Diego

Imagine you have to train a robotic hand to grab a cube using Deep rl: > Start by complex multimodal nn architeture, adding layers, lstms (nerve signals), CNN (cameras), every possible buzzword. Create a first rough reward function. > Not surprisingly, hand gets cramped and doesn't learn anything > Add batch normalization layers, gradient clipping. > Better progress > Try using different rl algorithms, PPO, DDPG, SAC, changing the architecture each time. > Reduce nn complexity, adjust reward function. > Still no significant results, partial progress. > Iterate like 25 times tuning hiperparameters, changing the archite ture. > Reduce complexity of model again, simplify reward function. > Finally, this happened

@XMihura Encima es como el 37% del presupuesto. Literalmente "Otras cosas"

Cohesion, Resilience and Values €75.8 billion

@EU_Commission We need data centers

We have proposed the EU budget for 2027! 💶 Almost €200 billion directed toward shared priorities: 💶 Strong and competitive economy 🇺🇦 Continued support for Ukraine 🛡️ Stronger defence 🌾 Agriculture 🏠 Affordable housing ⚡️ Energy transition 🎓🚆 Key EU programmes

@ylecun @ClementDelangue @Dan_Jeffries1 Is this the linux for llms?

@ClementDelangue @Dan_Jeffries1 Everyone, please join Project Tapestry thealliance.ai/projects/tapes…

Concentration of power, capabilities and economic wealth is the biggest risk in AI. We need open science and open-source more than ever!

@AliGrids A function that large shouldnt even exists

POV: Senior tells you to "just refactor line 6061" 🪦

@Gamingtronium What if you forget your passwrod?

Hot take: the username field is completely pointless. Change my mind.

@Samaytwt He apologized tho

Never touching cursor again 😭

@analogalok Imagine burning gemma4 12B weights on a FPGA board

Run Gemma 4 26B MoE on 8GB VRAM with 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the replies

@alvarobartt Maybe trying your luck on wallapop. Prices are crazy nowadays

It might be time to get my hands on a NVIDIA GPU myself, but pricing is crazy, not sure what to do, any recommendations?

Conciusness is the subconcious dream , the subconcious hallucinating a virtual RTOS. I believe It could change significantly in case It has access to immense computing power. Your thinking of yourself would evolve rapidly, until you are no longer recognizable for what you once were

Because it seems unreasonable that if you were replacing pieces of one's brain with electronic parts functionally identical their consciousness would fade out.

Yes, exactly. You could start with expanding memory packs for people with memory problems, that the brain could store and retrieve information. Then, if you apply this to the rest of the brain, you could even let the brain modify itself in order to acomódate for the new capacity, until the old brain was almost vestigial. Who knows, maybe your whole personality would change, since you Will be 100x faster and capable.

Amazing what 8h of sleep do to the human mind

@yacineMTB But you could increase the friction to get a succesful policy and then decrease It little by little

I refuse to make the task easier. I could increase friction or reduce gravity but that feels like giving up

still haven't gotten 6 pendulums but recently just got a top scorer

@_wallfacer @SorensonCorben @yacineMTB Precisely. Near the upward equlibrium there is also a guaranteed controller by kalman, as the linearization is controllable.

@didacum333 @SorensonCorben @yacineMTB Higher frequency resonant modes not being reached with the smooth action steps.

still haven't solved cartpole 6. i'm going to repeatedly bash my head through the drywall

@_nasch_ Cuanta VRAM tienes? O es en RAM?

190 tokens por segundo con Qwen3.6 35B en hardware local. A esta velocidad todo se siente instantáneo. Los que tienen suscripción de Anthropic o de OpenAI es porque quieren.

@Strife212 You could also solve step by step by hand all the molecules from your body

LLMs are intelligent, but they clearly aren't conscious? You could (incredibly slowly) run an LLM by hand by doing the matrix calculations with a pen and paper, would that be conscious? There is zero difference between doing that and doing it on a GPU except its faster.

@SorensonCorben @yacineMTB That could be interesting, increase the damping. I think one Big issue with this env could be the action steps, as you add joints, you should need snaller reaction times in order to stabilize It, since it's a much more unstable env

@yacineMTB Is there any resistance programmed into the joints? Perhaps even having a different resistance per joint. Then after it starts nailing it back off on the resistance and see how it does until you are back to ‘normal’.

@louis030195 This question is going to get harder in the future, but perhaps that is good news

probably the most important job question every company should ask in 2026 🥲

@Gonzalo_stba Ojalá hubiese más personas como tú en España. Estás recibiendo por todas partes solo por estar interesado en hacer cosas chulas

Pues primer día de prácticas finalizado y qué decir, estoy alucinado He estado de 9.00 a 18.30 y se me ha pasado el día volando, he aprendido más en un día que en los últimos 4 años. Mis jefes son increíbles y unos obsesos de la IA como yo, me dan libertad y me animan a desarrollar procesos y herramientas para cualquier cosa que se me ocurra Me ha dado pena tener que irme a las 18.30 para estudiar los exámenes, pero bueno, es lo que toca. En cuanto acabe exámenes no dudo que estaré currando de 9.00 a 20-21 (como ellos) porque es algo interesantísimo y el ambiente es perfecto Además puedo ir andando a la ofi, tardo 30 min. He tenido una suerte increíble con esto, soy un privilegiado de poder trabajar un par de meses con ellos, ojalá durasen más.

@yacineMTB What if you increase damping on the joints to learn a policy, then slowly decrease It again?

well it's getting one of them up..

There are two possible outcomes from this experiment: 1. There exists a policy st It stabilizes the 5 pendulums in the position 2. It is physically imposible due to a severe instability problem. Perhaps It IS the very action steps definition. Perhaps It is needed to reduce the time between actions to the maximum? Anyways, It could be cool seeing It work

2023, 3 pendulums 2025, 4 pendulums 2026: