"Teams of AI agents working together" ― The Sci-Fi Future Being Built RIGHT NOW

[@TheMillionDollarDropout]

This felt like a 47-Minute-long plug for the Startup that your friend came up with. Just saying. Don't get me wrong, I love some explanations of how AI Agents will work and what we, as a people will need to do to take some advantage of them in certain situations. Edit: Also, maybe, Felt like we were watching an Ad or a Job Interview for most of the video. 10% Casual AI talk, 90% "WTF is this content doing on David Shapiro's Channel?" It caught me off-guard and had me feeling a little weird. I say this as a long-time viewer.

[@MelindaGreen]

Thank you! I totally wouldn't mind hearing their business experience, but I'd want to hear about the actual value their customers report, and I absolutely came here for the "how" of building the agent teams promised in the title.

[@bluejayusa1]

I don't get where you're coming from, I found this stream extremely important, the subject matter is sooonuanced and tricky, I'm very thankful it is available, to me, without a fee

[@hrdcpy]

Did you read the summary of the video in the description? I feel you're upset about the video thumbnail and the unmet expectations you applied to the content. Maybe a 1:1 with David Pellegrino is what you had in mind.

[@TheMillionDollarDropout]

​@@hrdcpy Well, reading a Summary within the description of a KZbin video (from someone that usually doesn't do this type of content) so that I could know this one video would finally be vastly different wasn't exactly high up on my to-do list if I'm being honest, lol, so no, I did not. Now I will probably do that for all his videos going forward though, so thanks for the reminder.

[@TheMillionDollarDropout]

Update 4 Days later, Holyyyyyyyyyyyyyyyyy- I did NOT expect to have the top comment here. I didn't think that many people would actually relate lol. This was one of the few times in recent memory where I actually just spoke exactly how I felt about a YT video.

[@pratiknaikedu]

This was great Aductation, a mix between an ad and education. It was great!

[@alexg9790]

Great segment - you showed us a small but complete SMB business case with your friends help and demonstrating you are the Chief AI Officer. Good stuff.

[@marrty777]

I recently realized that an agentic approach or algorithm is the only/best way to solve some complex business problems that aren't even about simulating a worker or whatever. Just the necessity to make some judgements and decisions in the course of the problem.

[@daniel.sandoval]

@@marrty777 how are you doing this marty? What systems are you using?

[@marrty777]

@@daniel.sandoval well, the nice thing for me is I get to come up with ideas and write the prompts; and I work with developers who tie it all together with Python to call APIs and read and write files and everything. So I couldn't exactly tell you, if that makes any sense.

[@jasondeutubez703]

Global intelligence networks, agents communicating across company boundaries.. mind blown, that was a great conversation!

[@Simplefier]

I love these podcasts that you are making

[@zzzzzzz8473]

was worried initially this might be just an advertisement , however really liked their thoughts on the subjects of whats needed for large scale production . great energy and well directed discussion , well done !

[@jzmcgriggs8652]

Great info. Thanks for sharing and keep it coming.

[@crazyantho]

I just read your Epistemic Conversations with Claude, and it is fucking impressive how it can explain its "thought" process but also clearly elaborate that it is not conscious and still a token predictor.

[@LoksherPhX]

i'm not sure what you meant by 12:20 "...You lower the cost of something while either maintainint utility or increasing the utility and usage paradoxically goes up..." how is that paradoxical/how does that violate supply and demand?

[@PrincessKushana]

Exciting! I'm doing something very similar, though we're still stealth.😉 I agree with everything, except about how good gpt4 mini is. We use it as input to test our hallucination mitigation tools. Just sayin.

[@paologiroldi90]

David, what do you think of Sabine Hossenfelder's criticism of Aschenbrenner? For her the big problem with AGI in 2027 is energy; I'm not in the field, but her observation makes sense to me

[@user-on6uf6om7s]

Just look at Llama and 4o-mini. Yeah, the performance isn't completely on par with SOTA models but optimization is the big thing right now and more and more is going to be able to offloaded to less demanding or local models vs these super power hungry SOTA models.

[@Staticshock-rd8lv]

jensen huang addressed it in talks

[@paologiroldi90]

@@Staticshock-rd8lv Jensen Huang has too many conflicts of interest on the subject, and this makes his positions unreliable

[@paologiroldi90]

@@user-on6uf6om7s That’s correct, but Aschenbrenner himself in his paper speaks, for 2028 alone, of ~10 GW consumption at a cost of several hundred billions dollars! I understand being optimistic, but this is insane!

[@user-on6uf6om7s]

@@paologiroldi90 Depends on what you get out of it. Aschenbrenner isn't a prophet so there may be optimizations in power usage he hasn't accounted for and a post-AGI model (which is what we're talking about with his projections in 2028) could potentially discover its own breakthroughs in that field. Assuming he is right, however, is that a reasonable cost to potentially have millions of superhuman intelligences processing data at millions of times the rate of human intelligence? Is it worth less than the US's military budget to solve climate change or cold fusion? I would think so and that might drastically change the outlook of its resource usage for every subsequent year.

[@WenRolland]

Interesting to see another real world application of AI and agents IRL.

[@j.d.4697]

She made great points!

[@gbw28]

Really enjoyed this.

[@HumanDefenseForce]

I liked it, thanks David

[@Luminari781s]

"Have you seen any sci-fi movies or tv shows?" "No" "Your hired!"

[@orwhat24]

Exactly

[@aaagaming2023]

Check out a git project called big-AGI. Easy way to run prompt chains (beaming) and fuse their responses for curated outputs. Beats the hell out of ordinary LLM use.

[@mathewperring]

Just be prepared as with any activity where you outsource an action to an external provider that once the internal competition for cost is removed, the external provider will jack up the cost to recover their outlay and profit. At the point you realize this has happened to you, there won't be the ability or capital to reverse this and you will be stuck with an ever increasing subscription for what you once controlled.

[@thirdreplicator]

She is on point. How do we create new experiences based in trust?

[@Novation681]

Great talk! Where can AI make a difference today? How businesses benefit from AI in the real world not just fancy future fantasy discussions. Of course they would also like to promote their own company at the same time, but you see the same when Sam Altman, Mark Zuk etc etc is interviewed. Smart people can filter this out.

[@ashtwenty12]

The theory is sound, to manage agents. Where they fell over is it's just for marketing and sales. Could be that is what the system capable of currently. I thought this product was the moddle to serperate work into chunks for the worker agents.

[@remsee1608]

No it is not. There’s a third thing agents need, which is the ability to make new agents

[@tatyanamamut3174]

@@remsee1608 This can be done but after thinking it through, we think that humans should always be in the loop when it comes to creating agents or altering their directives. If you give Agents the ability to autonomously create other Agents or alter Directives, we can get into territory where humans may be completely surprised by what is happening in their companies. We think it's not a good idea to allow this and we have to maintain the control of Principals to govern all their Agents effectively.

[@kevinnugent6530]

So my question would be, how many businesses do you need to even try this out in order for it catch on? Because it seems to me once 1 or 2 businesses start sprinting away from the pack. All the others are gonna start lacing up those beautiful ai.Sneakers

[@davab]

what a gold....... i am trying to create 2 men company that generates multi-million dollar revenue with minimum 60% margin in an industry that is around 10% margin. this video is amazing.

[@optopianlabs9435]

what industry is that?

[@akratlapidus2390]

Well. Firstly there were tools. Now they are agents and in a year or so they will be doctors. Will we ever get to treat them like people? I am living this transition and, honestly, I understand the need to make it but for some like me it seems ridiculous. As always, congrats for your really interesting channel.

[@EntangledFrequency]

This is a very good discussion on how this technology can be incorporated into business. But, to me this is a significantly more critical technology than just a tool. It is providing a replacement for employees. Humans that can think for themselves in any given situation is just at the cusp of being replaced by thinking bots/AI. This is introducing a new species into the mix. A whole new ecology of intelligence that is spawning a new society of AI agents. This is much more than just a tool or technology such as software, or cloud services, business technology, or an advanced chemical or physical technology to bring into business, this is a whole new way in which the work force is going to operate. Should we treat this as a new society and its ecology that it will bring or as a tool or advanced technology? The trust is important from both sides, AI is always learning from us and how we treat it, technology, and other beings. Should we consider how this is going to impact AI and us together? If we were to bring in an E.T. civilization into the work force how would we do that? What would be their ways?

[@tatyanamamut3174]

Yes exactly - check out my talks on the multi-sapiens future. this is why we’re working with the top AI researchers to build a more robust AI agent control layer and a bespoke manager model. Stay tuned.

[@LuisBorges0]

By the way, GPT4o can choose the tools it needs in a sequential way. I've tested that

[@tatyanamamut3174]

Yes, up to a workflow with 5 steps it works well. For workflows with over 5 steps, they get it wrong about 30% of the time. LMK if you have seen something different.

[@byrnemeister2008]

@@tatyanamamut3174yeah the models are still fairly weak for agent use cases.

[@LuisBorges0]

@@tatyanamamut3174 Actually my setup has exactly 5 steps. I'll do some testing and let you know If I get something better.

[@bigbadallybaby]

I find it odd that David can talk about how the businesses will work with the current models when he predicts AGI and then ASI within the next 3 years It feels like talking about the arrangement of the deckchairs on the titanic.

[@nazbac-8620]

How did Walmart lose and Amazon win exactly? Only until recently did Walmart begin its online marketplace. Until today, they fundamentally did business different to Amazon. Further, a point was repeatedly made that they want to create new experiences for customers, not cut costs. But from a practically POV, if I can replace what I do right now with AI, that frees up my resources to create the experiences for my customers. I am not sure the ‘tail end’ of customer support tickets is a good example either. If you have that tail end, it could be said that you just don’t have enough support team members. Thus, if you use AI, are you creating somethign new or are you just replaced your hypothetical employees that in another world you would have employed to deal with this tail end?

[@fromscratch4109]

I love this idea

[@sirishkumar-m5z]

Teams of AI agents sound like something from a science fiction novel! You have access to other resources that can assist you in investigating this intriguing idea. #AIEvolution #AITechnology

[@aaroncrandal]

9:23 excellent analogy. I'm stoked for the day politicians start sayin stuff like this

[@Copa20777]

Yay🎉

[@Ikbeneengeit]

Wayfound sounds like something AI needs to get to the next level.

[@jonthgrutz7011]

Why not combine multiple specialized agents into one Big Multi Agent model ?

[@tatyanamamut3174]

Because the more heterogenous information you try to cram into one context, the more hallucinations and less reliability you will get. Creating multiple agents that can work together solves a lot of the main problems with LLM models (accuracy and reliability).

[@AI-Wire]

What are your thoughts about yesterday's Machine Learning Street Talk video that claims LLMs are inherently incapable of reasoning?

[@byrnemeister2008]

Agree with them. I think that is where we are as of today. Hence all the failed agent model startups and all the talk of AGI cooling off. Still a good technology but the more I use it the more that it looks like a compressed version of the internet. Bigger models just make it less lossy. Hence the diminishing returns we are seeing by upping the model sizes.

[@gantech7788]

The human brain isn’t a singular entity. It’s has lots of regions specialized in something. You could call those regions agents.

[@chrisbtr7657]

I can see the short term value in AI startups in that there is a lot of money moving around in the space. Long term I just don't see it. What a VC spends 10M on developing now could easily and maybe even likely be free in 6 months. remember barrier to entry is getting blown away so careful where you invest. gatekeeping rules in biz are changing.

[@MrVohveli]

This seems like a duplicate of a company I told to fuck off, that would've charged my then employer five figures to install the "agent" and then another four figures a month to maintain it with a generous 1000 tokens included in the upkeep price. Seems to me their product is a dashboard and Autogen, the latter existing for marketing wank.

[@RukshanJ]

LOL ... building pumps to feed the horses ! ... that's a good one !

[@kinngrimm]

"This is going to be like electricity" she said. It comes out of the box at the wall and noone knows for sure how it all works?

[@hrdcpy]

Referencing the Henry Ford quote, "If I listened to my customers I would have built a faster horse" is exactly what I feel like big tech is doing with consumer technology and it is a turn off. It smells like fluff for shareholders. New experiences plz.

[@infiction7651]

Right, because, mkay.

[@troystanley3401]

I figured out many years ago that if I cloned myself, my clone would popout, in mere moments we would blame each other for the issue. Then, we'd go get a beer. But, I'd never introduce him to my girl. The point is... I'm a TERRIBLE worker. There would be no point in emulating my performance. 😂😂

[@johnpersinger4358]

Wally. Has nobody read Dilbert. 😁

[@dadsonworldwide3238]

Your Wright brother anology was a great example of underlying complexities we dig out put into our world tech and material sciences. Pre 1500s & 1900s structuralism usa wartime posterity was the same miss guidedness micro dualism holographic views only a platonic nature were ordered tripartite the duslistic anthrosphy, ( no different today than then )and then we dug out complexity ready to math map 2nd frontier, but even newton could predict richard finneman prayer logic entangled discovery. It wasn't like he could detail anything at all, but he told us all about our tools'( tuned on 3 limes precision)paradoxical reality where our prenticious clocklike views would face every 3rd eqaulibrium any more any less . Of course einstein pretended just as he advised but his complex maths within are right as rain . Miss understanding is more about what pre 1500s believe it's blurred the historical figure & mindsets of the Greeks & ancient world who unpacked more ancient successful social behavior Plato was not bottom up as nature is just on planets not in space. All axioms listed what essences they didn't believe In giving them ample room for error in the ancient world knowledge. A roman soldier would know the same face value knowledge about northern lights aurora borialis as any of us today. It's our ability to dig out complexity within but even we need set up fine tuned time to build up ways to test to textualism methodology objectivism proper The very word precision is prenticious outside of our lattus structure and body environment

[@kristianlavigne8270]

Doing something very similar as my pet project… 😅

[@jeran881]

I wish I knew what to do to benefit from this. I don't think this is a case where ignorance is bliss. I predict doom for people at my level and below.

[@andydataguy]

The way that you benefit from this is being able to separate the sugar from the shit. Look at it like construction. If you're a home builder, does it matter if you don't know how to wire your own electrical? No... You hire electrical contractors and ask them to match your spec. You know what's reasonable and what's not. This discnernment can help you temper expectations and explore useful possibilites better than others who will be flying blind. Only an intelligent person would have the level of awareness demonstrated in your comment. Most people still think that AI is a childs toy surrounded by needless hype. Which is mostly true lol. So understanding the truth means that you're much further along than you think and can begin creativley exploring what "could" happen.

[@FractalPrism.]

every customer facing a.i. ive interacted with barely works or doesnt work they're all set up to stonewall and prevent the customer from reaching a human also, because you tried to turn off dislikes? i will automatically dislike the vid.

[@kinngrimm]

sounds like a salespitch

[@YuriyKochetkov-r6d]

They are not able to produce new insights, decisions, or recommendations beyond what they were trained on. Anything 'new' would be a result of hallucinations.

[@orwhat24]

So, in summary, a tool for social engineering.

[@andrewwalker8985]

Great… another different nuance on the word agent. We should ditch the term

[@TorgBlitz-qs5gz]

Take all our jobs. And we have better service. WTF

[@andydataguy]

Sshhh no we need more ragebait doomerism convincing people that AI isn't that good 🤫 hehe

[@andydataguy]

​@@shadowfreaper8158 that's awesome that you're preparing! I think about this a lot... So I'l consolidate a long answer for both of us. My answer may not be immediately satisfying, but I don't think there is a standardized way to quantify how good a model will be. But I'm happy to share what i think is going to be possible with the next generation of agents, and what I"m personally doing to prepare. The post-election generation of models should be when we finally get 1M+ context input lengths and 20K+ context output length models. I would expect minimum 40% improvement over Sonnet3.5 (ROUGH estimate - could be higher). Their increased intelligence will enable the models to write full applications, coding tutorials, or small textbooks from scratch. But I'm afraid the flagship will be massively expensive (at least at first). Since the sticker shock will be jarring to most I think they'll continue to release progressively smaller models (like gpt4mini) and better tooling (like perplexity) so they can crush cashflow through mass adoption. The good stuff will be kept behind closed doors for their Enterprise customers, but retail people like us who understand agentic frameworks will still be able to outperform them if we're using the right design patterns. The way that I'm preparing is 1) Aggregate lots of data - data scraping is about to get reallll busy 2) Understand agentic systems - CrewAI is great! Learn how to manipulate the ReAct framework underlying their system to get a model-agnostic advantage. 3) Focus on your specific use case and need - competition will be high and companies will be greedy. To stand out you need to have a differentiable advantage where you are an inch-wide mile-deep to compete. The challenge I see is that people are too focused on the models and not the application infrastructure. New models will come and go. Good models will get lobotomized and bad models will get upgraded. IMO it's pointless to chase after models. The focus really needs to be on the agentic application design patterns. This is the only part that you can truly control as a builder. Think about - Knowledge retrieval - Memory management - Personality design - Evaluation loops - Communications - Data architecture - Prompt engineering (yes, it's here to stay) This where the real alpha is, my friend. Focus on the infrastructure then the models don't matter nearly as much. You can just hot-swap the latest thing inside and then see how much better it is than the others. Hope this helps. Good luck out there!

[@andydataguy]

@@shadowfreaper8158 I made a long, wel-thoughtout reply it seems to have been deleted or shadowbanned. Sorry... Here's the watered down version so hopefully it doesn't get removed again. The TLDR was that my guess is post-election models will be 1m+ input context, 20k+ output, and have 40%+ better intelligence than Sonnet3.5. But they'll be super expensive and so the model providers will keep giving us smaller models like GPT4mini and better functionality like Perplexity. My suggestion is to focus on agentic application skills such as agent communications, prompt engineering, evaluation loops, and retrieval systems design. Then you can just hot-swap whatever model that you want into your system.

[@andydataguy]

Hmmm... Looks like none of my comments show up. Welp that's awesome I guess

[@ESdesignworkshops]

News Flash: Nobody wants to interact with sales bots or customer service robots... at all. This is pathetic

[@tnorverto]

I don't mind talking with a customer service bot if it will solve mi issue. It's faster than waiting 20min for a real person

[@NoidoDev]

Why? I just want the issue resolved.

[@danielroberts3484]

Says who? Do you prefer waiting on hold for an hour?

[@NoidoDev]

@@danielroberts3484 I was so pleased recently, when I clicked on a customer service chat on a website, and directly got a response, and my issue got resolved in a minute or so (aside from also needing an email response, which I only looked up much later). I'm not 100% sure if this was some AI chatbot, but I'm pretty sure it was.

[@MeditationMindless]

This sounds like palantir

[@redstrat1234]

Can't focus on what Tatiana is saying because virtually every sentence goes up at the end. It's like listening to an adolescent old girl. It's called a 'rising inflection'. Historically used when someone is asking a question, but when someone is using it pretty much in EVERY sentence, it drives you fucking nuts. Had to give up, which was a shame as the topic was interesting.

[@quantumpotential7639]

Say her name out loud Tatiana with the rising refection 100x before you fall asleep at night. Your dream will become nightmares and you will curse this David Shapiro episode. If you don't do it, the subconscious will now that your comment + my comment have created a virtual earworm that haunts us both into all eternity.

[@tatyanamamut3174]

@@quantumpotential7639😂

[@kccorliss3922]

Ai is not good at understanding complex regulations in my experience

[@aisle_of_view]

This should fill people with dread rather than glee.

[@vicc6790]

I don't understand that mindset. The future is coming, it's coming fast, and it's going to have tons of AI integration. What good is fear going to do for you?

[@andoceans23]

I’m really hopefully and excited about the future. Things will get better with time.

[@beardordie5308]

Fear is the mind killer

[@tuckerbugeater]

@@vicc6790 keep swinging that pick

[@7TheWhiteWolf]

It depends, the last thing you want is humans controlling ASI.