1. Debut of Code Interpreter

OpenAI announced the release of Code Interpreter:

https://twitter.com/openai/status/1677015057316872192

You can enable it by going to Settings:

This is currently available only for ChatGPT Plus users.

The name Code Interpreter is terrible. It’s not as self explanatory as ChatGPT. OpenAI has a technical explainer.

Logic and math have been persistent enemies of LLMs. Some doubt if LLMs can even be called AI because of this. On the other hand, LLMs can generate working code. So, how about using the ability to generate code to answer logic and math questions?

Google has applied this to improve Bard’s reasoning. Code Interpreter is OpenAI’s attempt at this.

OpenAI has further extended this ability. They added storage space (100MB). And now you can ask ChatGPT to work with the uploaded data. ChatGPT can write code on that data to do visualizations and data analysis and convert data between formats. You can download the converted/processed data. The best part is that the file is saved only for that session.

Why is it important? With Code Interpreter ChatGPT can:

1. Solve mathematical problems

2. Lower the hallucinations because it has to work with data and code

3. It writes and debugs its own code – In my limited testing ChatGPT found data conversion issues in its own code and then re-wrote the code to resolve such issues.

What’s next? Code Interpreter might be the best possible use of GPT right now. As a noob coder it takes me a week to implement a data analysis program. All the while I have to struggle with having proper data and types etc. While Code Interpreter can do the same analysis in minutes.

If you are a data analyst and worried about your job then fret not. Currently it supports only Python. So, you need to know Python to verify the code and ensure your instructions are being followed correctly.

To learn more, you can read Ethan Mollick’s guide.

2. GPT-4 API and other releases

The second announcement from OpenAI was the general availability of GPT-4 API. While GPT-4 has been available for ChatGPT Plus users, API was restricted to few people. With this change the API is now available to all paying users.

GPT-4 is still costly. The 8k version which was released is still 10x costlier than the 16-k GPT 3.5.

The second major announcement was OpenAI changing the focus from the original text completions and instruct models to chat completion models. The chat models have structured prompting with system, assistant and user roles. It is more conversational.

Why is it important? GPT-4 is more powerful than GPT-3.5. But many developers were hampered by the fact that an API wasn’t readily available.

The instruct and text completion to chat completion is a downer. The chat models are well..chatty. It needs to add more context and stuff like “As an AI..”. It takes multiple tries to stop it from outputting anything but the required text or code. The functional calling feature in chat completion might be added for this exact purpose.

What’s next? With GPT-4 API’s availability I expect more apps using GPT to be released in the coming months.

3. Copyright and Google Scraping

Copyright continues to be a challenge for Generative AI. This week saw some interesting news stories.

First, Steam responded to claims about banning a Redditor’s game. We discussed this story last week. The Redditor got banned from Steam for using AI-generated art. Steam’s response is that the policies were dictated by the current copyright laws.

The general feeling is that Generative AI games are going to be a copyright nightmare.

Secondly, there are a couple of lawsuits from book authors in both the UK and US.

In the US, Sarah Silverman and authors Christopher Golden and Richard Kadrey have sued OpenAI and Llama. The issue is somewhat glaring with Llama as it indirectly used data from shadow libraries.

While in the UK, it was by authors Mona Awad and Paul Tremblay.

While this is happening Google has updated its privacy policy. The policy now says:

For example, we may collect information that’s publicly available online or from other public sources to help train Google’s AI models and build products and features like Google Translate, Bard, and Cloud AI capabilities. Or, if your business’s information appears on a website, we may index and display it on Google services.

Google maintains a history of changes. So, you can view this change.

What’s next? Last week I noted that most of the Reddit data is free and maybe people cannot sue for it. But looking at these privacy changes from Google, I might be wrong. Things might be free to read but not reproducible or at least used for AI training. This is going to be an interesting copyright battle for the LLMs.

🗞️10 AI news highlights and interesting reads

1. While everyone wants their data out of ChatGPT, the performance has been taking a hit. An interesting discussion on the OpenAI forums.

2. And OpenAI disabled “Browse with Bing” to do right by content owners.

3. OpenAI says we need systems to steer and control AI systems. While RLHF is great, it doesn’t work for machines smarter than us. So, they are building an automated system to align and control systems. This reminds of the famous quote - Quis custodiet ipsos custodes? Or Who will watch the watchers? As in who will align these supposedly smart automated systems?

4. While ChatGPT’s rise might be slowing.

5. And the US military is using Generative AI for super secret stuff. Sigh!

6. 4 Tips for programmers to stay ahead of generative AI.

7. As a programmer, Langchain might be pointless.

8. New York City to regulate how AI is used in hiring. The idea is to stop the in-built bias AI tools might have.

9. Using GPT-4 to frustrate scammers. Though what happens once scammers can generate endless scam calls?

10. Everyone wants to build a chabot for answering questions. But what if customers don’t want chatbots?