The Shocking Reality of AI

I read all 154 pages of Microsoft Research's "Sparks of AGI".

Most people don't realise the shocking reality AI is going to bring.

Here's what I found:

— Alex Banks (@thealexbanks) April 16, 2023

1/ Tool use

GPT-4 suffers from serious limitations:

• Lack of world knowledge after September 2021
• Difficulty with math
• Can't execute code

However GPT-4 can use external tools like search engines, APIs and plugins to overcome these limits. pic.twitter.com/yBoVDI6vKK

— Alex Banks (@thealexbanks) April 16, 2023

For humans, the use of tools was a milestone in history.

It allowed us to overcome the limitations of our physical abilities.

The same can be said for GPT-4.

It can use tools with:

• No demonstration
• Minimal instruction
• Make appropriate use of the output pic.twitter.com/QjXOQi1xQb

— Alex Banks (@thealexbanks) April 16, 2023

2/ AI personal assistant

GPT-4 can use multiple tools at once:

• Retrieves information about a user’s calendar
• Coordinates with others over email
• Books a dinner reservation
• Messages details to the user

It can reason with free-form output to solve complex problems. pic.twitter.com/qpywrmNNbp

— Alex Banks (@thealexbanks) April 16, 2023

3/ Mental maps

GPT-4 is asked to explore a house and build a mental map.

It then summarises what it sees after navigating through the map. pic.twitter.com/JpzYRiC8VP

— Alex Banks (@thealexbanks) April 16, 2023

4/ Virtual handyman

GPT-4 is able to identify the actions a human needs to take in order to address a problem.

• Guides the human to find and fix a water leak
• Recommends the exact actions for the human to take

The human acts as GPT-4's real-world agent. pic.twitter.com/CTUj34TW2O

— Alex Banks (@thealexbanks) April 16, 2023

5/ Challenges and limitations

GPT-4 is a next-word prediction model.

It predicts future values based on past values.

It lacks:

• Planning
• Working memory
• Reasoning abilities

Whilst it can produce coherent text generation it struggles with complex problems.

— Alex Banks (@thealexbanks) April 16, 2023

This can be broken into two types of intellectual tasks.

Incremental tasks (GPT-4 is great at these)

• Solved sequentially or gradually
• Progress made by adding one word or sentence at a time
• Examples: writing summaries, answering factual questions, composing poems

— Alex Banks (@thealexbanks) April 16, 2023

Discontinuous tasks (GPT-4 is bad at these)

• Require a creative leap in progress
• Involves discovering new ways to frame problems
• Examples: writing jokes, developing scientific hypotheses or new writing styles

— Alex Banks (@thealexbanks) April 16, 2023

What I'm excited for:

• Augmenting GPT-4 and language models with external memory
• Implementing feedback loops
• Equipping language models with plans

With the recent addition of plugins to GPT-4 we have some exciting times ahead.

— Alex Banks (@thealexbanks) April 16, 2023

I feel the fire has been lit as we approach AGI.

Follow me @thealexbanks for more on AI.

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

Join 26,000 others and cut through the noise in AI each week:https://t.co/Kre1Uawx3B

— Alex Banks (@thealexbanks) April 16, 2023

Help everyone learn and retweet this thread: https://t.co/ix1fbCS7gZ

— Alex Banks (@thealexbanks) April 16, 2023

Read the full paper here:https://t.co/sqYRy5umIT

— Alex Banks (@thealexbanks) April 16, 2023