HNHacker News
TopNewBestAskShowJobs

kybernetikos

6,302 karma · joined July 11, 2012

[ my public key: https://keybase.io/kybernetikos; my proof: https://keybase.io/kybernetikos/sigs/yHgJmn4qufQOBXtQjZ-3EsFWRdtec9C5wGhHqz3ETYw ]
submissionscomments
kybernetikos··on Defeating Nondeterminism in LLM Inference
For fun over the last few days, I've built a compressor / decompressor that uses the logits from an LLM, for each token in the input, then takes the ranks and exponential goolomb encodes them. Then you work in reverse to regenerate the original

It took me ages to get the prediction for the second token after "hello" to match the same as the prediction for the second token when running the model on the string "hello world", despite the fact that I was using a causal model. I tried all kinds of things before discovering that `quantized: false` was the important setting.

kybernetikos··on The key to getting MVC correct is understanding what models are
I'm not saying it's not simple to write, I'm saying it's ugly and contingent and you can't really avoid that. It's exactly this reason that has led to a proliferation of MV* patterns, including the one you describe.

But to try to explain myself more clearly - in the architecture you describe, who is that it is implementing TableData and TableCallback? Is it your beautiful clean business logic classes that have no coupling to their representation - in which case that is weirdly coupled in an ugly way, or is it some other class that acts as the bridge between the two worlds, in which case, that's where your ugly code is living.

kybernetikos··on The key to getting MVC correct is understanding what models are
I've become convinced that the real problem is probably impossible to get away from.

Ultimately we want a nice set of reusable UI components that can be used in many different situations. We also want a nice set of business logic components that don't have any kind of coupling with the way they get represented.

In order to make this work, we're going to need some code to connect the two, whether it's a 'controller' or a 'view model' or some other piece of code or architecture.

However we choose to achieve this task, it's going to feel ugly. It's necessarily the interface between two entirely different worlds, and it's going to have to delve into the specifics of the different sides. It's not going to be the nice clean reusable code that developers like to write. It's going to be ugly plumbing, coupled code that we are trying to sweep into one part of the codebase so that we can keep the rest of it beautiful.

kybernetikos··on ReMarkable Paper Pro Move
I've been using a boox note air for many years and you definitely can zoom on that.

Android is great for this use case because it lets me syncthing notes and use sheet music apps and use both kindle and kobo and calibre library and offline wikipedia and my own tools. As far as I'm concerned if you try to use it as a generic android tablet you're doing it wrong, but android is a massive step above what everyone else is offering (i.e. none of that)

kybernetikos··on The Whispering Earring
This is associating the self with the thing that decides how best to achieve goals (the earring / the part of your brain that works out how to achieve a goal), while I'm saying that I think I would associate the self much more with the thing that decides what the goals are.

> they don't just want to solve them; they want to solve them theirselves. That's what is lost here.

I think in this story, the earring would not solve the crossword for you, if for some reason your goal was to solve the crossword yourself.

kybernetikos··on GPT-5
The wrongness is germane to someone who is doing their physics homework (the example given here). It's actually difficult for me to imagine a situation where someone would ask ChatGPT 5 for information about this and it not be germane if ChatGPT 5 gave an incorrect explanation.
kybernetikos··on GPT-5
So I worry that you think that the equal transit time thing is true, but is just one effect among others. This is not the case. There are a number of different effects, including bernoulli and coanda and newtons third law that all contribute to lift, but none of the things that actually happen have anything to do with equal transit time.

The equal transit time is not a partially correct explanation, it's something that doesn't happen. It's not a superficial explanation, it's a wrong explanation. It's not even a good lie-to-children, as it doesn't help predict or understand any part of the system at any level. It instead teaches magical thinking.

As to whether it matters? If I am told that I can ask my question to a system and it will respond like a team of PhDs, that it is useful to help someone with their homework and physical understanding, but it gives me instead information that is incorrect and misleading, I would say the system is not working as it is intended to.

Even if I accept that "audience matters" as you say, the suggested audience is helping someone with their physics homework. This would not be a suitable explanation for someone doing physics homework.

kybernetikos··on GPT-5
"air over the top has to travel farther in the same amount of time" is just wrong, it doesn't have to, and in fact it doesn't.

It's called the "equal transit-time fallacy" if you want to look it up, or follow the link I provided in my comment, or perhaps the NASA link someone else offered.

kybernetikos··on GPT-5
Pressure differential is absolutely one of the main components of lift (although I believe conservation of momentum is another - the coanda effect changes the direction of the airflows and there's 2nd law stuff happening on the bottom edge too), but the idea that the pressure differential is caused by the fact that "air over the top has to travel farther in the same amount of time" because the airfoil is curved is completely incorrect, as the video in my link shows.
kybernetikos··on GPT-5
ChatGPT5 in this demo:

> For an airplane wing (airfoil), the top surface is curved and the bottom is flatter. When the wing moves forward:

> * Air over the top has to travel farther in the same amount of time -> it moves faster -> pressure on the top decreases.

> * Air underneath moves slower -> pressure underneath is higher

> * The presure difference creates an upward force - lift

Isn't that explanation of why wings work completely wrong? There's nothing that forces the air to cover the top distance in the same time that it covers the bottom distance, and in fact it doesn't. https://www.cam.ac.uk/research/news/how-wings-really-work

Very strange to use a mistake as your first demo, especially while talking about how it's phd level.

kybernetikos··on The Whispering Earring
I'm not actually sure how horrifying this is. It sounds like it's just a better executive planner to achieve your goals. As long as they are still your goals, surely you'd want the best executive planner available. I would say it's the goals that are important, not the limited way in which I work out how to achieve them.

It would certainly be horrifying if I were slowly tricked into giving up my goals and values, but that doesn't seem to be what is happening in this story.

Perhaps if I were to put the earring on it would tell me it would be better for me to keep wearing it.

kybernetikos··on Open models by OpenAI
> Won't that work much better and scale easily?

Doing computation that can happen at end points at the end points is massively more scaleable. Even better, its done by compute you usually aren't paying for if you're the company providing the service.

I saw an interview with the guy who made photopea where he talked about how tiny his costs were because all compute was done in the user's browser. Running a saas in a cloud is expensive.

It's an underrated aspect of what we used to call "software".

And that's leaving aside questions of latency and data privacy.

kybernetikos··on Simon Tatham's Portable Puzzle Collection
Net can be done with reasoning rather than mindless iteration. You start by locking in end points surrounded by other end points except for one free space. if you have a straight line that can connect two end points then you lock it in the other orientation. If a line is locked next to a T pipe, the back of the t pipe goes against the line. If a corner piece is next to a locked pipe, you know that the side opposite the incoming pipe is empty, so it could be the back of a T or the side of a line piece, etc.
kybernetikos··on Major quantum computing advance made obsolete by teenager (2018)
Ancillary Justice is told from the point of view of a character of a culture that doesn't draw the distinction in language. It is occasionally remarked on, but generally from the point of view of the main character being always a bit paranoid that they'll cause offense by not referring to characters of other cultures and languages correctly.
kybernetikos··on AI coding agents are removing programming language barriers
Yes, I worry that we're in for an age of stagnation, where people are hesitant to adopt radically new languages or libraries or frameworks because the models will all be bad at them, and that disadvantage will swamp any benefit you might get from adopting an improved language/library/framework.

Alternatively every new release will have to come with an MCP for its documentation and any other aspects that might make it easier for an LLM to talk about it and use it accurately.

kybernetikos··on Font Comparison: Atkinson Hyperlegible Mono vs. JetBrains Mono and Fira Code
But that's the exact reason why it's problematic in contexts where the exact character is important in a way not clear from context. Atkinson Hyperlegible is my favourite for reading passwords or urls with codes in them etc.
kybernetikos··on What were the earliest laws like?
I think probably the distinction we're making here is just one of scale and acceptance. I don't see your distinctions as fundamental:

1. If 9 of the ten friends consent and one doesn't they can still be forced to follow the rule the nine friends decided on.

2. Although it's a smaller scale, there absolutely is a 'social fabric' that can be harmed in a group of ten friends.

3. I don't think laws necessitate the use of lethal force, but yes, groups of people e.g. in sports or school do sometimes use violence to enforce group decisions on their members.

In some early societies, there was a basic assumption that broadly speaking the head of a household should have the right to manage their affairs as they wished. Laws were there to mediate between households (rather than individuals) and the big benefit they give to justify the loss of freedom is the control and management of blood feuds, which otherwise can be devastatingly destructive.

kybernetikos··on Local LLMs versus offline Wikipedia
I've been carrying around a local wikipedia dump on my phone or pda for quite a bit more than 10 years now (including with pictures for the last 5 years). Before kiwix and zim, I used tomeraider and aard.

I do it both for disaster preparedness but also off-line preparedness. Happens more often than you'd think.

But I have been thinking about how useful some of the models are these days, and the obvious next step to me seems to be to pair a local model with a local wikipedia in a RAG style set up so you get the best of both.

kybernetikos··on How I keep up with AI progress
There are lots of careers that used to be widespread but are now extremely niche because of technological change. For example, most careers relating to horses.
kybernetikos··on How I keep up with AI progress
>> one doing the job ten use to.

> This would increase employment ceteris paribus.

This might be true, but if it is, the one "doing the job ten use[d] to" would not actually being doing the same kind of work at all, and so therefore might not be the same people or even same kind of people. Even if we do Jevons ourselves out of this situation, it might still spell employment disaster for mid level coders, while increasing employment for product managers.

kybernetikos··on How I keep up with AI progress
I think commercially there'll be a really strong pull for dev tools to resell access to frontier models.

It's hard to make money out of dev tools, but if you tie in a service people are prepared to pay hundreds of dollars a month for, then suddenly it looks easier to make money out of an otherwise unsaleable IDE plugin that accesses that service.

kybernetikos··on How I keep up with AI progress
The big question for me at the moment is whether to go pay-for-tool or pay-for-model.

If I pay for a tool that includes access to frontier models, then they'll keep the models up to date over time for me, let me use models from multiple providers, and the tool is carefully designed around the capabilities and limitations of the models it works with. On the other hand I can't really use the powerful model the tool works with for other applications or write my own.

If I pay for models, then I can only really use it with that manufacturers tools or tools that aren't optimised for the model but allow you to bring your own keys, and if the model provider I'm paying falls behind then I'm tied in for the duration of the contract. The big advantage is that there is a lot of innovation in tooling happening at the moment and you can avoid being locked out of that or having to pay many times for access to the same frontier models accross multiple different tools.

kybernetikos··on Engineer creates ad block for the real world with augmented reality glasses
The Steve Mann Version from a while ago replaced the ads with useful information, like maps of the local area or your latest messages, shopping list, to do list items etc.
kybernetikos··on Collections: Nitpicking Gladiator's Iconic Opening Battle, Part I
> Bret Devereaux isn't wrong. He's also not in the film business.

However I would be interested in seeing someone put to the test his thesis that a more historically grounded battle scene would be even more epic and visually interesting.

kybernetikos··on U.S. bombs Iranian nuclear sites
> The goals of israel are regime change and nuclear disarmament -- these cannot be achieved from the air. It might be that israel is content to lose on these objectives, and so be it.

This is key. The only way for this set of actions to go well is if there is regime change, otherwise the most likely outcome is that Iran's resolve to acquire nuclear weapons as quickly as possible has been dramatically solidified.

Like you though, I struggle to see any clear path for positive regime change to occur. The nearest attempt would be boots (but whose?) on the ground, but even that seems unlikely to work out well. Maybe there could be some sort of internal resistance, but I don't see how they could operate successfully while the country is under external attack.

My assumption with how things are at the moment is that the actions by Israel and the USA have all but guaranteed that Iran acquires a nuclear weapon in the next few decades, and so have dramatically increased the risk of Israel being attacked with one. One has to assume that radical Islamist terrorism in western countries will increase too.

kybernetikos··on Framework Laptop 12 review
I used to really like the keyboard on the old think pad t series, like the T60. But it seems that everyone is doing 'island' keys these days.
kybernetikos··on Framework Laptop 12 review
It's a bit surprising to find so little in these comments and the original review talking about the youth first laptop use-case. Lots of schools require a touch screen, and kids are going to break parts even on a fairly rugged laptop.

All these people talking about MacBook Airs are missing the point. None of the schools around me have MacBook Airs as allowed laptops for kids BYOD and even if they did, I'm not sure they'd have a long life getting the kinds of hits and knocks that will happen being carried everywhere in high school by a 12 year old.

This laptop is obviously for this use case. I know of no other laptop that really covers this use case well. Typically laptops aimed at this segment are cheaper, but not rugged, not easy to repair, and not really very nice. I strongly suspect that I'll only have to replace the screen or keyboard once before the total cost of ownership works out compared to a normal laptop.

kybernetikos··on Veo 3 and Imagen 4, and a new tool for filmmaking called Flow
All of the things you mentioned were available in phones before the first iPhone (assuming by ipod you mean mp3 player). In fact from a software point of view it was lacking a bunch of functionality and software ecosystem some competitors had.

In my view the reason the iPhone felt so new was almost entirely the incredibly responsive capacitive touch screen with a finger ui, everything I'd used before it did resistive and preferred pen for detail. Pen actually is better for detail so in some ways it was that more than anything else that turned the device from a creation device to a consumption device which was whole new way of thinking about smart personal devices.

Of course it was also sold in a decent package too where Apple did deals that ensured it was available with good mobile internet plans which were also unusual at the time.

kybernetikos··on Plain Vanilla Web
> On the one hand we have lots of people on here who are building full-featured web apps, not websites, on teams of 30+.

Photopea is a full-featured web app, not website, written without frameworks by a single individual. It is absolutely possible to build feature rich applications without the framework du jour.

kybernetikos··on Plain Vanilla Web
Normal HTML elements don't exclusively use HTML attributes either. Surely you've used button.onclick or element.classList or element.innerHTML?
← PreviousPage 4 of 34Next →