Live posting OpenAI DevDay for those not watching.

👇

Spending time thinking about AI, crypto rails, and defense tech.

Sam Altman announced GPT 4 Turbo, with 6 new key differences.

2/ More Control

Multi-function calls now supported.

3/ Built-in knowledge retrieval.

Rolling in RAG into the model.

4/ Rolling image generation, text-to-speech, and more into the API

5/ Releasing custom models as a supplement to fine-tuning.

Allows devs to modify every step of the model development, including pre-training and RLHF.

It will be rolled out with a select number of big enterprises to start.

6/ Less rate limiting.

2x tokens / minute

GPT-4 Turbo will be 3x cheaper for prompt and 2x cheaper for completion tokens.

Overall, 2.75x cheaper.

They're chatting about the collaboration between Microsoft and OpenAI.

Satya is excited about building training infra with OpenAI, using Copilot, etc.

Feels very sales-y toward the developers in the audience.

Moving on to talk about changes to ChatGPT.

First up, ChatGPT will no longer expose all the configuration params to end users. It will automatically choose to browse the web and more, closer to Bard.

Introducing GPTs, a seeming model marketplace.

Key elements - instructions, expanded knowledge, and actions.

OpenAI is becoming a true model marketplace, as rumored.

This is a HUGE move for OpenAI and one that will be bad news for many many startups.

Their first example, a custom code.org explainer.

Plugins -> custom GPTs.

Canva is an example. In practice, this is the next stage of plugins.

Now, Jessica Shieh, a Solutions Architect at OpenAI is showing off a new Zapier GPT.

A demo showing how Zapier can help navigate links between Calendar, Slack, etc.

Sam is now walking through a natural language approach to building your own GPT.

Helping startup founders think through their ideas.

Later this month, they're planning to launch a GPT Store. They're going fully into the marketplace world.

Huge news.

Also launching a new Assistants API.

Includes threads, retrieval, code interpreter, and function calling. Jumping into a demo with Ramon, head of developer experience.

Function calling now guarantees the JSON output with no increasing latency. This is an improvement on many post-processing approaches like Guardrails.

The demo is pretty impressive. It demonstrates how you can pull together multiple results into a single interface leveraging functions and threads in the API.

Threads allow you to save and view results from the API.

Code interpreter allows you to move from qualitative workflows to quantitative ones, i.e. math to split Airbnb prices between participants.

In some ways, code interpreter combines perfectly with GPT.

They're now showing how you can leverage Whisper for tts in a mobile app.

They just gave 500 USD in DevDay credits to everyone in the audience.

The system iterated through each account and granted credits on the spot to everyone.

Pretty incredible combination of Whisper, GPT, functions, etc.

This will likely go down as one of the biggest days in AI in the past year, which is saying a LOT.

So many different announcements in so little time.
* GPT 4 Turbo (including built-in novel retrieval)
* New Assistants API
* Whisper 3
* Multi-function calling
* API support for new modalities
* GPT Marketplace

1/ HUGE context length.

128k token context, way larger than the last 8k context.