Throwing shit at Mochizuki was not in my books for 2026.
But honestly? I do not respect him anymore.
He is a genius, probably top 10 brilliant minds in human history but he allowed a "trust me bro" with 0 facts (Sholze) to dictate the tempo and set the ground.
He allowed a step back in humanity just cause of ego; hes duty was to prove S is full of shit, if you cant prove it he cant either, trust me bro is not facts but hey, they give medals for that.
Most Critical: Layer 0 (99.87%)
Layer 1 (95.35%)
Layer 15 (83.06%)
Least Critical: Layer 12 (61.88%) ← Only "prunable" candidate
Layer 6 (66.23%)
💡 KEY INSIGHT
This 1B model has ZERO redundancy.
Every layer contributes 60-99% to cognition
Early layers are absolutely essential
Cannot be compressed via layer pruning
Fully utilized architecture
Layers 0-1: 95-99% impact
Handle tokenization and initial features
Cannot be pruned
3. Middle Layers are Distributed (62-78% divergence)
Layers 2-11: Ensemble processing
No single bottleneck
Distributed cognitive architecture
4. Late Layers Refine Output (62-83% divergence)
Layers 14-15: Critical for coherence
Layer 15 (final): 83% impact on output quality
Layer Impact Ranking
It took +2 days in 3 stages thanks to my pc rebooting... honestly I was expecting the architecture to have redundancy but this just proves: LLM's are one of the biggest achievements in human history and we actually build them properly (im building, just better).
Critical Discoveries
1. Layer 0 is THE GATE (99.87% divergence)
Removing Layer 0 = Total model failure
Cannot process ANY input
Absolutely irreplaceable
2. Early Layers are Essential (95%+ divergence)
jaccard index not fucking around taking 2 days to finish + 1 layer I finished yesterday but I had to reboot (checkpoints added hehe). I expect to find sauce once im done.
Ryiuk chat and UI/UX work is so motherfucking annoying... why I have to copy my GOAT steve jobs and be so perfectionist with UI? I dont know fucking app is free and open source but hey, do it right 1 time enjoy for life.
My genshin impact growing.
RE: https://mastodon.social/@daavilefx/116052310397727207
"We can build a better world."-- they had decades to build it why try now? 😂
Cause they are salty A.I is superior to whatever thing they used to squeeze.
I will also get salty if someone averaged 40% a year... oh wait no, anyone can profit in markets and thats irrelevant too me, I still get paid no matter what others do.
But hey, if they show facts who cares if its human, A.I or a rat. They will build a SaaS and sell it while we are trying open source route.
Facts over words.
Important to note, NTT is a shit company that threaths their employees LIKE TRASH, pays trash salaries and seeks countries like mine (mexico) to exploit us with shit salaries they will pay x3 times higher in any other country FOR THE SAME TASK.
I will never even consider working again for a parasitic company like NTT DATA.
Targeted scammers need better sauce, I built the typescript rust compiler microsoft wanted, I would might fall for that.
Bro scams getting wild:
1. Using a soccer player name (i love soccer)
2. Using 1 shit company I used to work for
Whats the purpose of this scammers for real? sad worl we live on.
I posted to chill some average tech i built, not even my prime rib sauce and I got nuked with SaaS parasites and now targeted scammers FUCK ME.
I hate internet, I want out asap.
