Nikola and Damien chew the fat about developments from AI giants Anthropic and OpenAI, comparing their new models: Claude 3.7 and GPT-4.5 Orion, and where they place in the consumer and enterprise markets. We also touch on OAI's specialized agents for hire, Amazon's new agentic AI division and the recent stance of the US Patent Office on AI-generated content. Follow PolyAI on LinkedIn Watch this and other episodes of the Deep Learning pod on YouTube
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
NikolaChanging versus just like waiting for grok to output something makes me not want to use grok. Even though it's a lot spicier, it's a lot less filtered. It's a lot more interesting. I think I, I got to the point where I was having a pretty deep conversation about how Nigel Farash campaigning for a ban on haggis would impact the prospects of Scottish independence.
DamienHi, and welcome to another episode of Deep Learning with Poly ai, the only podcast where finally, both hosting guests are sharing a continent and welcoming Nicola back to the podcast. Nicola, how are you?
NikolaI'm doing great. I think we were, I think we were on the same continent last week as well. But yeah.
DamienI am excited to talk to you today'cause one, I wanna find out how the West Coast is treating you. You're in the Bay Area now, right? Yeah. You're gonna be there for a little while. Oh,
Nikolaquite some time,
Damienyeah. Excellent. Good to hear. Opening up our office out there, which is great. San Francisco office. Yep. Yep.
NikolaCould have gotten a better background, but yeah,
Damienit's okay. It's all right. It just, it focuses all the attention. It's
Nikolaa great office, we've got three execs for in daily, so it's you always need those few kind of seed daily arrivals. So there others who might show up know that they're not gonna go to an empty office, it's writing a great morale.
DamienYou need those anchor persona. So I'm glad I'm glad to see that you're gonna be one of them, at least for the next few weeks. That's fantastic. But speaking of other like new developments and stuff that's going on in, in modern times. I wanted to chat with you. It's been a while since we've done sort of a news and review, and there's a lot that's happening in the general AI industry, of course, with some of the big names, but then also things like regulation and US policy. I'm excited to talk about one in particular that I noticed near and dear to us. So I thought maybe if you didn't mind, we take the next 15, 20 minutes and just chat through some of the headlines.
DamienHow's that sound? Yeah, let's do it. Let's do it. Let's go. First and probably foremost is announcement of two fairly significant model developments from our foundational providers in Anthropic and OpenAI. Claude three seven sonnet came out I think just a few days ago, maybe into late last week, GPT-4 or five Orion. comes out just moments before it seems, at least according to the announcements and the teasers. GPT five gets announced. So of course everyone in the industry is comparing the two models. There was a lot of hullabaloo coming out when sonnet three seven arrived.
DamienDescribed as more empathetic, more human-like, whereas GPT-4 five Orion described as a little bit more deeper thinking. Which approach, if we're gonna start to compartmentalize and take a focus area if there's a fork in the road. Finally, with these foundational model providers, which approach is gonna win out in the consumer market, do you think?
NikolaWhat I mean it is always hard to call anything. I think 3.7 sonnet is a reasoning model, right? Whereas 4.5 is not, so I think that's already not really a, kinda like apples to oranges comparison, but they're definitely always fighting for kind of like the news headlines and releasing just before the other, it feels like they're in a cold war. So it's always fun to watch. I think, getting to reasoning models, taking output things a bit more quickly. I dunno if you've used grok as well, I've been like trying to diversify into using all three and the UX for example, there where you wait and you don't see anything, there's a very stark difference between that and open AI and really just seeing the reasoning steps and the summaries changing versus just like waiting for grok to output something makes me not want to use grok.
NikolaEven though it's a lot spicier, it's a lot less filtered. It's a lot more interesting. I think I, I got to the point where I was having a pretty deep conversation about how Nigel Farash campaigning for a ban on haggis would impact the prospects of Scottish independence. And was. Like having pretty, it was a pretty good shot. It was some salient
Damiennuggets on the ha on the haggis divide. I,
Nikolaand it was hilarious. Okay. And it pushed back on I don't think it's likely to happen, but then, it would be quite a funny provocation, wouldn't it? And it was like, it was a lot less filtered and guarded, but just the user experience of not seeing it, thinking and summaries and it's just something they haven't worked on yet. Because I think they're just very focused on getting the model to perform versus the whole kind of a consumer angle. I think the consumer angle matters, right? So I think they're like opening ai, they have the Head Start and Satya was recently in a podcast where he talked about it and yeah, I think it's gonna be hard for them to not well to be disrupted by the other ones.
NikolaMeta is introducing their own apps, so I think they're like throwing their Yeah. In for just like an application view of it. And you can't really underestimate them'cause you already spent. Likely more than, a few minutes a day in one of their applications, if not several.
DamienDo you think that meta timelines was expedited because of the deep seek drop? I
Nikolathink so. I think so, but I think it's also something that just should do. Okay. Like when we spoke, right? I think that meta is the western, like leader in open source and they should continue to be, that's one like angle of attack. The other is just like. What are you doing with consumer? Do you have a consumer form factor? And look, it's a winner takes whole market. They know that better than others and Chad, GBT is the one for now. But yeah I don't know. I don't know. I'm not privy to any information, but that's your thing about reasoning versus not. Yeah, I think that reasoning as it stands, if you can get it to be very quick, the best user experience will be.
NikolaModel, deciding whether it needs to reason dynamically and offering up maybe a quick response, then going upon further thought, how about this? Could change the answer. There's maybe a good way to visualize it, but like you shouldn't have to go into the backend and choose which model you're gonna use. No, that's not a good like user experience.
DamienDo you think when it comes to the reasoning models, something that maybe makes them more compelling? On a consumer basis, but could damn them in the long run is a, this possibility of being over transparent in the reasoning. Wasn't that a pretty big criticism when reasoning models started coming out, that it was telling you too many steps in what it was doing as opposed to just delivering an answer or a response to the prompt.
NikolaI think, this can be fine tuned to be for that. It's also a design decision, right? Sure. I think the initial bid around like tokens, even though it was charging for at least an API usage, it was charging for the invisible tokens. I think there was a, there was an argument around not revealing it. 'cause it might contain like potentially sensitive or offensive stuff or god knows what. Okay. I don't buy that, but it's yeah, like in the end, at the end of the day, I think it's just a matter of ux, you should be able to see the trace if you want to, but you probably don't wanna see it. Yeah, you just want the better output as soon as you
Damiencan get it. So great. From a consumer perspective, the GRS of the world's a little bit spicier, a little bit more fun to use. What about on the enterprise adoption? I. Framework is that going to play against the applicability or the sense that something like grok is a model that actually has enterprise application because it, right now it seems like a toy, like a fun jerky friend.
NikolaWhy do you think that?
DamienI'm just going off the fact that you're typifying your grok experience by saying it was, a little bit spicier, a little bit more. A little bit more maybe flamboyant.
NikolaI liked it. I don't think that, I think that for every enterprise use case, you're gonna want to filter and adapt behavior. You're like cutting vegetables and I give you a blunt knife, right? Because I'm afraid for, Maybe I'm not confident that you're good at using a knife, right? You're actually more likely to hurt yourself with a blunt knife, right? So equally, if it's already been like, sanitized and made safer and, controls built in. I don't know. Do you use like lane assist in your car or, that whole it's on by
Damiendefault. My wife turns it off in her car because it annoys and frustrates her. I don't ever bump into any issues with it, but I also don't drive as much.
NikolaThere's that, especially if you're renting a car, and you're not just Yeah. And you wanna change lanes and. You go softly. You forgot to indicate you should indicate, but you forgot, right?
DamienYeah.
NikolaYou're getting to a point where you're like, oh, it's turning you back. You get confused. Like it takes a bit of time to, it
Damienstarts pushing against you. Seems a little bit more counterintuitive with regards to being an assistive device.
NikolaYeah, exactly. So I think like this is similar, right? If you don't know it's gonna happen, you might actually do something off, right? Or you might not get the full power of the thing because it's, like it's. Swapping some behavior and then you're adding extra measures in enterprise, you're gonna do it regardless. Yeah. So I'm not a particular fan of the, overt tune that was done for for open ai. And if you think about the direction they've chosen to take, which is probably the right one for them, which is have the application make, your money dominantly on consumer.
NikolaLike each company in the end develops a DNA that then. Dominates his behavior for, years, decades to come. If they're taking the consumer angle like they will need that safety and they will inevitably continue to put effort into it, right? So you know, the Enterprise Edition is likely to be less. Enterprise is a wonder, right? It's many things.
DamienYeah. So
Nikolasome people will need a thing that's sharp knife or a car without lane assist. And those people that are gonna do their own tuning and build their own kind of like applications on top will probably want something that's been interfered with less, right? So grok, I tend to like what they've done there, but, yeah, I mean it depends, like for consumer, you wanna do what opening height did, probably for the vast fast consumer. But we'll see. Maybe the other thing pays off.
DamienWe will see. And speaking of OpenAI, a little bit of specialization in some recent announced market moves. Information is reporting that OpenAI is starting to price out the potential for specialized AI agents skilled anywhere from, intro sales reps to PhD researchers with prices that are commensurate to their human FTE counterparts. We're talking$2,000 entry level per month, up to 20,000 for that PhD level researcher. Per month. That's 240 KA year. Is this indicating that Sam and company are taking a hard line stance that AI alone is greater than humans plus ai, which I think is the general consensus?
NikolaI think they're saying that kind of AI probably in tandem what humans is worth a lot of money, which will be true for a bunch of things. When Exactly, and to what extent we shall see.
DamienWhen you're pricing though as someone as at an FTE rate isn't it you're gonna have to pick or choose for a vast majority of organizations?
NikolaDamian, given what we do, I consider myself a bit of a specialist in pricing. I tell you what what I'll tell you is it's very hard to get people to pay like the same. Amount that they would pay a human. It's possible. We're the most expensive world all of our competitors, right? At the end of the day, people are like, AI should be better and I'd like to get a three XROI right? Over human labor. The good news is we can get both. We with AI for a lot of tasks these days, right? So you know that 20,000 a month, hey, who knows?
NikolaI think it's a bit wishful. It's the same way that, like Agent Force, which only finally announced. A second update and we're still struggling to find anyone actually using it other than all the people who you know, are on like the New York subway and the London Underground in the ads stocking, using it, but they came out with that whole like$2 per resolution and I don't know anyone who's paid that or anything remote for it. So outcome-based pricing for AI makes sense. If you really know what you're solving and what it's worth to someone and then you can justify the price coming out with list pricing ahead of actually defining a use case is nonsensical.
NikolaYeah. Now, yeah, if you've got a super intelligence that can do the work of any PhD level grade specialist for anything, which they don't have yet, but still yeah. I mean it might be worth that. It might be worth more, right? Sure. I think they're just. Waxing lyrical, being excited about the future. We'll see. I don't know the numbers. I don't know if anyone does like the kind of like the 200 a month subscriber range and like what they're doing. I'm mature customer of that in that I got very excited about operator, but honestly, it's too slow.
NikolaAnd if it gets faster, I'll pay that. I'll pay more. But until it does, I don't think so.
DamienInteresting. They seem to think it's there. It's their next significant moneymaker. So we'll see. And speaking of significant moneymakers, AWS recently announces or Amazon web services recently announces they're opening up an AgTech division, AgTech ai, taking it beyond the RPAs. They announce Alexa Plus, which obviously IBUs Alexa with more generative functionality. They're saying that this ENT wing, it's gonna drive the next multi-billion dollar market segment forum. Was this eventually, was it bound to happen or is it something where maybe Amazon's doing a little trend jacking and it's a marketing play?
NikolaI'm sure it's both. When you look at, agent is just a headliner for we're gonna commit ourselves to doing a lot of work with generative AI and realizing value from it. We're an agen company. We, yes. We genuinely build like, and deploy agents. But the agent workflows are exciting. And when you look at the Alexa plus demo, it's like a proactive thing. There's a Billie Eilish concert coming up. Do you want me to let you know when tickets are available? Yeah. Okay. All of these things. Have made sense and yet they've not led to that proliferation of usage for consumer voice assistance.
NikolaAnd I've thought about this my whole life, right? It was part of our company decision making and are we doing consumer? Are we doing enterprise? And the reason you, it's hard to do consumers that level where the thing is like. A superpowered EA for your life. It's oh, conflicts. Do you want me to call an Uber for your kids? 'cause you're not gonna be there in time to pick them up. This and that. It's a complicated orchestration of things where you need to have really high trust. And if you don't, if you get the same thing as autonomous driving, like people will not use it.
NikolaI think the demo there had reasoning steps for like. Turn on the lights like by the sofa and it figures out that it's like whatever your living room like. I, over Christmas I installed some smart lights like LEDs in the backyard. And every day I gleefully go and say, Hey. Hey Google, can you turn on the, turn the garden lights on right or off later, and it gets wrong half the time. Starts a blender. It's a nice demo. The real issue there is just will people use it? Will they trust it enough? Like it's great that they're doing it.
NikolaThey should do a lot more. This is exciting. It's
Damieniteration one oh of just like those personal assistance from back in the day. And you know this from your, early days over at Apple with Siri, and you like to tell the anecdote about how people, when they find that out, about you would say, oh, Siri sucks. And it's because it's the next iteration of going in excited about the potential about the hype. Then realizing that, it's still a fairly like new complex technology going on behind the works, and it's gonna be wrong a lot. Now the only thing that we have is a precedent of low expectations when it comes to those personal assistances. Being able to fulfill that to me, doesn't set up that trust requirement to succeed because we're already going in flimsy at best of the capabilities of these sort of devices.
NikolaYeah. No, look, I think that, the trust is not there yet. I think as a society, will we do more with these things? I don't know, man. I think that a lot of the people still using like GPT and stuff are like people like us, right? Who are like hyper excited, deep in it. The general population, I think there's still a lot more people who need to start using this very heavily, right? Sure. The trust and, the whole kind of accepting that it can do it, and then coming to use it. It might be generational, it might just come naturally to, gen Alpha. We'll see.
DamienYeah, that's fair. That's fair. I guess we'll see is the big modus operandi when it comes to most of these consumer developments, especially taking this huge step and saying, look, you may have been, you may have been wary about Alexa in, in the case of Amazon in the past, but we're taking this to a whole new level and we'll see if they can fulfill.
NikolaYeah. Look, I think it's just which tasks do you want to do and how are you gonna do it? Will you give them another shot like. It's good. They need to keep trying and eventually, we'll get there. It might be something that just happens gradually rather than overnight. We definitely have tools that are, a century ahead of where we weren't looking that year to go, it makes sense to take another crack.
DamienI it's gonna be a long time before I trust with anything other than timers and reminders, but we'll see. I'm open-minded. Alright. Last piece. Very recent announcement, the US Patent Office has taken an official stance on the copyrightable and trademark ableness of AI generated output, and it is mostly that. You can't patent this stuff. You can't protect the output. Essentially, the only part of AI generated work that can be protected, quote requires meaningful human authorship and creativity, not just AI generation.
DamienSo to me, that reads prompts, source material, et cetera, all that possible to copyright. Many cases it already, it is, but the generated content that comes as a result. Not copyrightable and I gotta be honest, this seems sensible. What are your thoughts?
NikolaI think so too, although, it's in a world where like you might get different outputs if you think of like novels and stuff like that, like what do you want it to be? Copyright If you worked for a long time and the thing assisted you, I dunno how this, I guess the prompt, the source prompt is still copyright right? It's complicated. I would just like for the whole thing to be, they're all training on everyone's data anyways, so I feel like there's almost like a, maybe this is my like underlying anarchism, but, I don't know. I don't know. It's what do you what do you think this changes anything?
DamienNo, I think we were already moving towards a world where, you know the sense of training data was already going into exclusive partnerships. So it was being addressed in court with some of those, like high visibility use cases or excuse me. Non-authorized use cases for training materials of like authors or of media publications. So then what happened is eventually, checks were cut and then things like perplexity was making exclusive or perplexity was refusing to get exclusives, but OpenAI was like working directly with some publishers and media outlets and saying, look, we.
NikolaThe first time that one of those models accidentally inhales like a website with, RO robots or TXT that says do not right? And then it's spit out similar content elsewhere and then other models train on it. It like by precedent it's in there and Right. Gen out of the box, you can't really do much to roll it back. Yeah,
Damienbut that's on the training side for the output side. I think a lot of it makes sense. One, yeah it's hard because you're building on for the most part, except for open source lms, you're building on a co, on a copyrighted model. And so if or a trademark model or some sort of protected model, if that's the case. The close yeah, the close, the data itself for the most part is gonna have some sort of protection, but then also just the processing model or whatnot. As far as a contributor towards the final output it's not an it's not an individual, it's not a person who is entitled it's a processing element so much as a computer can't file a copyright on, on.
DamienYeah I don't know. Anything that you create on it or a typewriter can't you need to have that personal element. The person is what's offered the protection. So if you're going to take a stance as a US patent office does, and you're gonna affect policy, it's going to be anything that shows significant human input. That's the piece that's copyrightable. And that, to me makes sense.
NikolaYeah. But there, there is a part which is IP address, poofing, right? Or where you're basically going and you train on some data, then you all put other data from that model, and then you've trained on some data. Like for example, deeps Seq was, I think most people would agree that they've heavily used like GPT-4. Yeah. To create this training data set. So then have you seen
Damienthe meme with someone phishing on the shore and then they're putting the fish in the bucket and the bucket says GPT? Yeah,
Nikolaabsolutely. Absolutely. Like at that point, like some money laundering operation, right? If they've done it, then they apologize, they do it, they remove it. But at that point, that model's created so much data that is no longer copyrightable.
DamienI
Nikolafeel like that's almost like a weakness to, and that genie is out of the box already,
Damienright? Sure. So
NikolaAll of the data that. Places like New York Times and others are trying to protect, they're not gonna be able to because, it's already been inhaled by all these models. It's getting outputted, getting put in other open source models. How you choose to use them or not. Like you, you can't really roll it all back. And it's and the modernization
Damiencomes in. The exclusive partnerships. Even though it's already out there, at least you can market the fact that this is the only sanction utilization of that data and.
NikolaYeah, potentially. I dunno. I dunno if they're gonna see a lot of monetization, those guys.
DamienNah, probably not. I think that time has come and gone. Like you said, it's already out there in the ether. Right?
NikolaYep.
DamienVery cool. All right that's all I had with regards to some of the headlines. It was great to chat with you, Nicola, as always. All right. Enjoy the West Coast and for the rest of you, we'll see you next week.
Three years ago, Simplyhealth's PolyAI deployment was only resolving simple FAQs over the phone. Today, they run voice, in-app, email, chat, and WhatsApp conversations on the platform. Nikola Mrkšić talks with Paul...
An agent that gets better every day compounds into a very different agent within months. That's the idea behind Wren, PolyAI's new release: an agent whose job is to make your other agents better. Some of what it finds...
Companies are pouring money into AI, but not everyone is getting the returns. Ian Jacobs, VP and Lead Analyst at Opus Research, thinks he knows why. In this episode, your host Nikola Mrkšić sits down with Ian to work...
When PolyAI built its Builder MCP, one client helped shape it: Vixxo Facility Solutions, one of the largest facilities providers in the US. In this episode, guest host Damian Sasso talks with Vixxo's CTO Derek Neighbors...
PolyAI just launched Dialog-RSN-1, its first audio-native model, and your host Nikola Mrkšić sat down with its builder, Matt Henderson, to unpack why it’s a game-changer for building voice agents. Most voice AI either...