Note on public speaking. I've been on the sales and conference circuit for many years. The trick is not to over-prepare a specific presentation but instead to just know your content really well and then get out there and improv and jazz hands your way through it. This will come off as much more fluid and natural. Sounds it ended up working out this way with you being too tired to go through your prepared speech.
Staying loose but knowledgable (ie, bullshitting) allows you to remain flexible and adaptive to the audience and the context. You also may come up with new ideas mid discussion and can incorporate those seamlessly. The presentations (if using a formal one) should just be a rough outline with key points reminding you of what comes next.
Kind of like being an LLM. The improvisation is good if it is based on solid training. Trying to over-engineer a presentation is like trying to hard code everything an agent does. It may work awkwardly but it is also much less flexible and adaptive.
With this year's fast moving excitement over all these agent tools, I've had the impression that a lot of the hype is a bunch of solutions looking for a problem.
I went through a bunch of Coursera stuff on Agents but then never really used any of it because it seems agents are just trying to badly automate what I'm doing in conversation with ChatGPT anyway: "Do this, then with the result, do that," etc. I didn't go the extra step of plugging into any databases. I just drop whatever file or content I'm working on into the chat then take the result and fine tune it myself. That works pretty well for a single user and shaves off 40-60% of the time for a complex task.
It seems the great dream and promise of agents is automating away the need for a human middle man to take idiosyncratic data, do some stuff to it and return a result. But the agents don't know what you're trying to do to the same depth that you do and maybe never will.
I think the best use cases are as gatekeepers in a complex process using limited small models at different points. This allows for better control over your process. If you have to spend more time finding and correcting their errors, then it was probably more efficient just to keep doing it manually.
Of course, speaking in vague generalities isn't very useful in an engineering context as every particular use case has its own unique complexities that need to be addressed.
100% agree on this "The presentations (if using a formal one) should just be a rough outline with key points reminding you of what comes next."
I started using this approach in my new videos and it works like a charm. You have continuity, while still sounding natural and wasting hours to repeat.
Re your agents idea. Yes, 100% agree. But that's why vertical AI Agents deeply rooted within your processes work. You need to put in a lot of effort into taking a generic LLM, such as ChatGPT, and transforming it into a specialist for a given area where it has access to the right context and reasoning to get the job done.
But until you know "the right process," I agree that trying to force agents with everything is just useless complexity.
“The main source of bad engineering decisions is often just following the hype.”
this rings very true. remember when everyone was tokenizing or putting every single product on the bloc chain a few years ago?
haha, yes! I wonder where they are now 😂
Note on public speaking. I've been on the sales and conference circuit for many years. The trick is not to over-prepare a specific presentation but instead to just know your content really well and then get out there and improv and jazz hands your way through it. This will come off as much more fluid and natural. Sounds it ended up working out this way with you being too tired to go through your prepared speech.
Staying loose but knowledgable (ie, bullshitting) allows you to remain flexible and adaptive to the audience and the context. You also may come up with new ideas mid discussion and can incorporate those seamlessly. The presentations (if using a formal one) should just be a rough outline with key points reminding you of what comes next.
Kind of like being an LLM. The improvisation is good if it is based on solid training. Trying to over-engineer a presentation is like trying to hard code everything an agent does. It may work awkwardly but it is also much less flexible and adaptive.
With this year's fast moving excitement over all these agent tools, I've had the impression that a lot of the hype is a bunch of solutions looking for a problem.
I went through a bunch of Coursera stuff on Agents but then never really used any of it because it seems agents are just trying to badly automate what I'm doing in conversation with ChatGPT anyway: "Do this, then with the result, do that," etc. I didn't go the extra step of plugging into any databases. I just drop whatever file or content I'm working on into the chat then take the result and fine tune it myself. That works pretty well for a single user and shaves off 40-60% of the time for a complex task.
It seems the great dream and promise of agents is automating away the need for a human middle man to take idiosyncratic data, do some stuff to it and return a result. But the agents don't know what you're trying to do to the same depth that you do and maybe never will.
I think the best use cases are as gatekeepers in a complex process using limited small models at different points. This allows for better control over your process. If you have to spend more time finding and correcting their errors, then it was probably more efficient just to keep doing it manually.
Of course, speaking in vague generalities isn't very useful in an engineering context as every particular use case has its own unique complexities that need to be addressed.
100% agree on this "The presentations (if using a formal one) should just be a rough outline with key points reminding you of what comes next."
I started using this approach in my new videos and it works like a charm. You have continuity, while still sounding natural and wasting hours to repeat.
Re your agents idea. Yes, 100% agree. But that's why vertical AI Agents deeply rooted within your processes work. You need to put in a lot of effort into taking a generic LLM, such as ChatGPT, and transforming it into a specialist for a given area where it has access to the right context and reasoning to get the job done.
But until you know "the right process," I agree that trying to force agents with everything is just useless complexity.
Thanks for the good 😊