¡NUEVO EQUIPO y MODELOS de IMAGEN! Se han juntado bajo el nombre de Black Forest Team varios investigadores de muchas de las tecnologías tras los modelos actuales y han creado un nuevo equipo, y publicado nuevos modelos: FLUX Y pinta bien!
RESEARCH
-
Gemini 1.5 Pro Updated Version Ranks Top Chatbot Arena
By
–
We have an experimental updated version of Gemini 1.5 Pro that is #1 on the @lmsysorg Chatbot Arena. This model is a significant improvement over earlier versions of Gemini 1.5 Pro (it cracks into 1300+ elo score territory). I'm really proud of the whole team of people that
-
Gemini 1.5 Pro Tops Vision Leaderboard Rankings
By
–
The new Gemini 1.5 Pro model is also #1 on the Vision Leaderboard!
-
arXiv Labs Community Project Integration and Institutional Authorship
By
–
Right, once you appeal authorship though, we can add your institution. Linking to arXiv accounts is also a great idea. We are hoping to join as an arXiv Labs Community Project so we'll keep you updated!
-
Epic Fall Winter AI Model Releases and Compute Innovations
By
–
8/ This will be an epic fall/winter of even more model releases. It will be interesting to see how differences in compute, data, and innovations (Q* etc.) will drive changes in the rankings going forward!
-
Post-Training and Data Strategy: New AI Competitive Driver
By
–
7/ We see strengths (Claude 3.5 w/coding, Gemini 1.5 w/Vision, Multilingual) from the models that are likely driven by post-training and data strategies. This is contrary to the prevailing beliefs 1 yr ago, where pre-training was believed to be the core competitive driver.
-
Post-Training Data Strategy: Key Competition Arena for LLMs
By
–
6/ Post-training data strategy is becoming a key arena for competition. Llama3.1 paper had 15 pg about post-training data (vs. 12 pg on pre-training), driving these capabilities: – Code
– Multilinguality
– Math/Reasoning
– Long-Context
– Tool Use
– Factuality
– Steerability -
Google’s TPU Advantage Could Dominate AI Compute Race
By
–
5/ Going fwd, Google's TPU advantage will be very important. Every lab is reliant on their NVIDIA GPU allocations for progress, except for Google, who can successfully train on TPU clusters. This could mean that Google will have more compute than the other labs in perpetuity.
-
Post-Training Data Strategies: SFT, RLHF, and DPO Approaches
By
–
4/We are also seeing remarkably similar data strategies for post-training from most labs at this point (at least what was published from Meta+Apple): – Hybrid data SFT, RLHF, & DPO setups
– Synthetic data on code and math
– Post-training data for most important capabilities -
H100 GPU Rollout Explains Timing of Recent AI Model Releases
By
–
3/The reason these are all so close together timing-wise is that every lab got their H100s at roughly the same time. They each struggled with early issues with the H100s last fall, and the big H100 clusters all started training this spring. Voila, 5-6 months later, big models!