Interesting - this seems to target a different layer than services like Tinker (https://thinkingmachines.ai/blog/announcing-tinker/). Monarch provides the infrastructure primitives while Tinker is a managed finetuning service. Could someone build something like Tinker on top of Monarch?
Nice, so the open source equivalent now exists. Meta basically commoditized Tinker's($12B valuation) value prop by giving away the infra (Monarch) and the RL framework (TorchForge). Will be interesting to see how a managed service competes with free + open source at this layer.
Can someone help me understand the pricing of zed? $10 per month- $5 credits for AI credits. This credits can be used for claude code / codex inside zed or should I manage different api keys for codex/claude code?
There are 2 modes of operation - an editor AI mode and a dedicated agent mode. For the agent mode like Claude Code or Codex, you don’t have to pay Zed, only the CLI tool providers. The zed subscription is for those who don’t want to deal with AI vendors, cli tools etc., and just use it in the editor
You can’t compare these with regular VM of aws or gcp. VM are expected to boot up in milliseconds and can be stopped/killed in milliseconds. You are charged per second of usage. The sandboxes are ephemeral and meant for AI coding agents. Typical sandboxes run less than 30 mins session. The premium is for the flexibility it comes with.
I think you can absolutely compare them and there is no added flexibility, in fact there is less flexibility. There is added convenience though.
For the huge factor in price difference you can keep spare spot VMs on GCP idle and warm all the time and still be an order of magnitude cheaper. You have more features and flexibility with these. You can also discard them at will, they are not charged per month. Pricing granularity in GCP is per second (with 1min minimum) and you can fire up firecracker VMs within milliseconds as another commenter pointed out.
Cloudflare Sandbox have less functionality at a significantly increased price. The tradeoff is simplicity because they are more focused for a specific use case for which they don't need additional configuration or tooling. The downside is that they can't do everything a proper VM can do.
It's a fair tradeoff but I argue the price difference is very much out of balance. But then again it seems to be a feature primarily going after AI companies and there is infinite VC money to burn at the moment.
I coud easily spin-up a firecracker VM on-demand and put it behind an API. It boots up in under 200 milliseconds. and I get to control it however I wish to. And also, all costs are under my control.
In my case, it is ignorance. I am not familiar with how to wield firecracker VMs and manage their lifecycle without putting a hole in my pocket. These sandbox services(e2b, Daytona, Vercel, etc.) package them in an intuitive SDK for me to consume in my application. Since the sandboxing is not the main differentiator for me, I am okay to leverage the external providers to fill in for me.
That said, I will be grateful if you can point me to right resources on how to do this myself :)
This is a pretty good use-case for an open-source project then.
For guide, just follow their official docs. I did those again today, literally copy-pasted shell commands one after the other, and voila.. had firecracker vm running and booting a full-fledge ubuntu vm.
It was sooo damn fast that when it started, at that moment I thought that my terminal had crashed because it's prompt changed. But nop. It was just that fast that even while literally looking at it I was not able to catch when it actually did boot-up.
By the way, two open-source projects already exist:
Lean has been under development over the last 13 years, part of that while chief architect Leo de Moura was employed by Microsoft Research (he's now at AWS). However, Lean is an open source project, not exclusively a Microsoft project. More accurately, see here: https://lean-lang.org/
There were other HN posts suggesting BMAD, ccpm, conductor, etc. I considered giving it a try. They were quite comprehensive, to the point where I was exhausted reading all the documentation they’ve generated before coding - product requirements, epics, user stories/journeys, tasks, analysis, architecture, project plans.
The idea was to encapsulate the context for a subagent to work on in a single GitHub issue/document. I’m yet to see how the development/QA subagents will fare in real-world scenarios by relying on the context in the GitHub issue.
Like many others here, I believe subagents will starve for context. Claude Code Agent is context-rich, while claude subagents are context-poor.
Unsolicited advice: Why doesn’t open router provide hosting services for OSS models that guarantee non-quantised versions of the LLMs? Would be a win-win for everyone.
Would make very little business sense at this point - currently they have an effective monopoly on routing. Hosting would just make them one provider among a few dozen. It would make the other providers less likely to offer their services through openrouter. It would come with lots of concerns that openrouter would favor routing towards their own offerings. It would be a huge distraction to their core business which is still rapidly growing. Would need massive capital investment. And another thousand reasons I haven't thought of.
I use Roo code with orchestrator(Boomerang) mode which pretty much has similar workflow. The orchestrator calls the architect to design the specs, and after iterating and agreeing on the approach, it is handed over to Code mode to execute the tasks. Google Gemini 2.5 pro is pretty good at orchestration due to its 1M context and I use claude sonnet 4 for code mode.
What else does Kiro do differently?
Edit: The hooks feature looks nifty. How is the memory management handled? Any codebase indexing etc? Support to add external MCP servers like context7 etc?
I’m exploring two different applications of AI for education and skill-building:
1. Open-Source AI Curriculum Generator(OSS MathAcademy alternative for other subjects)
Think MathAcademy meets GitHub: an AI system that generates complete computer science curricula with prerequisites, interactive lessons, quizzes, and progression paths. The twist: everything is human-reviewed and open-sourced for community auditing. Starting with an undergrad CS foundation, then branching into specializations (web dev, mobile, backend, AI, systems programming).
The goal is serving self-learners who want structured, rigorous CS education outside traditional institutions. AI handles the heavy lifting of curriculum design and personalization, while human experts ensure quality and accuracy.
2. Computational Astrology as an AI Agent Testbed
For learning production-grade AI agents, I’m building a system that handles Indian astrology calculations. Despite the domain’s questionable validity, it’s surprisingly well-suited for AI: complex rule systems, computational algorithms from classical texts, and intricate overlapping interpretations - perfect for testing RAG + MCP tool architectures.
It’s purely a technical exercise to understand agent orchestration, knowledge retrieval, and multi-step reasoning in a domain with well-defined (if arcane) computational rules.
- Has anyone tackled AI generated curricula? What are the gotchas?
- Interest in either as open-source projects?
I've been personally working on AI generated courses for a couple of months (probably will open source it in 1–3 months). I think the trickiest part that I haven't figured out yet is how to kind of build a map of someone's knowledge so I can branch out of it, things like "have a CS degree" or "worked as a Frontend Dev" is a good starting point, but how to go from there?
I really like how Squirrel AI (EdTech Company) breaks things down — they split subjects into thousands of tiny “knowledge points.” Each one is basically a simple yes/no check: Do I know this or not?
The idea makes sense, but actually pulling it off is tough. Mapping out all those knowledge points is a huge task. I’m working on it now, but this part MUST be open source
btw, feel free to email me to bounce ideas or such (it's in my bio)
