mayank

If you’re doing breakeven math on subscriptions, consider that your own rig can run 24/7 whereas you will get a fraction of that with sub rate limits. Even if you factor in PG&E residential rates, the breakeven is a lot closer to months for overnight long-running agentic coding a couple times a week.

And in terms of interesting use cases: recently pointed an agent at Blender and gave it vision. That setup can essentially iterate on a scene forever.

I assume this is a strictly B2B proposition in spite of "built for every side of the GPU market" -- meaning I'm probably not going to be able to buy/bid on small (e.g. 2-8) lots of used, matched RTX 3090s for example?

Close to $500 in overage for me :) I’m going to give GLM a shot next.

Oddly enough, I did the exact same thing this past weekend, for a couple hundred usd in Fable overage.

It was truly remarkably easy to build and package it exactly to my whims, in my case as a single docker container with a process reaper that runs llama, my Go code, tts, chat harness, browser in xvfb, and even a mailer daemon. With Gemma, it even runs on a RPi 5.

What’s absolutely wild to me is that over a couple hours, I could probably have it import parts of Home Assistant for my devices directly, and do other wacky stuff in what is essentially software for one.

I was curious about the cost angle too, i.e. how much "free" coding agent I can get for what cost. Here's the research by Fable if you're interested: https://claude.ai/public/artifacts/2c9a5001-0b7e-4944-beb1-9...

Spent nights over the last six months combining AI coding agents with containers, spec-driven development, and formal verification, so that I (eventually) don't have to manually review 10k+ lines of AI code a day, or worry about the latest model wiping out my home directory:

https://www.overplane.dev/ (Apache-2.0)

TL;DR: Overplane is an AI build system that pairs popular AI coding agents with containers, spec driven development, and automatic lightweight formal verification for a safer, replayable AI codegen workflow.

Like many others, I’ve watched AI code pile up at work, and get increasingly lighter reviews due to its size and maddening uniformity. On my side projects at night, I see the full speed and fragility of vibe coding. Neither practice seems sustainable, so I’ve been playing around for the last few months looking for a balanced, middle ground.

The key idea here is similar to chip design: if AI written software functionally does what you intended it to (by virtue of increasingly sophisticated testing against a codification of your intent), then it’s perhaps less important to manually review every line of generated code.

Overplane is a labor of love that brings that idea to life as an open-source experiment, combining some old personal loves: containerization, content addressed build systems, and formal verification.

The example I’d start with is rustdis, a partial, wire-compatible Redis clone in Rust with an empty [dependencies] section. Created with seven short specs and about $45 of Claude Opus, in about three hours: https://www.overplane.dev/examples

What sold me on the approach is that the one IR generated by Overplane from the specs paid for itself downstream: Z3 checks at build time, 48 generated proptest properties, and Kani proofs on the parser and arithmetic core.

It also holds up against redis-benchmark better than I expected, within about 90% of real Redis on my box unpipelined and a bit ahead on some pipelined workloads, which I mostly attribute to rustdis doing less than Redis does.

It’s v0.0.8 and rough. If you’ve tried spec-first workflows or lightweight formal methods in anger, I’d love to hear where they broke down for you.

Such a great comment, and I agree with all of them.

For me in a similar vein:

- mar ‘24: thinking about how to survey the field and implement a hard research task in Natural Language Processing, and then just approximating it well enough with a prompt and a completions api

- mid ‘25: Llama 3 being able to analyze a good sized codebase I was onboarding onto, and synthesize it into diagrams that matched the quality of ones I’d generated by hand with deterministic tools.

- dec ‘25: opus 4.5 basically generating multi-class modules and tests perfectly (syntactically). Finding that errors were my own under-specification of the prompt. Stopped writing code by hand, mainly because it was good enough and came with tests, docs, build scripts, and other goodies for free.

Indeed! It’s a form of a branch table: https://en.m.wikipedia.org/wiki/Branch_table

Got a source for this? Because it’s brilliant.

Agreed and flagged, this almost seems LLM written, not a bit of data.

This seems odd. If your scribe can lie in complex and sometimes hard to detect ways, how do you not see some form of risk? What happens when (not if) your scribe misses something and real world damages ensue as a result? Are you expecting your users to cross check every report? And if so, what’s the benefit of your product?

> They need to reword this. Whoever wrote it is a liability

Sounds like it’s been written specifically to avoid liability.

Very interesting! I wonder to what extent this assumption is true in tying completions to traditional code autocomplete.

> One of the biggest constraints on the retrieval implementation is latency

If I’m getting a multi line block of code written automagically for me based on comments and the like, I’d personally value quality over latency and be more than happy to wait on a spinner. And I’d also be happy to map separate shortcuts for when I’m prepared to do so (avoiding the need to detect my intent).

> aws s3 bucket needs to match the domain for website hosting.

This is outdated information, and not required anymore when using CloudFront.

And even in the past, you could use the S3 API to implement a reverse proxy without matching bucket and domain names.

> the minimal runtime is typically the best estimator

Depends what you’re estimating. The minimum is usually not representative of “real world” performance, which is why we use measures of central tendency over many runs for performance benchmarks.

This is a pretty standard measure called the Trimmed Mean: https://statisticsbyjim.com/basics/trimmed-mean/

Very debatable, if both are done right.

It is in industry, but you may be shocked if you read “research code”

> You don't actually write code with e2b. You write technical specs and then collaborate with an AI agent.

If I want to change 1 character of a generated source file, can I just go do that or will I have to figure out how to prompt the change in natural language?

> Because there are many services (each with their own readiness criteria), a cold boot takes a long time until all services stabilize (worst case I've seen was over 30 minutes). With hibernation they can resume where they started off within a couple of minutes.

This is exactly what we use hibernation for in conjunction with EC2 Warm Pools -- fast autoscaling of services that have long boot times. There's an argument to be made that fixing slow boots should be the "correct" solution, but in large enough organizations, hibernated instances are a convenient workaround to buy you some time to navigate the organizational dynamics (and technical debt) that lead to the slow boot times in the first place.

This is a wonderful article, architecture, and project. Can anyone from Clickhouse comment on any non-technical factors that allowed such a rapid pace of development, e.g. team size, structure, etc.?

All modern languages heavily borrow from each other’s latest iterations. In the case of Java though, playing catch up is by design since it’s intended to be a conservative/stable language.

Wow....confirmed!

`{"errors":[{"message":"Your current API plan does not include access to this endpoint, please see https://developer.twitter.com/en/docs/twitter-api for more information","code":467}]}`

Can you comment on a reply below that claims there was promotional pricing in 2019 at launch?

https://news.ycombinator.com/item?id=33929132

Exactly! Hashcash was proposed in 1997: https://en.m.wikipedia.org/wiki/Hashcash similar mechanics for a different use case

> - /organizations/:id

> - /blogs/:id

The pragmatic, large-company-only counterpoint is the narrow edge case where:

- :id must be human-readable for "SEO reasons"

- there are many competing organizations and blogs to the point where there may be a name collision.

Although in that case, I'd still suggest:

/:organization-name/:blog-name

There's a hierarchical modeling paradigm/tools called C4 that (while being boxes and lines) helps with the zoom-in/zoom-out nature of understanding systems: https://c4model.com/