This is really impressive, funny, entertaining, amusing, but hell dangerous. Do not use this. Do not let an LLM execute arbitrary shell commands on your computer without reviewing and comprehending them.

That’s a very interesting demo. Basically, right now what I can, I believe most people do, is to ask advices, aka script, to do a certain automatic job. As far as I know, most people will execute a function by himself and check the intermediate step before the critical steps. Let’s say to, you know, revise the file in a file system or do some requests to external world.

​

Right now, the bottom line to me is, we have more time gaps here. When you spend some time waiting the system to respond, in our case, waiting the function to execute, or even for the functions, for the tokens to be complete. So we are, you know, asking the computer to do certain things and right now it cannot do it in one shot with high accuracy. So it basically involves many steps to achieve a relatively complex task. I’m actually inspired or I was thinking about how is that possible to have a well-defined a sequence of things to do in advancve and we talk to an agent and let an agent to do the job and once our human feedback is given to computer, then we jump to the next one as long as the background information is clear or consistent, it won’t cost too much for this kind of parallel processing.

​

I would say it’s involving more planning than my peers’ workflow. So basically it requires one to clearly think about what exactly you want to achieve for a given section. It’s a challenge, but it’s interesting. I believe if we can do it in the right way, the results will be astonishing.

That’s a very interesting demo. Basically, right now what I can, I believe most people do, is to ask advices, aka script, to do a certain automatic job. As far as I know, most people will execute a function by himself and check the intermediate step before the critical steps. Let’s say to, you know, revise the file in a file system or do some requests to external world.

​

Right now, the bottom line to me is, we have more time gaps here. When you spend some time waiting the system to respond, in our case, waiting the function to execute, or even for the functions, for the tokens to be complete. So we are, you know, asking the computer to do certain things and right now it cannot do it in one shot with high accuracy. So it basically involves many steps to achieve a relatively complex task. I’m actually inspired or I was thinking about how is that possible to have a well-defined a sequence of things to do in advancve and we talk to an agent and let an agent to do the job and once our human feedback is given to computer, then we jump to the next one as long as the background information is clear or consistent, it won’t cost too much for this kind of parallel processing.

​

I would say it’s involving more planning than my peers’ workflow. So basically it requires one to clearly think about what exactly you want to achieve for a given section. It’s a challenge, but it’s interesting. I believe if we can do it in the right way, the results will be astonishing.