Yesterday we had an amazing product update by Sama, Satya and Jessica — total Steve Ballmer soaked-armpit levels of energy. I’m told the e/acc party in SF also had such energy.
Below, I give my thoughts on the update, please enjoy.
Two million developers
Visualize in your mind Ballmer on the stage saying “developers” 2M times and sweating more and more each time. By the time he’s done, we’re at the singularity already. That’s how large and powerful Ballmer’s army of developers has become.
Why This Matters: When you have 2M developers building products on your platform, working independently on independent problems, some of them are going to succeed big. The overall platform will expand massively.
In other news, Ballmer has just about as much wealth these days as his old boss Billy Gates. The only centibillionaire to not be a founder!
Product Update Recap
Many of the updates are competition-drive and reactive. The accelerated race to multi-modality comes ahead of Google’s Gemini, the increased context GPT-4 offering comes behind Anthropic’s general availability for Claude2.
Yet, many of the updates demonstrate genuine leadership in the space. There are innovations across capability, price, availability and SLA, DX, and importantly, business model & ecosystem innovations. That OpenAI is strongly innovating in all of these categories simultaneously is immensely accelerative.
I split the updates into three categories based on the proprietary AI-GPT scoring methodology (Advanced Innovation and Genuinely Powerful Technology).
There are three categories:
leading the pack — these innovations are meaningful and new even if minor
keeping pace — these innovations appear reactive in nature, even if major
mixed bag — these innovations appear experimental in nature, the purpose of these branches of the tech tree may become more clear over time
Leading the Pack
DALLE3 availability over API — huge, the delta between DALLE3 and DALLE2 is one of the biggest incremental deltas of all time. This new capability will unlock incredible new applications.
Price Rate Limit Improvements — doesn’t make for exciting headlines, does drive technology forward and improve margins for AI developers. 2x and 3x improvements on parameter values is amazing, and shows how OpenAI continues to advance their scaling capabilities. Thank you Sama!
GPT Store — The GPT store legitimizes prompt engineering as a discipline in its own right. Providing a way for prompt engineers to monetize their skills is BPE (Big Platform Energy), and will accelerate the development of this unique craft and ultimately the techno-capital acceleration.
Assistant API — building RAG into the OpenAI offering, giving these models infinite context by default will make it much easier for AI devs to churn out CRUD apps of all kinds. This update removes a big pain point for AI devs and will be a powerful accelerant.
Reproducibility & Seeding — this is powerful for API customers so they can validate precisely that the models haven’t dropped out from under them, a continued source of friction between AI devs and OpenAI.
Copyright Shield — yet another incremental reason to remain on the OpenAI platform, especially big corpo.
Keeping Pace
New TTS offering — a good business move on OpenAI’s part because it allows voice-based chatbot apps to stay fully within the OpenAI ecosystem. And the voices sound pretty good, so it’s good for consumers too.
Whisper2 → Whisper3 update. Whisper2 was already quite incredible, and this powerful update is exciting. Voice Recognition is becoming really mature.
128 context GPT-4 Turbo— A strong update, seemingly in response to Anthropic’s general availability of Claude2, that has the potential to unlock new use cases, but its ultimate impact remains to be seen.
Mixed Bag
We’re releasing several improvements today, including the ability to call multiple functions in a single message: users can send one message requesting multiple actions, such as “open the car window and turn off the A/C”, which would previously require multiple roundtrips with the model
Function Calling Enhancements
Not useful: It smells to me that OpenAI isn’t able to provide compelling examples about how and why one would want to use the functions calling feature. Nobody is building an application that controls your car with LLMs. I’m not convinced that there’s anything of interest with functions calling at the moment. 1
Useful: My belief is that there is some bigger technical vision with function calling, that it will enable us to build more powerful agents down the line. Interested to see how it evolves.
JSON-format guarantee
Not Useful: Being able to guaranteed JSON-format in the response without guaranteeing anything about the shape of the JSON structure is on the face of it absurd, like a gas station that guarantees what comes out of their pump is liquid, but not necessarily something that will work well in your car.
Not Useful: Anybody that is using JSON output in their applications is already validating the type in a more specific way (at least they really really should be). So the additional check is about as useful as checking that a shape is a rectangle BEFORE checking that the shape is a square — logically it’s redundant.
Useful: catching errors more quickly (at the OpenAI servers instead of consumer app servers) is likely to provide a broad-based performance improvement for consumer-facing applications.
Useful: as a “dog whistle feature” that OpenAI is moving in the direction of guaranteeing structured outputs from the API, this feature does have marketing/branding/communications value. I wonder if Microsoft pushed for this. It’s almost 100% messaging which feels like more their style.
I’m probably missing something?