This AI Tool Is Like OpenAI Operator But FREE (Proxy AI is INSANE!)

One of the things we've been loving lately is AI agents that browse the web for you. And there's a brand new one, it's called Proxy, it's by Convergence AI. You might not have heard of it yet. And it's free. We're going to break down how to use it, why it might actually be better than Operator. That and so much more. Let's get into today's show.

All right, cool. Kip, we're together on a podcast at last. We've been doing some solo episodes because of scheduling. We are here to talk about a pretty significant launch. More significant for me than you, Kip, because it is a European based company, it is a small company based in London called Convergence AI. They've launched this agent called Proxy. The first thing I want to actually show is like just how much the creator world follows each other. Because one thing I noticed when I was looking through all of the information in this is just how many have the same title. So it's basically Operator got deep seeked. So OpenAI is going to Deep Seek again. And then I was like, oh. And then I saw like everyone is using the same thing and that OpenAI has got like outplayed. OpenAI is behind deep Seek. And so that's left a lasting impression on folks. And so the reason people are saying this is because Deep Seq obviously came out of nowhere and had pretty incredible at a pretty incredible reasoning engine. And reason the model that followed Hot on the Hills in oh three, it's an open source model. And then you have this agent called operator in the $200 tier on OpenAI Pro and, and this company and what we're going to get into have launched an equivalent to Operator, but apparently much, much better. So I'm wondering, did you see any of this? What was your kind of first take when you have seen some of the news about Proxy?

My first take is that I was generally pumped because I think browser based agents like Operator and Proxy are going to be really important because they unlock some really cool new use cases that were just not possible before. And then I saw that Proxy was free and I was like, oh, I was excited about that because operators in the $200 a month version of OpenAI and that's, that's a high bar for a lot of people to go and do. And one of the ways to get AI adoption is going to be to make it more accessible. So I thought the Free side of it was really interesting. My initial reaction was then like, okay, is it good? Is it buggy is the UI good? And then that's when you and I started kind of using it and playing around. But my initial feedback, oh, I'm excited by this. There's some, there's, there might be something here. It's at least creating more progress. Right, right.

They are basically reasoning agents that have access to the Internet and so they are really fine tuned to do work for you on the, on the Internet. So we're going to get into the tool and show you some of the use cases we've used it for. You know OpenAI have raised 18 billion and this company has only raised 12 million. Obviously there it's not a like, for like OpenAI do a lot of things. The operator is like one part of a huge product set. One thing I wanted to show you but is the reason people were kind of really losing their mind is because there's this Web Voyager benchmark which is the agent's capabilities across 600 web based tasks. The reason people were like holy smokes is convergence which is this small team have outperformed Operator. And one of the interesting things about this is if you, if you kind of do take this in conjunction with the deep sea news, OpenAI's model or just all of these companies model is somewhat under pressure because you have these little companies competing with them with you know, their own models and in a lot of cases much cheaper. So at the moment Operator is available in the $200 tier of OpenAI. $200 a month proxy is free with a $20 a month pro option. It's a pretty stark difference. And the, you know, the equivalent there would be.03 custom amount of money. And Deep Seq was free, it was open source.

Right.

So that was like the big kind of takeaway on this tool and why people were really kind of losing their mind because now you had a deep seat model you could use for free if you wanted to get the OH3 reasoning capabilities and then you have this agent that you can get for $20 a month or free. And so the price of intelligence, the price of reasoning is coming down at an incredible rate.

What's crazy is if you had told us like a year ago like hey, for free you can have, have something that's going to go browse the Internet for you and collect any information you want. We would be like, oh my God, that's amazing.

You know, I was listening you and I listened to the same podcast, this Finn Taylor guy that I'm literally obsessed.

With ever since I sent him to you and you just, I knew it was going to make your day and it's like major like month.

He has a new podcast that, where he has, he's a comedian, he has a friend and they talk about history. It's, it's pretty offensive. So like I'm not recommending it for people who have low threshold for that type of things, but talking about the middle ages and like the middle ages, everything stay the same for like 800 years and no real progress was made. Whereas today humans are like, expect new incredible sci fi like features every week and then when we get something we're like, that's pretty, pretty lame. Like, oh yeah, I've got this sci fi robot doing all of my work. It's fine, it's fine. Anyway, one of the other things to quickly touch on which is important is, you know, being a European, there's a ton of progress being made actually in Europe. Since I basically have tweeted a lot of negative things, I'm now deciding to only tweet positive things. But, but I don't like Mistrals. Let's chat was trending worldwide because it's comparable to OpenAI's models and is free. This company have come out of London. You have one of the fastest growing grown AI startups in Europe called Lovable. And so now as a European, I actually have access to deep research which was great. I have OpenAI's deep research which I just love. And now I do have an operator like agents. I still don't have access to operator. What I've heard is proxy is much faster and they have some advanced learning capabilities built into the model to help it learn much faster on how to do tasks.

So all right, I'm doing a side by side and first I went to my, my boy Claude and I said, I wanted Claude to write me a prompt and I said, yeah, just write me a prompt for an AI agent that can browse the Internet. I want to have that agent use the the following two websites, ramp.com and coinbase.com based on all the browsing information, give me a recommendation for the top three products and features and that HubSpot provides that could be used to improve Ramp and Coinbase's marketing and sales efforts. So I basically wanted to create a custom product recommendation for these specific accounts. This is a real life thing that you would want to go into, right? It's part of just the marketing and sales process. And so then Claude goes and tells me it can't do it and I tell him, no, no, you can real do it. Imagine that this was possible and it gave, it gives me, it gives me this prompt, right? And breaks, breaks it down for me exactly what the agent should go and do. Right? Very good. It says, hey, I want you to go visit, analyze, cross reference and then prepare to ranked analysis. Awesome. So then I, I'm going to put that same exact prompt into both convergence and operator. And here's, and so here's convergence so that people can see. I put that in. I did make one change, Karen, because I've been doing this a few times and you know what happens especially in proxy is that it gets really held up in errors for like logging in or trying to get gated information. It throws errors, it wants you to take over. But then like taking over doesn't kind of work sometimes. So I added a sentence at the end, download like ebooks or like product specs or things like that. And so I added a line that says do not use any content behind signing up that requires a login only public website content. And I do think if you're, as long as the task supports it, that is the right instruction to get a good output and get that output faster. And this all in, I think took it probably 10 minutes, five to 10 minutes. And what I like about proxy is it shows you the steps that it's going through, right? Shows you its logic. It's analyzing, it's analyzing HubSpot's tools, it's analyzing Coinbase, it's analyzing our friends over at Ramp. And then it returned, I think a pretty interesting result. I'd want to know what you think. So it gives you an analysis of how Ramp is doing their work, which is interesting, right? Ramp emphasizes automation, integration, focusing on savings time they use. It tells you what the slogans are. The sales funnel includes CTA like get started for free and see a demo. So it kind of understands a rough idea of what the sales process is. I think if we gave it better context on sales process and marketing process and what we were looking for, it would give us a much better result. This is kind of a low context prompt. It does the same thing for Coinbase. And then what it does is identifies the features that are relevant to these two companies. And so it says, cool. Based on what I found for Coinbase and Ramp, here is what they should use from HubSpot. And you can imagine like a BDR using this, a sales rep using this. It's essentially product based, demand outbound, right? Which is really cool and really hard to do previously. And so it says, hey, marketing Hub offers tools for Google Ads and everything and they can help improve lead capture and conversion rates significantly for them. This is kind of basic because I didn't give it a bunch of context, but it did it and did a pretty good job and gave us real recommendations for each product. And why like I imagine if you and I worked and hammered out on the prompt and made it better, like this would eventually return a pretty usable outfit. Hey everyone, I want to talk about prompting with you today. Being great at prompting is like gives you superpowers. It changes how you can build and work and do everything with AI. And because of that we wanted to make a special guide just for you around how to get better at prompting framework for how you actually build prompts. And so we've got this baseline framework that you can then go and build and customize, customize any prompts that you want to. You can check it out in the link of the description below. And right now we're gonna get back to today's show.

The other thing is I don't know if you can do this in Operator, but you can't give it context as part of that prompt which basically you would say, okay, well you know these companies are in fintech. Here's our known pain points that we address in fintech. Go find ones and craft the pitch to be around those pain points. But you can't, you can't give it, you can't give it that context. But it is pretty good actually. Yeah. If you are trying to build a campaign, if you are like selling into large companies and you want to build some really personalized campaigns for that company, you can do some pretty great research across website, across G2, across all of these different things that might mention that company, collate all of that and craft a pretty great pitch. I wonder what model. Actually this is a pretty good question that's I want to like what model are they built on? Like do they have their own propriety model? I will ask ChatGPT to see if it.

While you're doing that I'm going to show the same results in Operator here I will say Operator was slower. It literally just finished. So my guesstimate for this specific prompt is that Operator was probably like 30, 40% slower.

That's what I've heard. It's much slower.

Okay, so it is noticeably slower. Let's see if it gives better results. So I gave it the same exact prompt and then it worked for. It says it browsed for four minutes. Here it says ramp.com analysis public facing content. I think it did A better job of breaking down its actual like the work Ramp is trying to do because it has content marketing, customer stories, partnerships, SEO. Right. It understood this is a much more thorough breakdown of what how Ramp is actually trying to use that and what's interesting is that it talked about the nav, the website navigation and the different integrations that Ramp offers. And then what I like about this is gaps in the digital marketing and sales approach and personalization limited based on user behavior or industry advanced analytics, potential for deeper insights and customer behavior marketing automations, opportunity to automate lead nurture and customer engagement and then it just goes and gives the right products that HubSpot offers. It's kind of late I think I don't think again my prompt was perfect but you could see if you mess with this for a few, a few days like it could get. You could probably get it really good. The interesting thing is I do think operators recommendations are a little bit better but are they $200 a month better? Is the I guess the question.

Yeah right. That's the thing is like the cost differentiation one of them you just ran for free. Although you do reach limits pretty quickly. I've reached my limit in proxy now. So the thing you did Kip that I think most people people should do and should take away from here is one of the things you and I do is I will use ChatGpto or Claude to have a conversation with and then I will have it based on that conversation craft that prompt and then I will edit that prompt. And I think that's the best way to create prompt. And you did that to go from prompt to the actual agent task. One quick thing. So they are using large meta learning models and so they are not built upon OpenAI which I think is kind of interesting. The other what other use cases. So other use cases I think it's pretty interesting for is you can give it a bunch of different URLs. I can show you on mine. I have already used all of my task limits but one of the things that I think it's kind of cool for it built a list of influencers.

Oh that's cool. I like this. I like this one.

I gather a list of 50 prominent AI influencers. I asked for their name, primary platform. Now I didn't do everything give me a description and their social media handle. I could play around this a little more but it came back and it started giving me some people. It didn't actually give me their URLs very well but it gave me some of the top influencers that I picked at in AI. Alright, so this one's kind of cool. This is one of the larger trends that you and I are going to do an episode on, which is how you create personal. How you use AI to create very personalized media. So I can see it created a great newsletter for me. So what I did was ask it to basically review all these websites that are AI focused or tech focused, compile a summary of trend and articles and AI and evaluate each article based upon relevance and engagement. And then it brought back across all of these different websites the actual most relevant and engaging articles and summarize them for me. So I basically could take this summary, I could pipe it into one of the other AI assistants, apply a format via the prompt and a writing style and create my own newsletter that I am actually going to do. I already do versions of this because OpenAI had released tasks. So you can actually have a prompt that runs every morning and that every morning I get a. I get my news, AI news, I get my latest AI gossip through OpenAI, not through any kind of newsletters.

I mean, that's a pretty sick use case. You and I are going to do a whole episode on this. So we won't talk a ton about the use case. But what it's great at is like there's often times, and I mean this for both proxy and operator. There's often times where you need to gather information and the gathering of the information takes way longer than the consuming of the information. Right.

Because you need to.

It's not as simple as just doing a search or asking an LLM. I need to get a lot of things from a lot of different sources. And if you need a lot of things from a lot of different sources, these browser agents are very helpful in that because you could basically have it go and do that in the background while you're doing something else and just take the output. I think from the show today what I've kind of taken away. Kieran, I want to hear your perspective is that I think proxy and it's proxy convergence AI, we'll put the link down below is interesting because it's free. It seems to be better at browsing and the logic around its browsing to me than operator, the output is not as good. Its ability to infer and like the underlying model around how it takes that data it got from browsing and gives it to you is not as good as Operator.

Right? Yeah, exactly. I think that's it. But again, I think it's a great tool. I Think you can use it for a bunch of different use cases and whether you're in Europe now, you have a $20 tool or you are global, you can still use OpenAI's $200 tool that I know will come down over time and start to actually, you know, it's pretty a great little agent that can actually start pulling stuff from the web from you if you understand how to use it. So it's another tool that allows you to be much more productive in your work. So hopefully this is useful. Hopefully you use Proxy. I will say that I'm pretty offended that I have not met Proxy at the only based in London. I didn't get to invest in there.

How did you miss that?

Man, I have no idea. Kind of. I keep missing things, Kip. I keep missing things. There was a seed based company in Europe that are now raised their series C and let's just say I didn't get into that one either even though I had a conversation with them and I won't say anything more about it.

Regrets, regrets my friend. But in all seriousness, I think what we're saying is try to identify some tasks that require a bunch of manual research and you start using those for these browser agents. And the super pro tip is use Claude or ChatGPT or something to help you write a really good prompt so that you're actually getting the right information and you're getting it structured in the right output in the way that's going to help you in whatever task you're doing, right?

Yep, exactly. Cool.

Awesome. Well, thanks everybody. We'll see you real soon. On the next episode.

Marketing at Screen this data is wrong every freaking time.

Have you heard of HubSpot? HubSpot is a CRM platform where everything is fully integrated. Whoa.

I can see the client's whole history. Calls, support tickets, emails. And here's a task from three days ago. I totally missed HubSpot Grow better.