►Recent Highlights from the Previous Thread: >>106181054

--Benchmarking large models on Apple M3 Ultra using MLX with unified memory:
>106182005 >106182020 >106182469 >106182033
--Qwen3-30B-A3B-Instruct struggles due to sparse activation and model version confusion:
>106184073 >106184096 >106184112 >106184124 >106184141
--PyTorch 2.8.0 and 2.9.0-dev show regression in inference speed vs 2.7.1:
>106182552
--LLM progress stagnates as multimodality becomes the new battleground:
>106181123 >106181186 >106181229 >106181354
--Local anime video generation improved with WAN 2.2 and new I2V LoRA:
>106181318 >106181331 >106181773 >106181992
--DeepSeek's influence persists through open-source model derivatives despite perceived decline:
>106182694 >106182704 >106182719 >106182737
--GPT-5 tops LMArena but leaderboard credibility questioned over style control and past inaccuracies:
>106182205 >106182236 >106182248 >106182254 >106182302
--GPT-oss 120B generates generic, warehouse-obsessed text with no creative spark:
>106182587 >106182701 >106182929
--Debate over AI companies engaging with 4chan amid distrust of corporate "safety" culture:
>106181983 >106182000 >106182040 >106182054 >106182174 >106182045 >106182056
--Skepticism over claimed novelty of unified AI architecture:
>106183618 >106183666
--GPT-5 minimal reasoning underperforms GPT-4.1 in early benchmarks:
>106182859
--Tesla Dojo team disbanded as members launch AI startup DensityAI:
>106182941 >106183041
--Anon claims GPT-5 is live in Copilot with screenshot evidence:
>106183385
--Benchmarks:
>106183826 >106183973 >106184375 >106184399 >106184529
--Logs:
>106182027 >106182335 >106183337 >106183359 >106183399 >106183422 >106183634 >106183699 >106183778 >106183959 >106183962 >106184282
--Miku (free space):
>106182747 >106182814 >106183862 >106184081 >106184207

►Recent Highlight Posts from the Previous Thread: >>106181065

Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script