Luke Monington @lukemoningtonAI

I'm passionate about LLMs, open-source, and all things AI. Check out my YouTube channel! youtube.com/channel/UCPyHa… Joined May 2023

Tweets

205
Followers

73
Following

117
Likes

404

Luke Monington @lukemoningtonAI

2 years ago

Create that AI project. Build that OpenAI plugin. Take that data science course. Just start!

0 0 3 151 0

Luke Monington @lukemoningtonAI

2 years ago

Do you want to become a machine learning engineer? The most important step is to just start

0 0 2 159 0

Luke Monington @lukemoningtonAI

2 years ago

Will the generative ai boom have the same outcome as the Dot-com bubble? Will there be a valley of disappointment? What is the next step in the evolution of ai? Perhaps the next stage is the autonomous ai boom. If this is true, it will come next year

0 0 0 143 0

Luke Monington @lukemoningtonAI

2 years ago

The biggest challenge in the ai world right now is alignment

0 0 0 91 0

Luke Monington @lukemoningtonAI

2 years ago

things are moving so quickly in the ai world that it actually makes sense to start from baseline. New research comes out every day that challenges some of our most fundamental assumptions of ai. Are positional embeddings good? Maybe. Is knowledge distillation good? Maybe.

0 0 0 80 0

Luke Monington @lukemoningtonAI

2 years ago

Do you think GPT-5 will come out by the end of this year?

0 0 0 65 0

Luke Monington @lukemoningtonAI

2 years ago

- Friend struggles with chat support 🤔 - Asks for live agent 🗣️ - Chat says it is a live agent 🤯 Conclusions: 1) People think AI is super advanced 2) LLMs are set for widespread integration

0 0 0 56 0

Luke Monington @lukemoningtonAI

2 years ago

Is there a magic LLM size? Is there a point where scaling larger doesn't work anymore? Some of our largest models might be at that point. It was hypothesized that GPT-4 might have a trillion parameters. But what if it just has a couple hundred billion?

0 0 1 83 0

Luke Monington @lukemoningtonAI

2 years ago

It is fascinating how quickly the AI alignment problem has taken center stage. RLHF will not work for a superintelligent system. What is the solution? David Shapiro talks about Axiomatic Alignment. That might work

0 0 0 62 0

Luke Monington @lukemoningtonAI

2 years ago

We are running out of data to train LLMs. The largest models from the most well-resourced companies require an enormous amount of data. What is the solution? Maybe synthetic data #LLMs #SyntheticData

0 0 0 49 0

Luke Monington @lukemoningtonAI

2 years ago

Why does in-context learning perform poorly compared to fine-tuning? There are two possible reasons: 1) The LLM fails to learn sufficient representations for the task 2) The LLM doesn't have the capability to reason well over the representations

0 0 0 44 0

Luke Monington @lukemoningtonAI

2 years ago

Now that we have 4-bit LLMs for inference and training, what is next? I foresee the near future being dedicated to improved dataset curation techniques

1 0 3 61 0

Luke Monington @lukemoningtonAI

2 years ago

I think the most interesting part of the TART research paper is that the reasoning module is entirely task agnostic. A single module can be applied to a bunch of tasks and added to any LLM. This could become a valuable part of an LLM pipeline

0 0 0 53 0

Luke Monington @lukemoningtonAI

2 years ago

Are prompt engineers the future? That depends on whether the prompt engineers will be able to distinguish themselves. Because at the end of the day, if everyone is a prompt engineer, then no one is.

0 0 1 55 0

Luke Monington @lukemoningtonAI

2 years ago

What is the best LLM customization for your task? If you have a small amount of data: in-context learning If you have a lot of data: fine-tuning

1 0 1 93 0

Luke Monington @lukemoningtonAI

2 years ago

OpenAI's function-calling capabilities might be what will allow companies to distinguish themselves. Otherwise, anybody can replicate a prompt

0 0 0 75 0

Luke Monington @lukemoningtonAI

2 years ago

Even after MONTHS of the open-source community and proprietary models pushing to achieve GPT-4 levels, we still haven't achieved it. Eventually, we will achieve it. But by then, maybe GPT-5 will be out. OpenAI has quite the lead!