fabian

Took me a while to realize that curiosity matters more than anything.

[fabian @Fabipar91]

GEM IN UKRAINIAN i started the story back when I was 16 and it's about the world where every person has a curse that makes their lives suck even more and ofc it has ppl with wings and is a relegion-critical piece so yeah I am very proud

[fabian @Fabipar91]

Reconocer las propias contradicciones no resta valor al mensaje, sino que humaniza el proceso de evolución. Nadie posee el camino completamente resuelto; avanzar consiste en sostener la intención de mejorar mientras se aprende a convivir con las propias grietas.

[ömer Ergün @merErgn86305811]

https://t.co/GRuzoGD7Pv
https://t.co/b6mceXHEvE

[fabian @Fabipar91]

@ClipsDrunk

[fabian @Fabipar91]

@NicoGarcia_IA this bothers me more than it should but look at her eyes

[fabian @Fabipar91]

@JustOneNic

[Tailer @0xtailer2]

THIS ROBOT PLASTERED AROUND A LIVE ELECTRICAL CONDUIT WITHOUT KNOCKING IT LOOSE ONCE. AN ELECTRICIAN CHARGES A CALLBACK FEE EVERY TIME A CREW DAMAGES ONE OF THESE.

A bare wire sticking straight out of the brick, right in the middle of the patch it's working, and the robot just plasters around it, smoothing compound up to the base of the conduit without ever catching it with the trowel.

That's the actual skill in finishing work nobody talks about, not the smoothness of the wall but knowing what not to touch. A careless pass catches a wire, bends a pipe stub, or buries something that was supposed to stay accessible, and now there's a second trade back on site fixing a mistake that had nothing to do with plaster.

The wall comes out clean either way. The difference is whether the electrician gets called back for a five minute fix or a full afternoon.

[fabian @Fabipar91]

@jnr_ben1

[Deepak kumar @Deepakk30081433]

आर्याणां देवता सूर्यो विश्वचक्षुर्जगत्पतिः।

कर्मणां प्रेरको देवः पूज्यो ध्येयश्च सर्वदा॥

🌞जय सूर्य देव जी महाराज 🌞

[fabian @Fabipar91]

Been busy with my new home 🎀🏡(This is pre-reno) Renovation and planning are honestly so much harder than I imagined. Between sourcing the best pieces, making countless decisions, and dealing with my current home going through HIP at the same time.

[Kotte @Gustafssonkotte]

A 671-BILLION-PARAMETER MODEL. 720GB DOWN TO 131GB. STILL WORKS.
A friend of mine, ML engineer, just published the actual playbook.

- Bit width allocated unevenly: first ~3 layers (0.5% of weights) and all attention kept at 4-6 bits
- Only the rarely-activated MoE experts pushed down to 1.58 bits
- Runs on 2x H100 80GB. ~14 tok/s single user, ~140 tok/s aggregate
- Uniform compression across the whole model would have produced garbage instead

Most people think quantization is one number you apply everywhere. The actual skill is deciding which 0.5% of the model you're not allowed to touch.

[martin ohiggins @eneasbahia]

https://t.co/2rkJpQPFZW
https://t.co/KbTAI5zjjR

[giorgiozabul@hotmail @aguiladecepcion]

https://t.co/gGez54PbT5
https://t.co/f3CyCKVYeR

[acabaenmiboca @rogeimi]

https://t.co/OEjFA3pBhg
https://t.co/jZjJYipu7h

[volkan diktepe @volkandiktepe]

https://t.co/vtwrNsaSl7
https://t.co/5hW9xGdYBu

[ANALBERT GUTIERREZ @anacoco51]

https://t.co/tyyNwKDLza
https://t.co/aaeiu1B8Wa

[tony2000 @elmanotas36]

https://t.co/IfHPUFZ3qH
https://t.co/IsuCEswVs6

[jeffersonmengo @jeffersonmengo]

https://t.co/6ePyk01pCt
https://t.co/8p0OsbINRK

[jesusalbertoherrera9 @jealher]

https://t.co/6HujiH5d5N
https://t.co/aZjP7bpGbU