Aurora, Meta's MMS, Ghost in the Minecraft, LIMA, AI Risk, Voyager
Created on May 31|Last edited on June 1
Comment
Aurora
Intel is planning to train a 1 trillion parameter LLM on their supercomputer, Aurora. They plan on training this model primarily on scientific data in fields concerning: biology, chemistry, materials science, physics, and medicine.
Meta MMS
Meta's Massively Multilingual Speech (MMS) project is focused on speech2text and text2speech. They extended this to over a thousand languages! What is MMS exactly? They combined one of their other recent works, wave2vec2.0, with lots and lots of additional data, with roughly 1,100 labeled and about 4,000 unlabeled. They've open sourced their project on GitHub.
Since most multilingual datasets cover only 100 of the 7000+ languages in the world, they had to find a way to find more data of rarer languages. And thus they turned to Bible translations to acquire an additional 1100 languages worth of data! They also employed a complicated method involving an alignment model to filter our bad data. Since, even with this new corpus of biblical data, it is not enough to train speech recognition models, they employed wave2vec2.0, the main model behind MMS.
Interestingly, they also compare with OpenAI's Whisper, which is another speech recognition model.
Future work in this ongoing project leans towards finding a singular model to solve a variety of multilingual tasks.
Ghost in the Minecraft
Ghost in the Minecraft or GITM is an example of an LLM controlling an agent, which in this case is a character in Minecraft. The authors employ LLMs to achieve a Generally Capable Agent (GCA), instead of one that is optimized towards a specific task.
One interesting diagram that caught my eye was their technology tree. Their agent was able to unlock a great deal more unique items than existing Minecraft agents.
The method they created uses 3 LLMs: LLM Decomposer, LLM Planner, and LLM Interface. One LLM is for decomposing complex goals like "craft a diamond pickaxe" into simpler subgoals. LLM Planner revises plans, leverages its memory bank, and generates new actions for the in-game character. They developed a structured feedback mechanism for the LLM Planner. The LLM Interface is the actual model controlling the in-game character and relaying in-game information back to the other models.
LIMA
LIMA, or Less Is More for Alignment, is a 65B LLaMA model fine-tuned with standard supervised learning without any RLHF for alignment tuning. They claim LIMA is able to generalize well to unseen tasks and demonstrates strong performance. They conclude that knowledge is more likely learned during pretraining and only a minimal amount of fine-tuning is necessary to urge the model to produce high quality generated text.
AI Risk
Similar in nature to the 6 month long pause letter, this letter addresses AI as an extinction-level risk that should be placed as a global priority next to things such as pandemic and nuclear war. Some of the signers include Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, and Sam Altman.
Voyager
Another minecraft LLM-RL paper, Voyager leverages GPT-4 to play minecraft!
The idea behind this paper is that the agent, powered by GPT-4, explores via an automatic curriculum. It develops a skill library over time, sort of like the memory bank in GITM. The iterative prompting mechanism controls the actual behavior of the in-game character. Across both Voyager and GITM, it seems that the primary paradigm is to develop a method to remember certain skills and leverage these for unseen tasks like combating zombies. Their ablation studies found:
- the skill library to be crucial to not plateauing in later stages
- the automatic curriculum is needed for consistent progress/exploration
- GPT-4 >> GPT-3.5 in code generation
- self-verification, the method to add and refine skills, is integral to new item discovery
GitHub Repositories
I've also come across a few interesting new repositories! LLM-As-Chatbot is a repo for people to use open-source LLMs as a chatbot service. All you need to do is pick one to use!
ToolBench is based on the paper Tool Learning with Foundation Models, and surveys the role foundation models and LLMs play in tool learning. They provide data and scripts to train a ToolLLaMA model, which, presumably, is a LLaMA model trained to interact with tools.
References
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.