A few weeks back, I surmised that now was the time for builders given the state of Generative AI. And, with yesterday’s Developer Day, OpenAI has completely upped the game on the entire industry on innovation once again AND reinforced the point that now is indeed the time for builders. In short, yesterday’s announcements evolved OpenAI into a fully-fledged developer platform. Let’s recap and analyze starting with the product announcements:
- Chat GPT-4 Turbo – This is the next iteration of the already available GPT-4 model. It adds two new versions, one for text analysis and one for text + images. GPT4-Turbo can now analyze up to 128K input tokens (equivalent to ~100,000 words), which is a 4X improvement over the initial GPT-4 model. Web-based knowledge (without having to resort to leveraging Bing) is now current as of April 2023 versus September 2021.
- No Code Bot Building for End-Users + a Future App Store – End-users can create their own chatbots. These can be connected to databases or knowledge-bases to import external information. But more importantly, they can be created and shared without writing code – catapulting OpenAI into the low/no-code game. These bots will be able to be shared with others in the future with a GPT app-store, which will include creations from “verified builders.” These bots will be available to any Plus or Enterprise Customers.
- API for Assistants – OpenAI debuted a new Assistants API to allow developers to build agent-like experiences. These can retrieve outside knowledge or call programming functions for a specific action.
- DALL-E v3 API – The recently debuted, very popular image creation API now has its own API, including built-in moderation tools.
- Audio API – OpenAI debuted its first text-to-speech API, with six pre-defined voices included.
- Copyright Shield – OpenAI is following other big-tech stalwarts such as IBM, Amazon, Cohere, and Microsoft in offering indemnification against IP infringement claims on content created by ChatGPT.
- Researcher Assistance for Model Creation – OpenAI will be initiating a program for strategic customers whereupon the company’s AI researchers will help customers build their own models utilizing OpenAI technology.
- Automatic Model Selection – The model picker feature will be going away in lieu of automatically picking the best model based upon prompts and data provided.
- Whisper Large v3 – OpenAI debuted the next version of its speech recognition model.
- GPT-4 Token Limits – These are doubled with immediate effect for all paying GPT-4 customers.
The first take-away is that innovation from OpenAI is coming at a blistering pace that shows no signs of slowing down. From initial debut of Chat GPT 3.5 on November 30, 2022, Chat GPT 4 came in March of 2023 – a timeframe of only 4 months. Chat GPT 4 Turbo came 7 months later, effectively becoming the first year anniversary update. Developers should plan for two major updates per year, approximately every 6 months.
The second key take-away is that OpenAI is serious about differentiating itself as an AI platform, versus just a large language model available for consumption. They are the first to market with a low/no-code solution, app store, and copyright indemnification. This is on top of having a fully mature and set of accessible APIs covering the breadth of their offerings. They clearly want people to come, build, and share. Their partnership with Microsoft will ensure that there are truly enterprise-friendly versions of these capabilities, as has been the case with Azure OpenAI to-date – though capabilities will inevitably lag by several months.
The third take-away is that anyone that has built a chat/agent/answer-bot is screwed. This is Chat GPT’s core use case and anyone investing here has just been commoditized. It’s time to push reset and find differentiation in another direction.
The fourth and final take-away is that costs (and potentially much higher) are coming. Buried in these announcements is the fact that Chat GPT Plus is no longer a $20/month subscription. Utilizing the text-only Turbo model costs $0.01 per 1,000 input tokens and $0.03 per 1,000 output tokens. It will cost $0.00765 per image to process with Turbo. This is inevitable. Speculative reports indicated that each query of Chat GPT 3.5 was consuming approximately 17 oz of water and the cost of the data centers was $700,000/day. Assuming all million active users were paying users, OpenAI would still be losing $1.7M on data center costs alone. This technology is going to get more expensive as it gets more sophisticated – particularly as customers start to care about things like data privacy (e.g. isolation; my data is not shared for model training) or guaranteed response times (e.g. dedicated compute capacity).
Regulation is on the horizon as well, but this is something that is (1) a topic for another day and (2) not something people have much to worry about, yet – as definitive legislation is still very much a work in progress.
If one was not convinced before that the days of building AI-enabled applications utilizing large language models weren’t here, this week’s announcement should be the definitive proof. It is time to build, and there is no excuse not to – especially with barriers to entry being materially lowered yet again. The only caveat is to find one’s niche outside of the path of OpenAI, which is focusing on the broad, generic use cases.
In closing, my good friend and colleague Arturo Toledo (@ArturoT), CEO of @Pixelspace, sums it up very well. A builder extraordinaire, his takes:
