notnotrishi

I built it because I was tired of typing out bug reports or constant prompting or wrestling with screenshots. Now I just circle what's broken and get a detailed ticket. Works with any app on macOS. Would love feedback on the approach and what integrations you'd want next. Thanks!

It’s bizarre that we can type or talk instantly on our personal computers but can’t scribble natively. IMO, freehand should be a primary input. Some thoughts on what I’m working on and why: https://notnotrishi.substack.com/p/drawing-the-overlooked-un...

not that i know, but that's something we def. need

agree!

wdym by 'screenshots of interviews'? like coding interview questions?

interesting, will check it out

curious what you meant by a knowledge system. thanks

haven't tried aider. looks interesting

yes, o1 def. seems to have a deeper 'understanding'

thanks!

> with Claude if you start having extremely long conversations I have noticed it allowing certain bugs it had already fixed to be reintroduced at much later times

i think this is a result of its inability to handle long contexts well?

My notes:

- Sonnet 3.5 seems good with code generation and o1-preview seems good with debugging

- Sonnet 3.5 struggles with long contexts whereas o1-preview seems good at identifying interdependencies between files in code repo in answering complex questions

- Breaking the problem into small steps seems to yield better results with Sonnet

- I’m using primarily in Cursor/GH Copilot and with Python

'AI/ML stuff' is a lot and if you don't know what interests you yet I'd recommend developing an intuition for all things ML/AI and building small ML models first:

1. Spend 6h going through this video: https://www.youtube.com/watch?v=1vkb7BCMQd0

2. Go through Google's intro to ML crash course and pre-reqs: https://developers.google.com/machine-learning/crash-course/...

3. Refer to other videos on Youtube (3B1B and StatQuest are couple of my favs) as you go through no 2 above

Spend about a month on the above and then see what you really want to dig into next. There could be a few different ways after the above, but one way is speedrunning through 1st sem coursework of any top ML grad school program

lol mind sharing an example?

also in case you haven't already, can try with other models too by changing it from the code to see if they do better

exactly! also, one can set parameters etc. if they'd like.

also, after this post, i was able to do some tinkering and run it outside chrome, which i think would be more useful. hope they release the weights and make it open

yea, it's a bit tricky. this came in from google team today:

> To get back on track, make sure you're using Chrome version 128.0.6545.0 or later. If you want the most up-to-date version, try using Chrome Canary or Chrome dev channel.

this document should have the latest changes from Google: https://docs.google.com/document/d/1VG8HIyz361zGduWgNG7R_R8X...

The screenshots included above are from real use. It's not 100% there but as mentioned on the form it's early access for people who are comfortable with bugs etc, and not just a landing page to collect data.

Also a bit of a background --

2018-19: I started working on the problem of automatically extracting all action items from all tools I use (email, slack, GitHub etc.), and getting them in one place. I’d built a few different solutions using NLP (using techniques available back then). Things sort-of worked, but there were always so many edge cases — early users churned soon.

2020-21: Graduate student at an AI lab — Mila, Quebec (dropped out eventually) — went there primarily to learn new techniques to build some thing in this space. Around mid-2021 I decided to put this on hold, and went on to work at Beeper (beeper.com) till recently.

good stuff! a detailed write-up w/ GH link available somewhere?