Dave Thompson

Focus on the hallucinations and serialization to disk issues - we see these regularly through multiple agentic interfaces. We use Grok 4 Fast extensible for spec writing, data model generation and other high level tasks. We don’t trust it for coding beyond demonstrations. You can’t trust the model because inexplicably it will sometimes not persist to disk or will alter things unexpectedly. My AI Christmas wish is Grok 4 Fast ready as a daily driver. For now we rely on GPT5 and GPT5-Codex for high complexity code, Claude Sonnet 4.5 for main tasks and GLM 4.6 for a range of dev. Grok Code Fast only for unit tests and small refactors. I feel like Grok 4 Fast could do it all but has to move beyond benchmarks. Others have said same in comments. Set up a problem, and run head to head - you will see. Fantastic potential and already our go-to for many arch tasks. Sharpen the code generation and stop the hallucinations!

What are the most important improvements to make to Grok 4 Fast? Critical feedback is much appreciated.

@thdxr Keep at it! Gets better literally every week!

on track for 200,000 monthly active developers crazy we launched 3 months ago and the product isn't even good yet will be soon though

Bringing up Omarchy (omarchy.org) today on a laptop. I hadn't used tiling on Hyprland before - it took a few minutes to get the hang of the keys. Seriously fast, crisp screen updates. Techies, give it a try!

I use it a fair bit also and get good results generally. Find responses oddly terse compared to interactions through Cline. Only issue I see is difficulty with editing sometimes - repeated errors on replace that I don’t see with Anthropic models. Results are very good though across refactoring, test coverage and complex design.

i've had grok code set as my default model for the past week it's the first time i've used such a fast reasoning model - most are too slow to bother with we've lost fast feedback loops with LLMs and it's a nice peek of how it'll feel when they come back

@_catwu I have two requests for your consideration in the Claude Code UX. We use a lot of windows at the same time, and switch models particularly at certain stages of development with our projects. 1) How about if we added the current model in the lower ribbon next to where you show the "auto" or plan mode? 2) What about a "/title" or similar command that would let you set a window with text or with a template variable? Claude Code is awesome!

Similar experiences - just create a task breakdown structure and a plan file that has assignments. You can start by just copy pasting between windows or using file as SSOT. The interplay is also fascinating. I effectively end up sometimes with a kind of simplistic specialization. Assistants working in tandem from shared framework will literally recommend delegating some tasks or bugs while taking others. Then I have the integrator give performance assessments to other channels. They respond in turn. Seems to have lowered some of the acceptance gaps we get at end of large tasks. Probably spending a little extra in tokens to drive the assessments :). My advice is don’t even worry about automating at first other than having shared plan file structure. Had two front end assistants running today and by end of phase one was self appointed mobile expert while other was e2e test expert. What matters is the code is better, sooner.

@_catwu Cat, when using Claude Code for parallel work, I found that instructing Claude to provide complete instructions to agents makes it an orchestrator, enabling complex task assignments. I'm stunned by Opus 4's capabilities. Thanks to you and Anthropic.

New in Claude Code: Plan mode. Review implementation plans before making changes. Perfect for complex changes where you want to nail the approach before diving in.

Some feedback on Teams. I ended up switching from Teams to Pro because you were rolling features to Pro ahead of Teams. With the change in Pro Ultimatr, how will previews and context and other aspects of the service differ between the two ? Seems like organizations should have incentives to use Teams.

We’re lowering the price of Windsurf Teams from $35/mo to $30/mo and almost doubling the number of credits from 300 to 500. Plus Teams users get access to features like centralized billing and analytics. It’s never been a better time to get the whole team on Windsurf!

Here’s a full breakdown of the pricing changes we announced 🧵 To start off, we’re getting rid of flow action credits. 🔽

Great updates - looking forward to the auto-context save. Have built some processes around compaction with task files as I think context is lost on the save to some degree (though recent updates help). Over last 3 to 4 days seems like token usage is higher. Have been using on a medium complexity data proc tool from first plan and getting pretty good results. Appreciate the rapid updates!

Three more features for Claude Code today: First, auto-accept mode. Press Shift+Tab to toggle on auto-accept for file edits. This reduces friction while keeping you in the driver’s seat—you can tab in/out of auto-accept mode anytime.

I recommend giving Claude Code a try even if you are a Cursor, Continue or Windsurf user. I have found a number of tasks that seem to work better in this shell-based interface. Even for those who are not experienced developers, the text interface is easy to navigate. The tool works equally well on really complex design tasks - I'll post some observations on those flows when completing some other work.

Then I used Claude Code on the same project. Work is progressing so fast in this space - things can literally change in a week's time. I found Claude Code actually better at this setup step for several reasons. By this time I had experimented with manually creating the folder first. When doing this, Claude Code automatically thought to look one level above. Running `/init`, the tool used my project file (one level up) and wrote a very good CLAUDE.md. It then sorted out the bootstrap without any issues, even adding some extra validation steps I had not documented.

For the techies and vibe coders working on projects, I have some observations on the agentic workflow for project setup. When you get beyond a one shot development and need a multi-step project setup, it takes a bit of work to get that environment up. What package manager? What process? etc. etc. This setup can be frustrating even for experienced developers. With some of our current work, I thought it would be good to test how well we could automate package / repo setup from a plan doc. Here are some initial results from Claude Code (@anthropic) and @windsurf_ai

@windsurf_ai Is the Claude Sonnet interface really hitting 20241022? Doesn’t seem like it. That would be good to remedy if not. And BYOK. Think you can work out a model for billing that allows this.

What could we improve in Windsurf? And what features should we add in Wave 4? Let us know and we'll ship them all asap

In complex coding work, I find it very difficult to keep context relative to other tools. I'll post / share files and then 2-3 messages later find out that the context is lost. Probably there is something I am missing. The actual interface is great - rules support refined well with recent changes. At present, whether because of my ignorance on what to do or because of actual gaps, I work through harder problems with continue.dev interface.

What are the biggest roadblocks you faced when starting on @cursor_ai ? Tell us so we can fix them all 👇

For developers out there, would like to recognize the @windsurf_ai team for great work over the last few releases. I've been using v1.3.4 in agentic mode to work on a small-ish Bun/Typescript project we are releasing at end of month. The interface is getting much better every week! Finding the diagnostic loops to work better with Claude. Have had to nudge a few times on unit tests and some syntax issues. Overall, though, this is great!

@ScottPresler Scott, you and your colleagues set the standard for diligence, vision, passion and commitment to your principles. Ignore the haters and keep your eye and your heart on what you know to be right!

Since I’m trending right now, I want to turn the attacks against me into something positive. New Jersey, Pennsylvania, Virginia, & Wisconsin, I’m asking for your votes this November. 🙏

@jarredsumner I am going to deploy some scripts using new S3 in the next couple of weeks. Great idea! More SSH and SFTP also good as others have noted.

i was a little surprised at the number of negative reactions to Bun.s3 is it because it wasn’t clear enough that there are many non-AWS S3-compatible object storage APIs? you haven’t had to write code that uploads or downloads files on a server? something else?

New threads, broader mission: helping transactional businesses and their partners wrangle their scaling problems into submission. We've leveled up our game to turn your business challenges into solutions. Want to see what we're cooking? Hop over to our new website at 3leaps.net. Call or message us - we would love to explore how we can help you! #3LEAPS #ScaleUp