We've seen a lot of new preference alignment techniques lately, but this is a really cool one combining model merging and RLHF.
LLMS
-
Decision Boundaries Analysis for LLM In-Context Learning Properties
By
–
Nice analysis of the properties and smoothness of decision boundaries under different conditions for in-context learning in LLMs by @siyan_zhao, @tungnd_13 and @adityagrover_. https://t.co/wpY9yfV3oB
— Jeff Dean (@JeffDean) 25 juin 2024Nice analysis of the properties and smoothness of decision boundaries under different conditions for in-context learning in LLMs by @siyan_zhao
, @tungnd_13 and @adityagrover_
. -
LLM Skepticism Sparks Range of Emotional Responses
By
–
Responses to my posts on LLMs not being all that, have moved on the grief scale: Shock *Denial **Anger *Bargaining Depression Acceptance Processing. Anger often alludes to me being too old to understand. But newly bargaining around how people are sometimes stupid too. Progress…
-
Parallel Response Generation Model Reduces AI Hallucinations Instantly
By
–
Imaginad un modelo que en un instante pueda generar varias respuestas "en paralelo" para minimizar alucinaciones, analizar y evaluar su propia respuesta, planificar y razonar paso a paso, etc Todo en un segundo o menos.
-
Specialized Transformer Chip Achieves 500K Tokens Per Second
By
–
Qué barbaridad. Un chip especializado sólo para ejecutar Transformers! Si nos parecía rápido Groq con los 800 tokens/segundo, esto llega a los 500.000 tokens/segundo ejecutando Llama 70!!!
-
Clarification on AI voice model availability and UI status
By
–
It is only the UI, voice model is not yet available
-
Data Wall Threatens AI Progress: Innovation in Data Abundance Needed
By
–
This is spot on from @dwarkesh_sp.
— Alexandr Wang (@alexandr_wang) 25 juin 2024
The data wall is a looming threat to continued AI progress.
Breaking through will require innovation on data abundance.
What’s become clear over the past few years is these AI models are a compression of their data, and that the only way to… https://t.co/836PSMUNo0This is spot on from @dwarkesh_sp
. The data wall is a looming threat to continued AI progress. Breaking through will require innovation on data abundance. What’s become clear over the past few years is these AI models are a compression of their data, and that the only way to -
How LLMs Acquire Factual Knowledge During Pretraining
By
–
How do LLMs learn new facts while pre-training? Excited to have authors @hoyeon_chang and Jinho Park
answer questions on their latest paper "How Do Large Language Models Acquire Factual Knowledge During Pretraining?" Leave questions for the authors: https://
alphaxiv.org/abs/2406.11813 -
Reverse engineering Gen-3 model access via frontend feature flags
By
–
The gen-3 option is hidden behind the feature flag from the public. It has been reverse engineered The UI is what is currently out there in the frontend code
-
User perspective on Google’s AI Search Summaries
By
–
MS not sure. Google released Search Summaries which are meh so not sure