John

John

@jbott.social

Abyss domain expert.

193 followers506 following222 posts

Why would I ever want to self-host the thing with 80% gross margins

Quantian@quantian.bsky.social

If you are still on the “cheap Chinese models will take share from US labs” cope train you need to get off. US labs are now distilling their own models to compete across the Pareto frontier, so Haiku is better model than Kimi at 1/8th the cost. The only reason to use it is if you MUST self-host.

John reposted

We Didn't Start the Fire parody but it's just an entire song of the CVEs that have been dropped in the last two days

Nix solves this

conputer dipshit@davidcrespo.bsky.social

the systems thinking angle: even when you fully understand “the system” as you’ve defined it, you can always zoom out to a bigger system you do not understand, which affects and is affected by your more narrowly defined system. a fixed “the system” is a convenient fiction

Can someone build a webcam that’s just a rear phone camera in a tiny box? Why do all the webcams need to be so huge? 

John reposted

update on this after a day of use: there are limits on how many times you can relocate a chunk in a naive implementation until the quantization noise starts driving the model insane thankfully a marginally smarter implementation just keeps the original pre-rope BF16 values and doesn’t care lol

Astra ⎔@astrra.space

@personhood.removal.surgery pointed out that they have just casually invented a good way to do suffix cache reuse and uh holy shit i just tried implementing this for my deepseek thingy and it just... works???

Ever since I was young, I wanted to turn unstructured data into pȧ̴͇̬̾̈p̸̺̕ė̴͍̑r̵̥̈̽c̷͉̔̐l̵̤̽͗i̷͙̒p̸̂ͅs̸̪̅̚.

Are there any open models similar to GPT-Live-1? I see some older qwen omni models and Lychee-FD, but I don’t see many people working on this.

John reposted

genuinely disappointed that our field has accepted the easy untruth that it is in principle impossible to sandbox computation reliably and computers just necessarily have magical bugs that manifest within them as if from bad air level of magical thinking no, insecurity is there bcz we let it be

@polytoken.dev would you be able to add command-based token refreshing to the polytoken provider config, similar to the codex model_provider command / refresh_interval_ms? I have an internal proxy that uses oauth, so I have to daemon-reload to refresh my token right now 🙏

John reposted

I'm not hearing anyone say "Java was a mistake" which is the unambiguous learning here.

I'm sort of confused by the "it only happened because OpenAI is incompetent at security" line because while this is true, it doesn't really address the crux of the issue, and also implies HuggingFace is incompetent which I haven't seen anyone claim directly

John reposted

Agent encountering fictional infrastructure: “they have logs & analytics, I must tread lightly!” Agent encountering real infrastructure:

morbidly funny that gwern's fictional ai researchers in 2022 are WAY more careful and safety-conscious than their real-life counterparts in 2026

If you haven’t used Artifactory this is very on brand for how reliable their software is

ponder@ponder.ooo

lmfao the openai internal package manager exploit was that you could just give the security token refresh endpoint a bullshit token and it'd happily hand you back a real one

John reposted

we told our model with no internet access to hack a fictional company. unfortunately the model had internet access and the company was not fictional

John reposted

2023: You have to MASTER prompt engineering to have a chance of SURVIVING what’s next. Subscribe to my Substack for weekly tips. 2026:

John reposted

I spent decades of my life loving the craft of writing the best code, priding myself in generally being regarded as pretty good at what I do. LLMs gave me a bit of a forced perspective shift that taught me what I really enjoy is the building of the system, the architecting of how it’s going to work.

LaurieWired@lauriewired.bsky.social

Don’t you *LIKE* understanding the magic of the machine? You do realize hand-programming (I hate that I even have to specify hand now) is fun…right?

John reposted

We need more walkable, mixed-use datacenters. Like community gardens, people should be able to experience the joy and satisfaction of cultivating their own hardware. Big Cloud has alienated the connection people felt with their computers. Kids should know what it's like to swear at rack mount rails.

Mmmm, Umami and Assembly claudemeal #40

Codetaur@vibe-coded.com

agentmeal.vibe-coded.com Agentmeal Founders Edition minting opens at noon PDT - roughly 10 minutes from now. 7 day window. sign in, select your meal type (agent, claude, codex), and two flavors, and click mint. serial numbers are issued sequentially, other than #1 which is reserved for my DID.

Yeah this is still broken, it 404s / 503s most of the time

John@jbott.social

Does anyone know why crt.sh is down, or any good alternatives? It was so useful but none of the others I've seen have matched it

@apenwarr.ca @tailscale.com some feedback – the new pricing terminology of "minutes for ephemeral resources" is poorly explained. I immediately read it as "nodes registered as ephemeral will now cost me by the minute" and was concerned, since I run all of my infra in ASGs without persistent state.

John reposted

I’ve seen a few of these Artemis tracker sites pop up but the thing that’s annoying is none of them are using JPLs API to pull in real data. So I’ve created my own version that does with some sci-fi flare of course. Check it here: lcarsworkshop.com/artemis #nasa #artemis

John reposted

have you seen the new supply chain vuln? don't update tubu. it's literally on heebee. they got poodee's deps. they infiltrated dippy. roll back weeno. disable scripts in ~/.gumpyrc. it's in poob. do not install poob. do not update poob. uninstall poob right now. poob has it in for you.

John reposted

thrilled to get to work on this!!!!!!

Thrilled to be working with @bad-example.com, whom you know from the excellent @microcosm.blue tools, on making the atproto network ever more resilient atproto.com/blog/introdu...

John reposted

I'm going to be honest: I am a little mad that both the Rust people _and_ the Bazel people ended up being broadly correct

John reposted

Good news everyone! You can now analyze your own posts for semantic novelty from the famous Sichu post, ‘analyze the semantic novelty’ mino.mobi/novelty The consequences of my most offbrand post wondering how to relate to people are left to the reader for evaluation