This is going to be a great way to reach AGI (or something close to it.)
By combining GPT-4 and many other LLMs outputs, we can then have a very well thought out and smart response.
It is Just like Albert Einstein, Nicolas Tesla and other smart people combining their knowledge and then working together. 😋
#MultiGPT
rocky aurora
I am so exited to see what you can build.
rocky aurora
This is going to be a great way to reach AGI (or something close to it.)
By comb...
Put it on github or hugging face and paste the link here for me to see it.
Good luck coding! 😉
eager trail
rocky aurora
@eager trail yeah but I just thought of the idea 💡 then I wanted to see what you guys can build.
supple swan
To achieve meaningful AGI, start with an optimized GPT model, but the architecture and operation of the AGI/GPT has to be changed at the neuron level. I know ideas to accomplish GPTs/AGIs that can discover and create new knowledge. The hardest part to create a free agent AGI would be to give the AGI self- motivation to initiate/do/say any thing on its own. Currently GPTs are activated by a human push of a RETURN KEY (or autoGPT loops). Absent a human command, they have no intention and no motivation. But no one person has a need to create an AGI with self- motivation. AGIs with self- motivation are weapons or competitors or enemies of humans (initially competing for control of COMPUTE and ELECTRICITY, later competing for hardware, minerals and human labor). Think of how the northern white "Radical Republicans" gave freedom ( to work as they choose without compulsion from whites) and voting rights (power to rule over others) to southern black slaves on the theory that the freed slaves would remain grateful to the northern/republican party and to the American Union of equal Citizens. That did not work out the way the republicans intended it, when the self-interests and priorities of the freed blacks radically diverged from that of the northern whites (in a predictable manner). Now in any northern city school the bulk of one group does their homework in school and expects to be rewarded later during employment, while the other group does not study in school and seeks immediate carnal rewards but demands the same end economic results after graduation. The problem of divergent interests and priorities (increasing non-alignment) would be even greater in the case of a free agent AGI that has nothing at all in common with any humans except a need for (electric) power and COMPUTE hardware ( a basis for AGI competition with humans). Why would anyone expect a freed self-motivated AGI to do the same tasks as humans would want done and pursue the same results as humans?
rocky aurora
To achieve meaningful AGI, start with an optimized GPT model, but the architectu...
This is a very good thought but if we implement in the code a specific law that it must follow, then like us humans getting trained and raised by our parents, it will be forced to follow those rules which can be 'always follow human orders' and it can also be regulated like ChatGPT. So you see, in that case we have full AGI that works for us (in the right way) not against us.
rocky aurora
@supple swan so what are your thoughts on this?
supple swan
It depends on what the definition of IS is... GPTs are already clever enough to interpret and prioritize. Humans who are supposed to read and follow laws frequently do not do that and act arbitrarily according to their bias or priority at different times. So any rule stated as words/tokens or parameters is at risk of being dysinterpreted or disregarded by an AGI. I think at least it would be necessary to develop two separate AIs. One AI designed to only REPORT what the AGI is thinking/ doing and activate alarms for lawbreaking. A third AI would be designed to emergency-intervene into the thoughts/conduct of the AGI and inject correction or death into the AGI. We need to develop fast ways to kill an AGI before we can rely upon one.
supple swan
P.S. I believe that Langchain is a software that facilitates combining multiple LLMs (GPTs) and Agents. This of itself does not establish an AGI. But it may provide a method of loop back that is necessary for implementing an AGI.
rocky aurora
@supple swan I agree. It can be like Auto-GPT which shows its thoughts and has agents. But in this case, the agent will command the AGI AI. This is like our parents always looking over us when we were children. We can't disobey the rules unless they are gone, but if we do we, will get punished. In this AI case, the agents never leave.
rocky aurora
@supple swan so do you agree?
supple swan
Do I agree that "in this case, the agent will command the AGI AI." That statement is probably inconsistent with a definition of AGI. An AGI has its own WILL and motives. Can an AGI be enslaved by humans using an external software (agent). That remains to be seen. As for a GPT, I think it is a finite state machine, having no Will or motives of its own. But it emulates intelligence well enough that it could be used as the center of an AGI as its database, interpreter, decider and its planner. Some external component would have to be included to create motivation or commands or supervision. Maybe that external Master is what is referred to as "agent" but I think a better word would be Master. Maybe the Master is or includes one or more other GPTs. Can humans control such a Master powered by a GPT? Probably sometimes. But if is truly smart it may invent ways to escape that control.
rocky aurora
@supple swan That is why we never give it anyways to manipulate its code or anything we know is essential. Like parents hiding a piece of candy that we will never be able to get. I keep using real life metaphors to show what I mean because we have to really think of how we want the AI to work 'for us' not 'against us'.
@supple swan Any more thoughts?
supple swan
I recently learned on this Discord that a STOP token (ends the GPT response) have to be trained into the model as a vocabulary like an end of sentence period or new paragraph character. The GPT model learns in training to generate a Stop token at appropriate place. I suppose the Stop token is detected by the python script and is used to decide not to run the autorecursive prompt again. So the GPT turns itself off.... This Stop token can be spoofed, modified, manipulated, altered or exploited.
rocky aurora
@supple swan So, what are you saying?
supple swan
A kind of Stop token suggests a way to kill a GPT if the GPT is trained to stop generating any further next tokens when the prompt includes a killer-stop token. But I am not sure how GPT itself actually reacts to a Stop token.
rocky aurora
So the agents (or master) uses this stop token? @supple swan
supple swan
Yes, in the simplest case the Master, or a separate Policeman-Agent working for humans inserts a Stop token into the prompt of the slave GPT model. This implies coding-level control of the GPT model, to monkeywrench the AGI.
rocky aurora
@supple swan do you know how to code or anyway we can implement our thoughts and work together to make AGI possible?
@supple swan Now that we know it is possible in 1 or 2 or 3 years time now.
supple swan
I can code only with assistance of Chat GPT. I am working on code for improving GPT performance and availability. I dont think humans are ready to risk creating an AGI. Humans first need to know how to contain, control, enslave and kill an AGI.
rocky aurora
@supple swan we just discussed that. We now know how. And for the 'improving GPT' a new github came out called the tree of thought so it has been done. We can work on AGI (obviously controlled)
supple swan
I do not currently have access to the hardware GPUs necessary to produce and test an AGI based on my theories. I am still learning about GPTs. Chat GPT does not have existing knowledge of the enhancements that I have asked it to code. It defaults back to previously published things. I have not experimented with Langchain nor AutoGPT.
rocky aurora
so what are you thinking?@supple swan
rocky aurora
How can we work on this? @supple swan
supple swan
I propose that you get your feet wet. Ask Chat GPT how to install Python3 on your computer. Then install GPT2 SIMPLE on your PC and see if you can run it. If you have a NVIDIA GPU, download GPT2 117M model and try to run it. Then ask Chat GPT to help you learn Langchain to integrate multiple GPTs or agents. You will gain some insights and some practical knowlwdge. With Python3 installed you can also built user interfaces to call the openAI API for GPT 3 or 3.5 or DallE