Exploring Innovations in Generative AI

Last week Ali Rowghani, Partner at Y Combinator talked with Raza Habib, CEO of Humanloop.

They dived into cutting-edge innovations in generative AI.

Here's a breakdown of what the future might hold:

— Alex Banks (@thealexbanks) March 7, 2023

1/ What is a Large Language Model (LLM)?

An LLM is a statistical model that can process and understand human language at a scale.

It learns patterns and structure from large amounts of text.

Then it uses that knowledge to predict what word will come next given previous words.

— Alex Banks (@thealexbanks) March 7, 2023

2/ What's the challenge of using a pre-trained model like ChatGPT?

They have a tendency to hallucinate.

This is when the model generates text that appears to make sense on the surface.

But underneath it confidently gets it wrong leading to misinformation—this can be dangerous.

— Alex Banks (@thealexbanks) March 7, 2023

3/ What does fine-tuning mean?

You adapt an LLM to a specific domain by training it on a smaller task-specific dataset.

The difference between GPT-3 and ChatGPT is fine-tuning.

Two types:

1. Adjusting the tone from data before use
2. Capturing data when the model is in use

— Alex Banks (@thealexbanks) March 7, 2023

4/ Problems when building apps using LLMs

This can be broke up into 3 parts.

• Prototyping
• Evaluation
• Customisation

Let's dive into each:

— Alex Banks (@thealexbanks) March 7, 2023

Prototyping

In the early stages, you have to get a good prompt that works well for your use case.

Building a chatbot for copywriting would be very different from one for interview practice.

— Alex Banks (@thealexbanks) March 7, 2023

Evaluation

This needs to be iterative to find the right prompt for you.

You need to evaluable how well your app is working with your end customers.

Conducting user tests and gathering feedback can help identify the most effective prompt that resonates with your users.

— Alex Banks (@thealexbanks) March 7, 2023

Customisation

Everyone has access to the same base models incl. ChatGPT's API.

But if you want to build something differentiated, you need to find a way to customize the model to your own:

• Context
• Use case
• End users

Make something your users prefer to the base model.

— Alex Banks (@thealexbanks) March 7, 2023

5/ The impact of generative AI on developers

In the short term, it augments developers.

GitHub Copilot has built an AI-powered code autocompletion tool.

It has over 100 million users.

Today, 46% of code is being built using GitHub Copilot across all programming languages.

— Alex Banks (@thealexbanks) March 7, 2023

In the long term, developers become more like product managers.

The grunt work is taken care of by the models.

Developers write the specification and documentation.

As we approach AGI, developers will see large fractions of their job be automated.

— Alex Banks (@thealexbanks) March 7, 2023

6/ The next LLM breakthrough

The model takes action for you.

AdeptAI is building a model that interacts with everything on your computer.

You can ask "Find me a refrigerator for under $1,000".

The balance will shift from us treating them as text-generation machines to agents.

— Alex Banks (@thealexbanks) March 7, 2023

7/ Will there be one model that rules them all?

This is unlikely.

The barriers to entry for training LLMs are capital and talent.

You need lots of capital to pay for GPUs and tap into a dense talent set.

But there's a lot of opportunity for building end-user applications.

— Alex Banks (@thealexbanks) March 7, 2023

8/ Has OpenAI reached its mission of building AGI?

AGI is when an AI system can perform intellectual tasks at a level that is comparable to or exceeds humans.

The median estimate for AGI is 2040.

This is remarkably soon for a technology that might upend all of society.

— Alex Banks (@thealexbanks) March 7, 2023

9/ LLMs for startups

Previously you'd require research teams to tackle tasks that once felt impossible.

Now you can just ask the model.

I believe it's your duty to use these models to place yourself in a position of leverage.

— Alex Banks (@thealexbanks) March 7, 2023

Follow me @thealexbanks for more on AI.

If you liked this thread, you'll love the newsletter.

Subscribe here:https://t.co/Kre1Uawx3B

— Alex Banks (@thealexbanks) March 7, 2023

Help everyone learn and retweet this thread: https://t.co/u9RKSGkkxu

— Alex Banks (@thealexbanks) March 7, 2023

You can listen to the full interview here:https://t.co/uvK4OgBS8u

— Alex Banks (@thealexbanks) March 7, 2023