: either personal, or domain-specific for sharing with users/collaborators, with Google Docs-style access controls.
For my personal registry, it works well as a knowledge base. I find it more intuitive than managing Markdown/Obsidian, plugins, and sidecar files for images/audio just to get the structure and feel I want.
Been working well for me so far, but would love to chat with anyone experimenting along these lines. https://t.co/HW2eZG9idM
@theo I don't actually believe you're this naive.. It was painfully obvious at the time that you were going to get rugged? I wish it had been on Polymarket, I would've bet on it...
@honjamesmoore This was inevitable from the very beginning. Vancouver gave the MLS fans, history and legitimacy, but the league’s economics were always moving toward stadium control and real estate upside. Now the value created here is being used as leverage: Vegas, or public concessions.
@bcherny How many more thinking tokens does it use? 1% more? 50% more? 100% more?
How much did you increase rate limits to compensate?
The way you guys communicate around tokens and rate limits is going to give me an aneurysm.
@lydiahallie Woke up a couple mornings to 20-30% usage. Computer was off...
Aside from that, have been hitting session limits a lot quicker than before.
@DimitrisSoudas@CandiceBergen_ There's also a provincial election in Quebec later this year. The CAQ, PQ and PLQ are trying to score political points off this issue.
@DimitrisSoudas@CandiceBergen_ Let's be precise:
There is a by-election next month in Montreal that will be key in giving Carney a majority. The Bloc and LPC are trying to score political points off this issue.