OpenAI Plug-In & Sparks of AGI
Created on March 25|Last edited on March 25
Comment
OpenAI Plug-In
What are these plug-ins? You can think of them as components that build onto ChatGPT. One example plug-in is Wolfram Alpha.
ChatGPT can use it to perform calculations instead of trying to compute its own, which may be much more error-prone. And many other companies including Instacart, Speak, and Expedia have added their own plug-ins (third-party plug-ins). In short, ChatGPT now has access to a wider array of tools, allowing it to be more dynamic, unlocking a rich set of capabilities and applications, and much more in the near future!
They have added 2 particularly interesting plug-ins themselves: code interpreter and web browser.
The code interpreter allows ChatGPT to run code! You could imagine how this would expedite the debugging process, especially for software developers who regularly use ChatGPT. The code interpreter currently supports Python.
As the name suggests, the web browser plug-in lets ChatGPT browse the web for more up-to-date information! This has been a long time coming and I'm confident it's been requested by the community for quite a while.
They also created their own open-source retrieval plug-in. The retrieval plug-in gives ChatGPT the ability to retrieve information from user-inputted documents or organization vector databases. I'd imagine it like Ctrl + F but on a grander scale.
Sparks of AGI
This 154-page paper argues that GPT-4 is an early version of AGI. They conduct a series of tests including:
- multimodal capabilities (working with image generation models)
- coding ability
- math ability
- embodied interaction
- testing theory of mind (interacting with humans)
- Analyses
- current metrics of measuring these models
- limitations of the autoregressive architecture
- societal influences like bias
Below I list some of the interesting examples I found:
- their approach to studying GPT-4 was more based on psychology (unlike how most models are evaluated on benchmarks)
- it passes mock coding interviews on LeetCode => could be hired as a software engineer 😂
- has stronger common sense than previous models
- able to integrate many different fields together (write a poem about physics, a play on discrete mathematics, draw an image with TikZ, etc)
- it can zero-shot a simple, but working 3D game, understand code better than ChatGPT, solve International Mathematics Olympiad questions
- though limited, it can reason about what external APIs or tools it might need to solve a problem
- it can produce a mental layout of a house given text descriptions
- GPT-4 can reason about the mental states of other characters better than its predecessors
- the model is autoregressive; it won't know what text it will end with (i.e. the model won't know what the last word or sentence it will send)
References
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.