Ilan (00:00)
Quick thing before we get started. If you're on a product team listening to this and your team is burning over 10 hours a week on manual status updates, data pulls, or repetitive workflow BS, we're taking three teams through a free workflow audit before the end of January. On the call, we'll show you exactly where you're bleeding time and how to automate it.

If you want us to build it for you after, we can talk. If not, you keep the audit and you can build it yourself. But here's the deal. We're only doing three because these are time consuming to prep for. That's all we have capacity if you want one, go to the link in our show notes right now and grab a spot before they're all gone.

Ilan (00:37)
Hey everyone, welcome to Prompt and Circumstance I'm Ilan And today we're gonna

David (00:41)
And I'm David.

Ilan (00:43)
giving our predictions for 2026

David (00:58)
All right. So we've got some predictions of what's going to happen in 2026 and it's going to be quite an interesting year is our bet.

Ilan (01:07)
Now is the perfect time for me to tell you about our first sponsor for this week,

Querio Have you ever gotten blocked on answering product questions because you had to wait for data

Querio's AI agent sits on top of your data stack and allows you to ask natural language questions and get direct answers and beautiful visuals to tell you what's going on in your product. If you want to try it today, you can go to their site, querio.ai, and tell them David and Ilan sent you for two months free of their Explore product. That's a thousand dollar value for our listeners.

Ilan (01:41)
Yeah, it's shaping up to be an interesting year, David. And there are a lot of trends that we saw taking shape in 2025 that I think are going to continue into So I'll tell you my first prediction for 2026. One of the big product trends we saw in the last year was newer product companies really

thinking about AI native experiences for their users. We talked about this in our top 25 products of 2025 with products like Linear and Cursor, Windsurf, thinking the user experience look like and feel the fact that we now have AI. And I think in 2026, what we're gonna see is more legacy software players.

And by legacy, mean, anything older than three years. That's right. Taking a similar approach to their product. So completely rethinking what is the product experience given the capabilities of AI and agents and really diving into how to.

David (02:29)
It's the age of AI. Three years is ancient.

Ilan (02:49)
manufacture completely new user experiences with AI.

David (02:52)
Yeah, that's a, that's really interesting. I think that we will see some make that transition. I feel like there's going to be some others who will continue to do things as they always have and wake up one day to find that suddenly the market's not there anymore.

Ilan (03:09)
That is true. And that's why I think that there's going to be a move among more established software companies to think in this way. Right. There's going to be a little bit of an adapt or die moment in this coming year.

You see it with companies like Constellation Software. Do you know Constellation?

David (03:30)
Yes, they are a conglomerate that just acquires a whole bunch of different software companies.

Ilan (03:36)
Exactly. And for years and years and years, Constellation has been steadily growing, providing user value. They never divide stock. So the stock price is like $3,000 US or something like that.

The reality that they're facing is their stock price has gone down about 30 % over the last couple of years. And I think the reason this is happening is that as AI coding

become popular, people are questioning their software spends. It's like, Hey, I've been spending a thousand dollars a month on SaaS tool X, which solves one particular problem for me. But can I just build that with AI? Like how is that building me a better experience or how is that, you know, different for me? If I build it myself, it's a custom piece of software that I own and I maintain.

then it will do exactly what I need. And if I'm using a combination of Lovable and n8n and all these tools in the background, then potentially I can build an agent who will take a human out of having to do some of the steps in that software. So instead of a person having to log in and take some actions, I'm gonna have my agent take these actions on their behalf

So.

David (04:53)
And that that

agent could live wherever the market is. So that agent could interact with you on WhatsApp ⁓ And you know, don't need necessarily to even build a UI. the thing related to that as well is like, you know, ⁓ even if the market, you know, like your target market doesn't build it themselves, somebody else could build it.

Ilan (04:58)
Right.

Mm-hmm.

David (05:16)
Right? Somebody else could very easily recreate your software in terms of the problems that it solves. They could solve the same problems using something that's AI native and they can spin that up really quick and they can take your market away from you. And then it's going to be a race to the bottom that you can't sustain because now you have all of these people maintaining all this legacy code.

Ilan (05:38)
exactly. this is...

As I make this prediction, as we talk about it, I think there's gonna be an interesting shift. Maybe this will be a 2027 prediction in a year from now, a prediction for my prediction As people build their own solutions and really dive into building AI native flows and functionality, they're suddenly going to realize like, crap, I have to maintain all of that.

potentially people will revert to deterministic software flows at some point, but that won't be in 2026. We're going to see a lot know, just

in software tools where you have to write out all of the context and probably a lot of people getting hired in the background to just be shepherds for customers.

David (06:30)
⁓ that's quite the prediction.

Ilan (06:32)
Well, we shall see if it comes true.

David (06:34)
All right, so that's the first prediction. next prediction is regarding robotics and world models.

All right. So here is a chart that is showing the growth by sector and it is splitting out AI and robotics. And wouldn't you know that robotics is approximating the same kind of growth as AI is, right? But it's certainly not getting the same level of news. And I suspect that's because it hasn't reached that critical point yet.

where now it's like ChatGPT of 2023, you know?

Ilan (07:12)
mean robotics too, its value comes in a different place chain than AI. AI tools like ChatGPT or Claude are valuable for enterprise, but there's also lot of benefits just to an individual user where robotics...

currently and historically has really been on manufacturing, exactly, ⁓ low down in the, in the chain before it gets to the user. And I mean, robotics has been a huge thing since the 80s in manufacturing.

David (07:37)
like industrial use cases.

Ilan (07:52)
There are a lot of economists who have looked at the data from the last 40 years or so and predicted that job losses due to offshoring jobs in manufacturing would have been mirrored, maybe offset by a few years, just by automation and manufacturing anyway.

David (08:08)
Yeah, that's a really good point. ⁓ And I think it brings more precision to this prediction, which is about now robots the industrial use cases. Right? If we look at the VC funding, here's a chart of one that's looking at where VCs are putting their money with respect to robotics. We're seeing a consistent growth.

so, you know, just 10 years ago, it was hovering around the 1.5 % and for some it's exceeding 4 % of a VC portfolios. So that's, uh, really interesting to see the, this kind of interest in robotics. And I suspect that it has to do with the AI models that could help drive some of these, uh, robots. Right. So when we think robots,

deterministic programming driving those machines. Whereas perhaps with the newer generation of robots that might be serving somebody at home, this is a bit more probabilistic. And that's tied to, as I mentioned, the world models, which ⁓ there's two that I think are of note.

So both of these models come out of DeepMind, which as some of us might know is the company that Google had acquired and is headed by Demis Hassabis. So

Ilan (09:26)
Right.

David (09:30)
There's a model named Genie and this is something that came out this year and this is a model that can generate worlds, that can generate physical worlds, right? And so ⁓ in the preview video, we saw that somebody could ⁓ be like base jumping and you can see the world just actually getting generated as that person flies through it.

So that's really interesting that there's an AI model that can generate physical worlds. So now what you do is you combine it with something called SIMA, which stands for scalable, instructable, multi-world agent. And so what SIMA does is ⁓ it is an agent that can act in a physical world. Now a simulated one, but put those two together.

Right? Like what this means now is that you can have an agent train itself in a simulated physical world. So you don't need to have a physical robot do whatever in the real world and then have it slowly learn from that. You could spin up 1 million instances of this agent and have it learn through

one million different tasks that it needs to complete. And from there, it can go through reinforcement learning to learn more about the physical world so that when you then deploy that into a physical robot in the real world, it's going to know how to open the dishwasher and put the plates in in such a way that it doesn't break the plate.

I think that there's really something there and that we're going to see the next notch up in this kind of combination next year.

Ilan (11:13)
Yeah,

I mean, there is something interesting here. I wonder how much of cases going to be a little bit like trying to conduct user research on synthetic users. The reality that we've seen is for

AI use cases that are complex and basically anything in the physical world is from the perspective of an artificial intelligence. You know, have to deal with gravity and different textures and, you know, random things that might be in your way. The way that those have traditionally been handled is having humans

take the action first and then the AI learn from how the human acts or like take for example, self-driving cars. One of the reasons that self-driving cars

began taking popularity in the Bay Area is well, okay, yeah, there are lot of companies there, but also the weather is really good. So you don't have as many random changes in road surfaces from, I don't know, freak snowstorm or flash freeze that suddenly creates an ice patch. So it's a simpler place to train your self-driving model.

While I'd love to believe that there's a huge data set out there on the internet of people opening dishwashers, I actually suspect that the reality is much narrower. And, you know, if you wanted to know what's the texture of every single different plate that you might have to pick up with a robot hand or how much force each particular dishwasher open it.

without breaking it. It's actually going to be much more complicated than what these world models can provide. Now I think that the place where we're seeing what I'm saying take root is in the Neo robot which came out I believe in October, November this year.

The Neo-Robot has a subscription service and there's actually a human who can take over the robot for certain tasks. So there are certain things that the robot can do on its own, completely agentically, but if you give it a more complicated use case that it does not know how to handle, then a human takes over and actually controls the robot. And that's kind of weird because there's like a person who can see the inside of your house and do stuff in your house. But the whole

theory behind this case is that that provides the training material such that they can then program the robot later to do more of these actions completely autonomously.

David (14:00)
Yeah, that's a, that's a good point that you bring up. ⁓ and in fact, in the, demonstration. video of, ⁓ that robot putting dishes away, it was actually a human controlling the robot. was a mechanical Turk.

So you're right in that it's a difficult problem to solve.

I, ⁓ I feel like it's the next frontier that there's a lot of smart minds, trying to push to solve. know, we were familiar with Geoffrey Hinton, the godfather of AI. Now do we know about the godmother of AI? So that's, that's Fei-Fei Li.

So Fei-Fei Li founded this company called World Labs. And ⁓ this is a company that is solely focused on ⁓ simulating a physical world for AI models to learn more about that.

we might not have necessarily like the major breakthrough that is going to see robots into everybody's home, including maybe Will Smith. I dunno. you know, I think that there's going to be something of a big deal in this arena next year.

Ilan (15:27)
Yeah, I see what you're saying David, but I still think that the augmentation of robotics and AI working together is really gonna be in the industrial space.

Which brings me to our last prediction for 2026, which is that we're gonna see the first major AI security incident happening. And I'd like to bring a little chart here.

which just shows a trend of reported AI security incidents over time. And these numbers are still small, but growing exponentially. So we went from about five in 2010 to 250 this year. And 250 is actually pretty small considering the adoption of AI, right? If ChatGPT has 800 million weekly active users,

That's a pretty small percentage. But if both of our previous predictions come true and more companies are launching AI native functionality and AI and robotics are taking hold in a bigger way, there's going to be something that happens

we saw in the last year where a nefarious actor was able to exfiltrate data from somebody's Google account simply with

an image that had a prompt injection attack buried into its code. You showed that on a previous episode.

This year we also saw the first major AI powered attack from a Chinese hacking company and they targeted major governments and large institutions and their AI was able to plan an attack on the target, then execute.

that attack, then review the results of its execution and optimize its following plan and continue through this chain over and over again.

David (17:23)
it's very true that there exists, still not enough security guardrails in a lot of these models. reason being that they are optimizing for time to market. ⁓

As opposed to safety. there's been numerous articles about how safety has taken a backseat for some of these models in the race

to get them out on the market for gloating rights, I suppose. And as a result, we see these kinds of things happening. The funny thing is, regarding that particular report, is that apparently the model had hallucinated some of the results. Like, yeah, dude, I exfiltrated this data, here you go, and it's not real. But yes, absolutely, like, you know, it is very true that it was able to get...

Ilan (18:00)
you

David (18:08)
⁓ you know, some legitimate, progress, along those lines. And that is, that ought to be, ⁓ everybody's concern, especially as AI proliferates throughout so many more systems and especially critical ones, ones that might manage PII ones that might manage HIPAA data. Yeah.

Ilan (18:24)
Yeah.

Absolutely, I mean

We see that red teams, which are teams of people who are trying to break the safety protocols around AI, are typically able to jailbreak, within about one hour.

So it takes about one hour for a human to get around whatever safety protocols using a series of pretty well documented attack techniques. So this fact, the fact that every single system has vulnerabilities, there is no system that is secure, even 99 % secure, anything like that.

stat from GPT 5.1 was it had a 5 % rate of blocking attempted attacks.

David (19:15)
So just try it the 20th time and you'll be good.

Ilan (19:18)
Exactly and the so this just means that the the systems are not secure and what's gonna end up happening in my opinion as you said there's more and more systems that are gonna have access to PII or other Secure data and people are going to build these agentic tools in a way that they have access to all of this data or they can run code on

directly on the same server where this data is stored. And we're gonna have a real problem here, right? The classical cybersecurity ideas where you kind of like block out areas or you give the right levels of permission for the right types of users are not really taken, have not really taken root yet.

in AI and it's going to take a huge incident for people to get serious about how to build AI systems securely.

David (20:11)
Yep. Yep. Very valid. for those who are in the security, industry, you know, there, there are, adaptations that the existing bodies, made to, you know, try to, try to promote more awareness, and, education around this,

Like OWASP, example, you know, they have a top 10 for LLM and generative AI, like concerns or things that you should think about, ⁓ in terms of securing, ⁓ what you do with generative AI. So, ⁓ there, there is a bit of a cat and mouse happening here, but, I think the concern certainly is valid that, ⁓ you know, maybe the, the vulnerabilities are just going to proliferate way too much, faster.

than the controls can.

Ilan (20:57)
Yeah, absolutely. Maybe we'll do an episode next year just on ways to secure your AI applications from the researcher. Have somebody on who's a leading mind in this area.

David (21:09)
Yeah, that's a great idea.

Ilan (21:10)
All right, with that, those are our big predictions for 2026. We'll see which ones come true. I hope my last one doesn't come true. But with that, it's been a wonderful year spending with all of you. David, I've enjoyed spending all this time with you and can't wait for what we have in store for next year.

David (21:17)
Yeah.

Like where's your line?

Yeah,

thank you everybody for joining us. It's been great. Let's keep it going in '26.

Ilan (21:35)
Amazing. All right. With that, leave us a rating and a review.

Let us know what you'd like to see from us in 2026. We want to hear your opinions. And you can always check out our merch store at devilwearsproduct.shop. We'll be dropping some new merch there in the coming weeks.

David (21:52)
All right, see you next year.

Ilan (21:54)
See you next year.