HNHacker News
TopNewBestAskShowJobs

nijave

2,500 karma · joined August 22, 2018

[ my public key: https://keybase.io/nickv; my proof: https://keybase.io/nickv/sigs/qMdHcggaKRYKr002ExXFme5MjZeOQbHt-PyAl_07HDk ]
submissionscomments
nijave··on TurboKV: Insanely fast Rust key-value store
I give a little leeway to distributed systems that replicate and don't flush since there's a bit of middle ground assuming they're in different fault domains. Garage object storage defaults to that

However, this doesn't appear to be the case here...

Unsurprisingly, performance goes to crap when sync is enabled.

This is pretty old now but has some useful fsync/sec numbers which can be completely divorced from other I/O performance https://www.percona.com/blog/fsync-performance-storage-devic...

nijave··on GLM-5.3-Flash
That is my understanding as well. Not a lawyer but recently looked into it since we have US, EU, and non-EU European employees and we were looking at token usage tracking tools (which would appear to potentially fall under employee surveillance or at the very least require explicit consent)

Not sure the exact legalese but legal put together a consent form for anyone wanting to do the PoC

nijave··on GLM-5.3-Flash
Z.ai doesn't claim ownership either.

>you retain all rights, title, and interest in the Prompts ...

However, it does grant them an irrevocable license to do, effectively, anything with the data

Not a lawyer but I think the practical difference is you can still license to others, sell, sue IP infringers

nijave··on GLM-5.3-Flash
Depends on the benchmark but yes. I think Opus is more heavily optimized for coding. On the usability side, its output is almost intolerable to read. It seems to code fairly well. Fable is more enjoyable to use for planning/interacting with
nijave··on Show HN: GlassBox – what the browser reveals, and how identifiable you are
Could be bad IP rep. Cloudflare isn't about uniqueness, it's about a reputation score. Presumably some overlap but not necessarily identical.

I started getting those more often when I moved 2 years ago and switched from Charter cable to AT&T fiber. The same hardware was transported 2 miles away...

nijave··on Show HN: GlassBox – what the browser reveals, and how identifiable you are
Also skeptical. Pixel 7 on Android 15 with patched Chrome incognito on T-Mobile CGNAT showing 1 in 11 billion.

The population on Earth is ~8.3 billion. I find it hard to believe it's that unique. Nothing in the data looked particularly interesting.

Maybe slightly plausible if you take into account the currently assigned IPv6 although that's largely time bound.

nijave··on GLM-5.3-Flash
Big "depends". Ali has "ships from China" and "ships from US" stuff. In a lot of cases, though, Amazon also tends to have cheap Chinese knockoffs for a comparable price (although sometimes their product ranking buries them and/or promotes the more expensive knockoffs)

Edit: Yeah I see an ONTi QSFP56 on Amazon for $45, 10Gtek QSFP112 for $62

nijave··on GLM-5.3-Flash
"For new hardware in 2026 with 128Gi of high-speed memory"

Checked a couple days ago and looks like we're at about 3.5x 2020 memory prices (looking at just $/GB).

nijave··on Behaviorally fingerprinting Ox Alpha's provenance
Error messages matching Z.ai GLM I think are the simplest/most compelling. I had Opus 4.8 poke it and it came back with a couple different errors than the article mentions.

Matching the tokenizer is interesting tho

nijave··on Rust Glancer: Rust LSP using 100x less RAM
I would think there has to be some decent middle ground like storing some structures on disk mmap'd and letting the kernel handle back pressure and caching
nijave··on AI companies destroy physical books – let's scan rare books before it's too late
>that is due to the publisher and copyright laws

Agree. While it certainly isn't the most environmentally friendly to render huge stacks of paper into waste, the real issue is copyright creating scarcity (inability to copy the thing).

nijave··on Vomit: Clean up Claude 5's token output with a separate LLM
>Why not just use that other vendor's model for everything?

On the SMB side, you can find yourself with enough money for a Claude subscription (which generally provides a really good $/token value) but limited other options (compliance paperwork, cost, finance, legal)

Personally I wouldn't bother with Anthropic at home but at work it's one of the most cost-effective options that keeps data in the U.S. (which our U.S. customers tend to want)

nijave··on Universal health coverage could save $1T and 114k lives a year: study
This is the crux. Capitalism doesn't work here and a fair bit of people like to pretend it does.

Then there's the people who bemoan "paying for someone else's healthcare" while completely ignoring that's what they're doing in the current system anyway.

nijave··on Universal health coverage could save $1T and 114k lives a year: study
I'm skeptical. It will be cheaper to hire a couple security guards and buy a couple armoured cars then turn away profits.

The reaction was short lived while the insurance companies figured out how to bolster security and return to the usual.

Look at what's happening in California with the proposed wealth tax.

>Page’s co-founder, Sergey Brin, has poured at least $45 million into the Building a Better California Super PAC specifically dedicated to blocking the tax

nijave··on Agent Reach: An open-source CLI that gives AI agents access to the internet
Currently using exa/searxng/scrapling with pretty good success. Has anyone compared these to agent research?
nijave··on Cloudflare OS: an open platform for agents, apps, and work
It has real business value. Letting non technical users run wild without the onerous layers of controls in traditional enterprise IT.

We're already contending with users wanting to hook up every SaaS MCP to every other SaaS platform and then slap AI on top. Having a controlled sandbox for that would hugely simplify things.

nijave··on Cloudflare OS: an open platform for agents, apps, and work
>This doesn't work in the cloud Software-as-a-Service model, because you are not running your own copy of the app.

Let me introduce you to scrapling+MCP--the closest I've come.

Stealthed, controllable browser gives you an API to arbitrary web things

You can't modify it, but you can extend it which imo is the next best thing

adb (Android) is close albeit a bit more tedious. It's fun watching an LLM take screenshots, read them, then send input events back to the phone.

If you have a rooted Android, you _can_ modify the apps. Xposed, ReVanced, et al subject to Play Integrity's ever growing pervasiveness

nijave··on Flowise is shutting down
n8n will do that
nijave··on Flowise is shutting down
I think you tend to need both. There's certain deterministic data and steps that surround the agent as well as certain agentic steps where the LLM is controlling the execution flow.

n8n had an LLM step, not sure if they had an "agent" step. For instance, your workflow could pull down Google Docs using their connectors and deterministic steps then run something agentic with the contents, then upload the contents to Google Drive or send an email.

nijave··on Cursor removed cost information from the usage page and CSV export
There's an env var you can set to tell it when to compact. It's especially helpful for models with big context that lose coherence much earlier.
nijave··on Cursor removed cost information from the usage page and CSV export
On Claude Code you can run /context and it'll explain everything
nijave··on Pgtestdb's template cloning approach to testing is fast
How do you ensure your SQL queries (either raw or ORM generated) retrieve the correct data?

A lot of business apps are mostly SQL with a thin layer of glue to integrate everything.

I think it depends on what the app does--is it data manipulation heavy (lots of complex relations) or is it compute/algorithm heavy (simpler relations, lots of logic).

nijave··on Pgtestdb's template cloning approach to testing is fast
Django has built in support for templates https://docs.djangoproject.com/en/6.0/ref/databases/#test-da...

I thought it had a way to do concurrent sharded tests on the same database instance by automatically creating a templated db per process then each process using transactions. I'm having a hard time finding the docs tho

nijave··on Ask HN: I still don't understand why AI agents need "skills"
To add to what the other person said, they're also loaded differently than arbitrary docs.

They're presented as a list of titles and descriptions to the LLM and it can pick which ones seem relevant and load them.

Markdown docs the LLM needs to make a tool call to read the files and potentially multiple tool calls to discover data if there's a hierarchy. Additionally, it now has to decide if it should read the entire file or cherry pick parts of it.

Skills organize that in an atomic and composable way.

I think ChatGPT and Claude skills will also package arbitrary artifacts like scripts as well. Not sure if that's standard or not

nijave··on Stacked PRs are now live on GitHub
If there's 1 developer working on 1 branch at a time: main is at commit A, you branch add B and C, you merge now main is at A, B, C. Your "integration" was just moving the code forward and if you ran tests on A, B, C commits and it passed, the tests would still pass on main.

Now say there's 2 developers that branch off A. Dev 1 removed dead code coolUtil and with it deleted module coolLib and commits B1. Tests pass, code works. Dev 2 adds new coolUtil2 in commit B2. Tests pass, code works. Now they both merge their changes back to branch A so it's A, B1, B2 or A, B2, B1 or A, B3 (merge of B1/2). Anyway, you try to run the code and it fails. Everything merged clean but Dev 1 removed the lib Dev 2 depends on. They didn't create any file level conflicts because they edited different things so the merging went fine however now the code is broken.

The real integration part comes from reconciling different (potentially conflicting) changes and making sure the end result still works. The server is doing the integration automatically and verifying it automatically

You could skip the server and merge everything on your machine and run the tests on your machine and achieve the same result. Still CI, no server, no automation

Hopefully that makes sense, if not I think the Wikipedia page "Practices" section does a decent job separating the pieces https://en.wikipedia.org/wiki/Continuous_integration

nijave··on Run Kimi K3 using 29 GB of RAM at 0.50 tok/s
Opportunity cost
nijave··on AWS Secrets Manager Terraform: Least-Privilege Access
If you can modify role IAM, I'd start with SSM Parameter Store + KMS encryption for secret values. That is generally much cheaper since typical use fits in the free tier.

They also recently added resource policies to parameters but I don't you have to opt in to the paid tier of offering--not sure how cost compared to Secrets for that.

nijave··on Stacked PRs are now live on GitHub
>Can't you run your CI locally?

If you mean automated tests and linters, sure.

Conceptually continuous integration is generally integrating 2+ different lineages of code together which is more common with multiple developers although I suppose it's becoming more relevant with agents creating a bunch of worktrees with different things.

In practice, CI has taken the same path as "DevOps Engineer" ie most people just mean "automated test server"

The opposite of CI is more-or-less merge windows or merge-fest like Linux where everyone mails in their changes and someone manually integrates.

nijave··on Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac
>So, if software knows in advance which data chunk (expert) it'll need for the next token, it can load that in parallel with computing current token.

Yeah, I was thinking MADV_WILLNEED might work there but not sure

nijave··on Keychron announces first open-source firmware for gaming mice
Iirc there are no release tags, they build off the tip of the branch

Actually it's their zmk fork, not qmk https://github.com/Keychron/zmk/tree/rtl8762g

← PreviousPage 6 of 34Next →