iiinova — where your digital self keeps evolving

Is there enough data for a pretty post lifetime vs probability of like graph yet?

Ok, *very* simplified, maybe possibly slightly over-simplified. I think it conveys the idea I want to communicate though, so it works imo Also, I blame the Bluesky post character limit and time of day

010

This is nowhere close to everything I want to write about LLMs and genAI, but it's kinda 4am... NOTE: *I'm not a machine learning expert.* The simplifications are what I arrived at based on other less simplified but also not totally accurate or full explanations. This might be ass and/or false.

000

LLM as a smart(er) auto-complete. No, I'm not joking. This is about the level of responsibility it can handle (pretty "ok" at it). Propagating a small change with a clear pattern? Do once, TAB for the rest - an actual time save and convenience. Is that vibe-coding? I don't think so. LLM use? Yeah.

But there are other uses for LLMs in a developer's workflow. Doing vibe-coding the proper way still takes time and the result is bad. Time waste for a dev who wants to make good. But what about code review? Almost no time required, yet it has the *potential* to make good discoveries/suggestions.

Of course, if the only thing that matters is shipping that code, to maximize shareholder value, and the business is pushing for "vibe-coded now, fix later/never", then it does make sense to vibe-code. Sacrifice ever achieving good for quickly achieving a working meh most of the time.

100

The value proposition for vibe-coding from a developer's POV just doesn't make sense. Vibe-coding discourages ever going in to do something manually, basically locking them into vibe-coding the codebase, unless a developer wants to waste enormous amounts of time to understand garbage code (hard).

100

And model context plays into the quality too. Asking it to continue low quality text obviously makes low quality output more likely because in training data low quality text is usually surrounded with more low quality text. Having a bad codebase nerfs a model too, no matter how hard they "align" it.

100

In training, model authors can feed it additional copies of what they consider good, making the output lean more towards that (simplified). But since a model is a black box, the approach has downsides. It screws with weights in unpredictable ways, usually decreasing model "stability" (simplified).

100

Average quality of all code ever? Even just all code on GitHub? It's shit. Not extremely bad but pretty damn bad. An LLM can only reproduce that, so the output is obviously low quality code too. Power law, few popular repositories with high quality code vs many many low quality repositories, etc...

100

That is why LLMs can for example reproduce popular Wikipedia articles pretty accurately, sometimes exactly. Wikipedia is frequently quoted and copied, so it occurs pretty frequently in training data, which carves it into the LLM's memory more and more accurately, gives it more "weight" (simplified).

100

So no matter how long and hard you train an LLM, the average output quality is going to hover around the average quality of the training data, maybe slightly above. Model weights are based on frequency (read "popularity") of input text (token chains) (simplified).

I myself have never vibe-coded, my biggest ever LLM usage in relation to code is generating a body of a small function. Vibe-coding from scratch gives shit results for obvious reasons - LLMs just predict the next most likely symbol (simplified) based on their training data. Garbage in - garbage out.

200

It's hard for me to fully comprehend that there are people who believe LLMs are strictly useless in programming and/or strictly harm the quality of code/product. Makes me want to write a blog post about it... Various scattered ideas and notes off the top of my head (*not fact checked*):

120

I want an M5Stack PaperMono... Went immediately out of stock after launch...

Essem :skeeter: @apub.essem.space · 03/09/2026

Reposted by IS2511

i don't think either of those are genuine or authentic actually

A screenshot of the CPUID Wikipedia page. It says that the Microsoft x86-to-ARM emulator reports its ID string as "AuthenticAMD" and that Apple's Rosetta 2 reports its ID string as "GenuineIntel".

713930

IS2511 @is2511.com · 28/08/2026

It could always be clone and delete (clone draft to post, delete draft) under the hood 🤷

100

IS2511 @is2511.com · 28/08/2026

Does ATproto support private data? I thought that was kind of the whole problem with implementing drafts on the protocol 🤔

200

IS2511 @is2511.com · 25/08/2026

Can't I just like the question? It's the like button, I press it when I see something I like 🤷

000

IS2511 @is2511.com · 25/08/2026

Every time I open my Bluesky feed it sends me down a thousand little rabbit holes, each no less than ten minutes long, with one eventually cascading into a full multi-hour web crawling session, just clicking on links and reading, reading, reading... It's 3am, again I'm so tired

040

Reposted by IS2511

Junferno @junferno.bsky.social · 29/07/2026

New video uploaded: What If Computers Were a Gorilla? youtu.be/63QlOBbl6pM

youtu.be

What If Computers Were a Gorilla?

YouTube video by Junferno

2558

IS2511 @is2511.com · 25/08/2026

Throwing seems pretty intuitive, catching seems to require training imo

010

IS2511 @is2511.com · 18/08/2026

The Dhall language has to be an elaborate shitpost. Yes, I probably do want semantic hashing for my configs. Yes, enums sound great. Yes, I might want imports. But the syntax... The syntax choices are just insane to me dhall-lang.org

dhall-lang.org

The Dhall configuration language

Maintainable configuration files

IS2511 @is2511.com · 23/06/2026

Turning on a VPN showed me the welcome banner for the DMs feature and added the appropriate setting in the Privacy menu, which I promptly turned off. So apparently it's a Russia thing, YouTube doesn't show ads at all here and apparently doesn't roll out new data collection features either, huh

000

IS2511 @is2511.com · 22/06/2026

Guess I'll stick to removing the si query param from the URL myself

030

IS2511 @is2511.com · 22/06/2026

If I don't have the option, does it mean the feature hasn't rolled out to me yet? Weird...

120

IS2511 @is2511.com · 14/06/2026

It didn't fix it :(

IS2511 @is2511.com · 12/06/2026

I wanted to test a website behind Cloudflare (kinda blocked in Russia) before posting that the strategy is general/working, so I picked a random domain from my proxy filter (reason note "RU CF ECH ban?") - mpv.io. For some reason it's crazy flaky, refusing connections. And then this... Dial-up fr

000

IS2511 @is2511.com · 12/06/2026

Still, it appears to not work for Telegram for example. A *general* strategy, against *general* blocking methods. I guess for Telegram I would need something more custom... Probably not worth it since my main zapret2 instance is for LAN, but Telegram's main use-case is communication from anywhere.

100

IS2511 @is2511.com · 12/06/2026

After reading the documentation manuscript, peeking at other configurations and lots of trial and error I think I might have finally stumbled into a general working zapret2 strategy. It's my 3rd or 4th try, yet another fresh look finally got me there. The manuscript btw: github.com/bol-van/zapr...

github.com

GitHub - bol-van/zapret2: anti-dpi software

anti-dpi software. Contribute to bol-van/zapret2 development by creating an account on GitHub.

120

IS2511 @is2511.com · 03/06/2026

Main router (OpenWrt) suddenly rebooted into recovery mode without me noticing 4 hours ago. Reboot didn't help. I so don't want to deal with this right now... I have a not entirely up to date configs backup, should've made one yesterday... Uptime was what, half a year without problems? Why now?

IS2511 @is2511.com · 01/06/2026

> But why do you have IPv6 enabled if you ISP doesn't even support it? I like having stuff enabled if it's supported, ok? My LAN is IPv6-capable, even if I don't use it. Maybe I should. Maybe 50 years from now we will all finally migrate to IPv6 and I'll get some practice with it...

000

IS2511 @is2511.com · 01/06/2026

Maybe I should set a static IPv4 address... The DHCP lease is static anyway. The IPv6 recovery seemed quick (SLAAC perhaps), especially compared to an IPv4 re-acquisition, probably bc of all the DHCP yap. I think the instant IPv6 re-acquisition made it skip IPv4, causing permanent connectivity loss.

100

IS2511 @is2511.com · 01/06/2026

I think I figured out a fix (pending confirmation). I'm not sure what is causing the link downs, probably some hardware reliability issues at 10G, but I'm fine with it as long as it can recover, which it failed to do. The fail state was link up with only an IPv6 address. I present to you the fix:

Screenshot of NetworkManager GUI, editing a connection, IPv4 Settings tab, Method set to "Automatic (DHCP)", mouse hovering over the checkmarked option "Require IPv4 addressing for this connection to complete", tooltip shown "When connecting to IPv6-capable networks, allows the connection to complete if IPv4 configuration fails but IPv6 configuration succeeds.".

Optimizing away the creative process ❤️

030

(bonus) Also, these cheap SFP+ 10G RJ-45 modules lie to the switch for some reason lol. They report being LC when they are clearly not. Wavelength? Bro, it's copper.

Screenshot of WinBox, crops of two Interface windows on SFP tab, sfp-sfpplus5 (PC) on the left and sfp-sfpplus8 (uplink) on the right.

Relevant differences between tables
Connector Type: "copper pigtail" | "LC"
Encoding: "" | "64B/66B"

It works now so I give up. Until this problem comes up again it might as well not exist.

And as I discovered in the original post, the switch only officially supports 4 SFP+ to RJ-45 modules (probably bc power), so swapping one of the modules for a DAC is a good idea all around. Now I'm at exactly 4 active modules. Was that the problem? This fixed itself before the swap, so no...? 🤷

(a day later) I was drafting this late yesterday and shelved until today. Now, no issues at all for like 12h+ already. Was it software? Was it an SFP module? I don't think anything changed besides the weather outside... Anyway, now I switched to an SFP28 DAC for better reliability, just in case.

So. Intermittent 1 second long connectivity failures. Affects something between PC <-> switch <-> test-host. Not PC port/NIC, not test-host, these were all swapped. Probably not cables/SFP modules. There's only one problematic part left to test... Yep, failure only happens to the PC. It's software.

Screenshot of two terminals, both show outputs of a ping command. Left is ping from 192.168.1.33 to 192.168.1.34 (via switch plus WiFi AP), right is ping from 192.168.1.10 (PC) to 192.168.1.33 (via switch).