EJ Campbell

@ivanburazin Not just sysbox? x.com/daytonaio/stat…

@ejc3 Still are

Here's the exact pitch deck that helped us raise $2M in pre seed funding. Almost 3 years later, I'm making it public. Fun fact: we actually went out to raise $1M but it was this pitch that over subscribed the round and doubled it. It was also largely composed of seasoned founders and operators of some really interesting companies, as opposed to big funds. Consumer & Social: Amazon, Spotify, Meta, Robinhood, Notion DevTools: Hashicorp, Postman, Netlify, Warp, Puppet, Kong, Grafbase, Datadog, Honeycomb, Sentry, Instabug, dbt Labs, Supabase, Benchling, Socket, Stack Overflow, Contenda If you're raising your first round right now, might be worth taking a look. Download it here: ordnl.link/gzd2tt6

@glcst Scale to N > 1 machine hard though for a single use case?

I keep hearing from people that one of the things they love about Turso is that "it just works". This is no accident. One year ago we embarked on this bold journey to go rewrite SQLite. We knew that the only chance this had to work was if our Cloud product was rock-solid stable and reliable and would generate virtually no ops. We switched cloud providers and abandoned scale-to-zero. Scale-to-zero is a very ops-heavy architecture, and was plagued with many issues. We now run a fleet of massively multi-tenant servers that takes into account the beauty of SQLite - being just a file. We almost never touch it, and have an update of zero nines. It just works, and reliably delivers per-agent databases that you can trust. Once Turso (the rewrite of SQLite) is ready for prime time, our multitenant servers will run it instead of SQLite. Then a whole new world begins.

@arvidkahl That’s not true agents treat work as ephemeral. It’s extremely rare to delete in progress work (though it sometimes does).

One of my most painful realizations using agentic systems is that they treat all work-in-progress as ephemeral. They’ll gladly reset your branch or dev database to “try something”, because they expect to write working code eventually. Never manually add code while using agents. I tried several times, and it wasn’t good. Work on your manual changes outside of the agent, always commit it prior to running the agent again. and ideally, have non-git backups: Time Machine on Mac, full db exports of your database with a proven way to restore. You can try to guard your agent from doing destructive stuff, but it’s smarter and faster than you.

@omooretweets The simpler explanation is that Google and Gemini will copy any good feature in anyone else’s browser, so an alt browser will never win.

As an early superfan of AI browsers, ChatGPT moving towards a desktop app instead actually makes sense to me. Perplexity Comet has been arguably the most successful product here - and while they have a real base of power users, it's been hard to maintain growth 👇 We've seen this in the past with other fantastic browser products like Dia / Arc - there are a few things that make building a mainstream new browser very hard: 1. It's an extremely high frequency product where users have little tolerance for changes. If even one workflow is disrupted or made more difficult, it's like a paper cut that the user then experiences 100x a day. 2. The browser behavior is so automatic that the physical act of switching and maintaining the switch is hard! There has to be something in the new browser that's so materially better such that you remember to use it. And, if you have to onboard users to the product, you’ve lost. 3. There’s not that much “space” to innovate in the browser. The most important thing is to not disrupt the core experience, and so much is available via extensions that unlocking a 10x for the mainstream user is hard. Chrome works decently well - it’s not a low NPS product where people are desperate to switch. In contrast, desktop apps have proven to be a very fruitful surface for AI-enhanced work - think Cursor, Cowork, etc. Now that you can give a desktop product browser access, the advantage is clear - especially when the desktop app also has native file access and feels more natural to set up recurring workflows in.

@Michaelvll1 @karpathy How is the quality of 80 loops with batch size 10 the same quality as 900 runs with batch size 1. Feels like you are just trying to push a you orchestration system.

Autoresearch from @karpathy runs 1 experiment at a time. We gave it 16 GPUs and let it run them in parallel. 8 hours. 910 experiments. 9× faster to the same best result. The most surprising part: the agent had access to both H100s and H200s. Without being told, it noticed H200s scored better (more training steps in the same 5-min budget) and started screening ideas on H100s, then promoting winners to H200s for validation. That strategy just emerged on its own. A human researcher can grab a cluster and run experiments in parallel. The agent couldn’t. It was stuck with 1 GPU, greedy hill-climbing, ~10 experiments/hour. We built a @skypilot_org agent skill that teaches coding agents to manage their own GPU clusters. The agent reads the skill, then launches clusters, submits jobs, checks logs, and pipelines experiments on its own. With that, Claude Code provisioned 16 GPUs on Kubernetes, ran factorial grids of 10-13 experiments per wave, and covered in one 5-minute round what sequential search takes six rounds to do. The biggest finding: scaling model width mattered more than every hyperparameter trick combined. The agent tested 6 width configs in a single parallel wave and found the winner immediately. Sequential search might have missed that entirely. Total cost: ~$300 compute + $9 in Claude API.

@nikitabier Why would you generate a summary on the fly?

We’re rolling out summaries for Articles now. Just tap the Summarize button if you want to know if it’s worth your time to read it (or if your attention span is 12 seconds).

@ivanburazin Wouldn’t they need this for a mobile app as well?

Recently met the head of product at a SaaS with a $100B+ market cap. They're building a headless version of their flagship product specifically for agents. Not the cloud version with a UI. Actual infrastructure level APIs that agents can call programmatically. Imo, this is a far more accurate evolution of traditional SaaS than the current SaaSpocalypse BS.

@rowantrollope Redis mount more secure then a folder?

Agents love the filesystem with MD files. But when you want to share your filesystem across multiple agents its tricky Check out redis-fs a POSIX compliant filesystem backed by Redis. Mounts as FUSE on Linux and NFS on Mac. github.com/rowantrollope/…

@mil0theminer Weird thing to lie about.

What a blatant lie. They’re sunsetting it because Apple is restricting vibe coding apps. Why lie?

@MerbTheDerb @MacRumors No app can modify code in random other apps. The idea is that through these vibe coding apps you can make a whole new app. And that new app runs without ever going through apple’s approval.

@MacRumors obviously no one in these comments read the article, it’s saying that apple doesn’t allow apps that modify the code of other apps, replit and vibecode’s updates were stopped after the new updates were found to be able to edit the code of other apps

Apple Quietly Blocks Updates for Popular 'Vibe Coding' Apps macrumors.com/2026/03/18/app…

@marcos_placona @svpino gh has extremely fine grained permissions.

@svpino I don’t think people fully understand the differences between them. And how MCP gives you really fine grained permissions and while an API can also do that, a cli can’t.

I’m so confused with people saying this.

@svpino Agents mimic humans. Humans use CLI’s + scripting to get things done. Therefore skills >> mcp’s.

@NelsonXLee @vercel @anything @supabase Why do you think they a vercel stays a wrapper for longer than it is beneficial to them?

Gonna flesh this out in a 2k-word article later this week. Companies like @Vercel, @Anything, and @Supabase are great $1B companies, but they’ll never be $100B companies. Their business model prevents any power law.

@johnloeber Way too greedy. Extracting rent from devs who just wanted full features read only views of documents makes me not sorry they are imploding.

I feel sorry for Figma. They pulled off a generational feat, maybe the best and most admirable of its era: 4-5 years of pre-launch toil on fundamental technology, using that to launch a highly disruptive product, doing what previously was unimaginable, creating one of the few truly good pieces of software of its time. But as so often, things get good just before the end: right now Figma is looking like a fitting capstone to the pre-AI software era.

@Antisimplistic @_jujucollins @mitchellh @h1th3sh And the suggestion in this thread was that ghostty should do what other terminal emulators do and default to xterm since it’s most compatible. Your google result doesn’t capture that and it’s dismissive to say “just google it”.

@ejc3 @_jujucollins @mitchellh @h1th3sh since you seem to want to die on this hill… its literally the top result returned and the ai summarizes it

why can't ghostty team fix this feature? everytime i log in to a linux VPS it gives this issue @mitchellh

@QuinnyPig @awscloud Waiting with bated breath for them to circle back, maybe. x.com/AWSSupport/sta…

i'm sorry @awscloud you're sharing what now?

@_jujucollins @Antisimplistic @mitchellh @h1th3sh Right, it was still useful to hear the full explanation why they steadfastly refuse the budge on something that would make the experience much less intimidating,

@ejc3 @Antisimplistic @mitchellh @h1th3sh I think the point he’s making is they could’ve googled it and it would’ve brought them to the same linked documentation

Aren’t you showing that 2048 simultaneous queries is less optimal than 64? A normal app doesn’t change number of parallel queries based on how many connections are established. So better test would be keeping client parallelism consistent and vary how many threads you spread the queries across.

This is why you avoid direct connections to Postgres. Benchmarked PG running on a r8g.2xlarge (8 vCPU + 64GB ram) with connections ranging from 8 → 2048. Clearly a sweet spot at 64 with degrading perf thereafter. Apps often need 1000s of connections. Scale with a proxy!

@Antisimplistic @mitchellh @h1th3sh It wasn’t a “let me google it” situation though. It was a useful discussion that wasn’t on Google, with both raising good points.

@ejc3 @mitchellh @h1th3sh you might even say it was humble and gracious, like I did