The AI Coach
AI meets human behavior in these fun and insightful conversations with Danielle Gopen, a top founder coach and advisor, and Paul Fung, an experienced AI founder. We blend AI breakthroughs with real-world insights on business strategy, industry disruptions, leadership, and psychology while cutting through the hype to ask the hard questions.
This is for decision-makers like founders and business leaders who want to harness AI's potential thoughtfully, exploring both the technological edge and human elements that drive long-term success. Join us to challenge the status quo, navigate the shifting landscape of an AI-powered world, and get new information worth sharing.
The AI Coach
Understanding AI Agents and Their Potential
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
AI Agents are everywhere right now. We explain what they are, break down the difference between traditional chatbots and agents, and do a real-time case study on how small businesses can leverage these technologies effectively. We also review Anthropic's groundbreaking computer use capabilities and what it might mean for the future of AI-human interaction.
Links and Resources:
Intercom
Zendesk
https://usefulai.com/tools/ai-agents
We love feedback and questions! Please reach out:
LinkedIn
Episode Directory
Mentions of companies, products, or resources in this podcast or show notes are for informational purposes only and do not constitute endorsements; we assume no liability for any actions taken based on this content.
New AI terms - Agent, Agentic, Agentify
Speaker 1Hi Paul .
Speaker 2Hello , I think agents are a really good topic to talk about because one I think you can't go anywhere in the Twitter sphere without hearing the term agentic being thrown around everywhere you go . I like how these new terms agentify , agentic just kind of come into our vocabulary out of nowhere . But , um , I myself have also been looking into agent .
Speaker 1I was just gonna say it's funny , because I said that to someone a couple days ago . I said something about agentifying gpt and they looked at me like I had three heads and this is someone in the ai space . So well , I said what do you mean ? And I thought maybe it's not everywhere .
Speaker 2I think it's also funny . I think one thing that would be worth talking about today for our audience is , like , what constitutes an agent ? Like , what is an agent versus a chatbot ? Right , like , why is an agent versus ? Like , is chat GPT an agent , which it's not , and I'll explain why . But I do think that's worth kind of talking about . But , yeah , I just think agents are really fascinating .
Speaker 2There was also a really big breakthrough in terms of what agents can do this week by Anthropic , so they released a new feature functionality called computer use , which we'll get into . But I think that is pretty game changing . So , yeah , I think it'd be a good conversation to have about agents and yeah , I would love to hear I know you were asking my experience with it . I mean , mine has more been on the research end , just looking into how these agents function , and our company is about optimization and accuracy , and so we're looking into , like , how do we correctly evaluate what these agents are doing ?
How to "Agentify" your GPT
Speaker 2But I'd be curious to hear about , like kind of what you're trying to do with your GPT .
Speaker 1Sure , I'll start with that because it's quick and easy , and then we can go into the more interesting things .
Speaker 1So we've talked before about the importance of prompt engineering , and we've talked about giving your chat GPT or quad or whatever LLM , you're using enough context so that it I don't know if maybe not the right word to use here , but giving it enough information that it knows it's the best way for it to respond . And so we talked a while ago about your friend who was a non-technical co-founder telling Claude , okay , you are an engineer , you are a CTO , and then it would respond based off of that . And so I do that most of the time when I'm interacting with an LLM and I'll tell it whatever I need for that specific request . But what I've realized is I keep saying the same thing over and over . And so I thought , oh , what if I just create my own custom GPT and then that way it's stored in the background and when I need it to be that persona or that quote unquote agent , I can have it pull from that , instead of having to give it the same level of detail in every prompt ?
Speaker 2Yeah , that makes a ton of sense . Is there a certain specific use case that you've started with in terms of , like , what kind of GPT are you trying to build to start , LinkedIn Thought Leader . Oh , I like that . Okay , is that for our podcast or for your personal brand ?
Speaker 1For my personal professional brand , of which our podcast is part of that . Yes , okay .
Speaker 2I like that , so tell me about yeah , how did you start this Give ? Me more on this .
Speaker 1Yeah , honestly , it's not working quite the way I want it , to which I will say is the case a lot with these LLMs . I feel like if anybody's heard us talk about this up until that point , or you know they're not so in the AI world , they might think that everything AI does is amazing . You know , based off the hype around it , that you submit one prompt and it comes back with everything you ever wanted and off you go . And you know , the reality is it's not always that so coming into creating it . So , just to walk someone through the actual steps , if you want to do something similar , the way you would do this is you open ChatGPT . You have to have the paid plus subscription for this , but it's only $20 a month and , honestly , I really think you get your money's worth . If you're using it even a few times a month , it's such a time saver .
Speaker 1Anyway , you open it and you go to the Explore page and on the upper right-hand hand side you'll see a button that says create and you click on create and it opens a window where it says , basically , tell me what type of persona you want , and so at first I typed a whole paragraph of all the things I wanted it to know and be in this persona , and it kept airing out . It kept saying like I don't understand , I don't understand . And so then I thought , okay , let me try this over again . So then I just typed one line . I said I want to create a LinkedIn thought leader persona , and then it seemed to like that . So it responded and said let's start with a name for this thought leader . How does LinkedIn insights mentor sound ? I said no , I don't like that . And so we went back and forth on finding a name For now . We landed on founders insight advisor , whatever that means . But the point is that this name is behind the scenes , so it doesn't really matter as long as I know that that's the GPT I need for that persona .
Speaker 1And then it got fixated on wanting to have a photo like an avatar for the GPT , and then it came up with using Dolly . It came up with a ton of bad options , and then I finally just said yes to one of the bad options because I wanted to move on to the real part , and so then it said , okay , great , because I wanted to move on to the real part . And so then it said , okay , great , now to make sure Founders Insight Advisor delivers relevant and powerful insights . How would you like her to express herself ? Concise and direct , more conversational , or perhaps a mix that adapts to the tone of each topic . And so then I told it , I want it to be conversational , personable , sometimes humorous but not over the top , still professional and direct , which I think are all attributes of what I write , naturally .
Speaker 2Did you type that or did it suggest some of those things ?
Speaker 1I typed all of that and that's how I like to think of my writing and my voice . And so it says okay , great , feel free to try her out in the preview . Let me know if any further tweaks come to mind . What's next ? So then I asked it does it help to see already written content in the preferred style ? And it says yes , absolutely , very helpful . So then I copied and pasted a ton of my LinkedIn posts that had enough content to them to show some of the style , and it said great , with this added guidance , founders Insight Advisor will be even closer to your preferred style . And it basically started to understand the types of things that I write and how I write it . And so then , after doing that , to the right , there's a little box where you can preview a suggested question or topic and then it will give you the response . So I previewed what's one thing founders should prioritize , and it came back saying one thing founders should prioritize is building the right team .
Speaker 1Your team isn't just a group of employees . It's a group of people with the power to make or break your vision . It's often said , the who drives the what , which is funny , because the who drives the what , or the who behind the what is one of my core philosophies that I have on my website . I have on my LinkedIn profile bio . I say often . So I thought , okay , it's really being on the nose here and saying this is what I should write . You might have the clearest roadmap and the best market strategy , but without the right people , aligned with your values , energized by the mission and skilled in the needed areas , execution can fall flat . And then it has a couple more paragraphs and I thought like , okay , that's pretty good for something that I just did in five minutes .
Speaker 1Am I going to post that ? I don't think so . But is it a good thought starter ? Is it a good thing of ? Okay , maybe I should have a post around . What is the one thing founders should prioritize ? Is it building the right team ? I'm not sure if that's what I would say . The number one , the one one thing is , but I think you know it got me several steps further than again going just piece by piece and saying you are a LinkedIn thought leader , you like to write in a personable tone , and so I feel like now having this here is helpful and we'll see what happens . What do you think ? But for what it's worth , everything I post on linkedin is ultimately my own word , like I don't . I don't believe in just like copying and pasting what ai wrote . I just don't think it goes that well and oftentimes I have to tell it that it's being too ai-ish . I said don't sound so AI . And they say you're right , sorry , okay .
Speaker 2I did have an experience this week where a friend texted me back something and he was kind of joking about something and I said did AI write that ? And he's the kind of joke that it had . But I'm finding myself having to ask that question and being a lot more unsure about the answer .
Speaker 2You know it used to be like oh , AI definitely wrote that and increasingly I'm becoming less sure when I see things about like is that my friend saying you know an actual response to what I'm saying ? Or did AI write that Like if it's a serious discussion we're having ? So I think that just goes to show that .
Speaker 1Wait hold on .
Speaker 2Like as text . He like he , he uses text on his like laptop and so I think he like typed something I said in a chat , gpt , and had it like generate some response about something , and then he copied , pasted it into the text back to me , but not like that's actually . Yeah , I think it's kind of funny actually , but or you know what it might have been . He actually is using the new iphone with apple intelligence . I don't have the new iphone Apple Intelligence yet . Does it allow for auto-responding or generating responses ?
Speaker 1I don't have the new one , but I thought Apple Intelligence wasn't . I know the iOS is prepared for it , but I thought it wasn't . Actually , I think he is .
Speaker 2If it's out in beta , I think he is installed . So he's the type of guy who always installs the iOS betas to get access to features ahead of time . So he's the type of guy who always installs the iOS betas to get access to features ahead of time . So , now that I think about it , I bet there's like an auto-response feature and I bet he used it .
Speaker 1Interesting ? Yeah , I don't know . I mean the news as of today is saying that Apple Intelligence arrives next week .
Speaker 2And that's the main . So I'm not sure the general release I bet . I bet there's a public or beta that's gone out Anyway .
Speaker 1It says yes . Right now , pieces of it have been available to those running the developer beta version .
Speaker 2Yeah , so that's what my friend likes to do . He's always at the cut .
Speaker 1Early early adopter over there . But also I don't want to text with AI when I think it's my friend .
Speaker 2You know it might happen . We'll just have to see . I hope not as well , but it is likely that some people are going to start using it for stuff . But what I was going to say , or a question I did have , was of the GPT instructions that you gave , what part of it is saved . So you gave it a little bit of a persona and kind of a style guide . Do you think it also saved the examples that you gave ? It was unclear to me if that was in the conversation , if that got saved into your quote-unquote agent that got saved into the agent .
Speaker 1Yes , and so now what I have , once I've created that . So I should say , once you do that and you see the preview and you feel generally okay about what you have , you can always go back and change it , add more . But you have to just hit create again on the upper right hand corner and then it will ask you if you want it to be available only to you , through a link or in the GPT store . I obviously picked only to me and then , once it processed , it's now on the left-hand column of the GPTs available to me . And so now if I come to ChatGPT to write a post or to refine a post that I've written , I can click on that specific GPT and feed it through that and it will already have the context . So I don't have to give it as much detail in the initial prompt what
Chatbot vs. Agent
Speaker 1I'm kind of curious .
Speaker 2So you know we're talking about agents , and what's a chatbot versus what's an agent is like something I mentioned earlier , and I would still argue that this is still kind of a chatbot , like it doesn't have a lot more than a persona and maybe some examples . And some of the reason I like talking to you on this podcast , which is really fun is to I lose touch with what the consumer-facing version of a lot of these AI looks like . So , like ChatGPT , I'm using the developer version , like the developer workbench , for a lot of these tools , and so I lose track of what you guys are seeing and have access to . So are there other features that you can use ? So can you upload files and have it do ?
Speaker 1what we call like rag . You know retrieval , augmented generation stuff . So you can do that . I can't . I'm not sure if it can do the rag side of things . I've never actually tried .
Speaker 2But you can upload files for a specific yeah , if you can upload files , then I think it's doing built-in rag on the back end , and I think I remember this basically because last year , during OpenAI Dev Day , the big stink that everyone made was that there was all these RAG companies PineTs and it turns out they built in like a very basic consumer version of RAG .
Speaker 2And just to go back , because I feel like this is a question you might ask me to clarify for our audience , rag being Retrieval Augmented Generation .
Speaker 2So what's happening there is it's allowing you to upload some set of files and I think the OpenAI GPT limit of the files is pretty small . I think there's only 20 files and they have to be pretty small files and so what it does is when you type in a question , it's doing a mathematical calculation on that question and then looking within the files for the most similar concepts or topic and then it retrieves those , which is the retrieval part of retrieval , augmented augmented generation and then injects those into the prompt for chat gpt or for gpt to reason about the information it's gotten as well as the query that you've given it basically . So it's doing like a basic built-in rack basically , and that starts to get a little bit more into bleeding the lines between like what is just a chat bot that just like sits there , like responds to you , versus like does it have more capabilities ? Can it like retrieve data ? And then we start to that starts to feel a bit more like what we'd call like an agent , if you will yeah .
Speaker 1So I think chat gpt in particular , creating your own gpt I do think it is a bit closer to an agent in that it is retrieving from itself the context that you've given it for how it should engage , as opposed to just talking , you know , answering something that has been pre-trained on or I don't know . I'm trying to think of how to explain this like it's it's . It's not like I gave it very specific data points . I said , okay , when somebody asks you about this specific thing , this is how you answer it . It's more general . When you get , when I come back to you and I have a question about you know , help me start a post on founders seeking fundraising , then it says to itself you know behind the scenes , it knows , okay , this is how I should approach this response . So then , how would you define that ? Yeah , so .
Speaker 2And I think the way that most people , when we talk about agents , at least on the developer side , is a lot of them are based on this framework called a React agent , so R-E-A-C-T agent , which , by the way , like the tech field this is just an aside needs to think of like better ways to name things , because there's like a React framework that's very popular in development and then they went and created a React framework for agents , which completely unrelated and very confusing , but anyway , just like tokenization also like tokens tokenization in crypto tokens and tokenization in ai means something completely different than tokens and tokenization in crypto yeah , we're running out of terms apparently , but anyway , so these react agents ?
Speaker 1are prompt query all of those two .
Speaker 2So React stands for reasoning and action , and so that's typically when people are talking about agents in the AI world these days . They're talking about agents that are able to reason about something and then take action on something , and so the difference between that and like a chatbot is like a chatbot is trained to either respond at every turn . So when I say a turn , I mean like a turn would be man . I really want to say it's an action that a chatbot takes , but it's like a step in a chatbot's thinking and so like in a chatbot .
Speaker 1Well , so I think of it tell me if this is helpful .
Speaker 1I think of it when you're on Amazon and you need to ask a question or you want to return something , and a little chat bot pops up and it gives you options that you can pick from and , based off the option you select , it gives you a new set of options . So it's kind of like a conversation , but it's not a free form conversation . It's basically saying here are the things that I can help you with , and you have to pick one of these and we'll keep engaging until we get to the thing that you need Like . To me , that's very much a pure job .
Speaker 2Yes , I think that's right . There's a lot more structure to it . It's very confined into the pathways it can take . I mean , even the way you program a chatbot on the backend is you kind of like you see what the initial user query is and then , kind of like a phone tree , you direct it down certain pathways that are like very structured pathways , and so , you're exactly right , that would be like an advanced chatbot , right ?
Speaker 2Whereas an agent has a set of instructions that allow it to reason about a problem or a question that you've asked it , and then also , I would say , the the act . Part of react is it allows it to take some action . And so what's really cool is there's a lot of actions or sometimes these are called tools that are being given to these agents that are allowing them to control a browser or search the web or , you know , take an action on a certain website or hit a certain API , like hit a weather API and return me the weather , and stuff like that . And so when I think most people are talking about AI agents these days , they mean specifically a type of AI chatbot , bot , whatever it will , that can both reason about how to solve a problem or answer a question and then also take action to potentially solve that problem .
Speaker 1So then , this GPT that I created I created . To me that sounds more like an agent so what do you know ?
Speaker 2I don't know if they consider gpt's agents I mean I guess the lines get a little blurry . I would say it can some , it can reason . It can't take a lot of actions . I don't think you're allowed many plugins or like tools . But I will say the developer version of gpts are called assistants . Don't know why they're called two different things . But assistants can use function calling or tool calling and so you can build more of a full featured agent on assistance which consumers don't have access to . Like assistance you can only build via the api interesting
Use Cases for Small Businesses
Speaker 2.
Speaker 1Okay , so let me give you a use case , and then you tell me what that would be sure that sounds good okay .
Speaker 1So on my dg executive advisors website , up until now I've never put pricing on there , because pricing is variable depending on what somebody's looking for and how it's structured , and obviously different for coaching versus consulting , versus advisory , versus all these different things .
Speaker 1But now , behind the scenes , I've had a couple of programs that have been relatively fixed in terms of what the offering is and what the pricing is , so like a founder fast track program and a pitch coaching program for founders who are going out and pitching , and I thought , okay , I want to put them on the website and then I would put those prices . But then is it awkward to have those prices listed but not other service prices listed ? So then I thought , well , maybe the cleanest way to do this , if it's possible , is to create some type of little chatbot or agent on the website that somebody can engage with to say , find out more about pricing , and they click on that and then they can start talking to it and telling it what their circumstances are and what they're looking for . And based off of that , I've supplied the agent or the chatbot with enough background information that it could give , if not the exact price , at least a small range of pricing . What would that be ?
Speaker 2I think it would still be a chatbot because it's pretty well-defined in the use case that it's working on and it has pretty well-defined answers that it would provide . I mean , this is all gray area stuff , so you know it could be . You could call it both . Realistically , I think I would call it an agent , at least in the current terms of how we're using the word agent if it could make decisions on whether or not it should go ahead and , like , book an appointment with you based on the conversations having with that person , so like so if somebody says it gives a range of pricing , the person says , okay , that sounds good , that's in my budget .
Speaker 1Then it responds saying , great , let's find a time to chat . And then it offers calendar time .
Speaker 2Yeah , exactly , and like kind of more specifically , I would call an agent if it like did some reasoning about whether or not it thought that person was a good fit for you . So if you gave the agent some agent instructions that were like here's the criteria for someone who should book a meeting with me they should be a founder , they should be willing to spend money in this budget , they should be looking for these types of services , and then if you let it reason about that and then make its own decision about whether or not to offer the booking , so like a chatbot flow might just be like always at the end , right , but an agent would be like giving it the power and the autonomy to decide , based on the conversation , whether or not it should offer sometimes , and I think that would be crossing the line into agent I'm laughing because I'm thinking wow , a new level , new level of rejection .
Speaker 1You get rejected by an AI agent . It's like sorry , I didn't like your answers . No meeting offer .
Speaker 2Yeah , I mean , that's the reality of it . It's sad but true . That's the world we're moving to .
Speaker 1Yes , Okay . So now I'm curious from your perspective . If you were advising a small business on how to incorporate this AI component , what would you do ?
Speaker 2Yeah , I mean probably honestly , there's probably some really good off-the-shelf options for this . So , like Intercom , fin is probably like the most well-known like customer success agent out there , whatever and when I say customer success it can also be kind of like sales oriented right that sits on your website . So , you know , probably Zendesk has a pretty good option for this , probably intercom has a good option . I think a lot of the sales tools have a good option . So I would be willing to bet you could find a pretty low cost because I'm sure , like sometimes intercom and stuff like that they'll you know the prices start at pretty high because they know that enterprises use these . But I bet you could find a pretty low cost chatbot agent tool that'll sit on your website and you could prime it with some of these criteria potentially and your pricing , and then you know it would route those things appropriately to you .
Speaker 1Those , yeah , there's a there's a ton of these things these days and so it's market research , like all founders and business owners should do before implementing new features . Yes , you as a founder . What would you want to see coming to the site and engaging in that way ? What would you find most helpful or engaging ?
Speaker 2Yeah , I think that's a good question . The one thing I like about having those little bots on the site is and I've been meaning to add one to our site we have like a little calendar that allows people to just book meetings . But I think what I find most useful about the sites that allow people to actually ask questions is you know , people visit your site and you'd like to think you have all the information that people need on your site , but the reality is you probably don't , and every time someone visits a website they have two or three questions that they want to ask and they're not going to email you , right ? There's too much friction to be like , oh , I have this question about your services . But if you offer them that really quick chat experience , that pops up and says like , hey , do you have any questions ?
Speaker 2You know , even myself , as a consumer of other services , I find myself asking a question in those things a lot and it leads to really rich and good experiences . Like I'm friends with a founder whose tool I was looking at . I thought it was really interesting and I just went in the little . I think they were using an account box and I was like , hey , do you guys do this , this and this . And he responded , like the bot responded at first , and then it routed me to him eventually and he was like , oh , I'm the founder , Like let's hop on a call .
Speaker 2And like we've maintained a friendship through this , actually he . So I do think those little bots in the corner whether they're agents or chatbots , you know , unclear , but I do think they're very useful from a business perspective .
Speaker 1so I think so too , and it's interesting what you're saying about the friction with email , because even if you have a contact us form where you can just submit through there , I do think there's some psychological barrier of the fact that , even if the person receiving the email responded within 60 seconds , it feels like a bigger lift than just typing a quick message into the little chatbot that showed up , because also there's more instant gratification , like we talked about last time . There's more there . When the chatbot or the agent , either way , is being offered right there in front of you , you think , oh well , I can at least get an immediate response and then decide what to do . Sometimes I wonder if it's too impersonal for certain businesses , but I actually think it's a nice thing , I think impersonality is that a word ?
Speaker 2I think the lack of structure or the lack of when you write an email , you immediately think like oh , I should sound professional or something right , and it doesn't really matter how professional your question is doesn't like determine whether or not you will get your question answered right . And so I think chatbots actually allow people to chat a little more freely , ask their question a little bit more openly and hopefully get the answer they're looking for . Because I think when people move to email , their mind moves into this mode of like I need to sound like a professional , and I think that can be . I think that can actually distract away from that actual question itself .
Speaker 1Such a good point To whom it may concern . I had a question about your product .
Speaker 2Yeah .
Speaker 1Please respond at your earliest convenience . Thank you .
Speaker 2Well , I think and like it's funny because it just sounds so robotic and it sounds like an AI agent .
Speaker 2But that's what we do when we move to email , right .
Speaker 2But I was also just thinking about how , if you put one of these little chatbots on your website , or if we were to do it on ours , I think I still would consider them chatbots , because I still think they're fairly limited .
Speaker 2I think what's going to be really interesting is , over the next year , they will move to more agents , and what I mean by that is you'll see small businesses start to give these chatbots in the corner of the website the ability to close deals for you , right ? So right now , at the end of these . For the most part , they will either close the ticket or route you to sales or something like that . But it would be really interesting to start seeing them be like oh , you would be a good candidate for our product . Do you offer them a free trial ? Or maybe even negotiate , give it some constraints in which it can work and say , hey , you can offer them a certain type of discount if they want to sign at the moment , and I think that's when these chatbots will start to become a little bit more agentic , in my opinion , behaviors based on like its own judgment .
Speaker 1so I know that is interesting and actually
Anthropic Computer Use Launch
Speaker 1I think there is a pretty big update from anthropic . Is it related to this ?
Speaker 2so the computer use thing or yes , where it's basically like it's making decisions on your behalf and using it so computer use is is wild , and so the thing that they released that I really wanted to talk about this week and I haven't used it myself , I haven't had time to look at it this week , but just from what I've seen it's really cool . So , basically , what I've been saying with agents is they can reason and take action , and then the way they take action is you have to define the tools that they can use , and so what that means is like you can , basically , from a developer standpoint , you can say here's , here's a Google search API . So like , if you give a query to this API , it can go search Google , it will return you 10 results and it's going to return data and what the data returns going to look like this . And so you kind of have to give it this like structure around what tools it can use , and so , like I said earlier , some of the tools that are pretty popular are like searching the web , you know weather APIs , like all sorts of different things . Any , basically any API that's out there can become a tool , essentially right .
Speaker 2What's really cool is anthropic release , computer use , and so the tool that it can now use is moving a mouse or a cursor around your desktop or a computer's desktop and clicking on things and like typing things in as if it was sitting at a computer .
Speaker 2And this is wild , because so many of the things we do on a day-to-day basis as knowledge workers are not exposed via an API right . So previously with plugins , it was kind of limited to things that were exposed via apis . But now you can tell an anthropic agent you know , open up chrome browser , click in the google like you don't even need to give it , click in the google , you can just say search google for this term , open up the first website you see and send me a summary of that website and all that stuff you could have done via apis before , but now you can do it via this computer use thing where it's literally moving a cursor around a computer and like clicking on stuff and so its ability to replace knowledge worker work or things that we do sitting at a computer , just like the game , totally just changed . So I'm super fascinated by this .
Speaker 1What does this mean for bot farms ? We talked about them a little bit last time , and part of the way that they crack down on the bot farms , or spam bots for email and things like that , is by recognizing and I'm not the technical wizard on this , so I apologize for any layman's terms here but basically recognizing if the actions being taken mirrored what a human would do , and so it was more likely that it was a human actually sending these things and not a bot , versus those actions were missing , and so then it was identified as a bot and then it was dinged . But now , if you have the ability to have computer use tied to bots , that what is ? I don't know . I'm just thinking out loud here . What do you think ?
Speaker 2Yeah , no , I think that's a really good point . So I think like a really good example for kind of you know you're saying how bot detection is all based on like are these ? You know , is someone browsing using like human behavior right , as opposed to that kind of programmatic behavior ? And I think the best example of that is this thing called captures . So you know , when you go to a website and it tries to say , are you human ? And they're getting increasingly complex about like the weird things it's asking you like , like what looks they're ?
Speaker 2so hard I get them wrong yeah , it's like what looks weird in this photo or whatever .
Speaker 1So a good example that went around either way , but can I just say my favorite captcha ever . I took a screenshot of it . It was the squares of dogs faces and it said which dogs are smiling I love that .
Speaker 2That's amazing . Um is also , as a side note , the best data labeling tool ever , so I think the way that and the reason they're doing Captcha is that it ends up labeling data that ends up being used for training data for AI models .
Speaker 1Yes , one day I'll see an image generated . I'll ask for it to generate an image of a dog smiling , and it will be the one .
Speaker 2That's what I was just thinking about . It knows what a dog smile looks like , because it asked you and 10,000 other people what does a dog smile look like ?
Speaker 1So anyway , as opposed to a bridge or a bike .
Speaker 2Yeah , exactly Exactly . Well , you know , listen , the more of these bicycles we identify , the the the fewer autonomous vehicles hit bicycles . So I think that's a good thing , because that's probably what that's been , that's true , except some of the bicycles , and the photos are like on their last wheels bicycles , like up against a wall , totally rusted out , and I'm like I think that's a bicycle , or at least it was at some point .
Speaker 2Does that still qualify ? But no , yeah . So getting back to the , the bot stuff is I saw a twitter thread of someone using computer use to solve captures , and so that's really interesting , because now we're gonna have to start to think of more innovative ways to do bot detection , because if they can solve captures well , then they can act as if they are real humans , and so , yeah , it's . That's why I think this is such a game changer . It's , you know , a lot of the things . Now , anything a human can do at a computer can , in theory , be done by an agent or a bot , which previously wasn't the case .
Speaker 1So crazy . I feel like both you and the friend you were just mentioning are brave in being super early adopters of these new technologies . I like to think of myself as an early adopter , but I think the reality is I'm not , because there's still a part of me that's always a little bit . I don't know if it's risk adverse or skeptical at the very beginning . I need to see it play out a tiny bit or spend time at least thinking about the long-term consequences and impacts and then decide what my risk profile is for that thing . But yeah , I feel like computer use is definitely a game changer .
Speaker 2It's going to be crazy . I've seen some really cool demos of it this week . It's funny , I'm blanking at it right now . But yeah , the ability for agents to control computers . And let me clarify here , I should say these demos what they're doing is they're setting up virtual computers that the agent is working against . So it's not like working against your laptop necessarily per se . It's like working against , like a desktop environment that has been set up so it has a chrome browser or whatever . I think you can set it up to work against your own home computer . But it's not as if like currently although this is probably possible in the future it's not as if , like currently , computer use allows anthropic to like connect to my laptop at home and like control my screen like . That's not what's going on here , although you can imagine a world in which that is possible .
Speaker 1Actually yes , and I think actually I can't remember where I saw it , but I do think I saw somebody show them doing it on their own computer .
Speaker 2Yeah , I think you probably can . And it's funny because the first thing I think about is man . It's crazy that , from a cybersecurity perspective , maybe in the future some virus will get installed on your computer and it will turn on computer use on your computer and then the agent will go click into things and like find your password file or like open up your bank account , and those were all things that could do programmatically . But again , some of these things we've built certain safeguards against , and those safeguards are keeping in mind that there's like a human intervention needed . So , like , here's a good example , right ? A good example is two-factor authentication , and so currently , with two-factor , someone would need to get a hold of my phone to see the codes and things like that .
Speaker 2But now , you know , my Mac is tied into my iPhone and so it's what's ? One of the really convenient things is when I go to fill in a two-factor . If I'm at a website and I say send me a two-factor code to my phone , it pops up in my Safari browser when the code is texted to my phone , right , and so now someone doesn't need my phone , they just need to be on my laptop and then they can get into whatever they do right , and so through computer use , they could like ask for a two-factor , and then it like pops up in my browser and then they like click on that thing and so , like I don't know , it just like does change the game on security and capability of agents quite a bit interesting yeah .
Speaker 2So I mean , it's it'll this , this computer use things that feels like it changes a lot . I think it's going to be one of those things that when you open it up to people's creativity , you're going to find a lot of really cool use cases that people can do with it . Also a lot of really scary ones potentially , but I do think it's going to be a game-changing kind of feature that they've released .
Speaker 1Yes , and , by the way , always it's two sides of the same coin . I think about that in life in general , general , but in any situation like this , the things that will be amazing , used correctly , are also the things that , if used incorrectly , you know , have now even more capability or power for that
Nvidia's Approach to LLMs
Speaker 1unrelated , but just to close out on this , I'm curious what do you think about what nvidia did ?
Speaker 2like a week or two ago . Wait what , what ?
Speaker 1did they ?
Speaker 2do they released a model ? I know that , nemo , is that what you're ?
Speaker 1talking about uh like . I've read a few things of like are they actually trying to get into the alum business or are they just trying to create , you know , capacity restraints on their product ?
Speaker 2right . Yeah , that's a good question . That feels like a huge distraction for nvidia , a hardware provider to like want to go down that road . But ultimately , the more open source models that are out there that are highly capable , people run these open source models on nvidia chips at the end of the day for inference , and so the more capable models that are out there , you know , the more nvidia's chips are being used and so I don't know if there's a way to like it would be a little bit beyond my knowledge to say if there's a way that their model is , like , optimized for NVIDIA's chips in some way , to say like oh , we released this model because it performs better on NVIDIA chips or something , or takes advantage of certain features of NVIDIA chips . That would be interesting , but I'm not sure .
Speaker 1I agree with you that it seems like it's pulling them away from their core business . And what's the end game with building open source LLMs ? From a business perspective , from a societal perspective ? I don't know if this plays off of what we talked about last week in terms of , you know , seeing the leadership exodus at open ai , if there are other players in this space who have the capability to deliver lms that compete and they're doing it in a way that they feel like is more to society's benefit in some way .
Speaker 2I don't know , but I guess that's one thought it looks like and this is maybe not exactly addressing your direct points but it looks like what they're doing is putting out what I would call like a reference architecture for how to train your own LLMs , because training LLMs requires NVIDIA chips and so , basically , the more they can say hey , hey , here's how we trained an open source model on nvidia chips , you can do that too . Now gives thousands of developers a reference architecture or a blueprint for how to train their own models , which means they're all going to go pay for nvidia chips to train these models . So I I think that's maybe what they're doing is like knowing that if they help push the boundaries and make it easier for people to train their own models , that they that basically in video chip usage goes up .
Speaker 1Interesting . So do you think that they're basically treating this as almost like a loss leader , you know , like the classic like razor , handle and blade model , where they're saying , well , okay .
Speaker 2Yeah , I think that's probably it . They're saying like , hey , it's really easy to train these models and here's how you do it , and , by the way , it just so happens that our chips are the best ones to do it on .
Speaker 1Yes , and they're willing to pay whatever it is to run this model on a daily basis so that they can deliver on that market Interesting . All right , well , with that , I appreciate your thoughts . Today I definitely learned more about agents myself . I hope our listeners did too , and we'll be back .
Speaker 2That sounds good . I think there's gonna be a lot more future conversations as we see agents doing more and more crazy things . I think we'll probably be talking a lot more about them over the next year , because it is the hot topic in in the valley these days definitely all right bye you .
