Very cool work direction but also fair question.
I wonder if ultimately is a little vision patch VAE the ultimate "tokenizer"? Unicode + UTF-8 is just too high description length.
RESEARCH
-
Vision VAE as Ultimate Tokenizer: Unicode Efficiency
By
–
-
Are People Unable to Detect AI Quality Anymore?
By
–
Maybe people are just not good at detecting quality at this point? (Based on my dive into the actual answers people prefer, certainly plausible)
-
LM Arena Benchmark Relevance Decline Among AI Makers
By
–
It is weird that leading LM Arena went from being the big benchmark every AI maker was aiming for to being not mentioned much in recent releases. Post-Llama 4 reputation hit? Post GPT-4o Sycophantic Apocalypse realization that arena scores were easily optimized? Temporary blip?
-
Implicit Feedback Mechanisms in AI Model Training Systems
By
–
TikTok fyp is a textbook example here. How fast you swipe, how long you watch…feedback cues you can get (and learn from) way more often than explicit thumbs up/down or those 'what did you think of this' post-video surveys
-
Universal Reward Function: Breakthrough in Reinforcement Learning
By
–
A *UNIVERSAL* reward function. Yep. This is one of the most shockingly simple + effective RL approaches I've seen. Kyle and the team are beyond special.
-

Open Source Non-Reasoning Model Achieves Competitive Performance
By
–
Nuevo super modelo de programación no-razonador, que logra unos números muy competitivos colocándose como el mejor modelo open source de este tipo! Eso sí, 1 trillion de parámetros así que olvidaos de ejecutarlo en local fácilmente. Aún así un regalo brutal para la comunidad
-
Handling Research Errors: Retractions and Code Fixes
By
–
i guess it depends on the nature of the issue. like how would that happen? if the conclusion of the paper changed, in the worst case, i guess pursue some kind of formal retraction more likely it'd just be a bug in the code in which case i'd rerun experiments then update arxiv
-

New AI Security Research: Quantization Defense and Runtime Analysis
By
–
they also did some interesting new experiments
– system-level runtime measurements
– pareto analysis
– quantization as a defense
– password-cracking test overall pretty nice work! thorough, detail-oriented, and fair glad to see my results held up http://
arxiv.org/abs/2507.07700 -

Model Training Results Improved to 94% Perfect Embedding Inversion
By
–
i was happy to see that the results held up! and i forgot that i trained the models a bit longer after the paper was submitted so they actually got somewhat better (94% perfect embedding inversion!)
-

OpenAI Embeddings API Changes Impact Model Performance Distribution
By
–
the other results, which use the OpenAI embeddings API got slightly worse. this also makes sense to me the API has likely changed a bit over the last two years they serve the models differently, use different quantization techniques etc. so there is a small distribution shift