Does this have use cases now? No, it's far too buggy.
However, the concept of this with refined and/or default chain of thought logic and better managing control over the sequence, it can do anything.
(E.g. Find me some good real estate investment properties in the San Francisco area and send a message to the seller's agent to schedule a meeting on my calendar) <> References internal data set of my investment criteria from long-term memory and prior responses, Browses MLS, Find Broker Email, Sends, Syncs with my Calendar.
An example is "create a table showing the current ratio of rental/sales prices (on a per square foot basis) for residential properties in the 10 most populated counties in Colorado."
I'm still yet to get a reasonable result from trying this.
"Find out the cheapest per unit Pampers New Born diapers from the ones available on amazon.in, flipkart.com, bigbasket.com, dmart.in, firstcry.com considering all the available card discounts, coupon codes, shipping and other charges etc. "
I see it's already ready to replace the sales team.
Customer: "Can your product compute infinite loops in finite time?"
SalesGPT: "Hell yes, our developers can do anything, would you like another vodka martini?"
It's also easier to do these days than you'd think. Let me know if you want help setting it up.
I recently used it to have it find a source of data and scrape it as a test, and it works. Took only 30 min or so.
Could have done it in under 5 min by myself, but taking 20s to write the prompt is still faster than spending the 5 min doing it myself
But this is a hacky fix, and will never be reliable enough for consistent use. For that, more actual research is necessary, on how to simulate and model goals and trains of thought and have them interface with the world model provided by an LLM.
Obviously we don't know what paths will be most successful. But a path where critical drivers of AI (like goals) are modeled in a transparent and comprehensible manner seems like a very attractive direction to take. I'd much rather be able to read my AI agents goals, plans, intermediate goals, self-analysis, etc., than have it all captured in a set of completely incomprehensible weights.
Part of the issue here is the massive amount of compute needed over what we're already spending. ToT is showing a likely 10 to 20x number of calls to get an answer, which when you are compute limited is going to be a problem for deployment in mass. It's very likely we're going to have to wait for more/faster hardware.
Even if the output were faked, it would help a lot to understand what this does. (I assume it’s “Geocities for AI agents” but couldn’t confirm.)
I wonder how profitable selling of non-products is with marketing like this.
Anyway, they don't allow email signups, either. You have to use a social media login. So the pie gets even smaller.
Edit: also the logo is not displaying properly on the signing page (Firefox, linux)
The first is that OpenAI has rate limits. They are especially small for GPT-4, which for anything complicated can be quite a lot better than 3.5.
The other one is that if you do leave it open, then you can be sure that a significant portion of your customers will be from countries that could not pass the OpenAI phone verification. Or just didn't want to identify themselves. For some reason.
Combine that with something that scares some people a lot like autonomous AI or connecting them to servers or the internet, and it feels like you might be on thin ice with OpenAI or some regulatory group. Especially if a bunch of Russian and Chinese users are finding you on some directory or post listed as a way to get around the phone verification.
Just based on experience running my own service.
Heck, even a video/screen recording would be a significant improvement.
And yeah, the initial "planning" query and the planned "tasks" are all separate GPT queries.
The idea behind these AutoGPTs is to give greater structure to the way the model thinks and acts, through a loop of planning, executing, and reflecting. Another goal is to provide the model with enhanced memory, similar to or identical to retrieval.
In short AutoGPT-like applications treat the LLM as a reasoning core around which a scaffolding of higher-order thinking is built.
It’s interesting, as a cognitive architecture. Imagine that your frontal cortex, responsible for planning, was very rigid and farmed out its actual reasoning tasks to your language-learning brain area.
There's nothing in the bottom right.
I wonder the same about "AI".