I tried to make it code, use twitter, and create a list of proper midjourney prompts as separate tests. Each attempt failed, it would get stuck in a loop, VERY off task, or would be incapable of properly assessing its own work. I think it really needs a gpt4 api key, but it also feels like tech in its infancy. It's there, just needs a lot of work.
I have a follow-up video in edit where I tried to get it to compare products, and it fell on its face three times in a row. The approach has promise and the "plugins" are there, but it needs to mature to be useful to a broad audience.
Maybe when it sees a repeat command / response it tries a few rounds with the temperature turned up? If this isn't already implemented (things move fast) maybe this is where I try my hand at contributing to the code base.
If 1 of the goals was to not get into a infinite loop or json file errors and so how it goes. Couple of your goals could be consolidated to allow the error goal
I have been trying this since the first version and, no matter the apis you add (pinecone, google, …), the problem is in the core code. It gets stack with relatively easy goals. I do not see major improvements in the last version 0.2.2.
My experience is the same, but I am hopeful that with system refinement and a lot of prompt experimentation, we can get real value. I'm looking forward to doing some framework experimentation of my own.
Great review. Isnt the limitation on AutoGPTs result due to your criteria: i.e. to buy one now? Thinking it might've left out stores where delivery is next day?
Excellent detail. I confirmed that that the prompt only used "for sale", which should have avoided limitation to "in stock at a store near me". I'm optimistic that future systems will properly detect meaning and ask for clarification to help the casual user bound the search well.
Try asking Bing Chat "find me an e-bike for a 5 mile commute in all weathers under $800. want a 20mph top speed"... ask some more questions, etc. Free and fast answers.
Always give AGPT a shutdown order such as in the last prompt: 'stop and summarize results of task after x comparisons" where x is the number you want. I also tried a power shopper like this and did ran into bugs, but it was done within 5 mins, just couldn't out put images and tables because I just don't know at this point how to enable those safely but the data was there.
Chrome on Windows? The short answer is no - it worked right out of the box for me. Might want to look at the .env file, I think I saw that that is where it is configuring the browser.
small tipp to reduce the cost, you can change the settings from gpt-4 to gpt-3.5-turbo, which is 1000x cheaper, then run would cost you some pennys and not dollars.
A good tip that anyone using these agents should be made aware of. In this case, I saw that Auto-GPT had a configuration for SMART_LLM_MODEL and FAST_LLM_MODEL and was excited - thinking it would choose the right one at the right time. Looking back in my API history - it did primarily use 3.5turbo and made at least some calls to GPT4. So looks like it was at least trying to balance usage.
It might fire up, but it won't be able to do the first planning step without an LLM to access (such as through the paid API). You may be able to point it to another free / locally run LLM though.
hey, I would be more concerned that it found a bike with fork improperly (inverted) mounted on an "official" advertisement ^^ this is a real fail for me :)) but thanks for the video!
Which bike - you had me curious, but I couldn't find it in the footage. Are you talking an inverted shock design? I had to Google what they were. Turns out I have that on an old hardtail I own.
@@ContextFound hey, at 15:10 :) and no, not inverted Design, only mounted with the crown backward at first assembly, which brings the disc mount on the right side and also make it somehow completely dangerous (it pulls on the legs instead of pushing on it) this and of course the fact that that swapping the fork direction violate the designed for "frame geometry's" rake offset :))) it's a recurrent "joke" in the E-Bike field. Chinese companies selling/advertising bikes and not even know how to mount them correctly (not knowing at all what they sell) --> not trustable at all. Again, was just an input :) your video is great and instructive :)
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
Love your thurough outline & framework great content
I tried to make it code, use twitter, and create a list of proper midjourney prompts as separate tests. Each attempt failed, it would get stuck in a loop, VERY off task, or would be incapable of properly assessing its own work. I think it really needs a gpt4 api key, but it also feels like tech in its infancy. It's there, just needs a lot of work.
Cracked me up when it tried to hire someone on upwork to help it clone a repo it hadn't made yet.
I have a follow-up video in edit where I tried to get it to compare products, and it fell on its face three times in a row. The approach has promise and the "plugins" are there, but it needs to mature to be useful to a broad audience.
@@ContextFound Agreed
yes it needs an anti-insanity patch asap.
Maybe when it sees a repeat command / response it tries a few rounds with the temperature turned up? If this isn't already implemented (things move fast) maybe this is where I try my hand at contributing to the code base.
God I would love to have this for some projects that I want to see come to fruition.
If 1 of the goals was to not get into a infinite loop or json file errors and so how it goes. Couple of your goals could be consolidated to allow the error goal
Excellent thought - I'll have to try that one out. Could work...
@@ContextFound did you give it a go?
I have been trying this since the first version and, no matter the apis you add (pinecone, google, …), the problem is in the core code. It gets stack with relatively easy goals. I do not see major improvements in the last version 0.2.2.
My experience is the same, but I am hopeful that with system refinement and a lot of prompt experimentation, we can get real value. I'm looking forward to doing some framework experimentation of my own.
Great review. Isnt the limitation on AutoGPTs result due to your criteria: i.e. to buy one now? Thinking it might've left out stores where delivery is next day?
Excellent detail. I confirmed that that the prompt only used "for sale", which should have avoided limitation to "in stock at a store near me". I'm optimistic that future systems will properly detect meaning and ask for clarification to help the casual user bound the search well.
Hey bro I’ve downloaded auto gpt and it seems to create some documents but I do not know how to access them
Do you have any idea how to access them
I found the output files in a folder under the Auto-GPT running folder. So for this Windows boy it was "C:\Auto-GPT\auto_gpt_workspace".
@@ContextFound they are in the workspace but they are not readable. I created a word document but it wasnt loading.
What a great video. Subbeed.
Try asking Bing Chat "find me an e-bike for a 5 mile commute in all weathers under $800. want a 20mph top speed"... ask some more questions, etc. Free and fast answers.
Solid point - you inspired me to check it out. Thanks! ruclips.net/user/shortszwl35yTWbME
Always give AGPT a shutdown order such as in the last prompt: 'stop and summarize results of task after x comparisons" where x is the number you want. I also tried a power shopper like this and did ran into bugs, but it was done within 5 mins, just couldn't out put images and tables because I just don't know at this point how to enable those safely but the data was there.
I just added this to an "Autonomous Agent Tips and Tricks" list along with other folks suggestions. I'm keeping that one in mind.
Mine doesn't open the browser... Any fix you might have?
Chrome on Windows? The short answer is no - it worked right out of the box for me. Might want to look at the .env file, I think I saw that that is where it is configuring the browser.
set headless to false
Great Electric Bike review video, thanks. 😆
People: "AutoGPT is Skynet!"
Auto-GPT: "Have you thought about buying a Swagtron?" :)
small tipp to reduce the cost, you can change the settings from gpt-4 to gpt-3.5-turbo, which is 1000x cheaper, then run would cost you some pennys and not dollars.
A good tip that anyone using these agents should be made aware of. In this case, I saw that Auto-GPT had a configuration for SMART_LLM_MODEL and FAST_LLM_MODEL and was excited - thinking it would choose the right one at the right time. Looking back in my API history - it did primarily use 3.5turbo and made at least some calls to GPT4. So looks like it was at least trying to balance usage.
will autogpt run if my open account is out of credits?
It might fire up, but it won't be able to do the first planning step without an LLM to access (such as through the paid API). You may be able to point it to another free / locally run LLM though.
If you've hit your credit limit the API call will not go through. It'll fail.
hey, I would be more concerned that it found a bike with fork improperly (inverted) mounted on an "official" advertisement ^^ this is a real fail for me :)) but thanks for the video!
Which bike - you had me curious, but I couldn't find it in the footage. Are you talking an inverted shock design? I had to Google what they were. Turns out I have that on an old hardtail I own.
@@ContextFound hey, at 15:10 :) and no, not inverted Design, only mounted with the crown backward at first assembly, which brings the disc mount on the right side and also make it somehow completely dangerous (it pulls on the legs instead of pushing on it) this and of course the fact that that swapping the fork direction violate the designed for "frame geometry's" rake offset :))) it's a recurrent "joke" in the E-Bike field. Chinese companies selling/advertising bikes and not even know how to mount them correctly (not knowing at all what they sell) --> not trustable at all. Again, was just an input :) your video is great and instructive :)