Hacker News Reader

smokelon GLM-5: Targeting complex systems engineering and long-horizon agentic tasks

> They are all just token generators without any intelligence.

Maybe, but "intelligence" doesn't have a clear, agreed definition. And calling them "just token generators" skips over how complex that generation actually is.

smokelon Slop Terrifies Me

Besides being a bit of a shallow comment, what exactly do you imply here? That capitalism logically implies that the rich become richer? I don't think this is necessarily the case, it just needs a stronger government than what the US currently has in place. (e.g. progressive taxation and strong antitrust policy seem to work fairly well in Europe).

smokelon Al Lowe on model trains, funny deaths and working with Disney

AI is a hype, Al Lowe is a legend.

smokelon France's homegrown open source online office suite

Work being done in offices is changing over time. I find myself writing less documents for printing and more for collaborating and sharing directly.

Even though many formal processes still require printable PDFs, we are slowly migrating to something paperless, or at least not paper-centric.

smokelon Claude Opus 4.6

I'll bite. The benchmark is actually pretty good. It shows in an extremely comprehensible way how far LLMs have come. Someone not in the know has a hard time understanding what 65.4% means on "Terminal-Bench 2.0". Comparing some crappy pelicans on bicycles is a lot easier.

smokelon Attention at Constant Cost per Token via Symmetry-Aware Taylor Approximation

What about n log n?

smokelon What I learned building an opinionated and minimal coding agent

ChatGPT is quite different from GPT. Using GPT directly to have a nice dialogue simply doesn't work for most intents and purposes. Making it usable for a broad audience took quite some effort, including RLHF, which was not a trivial extension.

smokelon Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant

> Pre literate peole could memorise vast texts

It seems more likely that there were only a handful of people who could. There still are a handful of people who can, and they are probably even better than in the olden times [1] (for example because there are simply more people now than back then.)

[1] https://oberlinreview.org/35413/news/35413/ (random first link from Google)

smokelon For me, Hacker News is probably the best community on the internet

Assuming that modifying the minimal user interface is unwanted, a simple confirmation interaction could fix this.

smokelon Ask HN: Is it still worth pursuing a software startup?

An interesting paper in this context is "Structural Inertia and Organizational Change" by Hannan and Freeman, 1984 [1]. A quote from the introduction: "...selection processes tend to favor organizations whose structures are difficult to change.".

In other words, it's typically a good thing that larger companies are slow to adapt. That's something that startups can make use of.

[1] https://www.jstor.org/stable/2095567

smokelon Drone Hacking Part 1: Dumping Firmware and Bruteforcing ECC

If they use unmodified Linux, then they only have to provide (a link to) the source code to that kernel on request. No source code is required for proprietary add-ons, unless they are kernel modifications.

The GPL also does not state that the source code should be easy to find. In the early days, one had to write a letter, send it by mail, in hopes of getting a tape or CD-ROM with the source code. For which you then had to pay as well.

smokelon Michelangelo's first painting, created when he was 12 or 13

I don't trust this for one bit. For the owners there is quite the incentive to label this as the work of a genius. But in reality, this is just pretty complex for a 12 year-old to produce by yourself.

Edit: as others have pointed out, and if I were to actually read the article carefully before commenting, the composition is not attributed to Michelangelo. So it is just a copy. Quite the achievement, but possible for a twelve-year old.

I once confronted a gallery owner who was proudly presenting a newly discovered work by Mondriaan [1]. An original black and white photo in an old newspaper [2] was shown as proof of authenticity. But many details such as the creases in fabric differ in the original and the new painting. No OpenCV required to see that. Mind you, the picture is already framed with Mondriaan standing next to it. Unlikely that he's still working on it.

Instead of responding, the gallery owner simply turned away.

[1] https://upload.wikimedia.org/wikipedia/commons/b/bb/Cavalini...

[2] https://www.vrt.be/vrtnws/nl/2022/03/02/nieuwe-werken-mondri...

smokelon Michelangelo's first painting, created when he was 12 or 13

That is slightly unconvincing. Durer is indeed a similar genius, but the complexity of that drawing is an order of magnitude lower than the painting.

Source: know how to draw really well.

smokelon Show HN: I built a text-based business simulator to replace video courses

Ha, this made me think of Viznut's Micro$oft - Simulaattori [1].

[1] http://viznut.fi/demos/msdos/ (pick mssim.zip) or on Pouet https://www.pouet.net/prod.php?which=21174

smokelon ‘ELITE’: The Palantir app ICE uses to find neighborhoods to raid

The core issue is not that people cannot think with nuance, but that nuance is costly and poorly rewarded.

smokelon There's a ridiculous amount of tech in a disposable vape

One of the more inspired design choices of the parking ticket devices in my area is the inclusion of a key repeat feature.

If you keep your finger on the touchscreen for just long enough, it helpfully repeats the keystroke while you're entering a license plate.

Given the inevitable hardware issues, this means that what should be a single tap frequently becomes a burst of identical characters.

The programmers who worked on this probably would've liked to be game developers instead.

smokelon “Food JPEGs” in Super Smash Bros. and Kirby Air Riders

And Castlevania [1] (also from 1986) and don't forget about Pac-man, who probably started it all :)

[1] https://castlevania.fandom.com/wiki/Food

smokelon Don't fall into the anti-AI hype

Agreed, but I'd say these would be artists in the "grey area". They are capable of drawing furry art, for example, and have the choice to monetize that, even though they might have become bored with it.

I was mostly considering contemporary artists that you see in museums, and not illustrators. Most of these have moved on to different media, and typically don't draw or paint. They would therefore also not be able to draw commission pieces. And most of the time their work does not sell well.

(Source: am professionally trained artist, tried to sell work, met quite a few artists, thought about this a lot. That's not to say that I may still be completely wrong though, so I liked reading your comment!)

Edit: and of course things get way more complicated and nuanced when you consider gallerists pushing existing artists to become trendy, and artists who are only "discovered" after their deaths, etc. etc.)

smokelon Don't fall into the anti-AI hype

I've given this quite some thought and came to the conclusion that there is actually no choice, and all parties fall into the first category. It's just that some people intrinsically like working on commercial themes, or happen to be trendy.

Of course there are some artists who sit comfortably in the grey area between the two oppositions, and for these a little nudging towards either might influence things. But for most artists, their ideas or techniques are simply not relevant to a larger audience.

smokelon One Formula That Demystifies 3D Graphics

> The rotation formula eludes me.

Interestingly, in a way, rotation is less mystical than the perspective projection. The rotation is linear: x' = Rx, but the perspective projection is non-linear.

This is where things become fun. Next up are homogeneous coordinates or quaternions. Takes a few years of your life to actually enjoy this though :)

smokelon One Formula That Demystifies 3D Graphics

It's a thing [1].

[1] https://en.wikipedia.org/wiki/Neural_computation

smokelon One Formula That Demystifies 3D Graphics

Artists are still struggling with the fact that human perception arises from binocular vision. Two distinct retinal inputs are integrated by distributed neural processes into a single, coherent 3D experience. This integration is neither a simple planar stitching nor a direct representation of the world, but an active construction shaped by neural computation and subjective awareness.

It is quite likely that artists in earlier periods struggled with this as well, and were less concerned with adhering strictly to a photographic or geometrically exact perspective, as we are. The adoption of the camera obscura probably influenced things a lot.

smokelon One Formula That Demystifies 3D Graphics

The formula is f(x, y, z) = [x/z, y/z], which does perspective projection of a 3D coordinate onto a 2D plane.

I can't really say that this formula demystifies things, but the video is nice if you're eager to learn about this.

smokelon Children and Helical Time

> This is a depressing thought to consider in (linear) middle age, but it is hard to escape the feeling that it is essentially true. Childhood memories have an intensity and a vibrancy that it is difficult for the rest of life to match.

Could anyone who is extremely fortunate and never had to work for money share their experience on this?

I find that the years that I spent on art (playing around, learning new things, not taking other peoples' orders) lasted longer than the ones I spent doing software development for money. Both were fun, but the remaining memories differ by intensity.

I personally don't find the logarithmic experience theory convincing. Why are the first three or so years excluded from this? It seems more likely that new experiences make more impact, or that repeated memories make them more intense. Or dozens of other theories.

smokelon Warren Buffett steps down as Berkshire Hathaway CEO after six decades

Still, slightly more than the 0.00026% who are billionaires.

smokelon Ask HN: Any example of successful vibe-coded product?

That's a bit dishonest. Obviously, vibe coding is only productive for engineers who actually know what they're doing. Perhaps it is best to consider it a multiplier, not an enabler.

smokelon Ask HN: Any example of successful vibe-coded product?

You could probably find a good answer anywhere, but the solution is in a more nuanced view.

Some types of programming benefit more from AI tooling than others. For example, prototyping seems to be the most fruitful area. Also, writing small utilities is much easier, to the extent that a two hour job would now take only a few minutes. That's where you get the multiplier posts from.

But working in a large codebase using proprietary libraries is not a solved problem for AI (yet).

It's just that the average engineer does not spend all of their time on things that can be sped up.

Speeding up 1% of your time by a factor 20 simply does not help very much. But for some roles, I'm sure that a 10% net increase in productivity is realistic.

smokelon Ask HN: What skills do you want to develop or improve in 2026?

Oh, I'm sorry, no nitpicking intended. Just wanted to share some enthousiasm :)

smokelon Ask HN: What skills do you want to develop or improve in 2026?

> The cost of all the paper I wasted in the last 5 years is probably less than the cost of a pizza.

Once you get into it, there is an amazing assortment of papers that can cost up to a pizza per sheet. Especially if you're going for the larger formats.

https://arches-papers.com/arches-range-of-papers/watercolor-...

smokelon Ask HN: What skills do you want to develop or improve in 2026?

I am pretty good at drawing, and would highly recommend starting with traditional media rather than digital tools.

Drawing on paper allows for a wide range of physical setups, such as using a notebook on your lap or on a table, large sheets mounted on a wall, or a board on an easel. Each configuration engages different muscle groups. Large-format drawing relies primarily on shoulder movement, whereas smaller, more detailed work involves the wrist, forearm, and fingers. I'm convinced that deliberately training hand–eye coordination at multiple scales (finger–eye, wrist–eye, or shoulder–eye), is beneficial in learning to draw better.

It is also a good idea to experiment with a variety of media: pens, pencils, chalk, charcoal, and different surfaces such as paper, wood, or canvas. The differing tactile feedback and resistance will improve your motor control. You don't need to spend a fortune on this, but don't limit yourself to the cheapest color pencils and toilet paper.

That said, if your primary goal is accurate photo replication, it's probably easiest to start with Drawing on the Right Side of the Brain [1], along with some YouTube tutorials.

[1] https://www.drawright.com/