Thread by @transitive_bs

Yesterday's GPT-4 launch was **crazy** 🔥

After discussing w/ colleagues and taking some time to collect my thoughts,

Here are the top 3 things GPT-4 enables that were previously impossible or extremely difficult to achieve – including examples 👇

1/ Image comprehension

⇒ Unlocks full-screen app and browser automation 🤖

We thought text-based co-pilot and writing assistants were cool, but now that GPT-4 natively understands images, automating app and browser workflows based on screenshots is going to skyrocket 🚀

Here's an example to get your creative juices flowing:

https://twitter.com/sharifshameem/status/1635750987272183808

Here's another example of this type of browser automation from @AdeptAILabs – they also raised $350M yesterday which mostly got buried in the news!

https://twitter.com/AdeptAILabs/status/1570144499187453952

⇒ Mobile adoption. Cameras were one of the killer features of phones in the early 2010s that led to Instagram, Snapchat, etc.

Now that GPT-4 understands images, we're going to see all sorts of crazy new AI-powered use cases emerge mobile cameras 🔥

https://twitter.com/mckaywrigley/status/1635740909383061504

Here's another early example of image comprehension + mobile camera adoption in action:

https://twitter.com/swyx/status/1635765117303521282

2/ Longer context length (up to 32k tokens!)

You can think of this as the AI getting more RAM like in the early days of computers. You can do amazing things w/ low RAM, but developers need to do al sorts of low-level tricks, and it's harder to scale.

With longer context, GPT-4 will have more memory, cohesion, and consistency for larger inputs.

⇒ With gpt-4-32k, apps can now ingest entire documents, websites, short novels, repos, transcriptions, etc, and it will "just work". This makes "training" the AI in-place much easier.

Apps will also have the capacity to generate much longer, more cohesive & consistent outputs.

With GPT-3, you could generate short poems and essays, and chain these together to generate longer text.

⇒ With gpt-4-32k, you can now generate entire short novels, stories, and entire apps.

This is going to massively boost no-code tools and help so many more people build / write / create without having to worry about the lower-level details 💯

https://twitter.com/AlphaSignalAI/status/1635747039291031553

Here's a great example of this in action: (the author also includes his prompts in the thread to explain how he created it)

https://twitter.com/felixbade/status/1635799243628892160

3/ Incrementally unlocking more reasoning and multi-language use cases

https://twitter.com/DrJimFan/status/1635694095460102145

Previous LLMs like GPT-3, ChatGPT, Claude, LLaMA, etc all struggled with different types of SOTA reasoning tasks. Same with cross-language tasks.

You can mitigate these shortcomings to an extent via prompt engineering, chaining prompts and custom logic, etc, but it's difficult.

⇒ GPT-4 will unlock many new use cases involving more advanced reasoning and cross-language tasks, without as much need for sophisticated developer tooling and custom logic.

A great example of this is it's ability to do taxes 🤯

https://twitter.com/JStaatsCPA/status/1635739660923330560

⇒ GPT-4 will unlock even more use cases in education. Better reasoning and consistency across longer users sessions greatly improves the UX for tutor / teaching scenarios.

A great example of this is yesterday's Duolingo release of their AI tutor:

https://twitter.com/duolingo/status/1635688521695633408

To summarize:

1.  Image support ⇒ app/browser automation and new mobile camera use cases 🔥

2.  Longer context ⇒ input/output entire stories, websites, apps and massively improve no-code tools 🚀

3. Better reasoning ⇒ solve tougher problems (taxes!) and education tasks ✨