Not familiar with this particular library, but similar libraries use S3 Express One Zone which has write latency <10ms, so you can use that for the WAL and compaction can move data onto other storage classes in the background.
Regular S3 has write latency 100-150ms, which might be fine depending on your workload anyway.
You nailed it! With some of the similar products I've seen, they either inherit the Kafka protocol and hence the KRaft and other complexities, or go the other way with single-node only deployments, commonly just using SlateDB's single-writer model for durability.
Frontier model providers (Anthropic et al.) gate new accounts to impractically low rate limits and spending caps until clients unlock higher limits with cumulative usage, make equivalent cash deposits up front, or talk to a Sales department to work out some other arrangement. They are handicapped by the postpaid billing model.
It's an administrative burden to sign up very every model provider, and there are many independent inference providers now that serve only open source models.
OpenRouter provides a useful service by allowing easy prepaid model access with much higher rate limits, and they also aggregate different model providers to route queries by price, latency, etc.
How many Kubernetes administration headaches trace back to the need for automated systems to surgically edit YAML? It’s absurd and YAML may be the worst choice for this use case.
If you are paying API rates (not using Max subscriptions) there's no reason to use Anthropic's API directly, the same models are hosted by both AWS and Google with better uptime than Anthropic.
How do things like prompt caching etc play into that? Would I theoretically have a more stable harness backing my usage?
Im seriously over the current claude experience. After seemingly fixing my 4.6 usage by disabling adaptive thinking and moving to max effort, it seems that the release of 4.7 has broken that workflow and Im 99% certain that disabling adaptive thinking does nothing even on 4.6 now. Just egregious errors in 2 days this week after coming back from vacation.
Im looking at moving to Pi and I like the minimal nature, but I disagree with a handful of decisions they make. So Id likely need to maintain a fork which is less than ideal.
What decisions is Mario making that you disagree with? My impression is Pi is minimal so any changes can live on top of Pi without needing to maintain a fork?
I started developing my own coding agent after using Pi for a couple months, so I’m curious what you don’t like about pi.
When I hear Mario talk about pi and his approach I find myself agreeing with a lot of it. But I also find myself agreeing with a lot of the points from this https://www.thevinter.com/blog/bad-vibes-from-pi
the opinions in question are that bash should be enabled by default with no restrictions, that the agent should have access to every file on your machine from the start, and that npm is the only package manager worth supporting. Bold choices.
To save others a click, though the article is worth reading.
He also mentions no subagents by default in pi as well.
That (and oh-my-pi) seem like an excessive swing in the other direction. Im all for the simplicity and minimalism of pi. There are just a few fundamental things that need updated (mainly subagent context and open-by-default security model).
yup thats mine. :)
i actually had some stuff layered into mono pi, and i frankly hit my limit in terms of architecture issues in monopi, omp aka oh my pi is frankly better architectured. if you pared back the fearure set to be minimal, you would full stop have a better designed minimal harness.
i do have a proper next gen no slop harness in the work.
amusingly , dog fooding existing tools with my improvements layered in, has repeatedly validated my design choices and if anything has reduced my tolerance for the errors that seem to happen in vanilla or first party harnesses
pi for the win, i have my own ai extend it when i want more specific features. vibe coded in 20 minutes shift+tab like claude code to add permission control.
I find it so funny that many of these harnesses sound like black magic and are completely mystical to me. I use Claude Code every day and yet i can't imagine the workflow of Pi. I also don't care to pay API rates just to experiment with them.
Largely though i'm happy with Claude Code w\ IDE integration, so i don't feel the need to migrate. Nonetheless i'm curious.
You may want to optimize the content serving a bit, since it's currently hotlinking multiple large (30MB) videos at 2K resolution from https://svs.gsfc.nasa.gov.
Yes, you're right about optimization, this what I'll do:
- Switch default to 1024p instead of 2048p so file drops from ~30 MB → ~8 MB (4× smaller)
- Proxy through existing Cloudflare Worker with edge cache
- Add a /video/* route that fetches NASA URL once, caches at the edge
- After 1st request per region, every subsequent visitor gets it from Cloudflare PoP
= NASA bandwidth: ~50 hits/day instead of ~10,000! NASA load drops by ~99%
Sonnet/Claude Code may technically be "smarter", but Qwen3-Coder on Cerebras is often more productive for me because it's just so incredibly fast. Even if it takes more LLM calls to complete a task, those calls are all happening in a fraction of the time.
We must have very different workflows, I am curious about yours. What tools are you using and how are you guiding Qwen3-Coder? When I am using Claude Code, it often works for 10+ minutes at a time, so I am not aware of inference speed.
You must write very elaborate prompts for 10 minutes to be worth the wait. What permissions are you giving it and how much do you care about the generated code? How much time did you spend on initial setup?
I‘ve found that the best way for myself to do LLM assisted coding at this point in time is in a somewhat tight feedback loop. I find myself wanting to refine the code and architectural approaches a fair amount as I see them coming in and latency matters a lot to me here.
2 minutes is the worst delay. With 10 minutes, I can and do context switch to something else and use the time productively. With 2 min, I wait and get frustrated and bored.
Context switching makes you less productive compared to if you could completely finish one task before moving to the other though. in the limit an LLM that responds instantly is still better.
> Sonnet/Claude Code may technically be "smarter", but Qwen3-Coder on Cerebras is often more productive for me because it's just so incredibly fast.
Saying "technically" is really underselling the difference in intelligence in my opinion. Claude and Gemini are much, much smarter and I trust them to produce better code, but you honestly can't deny the excellent value that Qwen-3, the inference speed and $50/month for 25M tokens/per day brings to the table.
Since I paid for the Cerebras pro plan, I've decided to force myself to use it as much as possible for the duration of the month for developing my chat app (https://github.com/gitsense/chat) and here so some of my thoughts so far:
- Qwen3 Coder is a lot dumber when it comes to prompting as Gemini and Claude are much better at reading between the lines. However since the speed is so good, I often don't care as I can go back to the message and make some simple clarifications and try again.
- The max context window size of 128k for Qwen 3 Coder 480B on their platform can be a serious issue if you need a lot of documentation or code in context.
- I've never come close to the 25M tokens per day limit for their Pro Plan. The max I am using is 5M/day.
- The inference speed + a capable model like Qwen 3 will open up use cases most people might not have thought of before.
I will probably continue to pay for the $50 dollar plan for these use cases.
1. Applying LLM generated patches
Qwen 3 coder is very much capable of applying patches generated by Sonnet and Gemini. It is slower than what https://www.morphllm.com/ provides but it is definitely fast enough for most people to not care. The cost savings can be quite significant depending on the work.
2. Building context
Since it is so fast and because the 25M token limit per day is such a high limit for me, I am finding myself loading more files into context and just asking Qwen to identify files that I will need and/or summarize things so I can feed it into Sonnet or Gemini to save me significant money.
3. AI Assistant
Due to it's blazing speed, you can analyze a lot data fast for deterministic searches and because it can review results at such a great speed, you can do multiple search and review loops without feeling like you are waiting forever.
Given what I've experienced so far, I don't think Cerebras can be a serious platform for coding if Qwen 3 Coder is the only available model. Having said that, given the inference speed and Qwen being more than capable, I can see Cerebras becoming a massive cost savings option for many companies and developers, which is where I think they might win a lot of enterprise contracts.
SlateDB offers different durability levels for writes. By default writes are buffered locally and flushed to S3 when the buffer is full or the client invokes flush().
Regular S3 has write latency 100-150ms, which might be fine depending on your workload anyway.
reply