Vasileios Valatsos :coffefied:
banner
aethrvmn.sigmoid.social.ap.brid.gy
Vasileios Valatsos :coffefied:
@aethrvmn.sigmoid.social.ap.brid.gy
ml engineer
#rl + #nlproc researcher
python enjoyer & nim appreciator

[bridged from https://sigmoid.social/@aethrvmn on the fediverse by https://fed.brid.gy/ ]
Me fixing this means that my thesis is even more buggy. I might revisit it some time, but I already got my masters in September so I probably won't
March 25, 2025 at 9:36 PM
anybody knows where I could possible get a cheap server to train models?

I dont think my setup is sustainable, and my gf is getting cranky (honestly I cant blame her)
March 25, 2025 at 1:34 PM
I might be blowing my own trumpet but the biggest issue (that is also unsolvable using #LLMs) when it comes to #nlp #decision #making is that #language #models have a distribution over #tokens not over #actions.

This is the issue that the #alectors https://apotheke.earth/docs/alectors tries to […]
Original post on sigmoid.social
sigmoid.social
March 24, 2025 at 11:49 AM
Fasting sugar for 3 weeks (only three weeks to go till Easter) and then drinking Coca Cola makes me feel better about fasting sugar.

Its so sweet and artificial. Even the "zero sugar" version
March 23, 2025 at 9:53 PM
Europeanists have extreme memory loss. #trump's position aince 2016 has been that the US is spending too much for #nato whilst the EU countries (bar Greece) are underpaying. He has also been saying that that the EU has been overly reliant on the US.

Now #EU politicians act as if doing exactly […]
Original post on sigmoid.social
sigmoid.social
March 20, 2025 at 3:26 PM
I wonder if encoders could be used to create search indexes. Like, you encode the contents of a website, store the [CLS] tokens or something like that, maybe a gated self-attention output, and store everything in a vector database.

Then you get the user query and compare it with the vectors […]
Original post on sigmoid.social
sigmoid.social
March 16, 2025 at 11:54 AM
The documentation is finished!!!

I am touching up a small couple of things on the repo and the code itself but until then I am extremely happy and proud to share...

Alectores! Cocks (Roosters)!

https://apotheke.earth/docs/alectores

It is a python library that enables #nlproc by a #rl agent […]
Original post on sigmoid.social
sigmoid.social
March 10, 2025 at 9:21 PM
another week, another evening spent moving things around on my site

I dont even keep any logs apart from nginx. I dont even know if anyone is visiting it, but I like it.

Probably will write a blog post soon too, one on the #Mozilla drama (my 2 cents) and one on free software and the free market.
March 7, 2025 at 10:48 PM
I want to get a good enough GPU to run a good enough model, and just feed it with mysticis litersture, like absolutely esoteric stuff, and build a wiki/knowledge base of all of the different things that come up, one paper at a time
March 2, 2025 at 7:47 PM
Champagne socialists are a specifically aggrevating type of hypocrite, because of their superiority complex
March 1, 2025 at 3:33 PM
Getting ready for Orthodox Lent by eating as much fast food as possible, so I have fat to burn through the fasting
February 28, 2025 at 7:51 PM
Reposted by Vasileios Valatsos :coffefied:
things need to be stupid for them to be reliable.
February 28, 2025 at 2:42 PM
a small prose on prompt engines.

When I was studying physics I heard the phrase "you dont understand something until yoh can explain it to your grandma.".

We also used to say "Make it simple but not too simple.", so I hope it's simple but not too simple.

https://aethrvmn.gr/self/words/engine/
prompt engines
Prompt Engines In order to trust that a large language model is able to complete tasks with relevant competency, it is important to equip said model with “blinders”, which try to make the model focus on the single task for which it is deployed, and hopefully prevent the derailment of the conversation by a user, whether maliciously or otherwise. The main way this is achieved is using prompt engineering, which is a fancy way to say that we are continuously reminding the agent/model what it task is, and what it should focus on.
aethrvmn.gr
February 27, 2025 at 10:13 PM
@matrix what if instead of talking about funding you stop hosting your servers and cut down on being an expensive metadata aggregator? Tell people to not use matrix.org for their account and let them diffuse.

#matrix is supposed to be decentralised.
Decentralise it.
February 27, 2025 at 1:54 PM
I think I might actually have to create a transpiler that takes a phonetic script like Greek and outputs a custom pictographic representation
February 25, 2025 at 12:00 PM
I was looking at my contract and I noticed that there is no mention of transfer of ownership, license, or copyright.

Assuming this means that the code and docs remain under my ownership, I'll probably upload a set of handbooks which are notes I've written about different topics regarding #LLMs […]
Original post on sigmoid.social
sigmoid.social
February 24, 2025 at 11:58 PM
Using mastodon as a rubber ducky: When it comes to learning #nlproc in a #rl setting, it seems to be very close to pathfinding; "Here is the sentence, here are the actions, find the correct next token out of all possible token".

The problem is that in this case instead of having four actions […]
Original post on sigmoid.social
sigmoid.social
February 20, 2025 at 5:20 PM
After more than a week of bugfixing my custom RL and Transformer implementations (this means that some of the bugs affected the performance of my thesis woops), I am finally ready to start overfitting my agent with openwebtext data. I'll try to build a truly open source NLP agent, so more info […]
Original post on sigmoid.social
sigmoid.social
February 20, 2025 at 11:11 AM
Reposted by Vasileios Valatsos :coffefied:
someone should feed these elaborate monologue models positive affirmations as part of the instruction for thinking. also, horoscopes.
February 17, 2025 at 3:27 PM
lol
lmao even
February 17, 2025 at 1:50 PM
What are some good libraries for #semantic mapping ? I want to convert a sentence (say "The sun is shinung") into a Python dict (say "Sun"-"subject" "Shine"-"verb")

I know #spaCy exists but that's all I know about the matter

#nlproc
February 15, 2025 at 10:10 PM
thinking about it, I would expect languages like pictograms like Chinese or ancient Egyptian to be much easier for LLMs to understand because they encode the semantics rather than encode the phonology. As far as I know they also encode context to some extent.
February 14, 2025 at 2:35 PM
@BrodieOnLinux asked ChatGPT about FUTO and it recommended your podcast.
February 12, 2025 at 12:32 PM
why are .ai domains so sought after but nobody uses .ml?
February 10, 2025 at 11:00 AM