Founder of @daertml. Training LLaMAs as a hobby (and no profit yet).
Madrid, Comunidad de Madrid
Joined April 2018
- Tweets76.1K
- Following3K
- Followers825
- Likes71.9K
Pinned Tweet
AGI achieved externally in the 4chan chat by miqudev anon, on 29th January 2024.
Here goes a 🧵with Miqu rocking everything I ask (from datasets, random things I find from the internet and more).
Feel the AGI!!
Using the Q5 (biggest model) version, with this llama.cpp config:
I loved this game as a kid, never finished a level, but the vibes were unmatched... possibly one of my fave games
2
238
J-lens from @AnthropicAI
Has been done before by @chrislhayes , who btw works for @IBM , you guys are learning from the wrong people 🙄youtu.be/8Ppw8254nLI?is=PqRU…
1
231
A new benchmark for the man @PINTO03091
more glitch exploration with my time slice nodes and Wan 12V + GPT Image 2.0 in @ComfyUI ✨
76
I got the 3rd gen book with all the mons back in the day (likely 20 years ago)... it felt magical when mum came back with it; she encouraged me to push in exams further with this kind of treats (also yugioh cards and the like). That appeared announced everywhere. Also had 1st gen
【全1025匹】『ポケモン公式ぜんこく図鑑 1996-2026』が予約開始
✅Amazon：amzn.to/4doUPnT
📅発売日：2026年8月7日（金）
✅価格：2,750円（税込）
・ページ数448ページ/オールカラー
1996年のシリーズ誕生から2026年までの30年間で登場した全てのポケモン1025匹を完全収録
#PR
77
Llevo sin escribir a mano una puta linea de código desde octubre de 2024. Esta semana van 1000+ líneas para hacer cositas tochas con IA. Lo siento, pero no vuelvo atras, por mucho que empresario empresariez quiera que pique las líneas a mano.
Arthur Mensch (Mistral) : "aujourd'hui les ingénieurs logiciels chez Mistral n'écrivent plus de ligne de code", ce sont des "managers d'agents" ; la productivité est énorme pour un individu, mais au niveau d'une grande entreprise il reste un goulot d'étranglement organisationnel.
109
My fav Pokemon game. The only one i finished the 721 Pokedex. 500+ hours played. ORAS hit different
1
73
At 130 bpm im panicking, hand sweating and leg tremors... He is at 130 bpm like any of us at 50 bpm rest
This Post is from a suspended account. Learn more
1
238
Continual Learning solved on a friday
Instead of forcing models to hold everything in an active context window, we can use hypernetworks to instantly compile documents and tasks directly into the model's weights. A step towards giving language models durable memory and fast adaptation.
Blog: pub.sakana.ai/doc-to-lora/
80
So, my take from last @polynoamial interactions: you can get a lot from old/outdated models (e.g.: llama3.2), and do things that new big models (e.g.: qwen3-coder) do by using the right multi-agent harness and looping/ensuring you keep the spark running in your agentic system
1
46
Apparently you can merge CNNs, classic ML models into LLMs, and your LLM will learn to do the things that such models were able to do... I did that with Blackjack, to let GPT2 play the game, using the knowledge from a 10 second trained CNN model:
github.com/DaertML/model_fus…
El dia que se de cuenta de que las zapatillas de 150 euros no son obligatorias llevarlas puestas. Se va a volver loca.
HazteOir descubriendo el capitalismo es algo que no esperaba ver este 2026. Imaginaos cuando descubran que las zapatillas que se venden a 150 euros ha costado 15 euros fabricarlas. Se van a volver locos.
23