HNHacker News
TopNewBestAskShowJobs

smokel

3,958 karma · joined December 20, 2011

Coder by day, artist by night.
submissionscomments
smokel··on AI Isn't Outthinking Mathematicians. It's Out-Remembering Them
Mathematics is typically concerned with "proofs" [1], which similarly to code, often allow for strict validation. Thanks to reinforcement learning techniques, it is now possible to train LLMs to perform very well on code generation, and mathematical proof generation.

Law and medicine are fundamentally harder fields to obtain decent training data for, and LLM results are therefore expected to be less powerful. Also, making mistakes in these fields is costly, but perhaps you were alluding to that already.

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

smokel··on AI by Hand
https://github.com/karpathy/llm.c

However, I wonder about the use of this. Typically, you'd want to use a GPU for inference, and programming those in C is quite a challenge. The educational value of learning LLMs in C/C++ seems a lot lower than in Python. May I ask why you're interested in this specifically?

Also, why ask this here, and not simply use Google or an LLM to find the information? It comes off as a bit lazy, especially given that you provide no extra context.

smokel··on I built a 500k-domain search engine for makers in a weekend for $10
I typically interpret it as an excuse, not a selling point.
smokel··on John C. Lilly on solid state intelligence and the elimination of man (1978)
> Its new plans called for traveling through the galaxy looking for entities like itself.

Why not simply set up a network with other planets and communicate over FaceTime? I don't see a reason why Earth should actually hang out with others of its kind.

Anyway, the article seems a bit anthropocentric to me. I'd much prefer a nihilist view, where our future AI overlords simply don't care what we humans do.

smokel··on AMD acquires Taalas to boost inference performance by etching models in silicon
The technical aspects of SOTA models are not publicly documented. How do you know if something is converging?
smokel··on AI Mania: From Tulips to Tokens
This is a pretty confusing article. It seems to string together some loosely related facts to make the current AI boom look bad.
smokel··on Software for One
Yes there is, at least to me. The reason being total freedom to tune the operating system, and not being limited to what someone imagined I would want to do in the future.
smokel··on Diátaxis
Perhaps add "A systematic approach to technical documentation authoring" to the title?
smokel··on Software for One
That would be the 1980s approach. I would be more interested in being able to read an email and then say something along the lines of:

  in context/mail:
    when today/evening:
      if content excites me:
        send alert
This should be available directly by pressing some code button, and there should be some kind of user interface that makes it trivial to enter these commands.
smokel··on Software for One
It's somewhat unfortunate that programming is still pretty hard to start with.

Early home computers would boot up with BASIC. Granted, the user experience was horrible, but at least you could start programming right away.

What is stopping us from creating an operating system that has programming as a first class citizen?

There is JavaScript in the browser, but that still feels as an afterthought, and incurs all kinds of security problems.

Why must we wait for AI to work around this? Is it really true that only very few people are interested in programming? Why not teach it in primary school, along with reading and writing?

smokel··on Gemini Robotics 2 brings whole body intelligence to robots
I think that credit might go to IBM:

https://en.wikipedia.org/wiki/TD-Gammon

smokel··on Book Corners: Community map of neighborhood book exchange spots
> If you want a good one you'd have to set up your own and aggressively curate it

Expect a community library with only books on homotopy type theory and quantum computing soon!

smokel··on Book Corners: Community map of neighborhood book exchange spots
Most of the books in these community libraries are really crappy. I would assume that people:

a. Toss out books at home they don't like and bring them to these libraries

b. Take interesting books from the libraries, and leave the bad ones

Any ideas on how to improve that?

smokel··on Flux 3
In the 1990s "Object-Oriented Ontology" was introduced by Graham Harman [1]. The name was borrowed from computing, but its meaning has little or nothing to do with Simula or Smalltalk.

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

smokel··on Flux 3
> Frivolous use of the term World Model

The term "world model" as it was once used in model-based RL can now apparently refer to anything as silly as linear regression. Then again, the RL folks probably borrowed the term from behavioral scientists before them. It's probably best to simply accept this :/

A similar thing happened to "object oriented" which has been misused by philosophers and visual artists alike.

smokel··on The unreasonable difficulty of time series forecasting
This article focuses mostly on point forecasting. There are many other things that are interesting to forecast.

Consider for example the use case of forecasting the average speed on a road segment with a maximum speed of 70mph. Forecasting whether that will be 69.8 or 70.3 is not very relevant. What is relevant is forecasting when the speed drops below a traffic jam threshold. But the exact timing of that might be impossible to forecast due to the inherently chaotic behavior of traffic. Forecasting the probability of a traffic jam occurring may be more interesting to practitioners.

smokel··on Is this the end of the once-mighty GoPro?
I'm looking at GoPro packaging here that says "Made in Thailand".
smokel··on GPT-5.6 used a prompt to close a 30-year gap in convex optimization
Lean is the Mizar here. For those who have no clue what this is about, Mizar [1] was an early automated theorem prover. Can't wait for HN to add AI features to explain concepts in the sideline, and autovoting.

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

smokel··on Kimi K3: Open Frontier Intelligence
> I’d buy these. Say $100 for a frontier class model. Maybe more.

Sure you would. Running frontier class models on current hardware costs in the order of tens of thousands of dollars. It is more likely that these custom ASICs will be priced competitively with that, and not with Super Mario Bros.

Oh, and energy consumption will be in the same order.

smokel··on Microsoft Comic Chat is now open source
There's an interesting easter egg in semantic.cpp, line 77:

  if (CheckWord(words, "OXio")) {
Apparently, if your text contains the word "OXio", it triggers the following riddle: What's round on the ends and hi in the middle?
smokel··on Measuring Input Latency on Linux: X11 vs. Wayland, VRR, and DXVK
Some things do indeed improve, but some other things require a massive effort that no hobbyist or small company can make happen.

This is most obvious in places where a lot of coordination is required, for example in supporting proper color correction throughout all applications, or decent support for advanced printer functions.

There are many incremental changes, but we often get stuck in local minima for years.

Still, I personally like that one can (relatively) easily watch what happens under the hood. It's not entirely clear to me why Windows and MacOS must remain closed source.

smokel··on Can We Understand How Large Language Models Reason?
It's probably helpful in this discussion to make a difference between two definitions of reasoning:

1. phenomenal reasoning, requiring consciousness and subjective experience

2. functional reasoning, transforming premises into conclusions using logic

I think you are attacking this using definition 1, whereas the article is obviously aiming at a different type of reasoning, and trying to formalize what is actually going on. It seems to be a genuine effort.

smokel··on Ask HN: Another "Hacker News" with less AI and more human-focused hacking news?
There are very few sites such as Hacker News around. Looking for one that is similar in quality, but with a slightly different focus is going to be disappointing.

Perhaps it helps to consider why you'd exactly want this? You may be better off not visiting any such site, and instead focus on real human-in-the-loop activities with a group of local hackers.

smokel··on 30papers.com – Ilya's 30 essential ML papers, in a beginner friendly format

  for x in 1611.02731 1511.06391 1811.06965 1512.03385 1511.07122 1704.01212 1409.2329 1512.02595 1706.01427 1410.5401 1806.01822 1706.03762 1409.0473 1506.03134 2001.08361 1405.6903 1603.05027 math/0406077; do curl -fL https://arxiv.org/pdf/$x -o ${x##*/}.pdf; done
  for u in https://www.cs.toronto.edu/~hinton/absps/colt93.pdf https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf https://www.vetta.org/documents/Machine_Super_Intelligence.pdf https://www.lirmm.fr/~ashen/kolmbook-eng-scan.pdf https://scottaaronson.blog/?p=762 https://karpathy.github.io/2015/05/21/rnn-effectiveness/ https://colah.github.io/posts/2015-08-Understanding-LSTMs/ https://nlp.seas.harvard.edu/annotated-transformer/ https://cs231n.github.io/; do curl -fLO "$u"; done
smokel··on Claude Design System Prompt
The GitHub page clearly states that it is reverse engineered.

Even if you rewrite an article in your own words, you cannot claim copyright for that. At least, not to my knowledge.

smokel··on Claude Design System Prompt
You cannot claim authorship of something that you were not even allowed to take in the first place. How would you then allow rights to somebody else?

Seems pretty obvious to me.

smokel··on Claude Design System Prompt
It does make a difference. If this text was generated by the LLM, then by those same terms you are allowed to publish the data. But given that it is reverse engineered, you are not allowed to do so.

Note that even in the normal case, there are restrictions. You'd have to validate that the generated output is not copyrighted. Which in this case may not be trivial.

smokel··on Artful Cats: Feline-Inspired Art and Artifacts
Can't have a discussion on art and cats without mentioning Louis Wain [1].

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

smokel··on Claude Design System Prompt
If this is regular output of the LLM, I'm not sure, but given that the author proclaims that this is reverse engineered, then they are not allowed to redistribute it under their own license terms. The terms of service are also pretty clear on this not being allowed, which makes it extra hard to defend (section 3.3):

> You may not access or use, or help another person to access or use, our Services in the following ways:

...

3. To decompile, reverse engineer, disassemble, or otherwise reduce our Services to human-readable form, except when these restrictions are prohibited by applicable law. [1]

[1] https://www.anthropic.com/legal/consumer-terms

smokel··on Claude Design System Prompt
> Open source, MIT licensed.

I don't think that is how copyright licensing works.

← PreviousPage 3 of 34Next →