Big News: OpenAI Unlocks Powerful Coding Workflows with "GPT Mentions" Feature
Agentic AI - Edition 8
TWO pieces of exciting news this week!
Item 1: Awesome New ChatGPT Feature
OpenAI released a huge new ChatGPT feature on Friday with absolutely zero hype. The feature is called "GPT Mentions", and it allows ChatGPT users to include multiple custom GPTs into the same conversation.
This may sound innocuous, but it's actually another big step toward publicly available agentic AI—AI that can think and act on its own. Here are the facts:
ChatGPT is currently unable to act agentically because it is stuck in "conversation mode". It can participate in a back-and-forth conversation, but it cannot continuously act on its own. This is partly for safety/oversight/accuracy and partly a limitation of the models.
ChatGPT is currently unable to work through difficult problems because it doesn't have the ability to "take time to think". This is something Ilya Sutskever talks about. We need some way to let ChatGPT take more time.
If two or more GPTs can be setup to talk to one another, we can overcome both of the aforementioned issues. Two GPTs communicating with each other would produce an autonomous system that could think and act indefinitely, taking as much time as they need to solve any given problem. The AutoGen open source project already leverages this concept, using the OpenAI API to create autonomous AI agents.
How does "GPT Mentions" get us closer to having multiple GPTs communicating?
We can now bring multiple GPTs into the same conversation, sharing the same context. As the user, I still need to prompt each GPT to talk, but they are now talking to each other. I can literally just sit there and type:
Hey Tech Lead GPT, create a new task.
Hey Coder GPT, implement the task.
Hey Reviewer GPT, review the code.
Create a new task.
Implement the task.
Review the code.
...
For now, OpenAI is keeping humans in the loop, but with GPT Mentions, we can now take a backseat and let the GPTs drive.
So does this actually work?
That brings me to the second big news this week.
Item 2: My YouTube Debut
There is no way for me to explain in a Substack post the potential impact of "GPT Mentions", so I've put together a video showing you my exact coding workflow. Please take a watch! It was a lot of fun to make, and I think it's really going to open eyes to the possibilities for these tools.
And this is just the beginning.
People think that OpenAI is creating hype to drive subscriptions. From what I've seen so far, I think OpenAI is hiding the true capabilities of their tools in the name of “gradual rollout”.
Yes, Sam Altman is perfectly open about the fact that they're trying to create an AI to replace everyone's jobs. He's not subtle in that respect.
But how close are the current models? He states every time he's on camera that there will be a slow rollout of AGI to give people time to get used to it.
This new feature is something we will need to get used to.
The only AI creator that I follow who has put-out a video about "GPT Mentions" so far is Wes Roth. There was no hinting from OpenAI ahead of time. This feature was practically stealth released.
OpenAI is still downplaying the capabilities of the already-released GPT-4.
Google's Gemini Ultra is reported to be similar in capabilities to GPT-4, and Google has now stated that they're laying-off a bunch of people to focus on AGI.
Does AGI seem far away? That AI coding workflow in my YouTube video is more powerful than a lot of human coders out there. It still needs guidance from an experienced engineer, but its speed and accuracy far surpasses what I could do on my own.
ChatGPT is not as good at updating large, pre-existing codebases, but does it matter? If it can make programmers 100 times quicker, it can re-write any software out there in a fraction of the time. With the right guidance, the re-written code would probably be much higher quality.
I know ChatGPT cannot write good stories yet, but do we really need it to?
What happens when ChatGPT can start coming-up with better AI algorithms? How long until it’s able to make the GPT architecture obsolete by creating something better?
They won't be able to call it ChatGPT anymore.
What did Ilya say a couple months ago when asked about when AI models will be able to solve complex problems?
"How do you know that the current models cannot do it?"
Perhaps the only thing holding us back right now is those GPT-4 usage caps, or, if you’re using the API, the prohibitively expensive cost of GPT-4 usage.
When the GPTs start doing all the work, 50 messages per 3 hours just isn’t going to cut it.
Have you tested any meta prompt frameworks, conditional or dynamic prompt flows with the GPT Mentions feature?? I'm super interested and the only content on the web about using this feature to make multiple GPT communicate and work together is yours. Please I'm very interested and think this feature is capable of so much more, trying to figure out to tackle the issue of contextual awareness. Do you think ICL or making use of the knowledge base is more optimal? such as SOP, or a user guide but for each GPT to understand the mission, what they are part of, and their responsibility, etc.