IIRC this Verdagon guy had a language called Vale... is Valen a rename or a new project?
Edit: tfa makes it clear it's a separate project
151 karma · joined November 8, 2023
I like algorithms and abstraction
Edit: tfa makes it clear it's a separate project
I am confused
I = 2 (A - B + 1)
In my case I see Claude produce code much worse than I would, but it's certainly much quicker and, even after reworking, it makes me finish tasks in less time.
From a space perspective, this is actually better because tokenization tends to compress text quite well. For example, common tokens in English text take up ~4 characters on average (expands to 32 bits), but only take up a fraction of that to store (15-18 bits/token depending on vocabulary size)
In fact it appears that designing the tokens as a text compression encoding is a decent approach, since it's roughly what some LLMs do. For example, early GPT tokenizers followed byte pair encoding to create the vocabulary, which is a text compression algorithm from the 90s.
The normal behavior for link posts in HN is that the title takes you to the link address, not to the post address.
To enter the post address, you need to hit the "discuss"/"x comments" button.
So, what you describe is HN working as expected AFAICT, it's just a bit counterintuitive because it's an HN post linking to another, which is uncommon
Also, there might be other new 3D software with better UX. I am not a Blender fanboy, but I do love 3D art and graphics programming and want as many people as possible to get into it :^)
I think you just want the illusion of safety :p
A big advantage of verified code is that it enables you to write the sketchy and dangerous-looking code BECAUSE it's proven correct
In fact, skipping as many safety checks as possible is highly desirable. For performance, yes, but also because it's less code to maintain.
Our tools already do this to some extent, for performance. E.g. compilers that remove your bounds or type checks in the generated code when it can prove it's not needed.
JPEG compression can only move information at most 16px away, because it works on 8x8 pixel blocks, on a 2x down-sampled version of the chroma channels of the image (at least the most common form of it does)
But understanding your weaknesses and working on them is huge, and I think most people just don't try to do it.
Being stuck for days is something to be overcome.
The next step would be being slow because you are trying out many different ideas and have no intuition for what the right one is.
As a beginner I often thought about a problem for days before finding a solution, but this happened less and less as I improved
I got better at exploiting the things I knew, to the point where I could be pretty confident that if I couldn't solve a problem in a few hours it was because I was missing some important piece of theory
I think spending days "sitting with" a problem just points at your own weakness in solving some class of problems.
If you are making no articulable progress whatsoever, there is a pathology in your process.
Even when working on my thesis, where I would often "get stuck" because the problem was far beyond what I could solve in one sitting, I was still making progress in some direction every time.
I was a teenager with focus issues when a teacher introduced me to Pomodoro. It was great because I was finally able to get my schoolwork done (I would regularly fail 3-4 classes every year before this). The general structure he gave was as follows:
- You split time into work/rest intervals
- At the start of each work interval, you set a 15-minute timer and plan out what you are going to do in the interval
- When the timer goes off you HAVE to stop
- Then you set a 5-minute timer and rest until it goes off
- Repeat until you are done
Back then, the "planning every interval" part was hard for me but just forcing me to rest was enough to do what I needed to do. In reality, 90 minutes of homework wasn't enough to make me tired (I would go on 10-hour long programming sessions with no breaks on the regular), just emotionally tired, bored, or frustrated. The breaks solved that.
I think the timings he suggested are adjusted for younger people (15 work minutes instead of the more common 25, and no long break after N cycles). As I grew up, I slowly lengthened the work and rest periods up to 25min work and 7min rest.
The benefits of Pomodoro are Three-fold as far as I can tell:
- You plan out your work, so you spend a higher % of work time doing stuff that matters.
- You rest regularly, so you are able to keep working for longer.
- You limit rest time, so you spend a higher % of time working.
Recently, after I mentioned I use Pomodoro, one of my mentors made the observation that the interest of a paid worker is not to work as long as possible during your work hours. Instead, it's to get your work done as easily and effectively as possible, get paid (hopefully make your boss and coworkers happy, get a raise and promotion every once in a while), and get out.
Doing the math, with my flavor of Pomodoro (25/7, no long breaks) you end up working 5:30-hours out of an 8-hour workday (it's roughly 5:10 with the technique taken straight from Wikipedia). I wonder if he meant that it's possible to be very effective while working significantly less than 5 hours a day. I wonder how much I could get done in just 2 or 3 hours of fully focused work, and how I would feel about it at the end of the day.
> But I think this is an interesting perspective difference to discuss and not a blatant ad hom argument used to ‘fight’ Rust users on the internet.
(It is already clear that you share your perspective because you think that it's interesting and not a fallacy.)
Then, there are some paragraphs close to the end that seem to repeat ideas.
But IMHO it's a good, nuanced comment that both addresses shortcomings in current discourse and adds ideas to the discussion. This is obviously difficult to do in few words.
A few concepts that come to mind from your comment, but I missed from others:
- Are Rust-style ownership semantics complex or just hard for me? Is it an issue of familiarity only? Is it ergonomics?
- Are they hard in general or just due to the way I like to program? How can I change my style to better fit the model?
- Are they even fit for the software I write? i.e. is it worth it to change how I program to better fit the model?
- What other tools are there to deal with resources?
- What could a programming language do to offer multiple of those? How can we mix GC, region types, linear/uniqueness types, manual management, etc. in a single language?
- A bunch of stuff about discourse in this thread and HN/the internet in general (which is maybe not the point of the comment and could be omitted?)
The author defines American style intelligence as "the ability to adapt to new situations, and learn from experience".
Then argues that the current type of machine-learning driven AI is American style-intelligent because it is inductive, which is not what was supposedly (?) being argued for.
Of course current AI/ML models cannot adapt to new situations and learn from experience, outside the scope of its context window, without a retraining or fine-tuning step.
So it's two records in the landlord's own books which u dont necesarily know about, and two records in your books.
This is good enough for rendering via sphere tracing, where you want the sphere radius to never intersect the geometry, and converge to zero at the boundary.
A particular class of fields that have this property is fields with gradient not greater than one.
For example, linear blends of SDFs. So given SDFs f and g you can actually do (f(pos)+g(pos))/2 and get something you can render out the other side. Not sure what it will look like, or if it has some geometrical interpretation though.
Note that speed of convergence suffers if you do too many shenanigans.
> Itheve whe oiv v f vidleared ods alat akn atr. s m w bl po ar 20
Using pairs of consecutive characters (order-2 Markov model) helps, but not much:
> I hateregratics.pyth fwd-i-sed wor is wors.py < smach. I worgene arkov ment by compt the fecompultiny of 5, ithe dons
Triplets (order 3) are a bit better:
> I Fed tooks of the say, I just train. All can beconsist answer efferessiblementate
> how examples, on 13 Debian is the more M-x: Execute testeration
LLMs usually do some sort of tokenization step prior to learning parameters. So I decided to try out order-1 Markov models over text tokenized with byte pair encoding (BPE).
Trained on TFA I got this:
> I Fed by the used few 200,000 words. All comments were executabove. This value large portive comment then onstring takended to enciece of base for the see marked fewer words in the...
Then I bumped up the order to 2
> I Fed 24 Years of My Blog Posts to a Markov Model
> By Susam Pal on 13 Dec 2025
>
> Yesterday I shared a little program calle...
It just reproduced the entire article verbatim. This makes sense as BPE removes any pair of repeated tokens, making order-2 Markov transitions fully deterministic.
I've heard that in NLP applications, it's very common to run BPE only up to a certain number of different tokens, so I tried that out next.
Before limiting, BPE was generating 894 tokens. Even adding a slight limit (800) stops it from being deterministic.
> I Fed 24 years of My Blog Postly coherent. We need to be careful about not increasing the order too much. In fact, if we increase the order of the model to 5, the generated text becomes very dry and factual
It's hard to judge how coherent the text is vs the author's trigram approach because the text I'm using to initialize my model has incoherent phrases in it anyways.
Anyways, Markov models are a lot of fun!
Very high temperature blackbody radiation perhaps?
git rebase --onto squashed nonsquashed next-pr