Data Machina #173
Limits of LLMs. MLOps for foundation models. Interoperable transformers.TESLA's New NNs. Precision ML. GPT for robots.
The Limits of Large Language Models (LLMs)? So there are lots of researchers out there asking: Can LLMs do this? Can LLMs do that? and so on. A few notes.
A team @Deepmind, investigated whether explanations of few-shot examples can help LLMs, and concluded that explanations can improve LLM performance - even without fine tuning. Weeks ago they published their findings in Can Language Models Learn from Explanations in Context?
Formalising math proofs is extremely difficult. A collab team from MetaAI, Google AI, Allen AI and several top unis, published a paper in which they show that LLMs can write informal proofs, translate them into formal ones, and achieve SoTA performance in proving competition-level maths problems. Paper: Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs
In Large Language Models Can Self-Improve, a team @IllinoisUni @Google, demonstrate that an LLM is also capable of self-improving with only unlabelled datasets using Chain-of-Thought prompting and self-consistency. They claim their approach achieves SoTA performance, without any ground truth label.
An issue with generative text2image prompting is that it takes time for a human to get a decent prompt that generates a decent image. In Large Language Models are Human-Level Prompt Engineers, an anonymous team :-) proposes an LLM as an Automatic Prompt Engineer (APE) that generates prompts on par or better than humans.
Having fun with LLMs. Yesterday I came across an app that uses LLMs to explain papers. Upload your paper and start instantly getting explanations! Try ExplainPaper here. Pretty amazing.
LLMs are an awesome realisation of transformers, but I guess alone by themselves aren’t enough to create useful apps. Harrison just released LangChain, a Python package for building LLM apps that helps you stitch together all the bits & pieces you need.
Jacopo takes LLMs stuff to a more enterprise -less academic- practical level. You know that entity matching, resolution and NER are an absolute pita in the real world. Inspired by Can Foundation Models Wrangle Your Data? he’s built an awesome project: Foundation Models for Entity Matching with Python, OpenAI GPT-3, AWS Lambda, dbt, & Snowflake
Have a nice week.
10 Link-o-Troned
A Pythonista *Experience*
Scripting aRt
Deep & Other Learning Bits
ResearchDocs
Algorithmic Potpourri
El Robótico
data v-i-s-i-o-n-s
DataEng Wranglings
startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.