AI Agents are rising up, don’t get left behind

RESOURCES:

Please Subscribe.

This video is a deep dive into AI Agents and how to design them.

[Gee]

I agree with every point you make in this video but every innovation in AI, no matter how extraordinary, will soon become child's play for the masses to achieve. Yes, some people will make some huge, short term gains in deploying agents to perform "next level" tasks, but that next level will become the current level for everyone soon after. Also any task that has a high return on investment will become generally cheaper until there's little profit remaining. The reason for this is we can train AI from AI and there's very little any AI developer can do to prevent this (other than to keeping innovations secret) - this was the meaning of Google's leaked memo about Google and OpenAI having no moat - this is true for all institutions with public facing AI and open source AI. Narrow task field agents will be amazing though and then conglomerations of these agents will be available as a service for the price of electricity and a reasonable markup. There's certainly a viable future in these conglomerations, manufacturing chips and selling electricity and connectivity (data centers), but everything has a short term viability - a year at best.

[tropicaltaffy]

I worked for a client that uses AI agents for lead generation, personalized multi-touch point cold email outreach, and chatbots that answer questions and book appointments. I would love to build a system like this

[elexceer]

Agent Smith in the Matrix is becoming real.

[bdown]

Dang my good guy, your flipping awesome videos are valuable bangers

[CalebDietz-kuxx]

Man I wish I had more of the technical knowledge of these things. (working on that). General intelligence doesn’t seem to be this complicated bar. The pieces are in place, just need a few minor adjustments and they fall into place, IMO. When it does. Dominoe effect. Just one or two specific cases of adoption, and it’s game over. It won’t be a matter of budget, knowledge/experience, it’s not going to be a breakthrough in tech capabilities. Literally just a matter of happenstance. The building blocks/pieces are there. Someone’s got the puzzle pieces laid out and is just going to ‘bump’ the coffee table this puzzle is laid out on, and the picture is complete. Mad dash from there. Specifics: Human in the loop will be crucial. (Until it’s not) But initially it’s going to be component that gets looked at, considered, and implemented in slightly different or novel ways. Don’t just think in terms of: checks and balances. Think in terms of the antithesis of a LLM. Longer run training/interactions between the same/specific individuals and training model. Think mentor/mentee dynamic type coaching. Focused, intentional, and consistent engagement between the model and the ‘trainer’, aka ‘boss’. My theory: a practical/working general AI is basically this infrastructure. Layers: It’s the tacticians, the project managers, and the visionary. Similar to a CEO. Memory storage/retention of desires or outcomes? Well that gets smaller and easier to incorporate when it’s essentially a moving subset of the process. Like a capsule or buffer surrounding the forward movement of the model into higher level of capabilities and cognition. Have the last trailing ‘x’ number of decisions stored as a rolling set of values. Factor in and account for a standard level of deviation from any two linked/subsequent choices in this sequential thread of trailing decisions. Like a trailing V shape following the wake of a boat. Nimble enough it becomes a light evolving resource/reference that allows for adaption, change, etc. Anchored (dynamically) in the current field of options/time. Insuring a certain level of reason involved just by the nature of iteration and logic based decision making. This same type of concept/mechanism effectively could serve to function and aid in solving the dilemma of model degradation. 1. As a hedge unto itself. But 2, if used in conjunction with: a feedback loop that compares the %deviation in current decisions to %Deviation in responses at prior historical points. (something like a capture or self audit. Taken at fixed/set intervals. Treated more like metadata on the models decisioning. And Stored in ancillary memory. Available to the model, but succinctly different from the pool of evolving data that is used as working memory for processes/decisions) Where the results are stored from past ‘audits’, and available as a parallel/tangental data source. These period audits could essentially be a combination of two things: injection of a the exact same inputs/outputs for every single audit, along with a larger portion of rotating/fresh training data to keep the model true to original training data. Basically combining a process for maintaining model integrity (via fresh/new training data injections) and at the same time having a portion of the answers ‘rigged’. And capturing the change/deviation in the responses. Keeping it within certain perimeters as part of the models fundamental programming. Basically the hierarchy you described. But an added layer of tools, in the form of additional data. Tools specifically for the boss. Like a tool belt. That it can pull from. And perhaps the boss would benefit from some other tools (also data sets) I. The toolbelt. Data that’s not directly related to its training or decisions. But general knowledge base type data. Like the allWikipediaData set. But probably makes sense to lean on the kinds of things human intelligence has relied on. For instance? One of the tools in that belt pocket being the data of what’s taught over the course of a persons educational experience. Like all the school info from k-12+. And these types of resources the AI could have as its general working knowledge of the world it’s operating within. Ethics? Same sort of thing. Touch of game theory, swarm theory, mass psychology, and ultimately crowdsourcing constraints on ethical limits based on a general consensus of humans as a whole. The individualized training I mentioned earlier with more specific/persistent subset of small # or even, single, individuals could play an interesting role in adding an additional layer in terms of socioeconomics. Could even result in interesting/unexpected positive unintended consequences like how just the right, small/minuscule amount of inherent level human bias made its way in there. Gotta believe personally: the final touch to a true generative ai (that can be mistaken for just human enough) is going to be a like a little drop. Some tiny nugget that gets infused into the system or process. Somewhere, somehow. And that little unknown spec, of whatever, that finds its way in? Might just be the secret ingredient. The thing that makes our: general knowledge autonomous AI, Frankenstein-esque monster come to life. Went off on a little tangent at the end there. Could have gone into more detail on other parts. But basically, we’re past the technological threshold to do it. I think we probably have been. Just a matter of when/how/who brings the real deal to life. Paradigm changing. Faster than the speed of sound. It happens, and the impacts on markets and industry will be seen before the alarm even gets out. A handful of people will probably be paying attention to the right thing at the right time. And they’ll get to see it coming. Everyone else will just wake up one day. Different days. But on each of those different days, it will have happened. The early adopters of the real thing? Might be will be the only adopters. It’ll get figured out, and then it’s down to the ‘where’. Where it’s applied will just be zero to monopoly. Anyways. If you read the esssy hope you liked it. Liked the videos and energy. I agree with you. Really is that exciting

[nithingeorge]

Hi David, I've been following your channel closely and tried to sign up to your community page. When I checked about a week ago, it was 37$ a month and it was mentioned price will increase once number of members cross 300. I checked today and it is saying 77$ per month even though number of subscribers is 243 (<300). I was disappointed seeing this.

[vulcanlancer]

Great informative video it's appreciated. I didnt know AI is advancing so damn fast I can't keep up.

[Koryogden]

Im convinced this is baby-AGI already.

[thehari]

U have great content man im gonna join ur ai agent group on skool, i wanna learn how to use ai agents to test them on my spefici use cases, i work for a ai education tech startup built on gpt4 - a math tutor

[Riskninjaz]

Great content again man. Time to get building!!!! Will be exploring the gpt visual api.

[lilith]

May i ask, why not just use zapier or make for simple tasks, over crewai or autogen?

[jarad]

Thanks for the videos man! In the last building one with crewai you mentioned another agent platform to try that had another feature (i cant rem what), you said youd make another video about it. I cant afford to join community yet but hopefully soon but could you share that tool to check out in the mean time? Thanks!

[JumpDiffusion]

Do you think Devin AI uses OOTB LLM or fine-tuned models specifically for better long-term reasoning/planning?

[shimblypibbins]

it's funny you say claude 3 is worse than gpt4 - it beats gpt4 in every benchmark right? could you expand on that?

[xlrusa]

We need open source LLMs that run consumer gpus in parrallel and combine the vram from the cards. One 4090 only has 24gb of vram. Im looking into building digital agents that will help build this. Also someone must be working on it or should be as soon as possible. Also we should be using AI to help develop video games that can support gpus in parallel.

[wildheartc]

3:30 I understand why you want us to use cheaper models here. Sort of like building a proof of concept then going for the more expensive (and therefore) more smarter ones but based on limited experience, i found that I had to redo my system description/role when switching between "smart" and "dumb" models. What do you think?

if we join your community will we have access to all these amazing modules? just join and watch?

[dankay]

I wonder if there's a way to have an agent controlling your computer through mouse and keyboard by looking at the screen like a person

[RowenHolliday]

Yo AI is good teacher too, taught me some things. Also good vid!

[dylan_curious]

Woah. David Shapiro thumbnail. Midjourney? What was the prompt? I do like the style. Great info.

[Techtalk]

Imagine the vast amount of misinformation via videos, blogs and comments. Going to be very hard to trust anything.