PR suggestions no longer work in GitHub...been completely down for days now. What is going on over there

The hours of my life that I've spent waiting for this...

I have an email chain going back months asking Azure to give access to o3 deep research (released in February) Today the Google team reached out to me asking if we wanted to try the Gemini 3 preview in the API (released last week)

We already have a lot of AI agents in our Slack, excited for when I can do something like: @ linear create this ticket, ask @ reddyai for any details you're missing, be sure to include docs links that you find with the @ kagi api, and then @ cursor @ amp @ devin build it

The simple version of this is commits formatted like: > Commit Summary - derived from Code Diff + Prompt Thread(s) + maybe manual tweaks > Prompt thread dump

this leads to interesting search options as well, e.g., write to chromadb instead and query for related code changes with a semantic embedding model embeddings work better on english than code anyways

git blame already lets you go through the history of commits for a given line of code, so you could just enrich that by including the prompting pieces (not to mention git commit messages can be generated much better from prompts than code diffs)

it would look something like in a post-commit hook we could write the prompting that went into a diff into a file-based DB like sqlite which would then be queryable by commit ID. Then make a follow-up commit to save sqlite

We need to create another dimension for code repos. Lots of prompting work now gets the "engineer's brain" into text like we've never had before, yet we throw away all that data as soon as we commit and push. Imagine if you could see WHY a line of code is what it is.

this is accidentally a great browser agent test: "go to [website] and come up with a signature that makes a 5-point star" that should be sufficient, but even cursor struggles with this quite a bit it should be able to find 'zecafz' or similar

Code is bad and features are good. But code is less bad now than it was 5 years ago. And even less bad if it is isolated from other code. I think we'll see an industry shift towards new mechanisms to isolate (more) code. At extreme levels, this has meant features that no

I 100% believe the crazy numbers for LOC per day Mostly because we are massively outdoing them. Starting to make a distinction between stable/volatile code and guardrail/implementation code. While this thinking is still somewhat experimental (and we've isolated what we apply it

I think concurrent AI coding is fundamentally changing the best practice of keeping PRs small and isolated. When you can have 3 agents working on the same code at the same time, its a lot more work to maintain separate environments, track across all of them, etc.

Learned so much already from Hamel and Shreya's evals course and still behind half the lessons. Great sampling of the main problems and solutions with LLMs in real world products. Even though we live this every day, it's tough to cross-pollinate ideas from other teams with such

One of our engineers showed me vs code can open multiple repos in the same workspace And cursor can use that to make edits across repos...quite well actually Been using this so much the last few weeks as we've started splitting our monolith

Claude 4 sonnet >> Claude 3.5 sonnet for coding

When do we get some form of background agents in local environments? E.g., why can't I turn off CLI tools and get unique code edits applied in multiple concurrent chats? @cursor_ai

Cursor has really run away with it...pitted it against several competitors lately and its not even close even though they have the same models

why haven't they done this already? my best guess is not every location has easy or free parking and they'd have to track which is which but even that could be done en masse by adding user-driven reporting like Google maps

uber needs a chauffeur mode pay the driver per minute I'm in a location to wait, so i can get an immediate ride to the next place sometimes it's a 2 min stop, and then I have to wait 5-10 minutes for the next uber to show up