I’ve been building stuff w agents for ~3 months. Some scattered thoughts on the current agent hype:
On agent reliability:
before you get too scared or hyped about autonomous agents, try running one yourself
they have flashes of clarity and then tend to rabbit hole / get stuck / go off course. they need intense guardrails to reliably perform a task. They can be very random
They are looping computer programs that have a 5-15% failure rate every time you execute a new loop
People are assuming this will be trivial to fix but I actually think it’s quite hard because there is inherent non determinism in the LLM which is multiplied at each loop of execution
On agents seeming alarming:
agents have as much power as you give them w tools / API access, if you give them access to search the internet you’re building yourself a web scraper, which is cool, but not dangerous. If you give them REPL access, maybe they’ll run rm rf, so probably don’t do that. In general running agent-generated code without serious safety checks is not a good idea for a production app, just like you wouldn’t run random code a stranger sent you without reading it
The agent doesn’t have goals / abilities of its own beyond what’s baked into the LLM and the APIs you provide it. It’s as if you are running a python script that continues to generate as you run it. It can do what a python script w access to specific APIs can do. It does not have superpowers to do anything beyond that.
Not saying it’s not cool / exciting to see the agent programs self assemble! I like playing with them too. :)
@rationalaussie Probably most b2c products will have a presence in Chatgpt. But businesses will be cautious to use it entirely due to platform risk. Many existing systems will use it as a channel alongside the other channels.
I've been seeing various experiments on ChatGPT.
I just created my first ChatGPT plugin to get the Price of Bitcoin.
The integration is fairly simple and you just need to build the right API.
Lots of consumer-facing apps will quickly build their ChatGPT plugins.
At @MSFTResearch we had early access to the marvelous #GPT4 from @OpenAI for our work on @bing. We took this opportunity to document our experience. We're so excited to share our findings. In short: time to face it, the sparks of #AGI have been ignited.
https://t.co/2DucYCvEoA
GitHub is releasing Copilot X. We’re starting to see a shift in user experience of the GPT powered software.
Soon the software we use today will feel like software we used to have in the 90s.
https://t.co/7dOGJNnSSY
The new AI wave can be compared to the period when the Internet emerged. The generations split into before and after Internet users.
Now we are the before AI generation. We're gonna need to adapt if we want to keep up with the AI native folks.
The next paradigm in programming interfaces would be Language Programming Interfaces or LPIs. They'll be the young siblings of APIs.
You can simply tell any service what to do in natural language.
It is already possible to build that on existing APIs using LLMs.