Language models (LMs) using tools is one of the most exciting research & engineering directions in AI. Philosophically, it’s the extended mind hypothesis (
https://
en.wikipedia.org/wiki/Extended_
mind_thesis
…). Practically, it is likely to lead to LMs that are more accurate and factual, an to many cool apps.
@nandodf
-
Language Models Using Tools: Extended Mind and Practical Applications
By
–
-
Senior Researchers Mentoring Students at Khipu AI Conference
By
–
I’m looking forward to attending @Khipu_AI – I hope all speakers do. When students at @DeepIndaba and @Khipu_AI interact with senior researchers we all learn, and it does wonders for their aspirations.
-
DAGGER vs RL: Feedback Methods for LLM Training
By
–
While DAGGER is a great idea to enable Feedback for LLMs (eg chat) it is not a replacement for RL because RL opens up room for different forms of feedback (eg preferences). However, as a teacher I would advise careful measurement of the contribution of each to the final metric.
-
DAGGER Counterfactual Teaching Method for LLM Training
By
–
DAGGER is a form of counterfactual teaching as explained in https://
arxiv.org/abs/2110.10819 – Note that it is the student who always acts. The teacher only provides corrections, which are used to minimise the LLM loss directly. Note however that this imitation IS NOT supervised learning. -
DAGGER: Imitation Learning Alternative to Reinforcement Learning
By
–
People are asking if there are alternatives to RL in RLHF. Yes, imitation with DAGGER (tutorial: https://
ri.cmu.edu/publications/a
n-invitation-to-imitation/
… ). The user provides feedback with corrections, e.g. when the agent says “that” the user tells the agent that instead of saying “that”, it should say “this”. -
Chain-of-Thought Prompting Enhances Language Model Reliability
By
–
Fascinating work combining chain-of-thought prompting and tool use to improve the reliability of language models.
-
Computing Performance and Algorithm Innovations in AI Benchmarking 2022
By
–
How are we doing in computing wrt this summary in June 2022? On algorithms, any great new ideas like FlashAttention was back then? https://
spectrum.ieee.org/mlperf-ranking
s-2022
… -
Nvidia’s Megatron Language Model: Foundational AI Research
By
–
Nvidia released the megatron language model before the pandemic. It’s amazing how influential this paper became. A must read for people wanting to learn about AI.
-
Efficient Checkpointing Challenges in Distributed Systems
By
–
Haha don’t give all the answers away! Though efficient and fast checkpointing in a distributed system is a challenge in itself. Would love it if you could share some of your exercises with the rest of the community
-
Distributed MLP Training Across Multiple Machines: Architecture
By
–
One possible exercise: You have a one-hidden-layer MLP, which is so big that the weights can only fit across 4 machines. Describe how you’d do the forward-backward passes. What would you do if each machine had a small probability of breaking down?