HNHacker News
TopNewBestAskShowJobs

xlii

3,508 karma · joined August 23, 2017

https://xlii.space

e-mail me: alexander (at) kaminski.se

submissionscomments
xlii··on What happens if an entire class of workers loses faith in their careers
Long long time ago people in tech were in tech because of the tech.

Then, at some point, being in tech became a gilded profession and everyone: salesmen, artists, clerks etc. joined in. Not because they wanted to be in tech.

They wanted gold.

For years it was working, but right now things shifted. You can no longer just do the thing you learned for a job and continue doing so, because landscape shifted. So they're miserable and they want to get back to knitting, pottery or whatever thing they want to do but not tech.

For people I know who are in tech for the tech nothing changed. There's new tech in town, one has to adjust, and it's still possible to discover new things and pave new paths. Tech was always shifting and I suppose people who worked in it for it always knew it, and it doesn't derail them.

xlii··on Humans missed 1 in 3 threats approving AI agent commands across 40k game runs
I implemented few agent harnesses (and rik! advertising time: https://rik.axk.sh), and once doing that I noticed one thing:

Context-less self-approval is working well. The failure mode is usually false positives (i.e. safe commands being rejected), not the other way around, with root cause of requesting agent underspecifying context (e.g. not mentioning in the request that it's made on behalf of user etc.)

Thus, I'm running self-approval YOLO modes on state-of-the-art models for quite some time and it didn't bit me. It might, but hey, we're long gone from the age of predictable software development.

xlii··on Zed DeltaDB
I start to sound like a Jujutsu fanboy but...

Jujutsu already does this. I had agent screwing up changeset over couple hour long session and it took one shell command to split on pieces with longer pause.

The only downside is that because of Jujutsu snapshots it can't be used inside sandbox (since all jj commands are mutating)

xlii··on LLMs reward expertise
I consider myself senior engineer. When talking with junior colleagues they often are surprised how little I care about some things and how much I care about others.

These internal "attention weights" are highly influential parameters of how I work with LLM.

E.g. when working with Rust I often hold strict control over structures and lifetimes. But when lately I've been doing token-based bind generation I didn't care about anything outside of high level patterns like RAII and ultimately - API ergonomics which was verified in consumer app.

I've been in position of porting real-code to vibe-code platform and seeing non-technical people prompt-stream (they were shared across accounts) I know why they engaged engineer to run this work.

Their efforts took 6 weeks, I ported app within 4 days and (to be honest with myself) without LLM I that'd be 3M+ work pre-LLM.

In short: I observed same effect as claimed.

xlii··on Agent Skill to Force Docs in ASD-STE100 Simplified Technical English
Ccn't say, though since Opus 5 I'm much more often staring at the wall of text and I have no idea what it's telling me.

WTFs per prompt metric raised sharply for me :)

xlii··on Agent Skill to Force Docs in ASD-STE100 Simplified Technical English
New models (especially Opus 5, released just last week) produces unintelligible output, so people are trying to fix it.
xlii··on French musician Kavinsky found dead
https://youtu.be/MV_3Dpw-BRY
xlii··on Our position on open-weights models
"Bottled-water vendor position on tap water drinking".
xlii··on Decathlon Germany adds Wero payment option to decathlon.de website
There's an interesting angle to BLIK vs Wero. While I believe QR codes are more convenient it is actually possible to give BLIK over the phone. Which means that youngsters (or fraudsters ;)) have access to their guardians' money over the phone line.
xlii··on Decathlon Germany adds Wero payment option to decathlon.de website
They most likely don't charge fees on consumer.

Usually the fee is on merchant (and it's few % depending on region regulations etc.) Payment providers such as Wero are doing direct bank transfer, which is usually fee-less, but then they can afford to put a on-merchant fee that's much lower than what credit card companies are charging.

Thus the model is win-win-win: Consumer doesn't pay more, Merchant pays less, and Wero gets their share.

xlii··on The US is charging an American citizen for wiping his phone at the border
> ...this is not unique to the US by any means

I've searched and found that Australia has similar concept. According to EU laws person on-territory is under jurisdictiction of said territory (even though it might bar from entering, but still EU/local laws apply there) - this is probably why I'm so surprised.

Also regarding this exact nuke-phone-through-distress-password: in many European countries it's perfectly legal to protect oneself from law. While the law is nuanced (i.e. one might not be obligated to provide password, and responsibility for destruction can be put on forces as they could be found at failure when securing device), usually it's fair game.

In the end I just find this terrifying. The idea that I'm getting back from vacations and then I get myself as probed like that. Here, the only things I'm afraid of are... queues.

xlii··on The US is charging an American citizen for wiping his phone at the border
There are so many peculiar things here, I - as not an US citizen - don't understand:

If the border isn't "US soil" then what jurisdiction is it?

How come American forces can seize goods if that's not US? What prevents them from trying the same, e.g. in Europe or Australia?

Also I find the law penalizing people who might deprive state of evidence especially odd. In (AFAIK) most of the Europe there's no self indictment, so destroying any kind of evidence that might get you indicted is fair game and you can't get penalized by such fact alone.

And of all of it the fact that US government is pulling this stuff on US citizens is unbelievable.

No idea, might be only me, but I start to perceive US more and more as a totalitarian state.

xlii··on Opus 5 is currently #1 on Artificial Analysis Intelligence Leaderboard
Not all HN visitors are native English speakers and in some languages "it" doesn't construct well with verbs, thus thought frameworks forms through usage of him/her. Nothing more to see I suppose.
xlii··on ascdraw: Editor for ASCII/UTF-8 diagrams (in 144FPS)
Good idea, haven't thought about this before.
xlii··on ascdraw: Editor for ASCII/UTF-8 diagrams (in 144FPS)
It's infinite canvas, so I'm actually drawing diagrams for code, for LLMs (those are much more receptive to text diagram than screenshots).

Scratchpad is default mode but it can edit text files, or have own files for editing. One more magical use case is filter mode when using CLI agents. Switch prompt to editor, select some text I want to be laid out, I filter through `ascdraw -` which opens editor, let's me move things around, align, add arrows & stuff and then after closing I get selected text in a new form.

xlii··on ascdraw: Editor for ASCII/UTF-8 diagrams (in 144FPS)
Well it isn't that simple :)

It mentions 144FPS, because of scrolling/zooming events.

If there's a key-by-key panning then probably you need 1 frame and you're good. Macbook's Pro touchpads generate events with 150hz - and in order to maintain "butter smooth" FPS has to follow.

Number of cells on screen matters: rendering 80x80 is not a problem. After zooming out and 1000x1000 on screen panning with 150FPS is straightfoward.

In the end there is a deterioration of user experience when jank in scrolling, zooming and even typing happens.

Thus ceiling right now is around 144FPS, mostly because ~6ms budget (on M2 Macbook and 4K 144hz display). 60hz refresh rate 60FPS is an easy mode ;-)

xlii··on ascdraw: Editor for ASCII/UTF-8 diagrams (in 144FPS)
effect of accidental development of a framework for GUI apps and personal sympathy toward diagrams
xlii··on Decoy Font
I've seen the other lately submited text-effect-claim-to-be font and I must say I find this much more amusing.

Humans can squint, AIs can't squint!

xlii··on Ask HN: Has single-task focus become outdated in the AI era?
It's possible to single task using Round Robin strategy.

Have a list of things you do, go one by one over them. If new task appears add at the loop "gap".

I found that the worst thing is chasing the bunny: Oh wow. This finished. I should chboing oh another agent that's probablboing... - completely unsustainable.

The loop is very similar to old way of things. You just don't pay attention to notifications. Also it's worth grouping loop to minimal że context switching or ease context switching (e.g. task on project a, task on project a+b, task on project b, task on project a+b, etc.)

In order to get to looping part I use self-approval modes for agents. It's slightly uncomfortable but I built own agents with own permission reviewers and they are quite good. These can be used to run agents in the background. And if you do a loop and find agent still spinning - it's a good moment to take a five for yourself.

The deep focus is still there but somewhere else. Usually in coordination and integration.

xlii··on Your code is fast – if you're lucky
I ran 10k test locally on 2e5 and I'm seeing 4 orders of magnitude instability, but very high local stability (i.e. runs within specific second are very stable, showing almost no deviation, runs couple second later are the same, but results are within 1 OoM of the prior results (smaller batches, 500 tests).

I'm not saying that optimization isn't valid, what I'm saying is that Quicksort shouldn't be optimized over randomized per run data set.

xlii··on Your code is fast if you're lucky
Is it only me..?

Quicksort is supposed to be an algorithm that has O(n) to O(n²) performance and O(n log n) being only an average performance case. Test was made on random data coming from different archs (so I doubt it's characteristic would be remotely identical).

Given input size of 50M it means that performance could be between 50M (5e7) up to 2.5e15. That's like performance instability of 8 orders of magnitude.

I'm not sure here if we can't write instead that "Your code is fast if you picked fast case for it" especially since fix of 6 OOM is smaller than algorithm's performance range.

xlii··on Good Tools Are Invisible
Obviously Kakoune is much simpler than Emacs, so face variability, very interactive UI is not there (though there'd be nothing preventing writing it, but I doubt anyone wants it enough).

As for transition - I always was somewhat of an UNIX guy, so I replaced Swiper/Occur/Consult with delegating to shell. Kakoune has just enough utilities to create a on-keystroke-updated-buffer so I'm happy with that. In some languages I go as much to create "find functions" special mode - composition with shell is easier than Lisp - I rarely have to read documentation.

For Git I use Jujutsu (so I stopped using Magit long time ago) but Kakoune has a very nice "!commmand<ret>" utilities. It's nothing more than a "C-u M-!", but positioning of feature differs.

So the transition is mainly about delegation, not sticking to one application, but instead finding utility that does it and use that instead.

xlii··on A font that humans can read but AI cannot
Oh, cool I was wondering how can I get to see that decoy!
xlii··on Ghost Font: A font that humans can read but AI cannot
Technically it's not a font, because font needs to be still. Analogy: if I took photo after book was closed would we say that font cannot be read by a camera?

Took a picture (only a single frame) and a 1s movie and threw it toward GPT 5.6 Sol (High):

Frame took 9m30s to decyper and GPT 5.6, it returned: WRITTEN IN GHOST FONT. Weird because I can only see "GHOST FONT" on the demo... but extracted data from image (I saw the highlited one) definitely looks like the "Ghost Font".

--

Video is more amusing, because after 3m GPT 5.6 figured it's motion-defined and asked to run QuickTime. At one moment I got:

> The animation is a motion-defined illusion. I’ve confirmed there’s no readable static OCR layer; I’m decoding its optical-flow field so the letter shapes become explicit.

At 4m it got extracted motion image that was in shape of letters but analyzed for 9 more letters and returned (at 13m36s) "GHOST FONT"

--

So:

    a font...             - FALSE - not a font, but video effect
    ...humans can read... - FALSE - I can't read it from image (but AI can!)
    ...but AI cannot      - FALSE - it can
:D

Edit: https://imgur.com/a/SHlGu4O - work-in-progress images

xlii··on Documentation is still in your Mum's filing cabinet
In this space I think I'm becoming LLM advocate.

Finding, reading outdated, writing, updating, grooming documentation is so much more expensive than just throwing LLM on repository (or multiple ones) with a "Go Fetch" quest.

Sure, it's ephemeral, but with assumptions of $50/h earn and assumption that non-naive piece of internal documentation takes 8h of work time and will be read ~100 times that comes in at $4 at read which I think is much more expensive than straight token-API costs (and probably much much more expensive than subsidized subscription costs).

And this is generous. Looking at Jira's stats from my past work, many long documents in small (but specialized) team were read 10-20 times boosting the (assumption-average-costs) to $20-$50 per read.

But wait, there's more (;-))! That's all assuming that knowledge absorption is 100%, what if only 50% of document is relevant. What if it's 25% etc.

In the end today there might as well be no documentation for code and LLMs could extract it raw from code. Raw - because I don't think indexers/RAGs/compressers are useful. I found that spending time on building such is fruitless: Indices can (and will) go stale just as materialized documents and lower context saturation results in more hallucinations in the end.

(There are out-of-code documents which obviously have no other source than then themselves so there's no other around it, though)

xlii··on Good Tools Are Invisible
I agree with the premise of the article.

For example I've been using Jujutsu exclusively (as a Git frontend) for years and I don't think about it, I just use it. I reflected on this couple of years. It's existence is completely transparent to me.

I, however, don't agree with sibling commenter that it's a function of time spent with X though. As a counter example: Emacs was my go to editor for 15+ years, last 2 years - because reasons - I was switching between Neovim, Helix, Emacs, Kakoune. 6 months ago I settled with Kakoune.

Even with many years in Emacs, I still tweaked and tuned it. There was always something to do, change, understand. I actively thought about Emacs.

With Kakoune after initial "set me up" phase, it's just as transparent as Jujutsu. Sure, I made complex plugins (for searching, highlighting unbalanced parenthesis and even a GUI wrapper called Kakvide). But the difference is that in Emacs the driver was the tool itself and in Kakoune it's always "I wonder if I can do X".

And so I believe that Kakoune is better tool than Emacs as it's more transparent to me even with a big time difference in usage.

xlii··on ChatGPT Work
Not very thrilled mainly because I used chats for ad-hoc search queries. Now it's hard to reach. Not to mention private chats being removed.

Regarding "ChatGPT Work" vs "ChatGPT Codex" - In Work diffs are gone and in new chats there's a popup proposing creating docs or sheets. That's it I believe.

xlii··on How to Write an Email
It has spoilers in it :)
xlii··on My thoughts on the Bun Rust rewrite
For me, using Fuzzilli for testing a Zig code is not fuzzing, it's integration testing. If you're running code externally (e.g. wrapping binary) you cannot guarantee that side effect isn't caused by IO. I consider fuzzing a low level activity with many external variables removed.

Depending on where you are and how you communicate semantics matter more or less. It's very similar to compiler/transpiler. E.g. TypeScript "Compiler" is called compiler but in fact it's transpiler (it emits other high-level language as a result).

xlii··on Opinionated and easy Pi.dev configuration
Yes, because neovim != vim :)
← PreviousPage 2 of 22Next →