bhl

gh pr diff [num]

also works if you have the GitHub cli installed. Would setup an AGENTS.md or SKILL.md to instruct an agent on how to use gh too.

It does not matter what 80-90% of developers do. Code development is heavily tail-skewed: focus on the frontier and on the people who are able to output production-level code at a much higher pace than the rest.

Multi-agents.

Not for us to question or answer though.

What’s the solution here, reward code that works without try catch, reward code that errors and is caught, but penalize code that has try catch and never throws an error?

Cursor will pivot to a computer use company.

The gap between coding agents in your terminal and computer agents that work on your entire operating system is just too narrow and will be crossed over quick.

> Originally, it was all about quality data sources.

It still is! Lots of vertical productivity data that would be expensive to acquire manually via humans will be captured by building vertical AI products. Think lawyers, doctors, engineers.

Not all will become OSS. Some will become products, and that requires the best people.

The moat is people, data, and compute in that order.

It’s not just compute. That has mostly plateaued. What matters now is quality of data and what type of experiments to run, which environments to build.

Hiring is always a sh*tshow. The only thing that matters is survival: keep applying, keep grinding, keep growing.

And if there's any opportunity to show off, don't be shy :)

Stacked diffs is a huge one, and also where improving git would also improve LLM workflows. The bottleneck after code generation is PR reviews, and stacked diffs help break down large PRs into more digest-able pieces.

If you help humans collaborate better, you help LLMs collaborate better.

Mobile has really strong offline-primitives compared to the web.

But the web is primarily where a lot of productivity and collaboration happens; it’s also a more adversarial environment. Syncing state between tabs; dealing with storage eviction. That’s why local first is mostly web based.

IIRC, Kleppmann built a prototype for it but it’s not included in Automerge yet.

1. You can use TipTap with it: just have to wrap your existing schema with automerge attributes. Undo redo would also swap out.

> The Codex sprint was probably the hardest I've worked in nearly a decade. Most nights were up until 11 or midnight. Waking up to a newborn at 5:30 every morning. Heading to the office again at 7a. Working most weekends.

There's so much compression / time-dilation in the industry: large projects are pushed out and released in weeks; careers are made in months.

Worried about how sustainable this is for its people, given the risk of burnout.

Cursor has custom tab and embedding models. And has a lot of distribution / paying users already.

Arguably they have the strongest product moat, and I wouldn’t be surprised if they beat OpenAI in a vertical coding model from that. Easy for them to have users generate evals and have model product feedback loop here.

I don't know anyone who heard or used Windsurf outside the Bay Area. Even Cursor feels very Bay Area bubbly (although that is the market to go after if you're in ai dev tools).

I always referred it more as context degradation, but rot is more visceral.

It's not a 2 year cliff: it's 6 months before vesting, then 2 years before you can sell.

Ecosystem for local-first and mobile is pretty immature, at least for Swift.

In comparison to the web where there's so many libraries e.g. Zero, LiveStore, LiveBlocks, I've yet to find a good GRDB (sqlite abstraction) integration / client.

Offline-first is definitely very strong, but now how do I get data into a remote database with conflict resolution support?

Slightly weaker, but cheaper models mostly good for Windsurf only. As a developer, I would rather have stronger models I can throw more money at.

I've been contracting with a startup. The bottleneck is not the lack of tools; it's agency. There's so much work, it becomes work to assign and organize work.

But now who's going to do that work? Still engineers.

Cursor has enterprise mode which forces a data privacy feature.

Can you open source the code used to produce the evals?

Would love to see how its setup: the questions you linked to a ChatGPT chat, but the system prompt, tool calls would all be useful.

> The second is that LLMs provide the last-mile intelligence needed to verify every result. Each result and piece of data is backed with supporting references that we used to validate that the result is actually a match for your search criteria.

Evals on this would be great to benchmark the gap between using websets versus a generic web search tool. Otherwise to a developer, it's just marketing.

Building an app to reverse geocode all the photos I've taken and screenshots of places I want to go from TikTok/Instagram.

Use cases are extracting the itineraries I took from prior trips, then using that to ground LLM search recommendations: I have a set of bookmarked places, say in NYC, can you make me a week of plans given what I enjoyed in Taipei?

I'm surprised the chatbot on Instagram doesn't already have that. Giving an LLM access to a highly personalized search tool (the scroll algorithm) would make asking for recommendations so much better.

Just give your tool call loop to a stronger model to check if it’s a loop.

This is what I’ve done working with smaller model: if it fails validation once, I route it to a stronger model just for that tool call.

Harder than GeoGuesser is tagging a unique place from a photo even with gps exif data. Try this with bakeries in Seoul or sushi in Japan where there’s high density of same places.

What's the hard technical aspects of building a product like Windsurf or Cursor?

I've only seen codebase indexing or generating embeddings with Turbopuffer. There has to be more magic to that, right?