noosphr

9 ms·

181.
by noosphr 4mo ago
I'm going to get flagged to hell but 'we' do not. Women do. Men can happily wait till their 80s to have three children. We need to get over the fantasy that biology doesn't drive survival. If anyone is upset by that plea
182.
by noosphr 5mo ago
>Integers in normal programming represent data or instructions; instructions are hand coded, have rigidly defined semantics, are not differentiable and have no redundancy. I can, and have, written programs using an evolutionary algorithm
183.
by noosphr 5mo ago
As opposed to integers in normal programming. The 'magic' in weights is that the rules are spread through the whole model and you can't point to one place which encodes them. The grokking paper shows that this stops being the
184.
by noosphr 5mo ago
The point that when the rules the model learns are simple enough they stop being spread out over all the layers and become as easily interpretable as any expert system. It's just that the rules we feed in the model are extremely poorly
185.
by noosphr 5mo ago
The Chinese alphabet is very much a dictionary. All the major tokenizers are far larger.
186.
by noosphr 5mo ago
I'm kind of stunned that someone is using my work to tell me I'm wrong. I wrote the code for the dish brain pong and encoding information was a huge part of what that experiment was about. So when I way that the grok paper and the
187.
by noosphr 5mo ago
It's not often I see something that's fractally wrong but here we are. There is a dictionary, it's called the tokenizer. There are grammar rules, they are just very weak because the structure of human language is generally qu
188.
by noosphr 5mo ago
Scientists have been pissing in all our cups for longer than I've been alive. That you're arguing what concentration of urea is acceptable is part of the problem. There comes a point at which an institution is unsalvageable and th
189.
by noosphr 5mo ago
The crisis has been ongoing since the 80s and isn't slowing down. If the cure for it is to wait till the researchers die so new better ones take their place you just reinvented feudalism with tenure.
190.
by noosphr 5mo ago
Like leeches.
191.
by noosphr 5mo ago
Clinical trails have produced medical breakthroughs, peer review gave us leeches.
192.
by noosphr 5mo ago
It wasn't the party that made the replication crisis. Right now universities have all the trappings of science with none of the substance. Do I particularly care than they are going to be bent to the ends of those in power? Kind of. Th
193.
by noosphr 5mo ago
It's been 40 years, surely someone could have done something to fix it if they wanted to.
194.
by noosphr 5mo ago
That you link to a github is ironic in the worst way possible.
195.
by noosphr 5mo ago
Why not ask about the mismanagement of the other 60 billion?
196.
by noosphr 5mo ago
>California’s public universities spent $16.9 million on A.I. during a financial crisis, and the result has been chaos. So peanuts. The public universities budget in California is something like 60 billion. This isn't even a roundin
197.
by noosphr 5mo ago
Become active investors.
198.
by noosphr 5mo ago
You do realise that you need to store arbitrary binary blobs which don't nicely align to memory words? And that once you can store them you need to write custom functions that do bitwise manipulation on those arbitrary blocks of memory
199.
by noosphr 5mo ago
I'd count it towards the daily cardio target as well. Every minute counts as three. You'd be crazy not to use that productivity hack.
200.
by noosphr 5mo ago
Balanced ternary fp is not a reduced precision type of binary fp: https://arxiv.org/abs/2512.10964 >Unlike their binary counterparts, posits and takums, tekums simultaneously accommodate both ∞ and NaR, while retain
201.
by noosphr 5mo ago
I once replaced IEEE 754 floating point numbers in a model by balanced ternary floating point numbers. It took me 20 minutes. Tell me how you'd do that in cpp?
202.
by noosphr 5mo ago
Well yes, you have to use one of the newer mnist variants these days if you want to get anything meaningful. A linear classifier gets something like 87% on the original one.
203.
by noosphr 5mo ago
Sounds interesting but I'm using very spare very high rank tensors, e.g. rank 3 neuron equivalents. As such pretty much all numerical optimisations are useless for my work. Racket however chugs along happily, if slowly.
204.
by noosphr 5mo ago
I'm doing the learning in racket because the bottleneck is human understanding. That mnist takes 30 minutes per epoch isn't a worry when I don't even know what vector addition should look like.
205.
by noosphr 5mo ago
Racket is an amazing language for prototyping ideas that you don't understand yet. At $dayjob I'm using it to test what novel geometries of deep learning models would look like. Being able to redefine any part of the stack for any
206.
by noosphr 5mo ago
This is more of a utopia than expecting the average office drone to learn emacs.
207.
by noosphr 5mo ago
You should be aware that pandoc markdown has extensive non-standard additions to the language to make it usable. If you want a pure markup language that is simple, plain text readable and able to produce text more complex than what a type w
208.
by noosphr 5mo ago
Ah you should have just said you use arch. But it's a very weak programmer that needs a test suite. Real ones prove the consistency of their programs with goose feather and vellum.
209.
by noosphr 5mo ago
If you get deeply excited about config files, sure, go right ahead and set up the dev env without help.
210.
by noosphr 5mo ago
Ask an llm how to setup. This is one area where they excel at with no caveats.

More ›