Stratten Waldt

Hi. I'd like to introduce you to Basil. It's a project I've been working on and working with for over a year. It's an AI assistant for your Mac. You press a hotkey or say "Hey Basil" and tell it what you need. It sees what you're working on, understands the context, and handles it. I wanted to create something that was a joy to use. Something that made AI feel magical, but also like a seamless extension of my own capabilties. Something that fit into my workflow and was always there when I needed it, without being intrusive. When I started building this, it was called "Jarvish", because I was genuinely trying to build Jarvis. And because I've always wanted a cease and desist from Disney. That was a year ago. Since then it's grown into something much bigger than what I originally had in mind. It does voice- (or keyboard-) initiated agentic workflows from anywhere, context-aware content generation, live meeting transcription, direct application interaction, file management, shell commands, AppleScript automation. And it can do pretty much everything using local models. It also does regular ol' transcription, if you want to be vanilla. It really seems like AI has lost its shine for a lot of people. Every app tries to cram a button into every surface whether you asked for it or not (seriously, fuck you Copilot. I downgraded to an older version of word just so I don't have your obnoxious icon every time I start a new line). It feels less like a tool and more like a pop-up ad you can't close. I designed Basil to always be one keystroke away, but never shows up uninvited. That distinction was really important to me. And the more you use it, the more it has to work with. It remembers your conversations, tracks what you've been working on, and can learn your writing style so that what it creates sounds like you, not a robot. It started as a local-first project on purpose. I figured if the core capabilities worked without an internet connection, then cloud models would only make it better. Everything runs on your Mac, your data stays on your machine. If you want to use cloud models, you can use Anthropic, Google, OpenAI, or any custom endpoint you want. Bring your own API keys, bring your own models. But nothing leaves your computer unless you decide it should. The current version is free. No email, no credit card, no trial period. Just download it. I've spent a year building this mostly in isolation, and at this point, what I want more than anything is for people to use it. I use it every day. It saves me hours. It's genuinely the most useful thing on my computer. But there has never been a moment where it felt *done*, and I'm realizing that moment is probably never going to come. I've learned more about architecture and engineering building this than I did in the five years before it, and I'm really proud of where it is. If it's as useful to you as it is to me, that's enough for now. So here it is. I made a thing, and I think it's pretty cool. Take a look. See if it's for you. You can see it doing a bunch of things here - basil.ac And you can download it directly here - basil.ac/download P.S. I hope you can feel how much care has gone into this if you use it. Every part of it has been intentionally crafted; down to the window drop shadows, which are custom because the native ones didn't feel quite right. P.P.S. Here it is cleaning up my downloads folder in less than 4 minutes (for me. for you it's only about 45 seconds). If I'm being honest, I probably would have just never gotten around to it otherwise.

@icanvardar And if you leave your tools out in the rain, they'll rust, especially if they don't have a finish.

zig isn’t better than rust, people just like arguing about unfinished tools more than using finished ones

@plainionist You can still use a moped without a license. So I guess anyone should be able to use Haiku?

Most countries require a driver's license to operate a car. Do you think developers should need a "vibe-coding license" to use LLMs? 🤔

I mean if we're doing a wishlist - -Sub-millisecond latency at scale - Not just p50, but consistent p99.9 performance under load. - Tiered storage architecture - I want it to provide cache-level performance with sub-millisecond data access together with built-in persistence is crucial. It should be intelligent about data placement between memory, SSD, and potentially cloud storage without sacrificing speed. - Predictable performance under concurrency - Single-threaded architectures can reduce CPU context-switching overhead that slows down multi-threaded databases, helping maintain predictable latency as load increases.

what would a database need to do so you don’t ever want to use a cache service (e.g. redis, valkey, memcached)? sub-ms latency, predictable perf, low memory / disk usage, redis like kv interface?

@ConnerBean Well yeah, if you know what you're doing, it just makes applying good fundamentals less of a pain in the ass. If you don't, it lets you skate by without knowing they exist in the first place for waaaay too long.

AI is making the average engineer worse but the best even better

@theo Deepseek, more like Deepsick, amirite? (But seriously, someone let me know if they're doing okay)

Anyone check on Deepseek recently?

@sandislonjsak You have no idea. They're about to sound so much like a real person.

@strattenwaldt Emails are about to be lit 🔥

Seems like Opus 4.7 has improved on everything except coding.

That's why part of the process of building your application should involve periodically reviewing and consolidating tests like this so that you can identify the underlying patterns and redesigning tests for those. If you do it correctly, this will still cover those individual edge cases, but prevent the test suite from being purely tactical. Instead, you'll have a strategic test suite that validates the fundamental behaviors and invariants of your system, making it easier to understand, maintain, and evolve your codebase without being held hostage by brittle, incident-driven tests.

the strategy of adding a specific test whenever a bug is found in prod is a terrible strategy for testing you end up with a huge suite of one-off edge cases, often hardcoding specific behaviors that you want to be able to change but can't because tests fail in all kinds of weird ways it's fine temporarily if you want to make sure it doesn't happen, but you should revisit it and take a more cohesive approach of testing all behaviors and combinations of them you wouldn't expect. reading tests shouldn't read as if you're reading a journal of incidents

@ChShersh That's so dumb. I love it.

God is 66.7% Go and 33.3% D

Hey there, working on Basil. It's an AI assistant that's always there when you need it, embedded in your Mac - you can engage it with a hotkey or a wake word ("Hey Basil"). You talk to it naturally, tell it what you need in the moment, and it handles it - emails, files, browser actions, research - without you having to stop what you're doing. It's less of a tool you switch to and more of a presence that sees your screen and executes, so you can stay focused on the problem in front of you. basil.ac In the demo here, it doesn't actually have built-in tooling for this set of operations. It figures out how to do what I'm asking on the fly. The current version is 100% free for pretty much life if you download now, because I just really want people to be excited, and get their feedback about what isn't doing it for them. And because let's be honest, I'm never going to patch the build pipeline to exclude broadcasting updates to early version.

If you don't promote your product, nobody else will. 📢 Share your App, SaaS, or Product below 👇

@brankopetric00 The real hack to perfect security is having few users. 😂

SSH keys expire. Passwords get leaked. The only secure authentication is hoping nobody finds your server.

@thdxr @LukeParkerDev So I understand you're the owner of a new kidney.

@LukeParkerDev ??? i'm pretty sure i paid for it

i had to sell my kidney to pay for opus 4.7 today

Hey there, I'm working on Basil. It's an AI assistant that's always there when you need it, embedded in your Mac - you can engage it with a hotkey or a wake word ("Hey Basil"). You talk to it naturally, tell it what you need in the moment, and it handles it - emails, files, browser actions, research - without you having to stop what you're doing. It's less of a tool you switch to and more of a presence that sees your screen and executes, so you can stay focused on the problem in front of you. basil.ac In the demo here, it doesn't actually have built-in tooling for this set of operations. It figures out how to do what I'm asking on the fly. And right now I'm working on giving you the ability to schedule agents, either one time or on a recurring basis, to do things like clean up your emails every day.

What are you guys building this week? Let's drive some traffic 👇

@icanvardar Can, what you think we're all doing here?

selling is much harder than building and most builders only realize it after nobody shows up

@buildwithshyam basil.ac - Basil - Meet your new minion.

It’s Friday 👀 What are you building right now? Describe it in 5 words max + link I’ll rate it out of 10 😀

@plainionist Because there's no bigger hit of dopamine than going into a demo blind and having a brand new, untested functionality work. (jk - obviously TDD is correct)

Why aren’t you using test-driven development (TDD), knowing it results in fewer bugs and better design?🤔

You're 100% right that we need to avoid "agentic everything" hype and apply good engineering judgment. And those two questions are excellent filters. I'd add that I think the most powerful applications will be those that intelligently use both approaches in complementary ways. I've been building my Basil on this exact principle: make AI available when needed for tasks that genuinely require reasoning and natural language understanding, but default to deterministic code for tightly defined operations where the logic isn't ambiguous. The key is knowing when to use which tool. Code provides the scaffolding-reliable, predictable guardrails that keep operations on track. AI agents fill in where you need flexibility, judgment, or natural language interaction. When you combine them thoughtfully, the structured code acts as a framework that channels agent behavior productively rather than letting it drift. So I completely agree with your filters, and I'd argue the real engineering challenge isn't choosing "code OR agents", but rather architecting systems that use both appropriately so they amplify each other's strengths.

"Agentic everything" is the new "microservices everywhere". Now, everyone wants everything to be agentic. (I've even seen people call a for loop "agentic"... like, seriously, dude?) We'll need to get back to common sense and good engineering principles. Two questions that help me differentiate the honest from the hype: 1. Is English the core interface of your app? If you can rewrite the entire flow using code, then you should use code. 2. Does your app require reasoning, or just execution? If you need execution, you should write reliable functions, not use LLMs.

Totally get the frustration, but there are actually better options available now. Postgres does support binary protocol (not just strings), and embedded databases like SQLite or DuckDB run in-process with zero network overhead. There are also solutions like Apache Arrow Flight SQL that use columnar binary formats for high-performance transfers. You could also explore dbs with native language bindings that eliminate serialization costs entirely. The tech exists-it's just about choosing the right tool for the use case. But, like, the data does at some point have to be stored somewhere. 🤷🏼‍♂️

I never liked databases as an idea. You literally send a string query over TCP to postgres and it returns you data over tcp as strings. There is so much potential to make this whole thing better …. but everyone seems to be just fine with it.

Schrodinger's Pipeline: it exists in a superposition of working and broken states until observed in production, at which point it collapses into 'broken' with 100% certainty. (But jokes aside, this is why IaC exists. Document your infra, use Terraform/Pulumi/CDK/Ansible/whatever, and treat prod config like a crime scene: preserve it, version it, replicate it. Analyze it in a lab later, if you're into that.)

The pipeline passed locally. The pipeline passed in CI. The pipeline will fail in production for reasons that defy the laws of physics.

@xeophon @htihle Hey hey hey. Slow down there. Vision instruct.

@htihle adaptive, max, xhigh, no thinking maybe we should just go back to instruct models. simpler times. was just CoT or no CoT prompting

Claude Opus 4.7 (no thinking) scores 76.4% on WeirdML, right behind gpt 5.4 (xhigh) at 77.7%, Opus 4.6 (adaptive) at 77.9% and gpt 5.3 codex (xhigh) at 79.3%, using an order of magnitude less tokens. This looks like a major step forward, things are moving fast now! Results with thinking next week.