AI year in review: Trends shaping 2026

It feels like you were wrong on your prediction. Do you agree [music] with

prediction. Do you agree [music] with that?

>> User error, Tim. Hey, I don't know where I [laughter] don't know where you buy your dogs from, but my dog is barking fine. All that [snorts] and more on

fine. All that [snorts] and more on today's Mixture of Experts.

I'm Tim Hang and welcome to Mixture of Experts. Each week, MOE brings together

Experts. Each week, MOE brings together a panel of the smartest minds in technology to distill down what's important in the crazy world of artificial intelligence. [music] We've

artificial intelligence. [music] We've got a tradition here ate where at the end of the year, we record short clips with all of our experts talking a little bit about what happened in the year that's ending uh and where we're going

to be going into next year. So, without

further ado, here is the holiday clip episode for Mixture of Experts.

Chris, thanks for joining us for our end of year episode. So, this is a little bit of a kind of clip episode that we do at the end of the season. We're talking

with all of our experts about the predictions they made last year and whether or not they were right and where we're going into 2026. And you know, I was just catching up on some of your

predictions and you were a big agent booster at the end of 2024. You're like

2025 is going to be the year of the agents. It's going to be the biggest

agents. It's going to be the biggest thing, best things sliced bread. But I

gotta say, I feel like we're now, you know November December uh 2025 and it seems like agents are like the dog that didn't bark this year, right? Like

I'm not I'm not using agents to go book travel. I'm not using them to, you know,

travel. I'm not using them to, you know, I'm just not using them in the way that was promised. And so I guess I want to

was promised. And so I guess I want to say like it feels like you were wrong on your prediction. Do you agree with that?

your prediction. Do you agree with that?

>> User error, Tim. I [laughter] don't know. I don't know where you buy your

know. I don't know where you buy your dogs from, but my dog is barking fine.

So, >> all right. Agree. Well, tell me about it. I mean,

it. I mean, >> yeah, >> I I think what I said, Tim, I went even further. I think I said 2025 is going to

further. I think I said 2025 is going to be the year of the super agent.

>> Agents overhyped in 2024 or underhyped in 2024.

>> Underhyped, not hyped enough. Agents are

the world. Agents are everything. And in

2025, we're going to have super agents.

That's what's coming in 25.

>> [snorts] >> I don't even know what that means, you know.

>> I knew you didn't at the time.

>> And when you [snorts] say super agent, what do you mean exactly?

>> I just made up the term, Tim. So,

[laughter] >> you heard it here first on.

>> And I think I said at the time it was just going to be like we're going to have, you know, we combine reasoning and we combine tools and then the agent is going to be able to do multiple things.

And and and and I think that's true actually. If if if I really think about

actually. If if if I really think about where we were in 24 agents were very small and specialized and you would have an individual thing. This is the you know the writing the email writing

agent. this is the uh the research agent

agent. this is the uh the research agent whatever right and they they did one specialized thing and of course we moved into this multi- aent world but actually

reasoning has came into the models and now what we're seeing is with reasoning with um test compute the the models are able to think much longer and they're able to plan and they're able to use

much more tools and actually say okay I've got this tool and I've got this tool I'm going to call this to be able to achieve uh that goal so I I think I think I I'm I am defending my calls but I think I was right right which is

reasoning was going to make a big impact there and then from a super agent point of view I do think we've moved away um from specialized agents I I I think

there's still a place for them but but if you think of things like Manis which we talked about earlier in the year where I submit one job and then Manis goes off and then you know about 30

minutes later it's created my entire presentation a multitude of that that really is kind of the super

agent that I was talking about and I do think that is going to continue um and I think it has taken off. We've seen

Manises, we've seen things like um you know Claude Code come out and you know which we're using from a coding assistant perspective. Um, we even think

assistant perspective. Um, we even think about chat GPT itself and Claude, you're able to put your connectors in and then Claude is going to be able to work with the different connectors and and go talk

to different systems. ChatGpt will go off it, you know, will create your PowerPoints, it will create your code, etc. on the canvas. So, so I think chat GPT, I think Claude and I think Gemini,

they're all evolving themselves into these super agents. Um and and I think that is that is the space where everybody wants to go because as much as

we may want these specialized agents, the reality is, you know, the major frontier AI players are are going to want you to come to their one agent. And I think it's I

think it's too hard. I mean, I I think about the marketplace of agents and I think it's too hard to ask a user, oh, you need to write a PowerPoint deck.

Well, you need to add in the PowerPoint agent. Oh, no. No. Now you want to write

agent. Oh, no. No. Now you want to write Word. No, you need to add in the Word

Word. No, you need to add in the Word agent. Oh. Oh. Oh, you're doing

agent. Oh. Oh. Oh, you're doing research. No, you need to put the

research. No, you need to put the research. It's uh it's too much. It's

research. It's uh it's too much. It's

too much agents, too much tools. Right.

What what you want to do is just speak to that orchestrator and say, "Hey, I want you to do this task. You go look up the tools that you need to use." And

then if you've got some buddy friends who are agents as well, you you go and deal with them and come back to me with the answer. Um, so I I I think that I

the answer. Um, so I I I think that I think that has happened. It's maybe not the way that we want want it. I think

there's much further to go, but I but I do think that that has happened and and that is the fight because whoever is that front door to the super agent is the one that's going to control the

market as much as possible.

>> That that is a good point, right? Like I

mean, if you're using reasoning and planning as kind of one of the distinctive agentic features, if you will. It's true. I mean, I use like deep

will. It's true. I mean, I use like deep research all the time, right? That's

like that's used agents. It's almost

like the old adage, you know, like the adage of like, oh, if it works, it's not AI anymore. It's almost like if the if

AI anymore. It's almost like if the if the agent works, it just becomes AI, you know? I don't even think about it as an

know? I don't even think about it as an agent anymore. Uh, it just like

agent anymore. Uh, it just like disappears. It's just AI basically. Um,

disappears. It's just AI basically. Um,

>> but but do you remember that? Do you

remember that though, Tim? I mean

because if we go back a year to like November last year, I don't what I don't I think even things like web search within chat GPT was experimental at that point. We certainly didn't have Google

point. We certainly didn't have Google AI overviews and stuff like that. So so

agents has just been brought into our workflow and just part of our consumer products but we don't think about it in that way to your point anymore. We're

just using the tools, right? You

remember chat GPT when you did a search and you know without agent tools, you know who is Chris Haye of IBM? I'm

frequently without web search I'm the chief finance officer for IBM right you know which is you know you wouldn't want to put me in touch of IBM's money we'd

be broke so um you know so I I I I do think it has happened we're just we we just don't think about it in that way >> yeah that's right and I think the other point that you raised which I think is

really interesting is this kind of coming battle between who's going to be the like front door on these agents because I think the joke that I always had with agents is like, yeah, if you have an agent for PowerPoint, an agent for writing, and an agent for research,

it's like agents, do you mean apps? Like

just like normal software on your computer, like it starts to resemble basically like what we had before. The

thing that does really make it genuinely feel new is if there is this kind of orchestrator agent, which is just like the front door for all of these requests. And yeah, I can see obviously

requests. And yeah, I can see obviously there's a big coming battle between all of the big, you know, kind of frontier AI companies. uh on this front I I guess

AI companies. uh on this front I I guess kind of question for you is like if you feel like any one is more well positioned than the other to kind of like you know win in this space or if it really will be kind of just like everybody wins you know

>> I I hope it's everybody wins. I just

think that if you're you know JT GPT is obviously fighting for that space especially with things like commerce and deep research etc. Google's fighting for that space for search so they're all

going to be fighting against each other um for that. Claude is really their face they're fighting for the enterprise and they're fighting for the kind of the coding side of things. So I think everybody's fighting against each other

in that sense. Um I imagine Amazon's going to do the same for shopping and stuff like that as well because Shopify will release their thing. So I I I I

think that the the space should be open.

I hope it will be open but you know who knows. Um but we are we are seeing that

knows. Um but we are we are seeing that channel become important because again the other thing that we've seen this year is the agentic browser. Yeah. So

we've seen Perplexity release the comet browser. Touch EPT's release their

browser. Touch EPT's release their browser. Um you know I'm quite sure

browser. Um you know I'm quite sure Google's put Gemini in the in Chrome if they haven't already. Um everybody's

fighting after that browser space. And

then you're seeing the same thing on the mobile app as well right with you know Apple extending Siri at some point whenever that will be. Um um so I think

you know and then Gemini get embedded into Android. So I I think that digital

into Android. So I I think that digital you know front point and then all the voice models I mean you know um chat GPT for example. So I I I think the fight

for example. So I I I think the fight for the browser fight for the mobile again is part of that kind of front door super agent space as well. So we we are seeing that this year. Again, there's

much of a muchness between all of the providers if if we're truly honest. Um

if somebody does something truly breakthrough, you know, where it's just different level from any anybody else, then then I think then I think that's going to get super interesting. But I'm

not sure that's going to happen next year, though. I mean I think it's kind

year, though. I mean I think it's kind of yeah that that's that definitely opens up the sort of front door discussion in a way that I hadn't thought about is like you know I was very much thinking about like the chat window right for you know but partially

part of this battle is like what existing software paradigm or platform can you bolt the agent onto as a way of bootstrapping this right they're like oh you you know browsers well now there's

agents in it or like you know Amazon from buying things like now there's agents in it and it's kind of like everybody's always trying to figure out like if that's the best way to get agentic uptake is like take an existing paradigm that people understand and then

just like make it better with the agent as a way of kind of strapping this up rather than having oh you know there's this new thing called the chat window and now we can do agentic things basically so >> and and and and I I think you're right

there I mean we are seeing that already is like all these front doors you put something in there so that's what's happened in the code editors right every single coding editor now has got an agent in in there I think what Google

did is recently is quite is quite interesting with Andy gravity, right?

Because although you you know it's obviously a foro VS code and you know it's got its own issues etc and it's agents in the editor but what they've done that's quite interesting is they

also have the ability to run the agents outside of that as well. So you can run your agents in the terminal and they've got the inbox etc. So you'll be able to control your agents from the outside and and again that's kind of back to that

manis style um you know here's my task workflows and then it's going to run in the background. So if I if I start to

the background. So if I if I start to think about 26, see what I did here, we moved from 25 to 26. I I think if I

start to think of 26, I think being able to control your agents um centrally um

either anti-gravity or or um or manis style, but have those same agents be able to operate in different environment. I can see that being a

environment. I can see that being a thing, right? Um, so, um, and they're

thing, right? Um, so, um, and they're all going to start fighting across that space. So, I I I think that sort of

space. So, I I I think that sort of agent control plane type thing and multi- aent dashboard, I I I I can see that happening and and we go and like we would look at our inbox, go kick off

this agent to do this, go kick off that agent to do that. I I see I see that being a thing.

>> Yeah, for sure. I'm going to return to that like a lot more in 26. uh is like the kind of agent control dashboard. How

that looks like, how do we go about doing it is like a really interesting area of development I think in the next 12 months. Um well Chris, that's all the

12 months. Um well Chris, that's all the time that we have. Thanks for joining us. Uh appreciate it and uh we'll see

us. Uh appreciate it and uh we'll see how these predictions turn out in uh in 26.

>> Fun fun.

[music] >> Gabe, thanks for joining us for our year-end episode. Uh we obviously we

year-end episode. Uh we obviously we want to have you on because in some ways everybody was saying that 2025 was the year of agents and it kind of was but in my mind it was really the year of open

source right we saw these models just get better and better and better and like the gap between proprietary so-called state-of-the-art and where open source is just has continued to narrow and narrow and narrow and

arguably with Kimmy K2 thinking we're now in a world where even open source is now like ahead of where these kind of proprietary models are. And so I kind of want to start with like a note of kind

of um sober assessment. Uh it's very easy to be like open source has won, you know, open source is the best and now has conclusively destroyed, you know, anthropic and open AI on its way to

world conquest. Um, but it strikes me

world conquest. Um, but it strikes me that there's still we should have a little bit of hesitance and talk a little bit about like what remains to be done. Like what are still the frictions

done. Like what are still the frictions that are getting open source preventing open source uh from being as big of a deal as we we hope it will be.

>> I love hearing that 2025 is the year of open source. Um, and you know, when I

open source. Um, and you know, when I hear open source in the AI context, um, my thought process is always to try and figure out what the person using that

phrase actually means by open source. I

think AI for better and worse is still being talked about primarily from the context of models and models being the kernel that makes it all tick. Uh, also

suck up a lot of the oxygen in the discussions. Um to your question about

discussions. Um to your question about the sober reality of it, open source uh has maybe a branding problem in this uh

branding is maybe not the right word, a packaging problem. So if if you go to

packaging problem. So if if you go to any of the uh package solutions out there that are available for purchase or free trial, you're interacting with a whole stack of software. And I think I

mentioned this on an earlier episode of uh Mixture of Experts this year, but um the real things that are driving

user success and joy are actually in my opinion mostly at the software and user experience layer and the quality of the model is enabling those. So the models

need to keep coming along, but what you're really getting joy out of is the things that the application is doing with that model. Um, you know, the the

example for me in my not professional life is my my son loves drawing his own Pokemon cards and the fact that in Gemini I can go take a snapshot of a

sketch he did. he's seven and a half in second grade and it can turn that sketch into a beautifully rendered Pokemon card for him is so cool. And uh you know it's

getting my second grader hyped about using AI uh that has very I mean there's a great model behind that. That's the

nano banana model. However, the UX of being able to do this all from my phone while you know waking up on a Saturday morning um and presumably with a whole bunch of guardrails in the middle and

actual software that's making sure that that Nano Banana model is running the right number of generation loops and etc etc. Um it's a whole system at play there. So bringing this back around to

there. So bringing this back around to your question um opensource has all the bits and pieces.

It has probably better bits and pieces in almost every single category of the logical architecture of these systems, but there is not yet a de facto standard

way or even a collection of deacto standards ways to pull them all together in 5 minutes to something that is going to achieve that same level of user

experience and joy that you're going to get with a closed source solution. Um so

you know my view of what's been happening over this year uh if you call it the year of open source is that those patterns and those slots in the logical architecture have been getting refined

right uh you know we used to think about an agent at the beginning of the year as just like a bunch of code that uses an LLM. they have a much crisper view of

LLM. they have a much crisper view of what an agent is today and what a tool is today and how they work together um and where uh the you know the slots are

around that tool that need additional you know code. So that it's a complicated software architecture to put together a a gentic AI system and uh a lot of people have started to try to

take stabs at how do you actually instantiate that with pure open source.

Um and it's just it's not quite there yet and settled like it's not a done deal yet. So I think uh there's still a

deal yet. So I think uh there's still a bunch of catching up to do in terms of just the the finalizing of this is this is what it means to build an agentic AI system with open source parts. this is

how it looks in closed source parts. Um

they're logically the same thing and you too can have this on your phone running in a matter of minutes, right? We're not

quite there yet.

>> And I think that orchestration bit is uh is really interesting because I guess it's also been in some ways kind of like a traditional open source weakness in some sense, right? Which is basically

that like lots of people in open source don't like to have to be forced into some kind of architecture. They want to kind of take all the components and piece it together themselves.

>> Absolutely.

>> Um, yeah. And also the other one is all the other all the various bits and pieces are designed by different communities with different types of standards. So it's also really hard to

standards. So it's also really hard to build that as well.

>> Yeah. As an open source tinkerer, this is super fun, right? I I enjoy that stuff, but that's also not going to that that's a me problem. That's not a the rest of my [laughter] family, let alone the rest of the world problem, right?

>> Yeah, that's right. And which has I think accounted for like open source, right? obviously being like very

right? obviously being like very successful in certain domains but but not others. Do you think that'll kind of

not others. Do you think that'll kind of be the case here as well where it turns out in 26, right, we're going to start to see like what are the kinds of, you know, alignment of stars that allow you

to actually have that kind of like common, you know, call it middleware or whatever, right? Like these slots where

whatever, right? Like these slots where all the bits and pieces can plug in. Um,

but like that will only really happen in some domains and it might not even happen in say consumer, right? or or do you feel like you're more of um an optimist like you feel like oh yeah we actually will be hitting on these across

the board across the use cases where AI might appear.

>> A couple of answers to that right so I think um from a consumer standpoint I suspect that we will see a high quality

open-source generalist agent of some variety emerge that is paired with open- source models and sort of a full open top bottom stack. I suspect that that will happen next year. Like I don't

think we're very far off from that. Um

and in fact I suspect that that will probably be something that you can almost pick and choose like the level of quality you want that fits on your device that you want to run this thing

on. Um but uh it still won't hit par

on. Um but uh it still won't hit par with the closed source labs just scale right like uh you know it's hard to run those big models even the open ones that

are extremely big. I couldn't run one of them. uh and I'd need a whole, you know,

them. uh and I'd need a whole, you know, cloud to run them on. So I might as if I'm already, you know, delegating out to a complicated piece of infrastructure for a model of that size, might as well

just use one that's already hosted for me. Um unless there's a security

me. Um unless there's a security sensitive nature that that has to hit there. But but I suspect we'll see it

there. But but I suspect we'll see it all come together. The other thing that I think we will see in the open is that

the developer patterns will refine, right? So we will we'll end up with um a

right? So we will we'll end up with um a much cleaner story when told to go build an agent that does X Y and Z and solves this problem. Um you know we already

this problem. Um you know we already have a handful of the most popular things that sit at that framework level.

Um but we don't necessarily have good instantiations of how to plug that together with the guardrails at the right points and the tools at the right points with the right security models with the right you know it's getting

there. Um I think that those things will

there. Um I think that those things will will shake out to you know a collection of de facto standards both like interface standards but also implementation standards for different

domains. Um so I think we're going to

domains. Um so I think we're going to see progress here. Um and I I I really think it's a it's a it's a two-sided world. Like I don't think one is going

world. Like I don't think one is going to win against the other. I know

hearkening back to you know operating system wars. Um, you still don't see

system wars. Um, you still don't see very many everyday users walking around running Linux on their laptops. Um, but

Linux is everywhere. It's running all of our servers. It's underpinning our

our servers. It's underpinning our Android devices. It's uh, you know,

Android devices. It's uh, you know, Unix. Uh, variants thereof are

Unix. Uh, variants thereof are underpinning our our MacBooks. Um, so I think many of these frontier labs are almost certainly building with open source components to build their

perfectly crafted agentic systems that they're exposing as models. Um, so I think it's going to follow that same trend where we have the bits and pieces out there in the open. We have really

solid solutions purely in the open. Uh,

and then we have better solutions in the closed that are people are making money off of. Um, and those things will

off of. Um, and those things will continue to evolve.

>> Yeah, I like that a lot. Well, cause for optimism. Uh, I was saying that uh I

optimism. Uh, I was saying that uh I think 2025 was the year of open source.

Maybe it will just turn out that 2026 is also the year of open source. Um, so,

uh, Gabe, thanks for joining us.

>> Always the year of open source.

>> It always Yeah, it's always been the year of open source.

>> Nothing exists without open source. It's

just a question of how you get it.

>> That's right. Um, but, uh, Gabe, thanks for joining us and appreciate, uh, your predictions.

>> Absolutely. Thanks.

>> Thank you, [music] >> Kar. Thanks for joining us for a few

>> Kar. Thanks for joining us for a few minutes of chat. You know, as we get to the end of the year fore, we're asking all of our guests to do some predictions. Maybe just to start, let's

predictions. Maybe just to start, let's do a little bit of a year in review. I

mean, from your perspective in the world of AI hardware, what were the big standout stories for you this year? Like

what really, you know, when you look back on the last 12 months, we'll say like, oh, you're going to remember 2025 for this.

>> Yeah, that's a great question. I think

there's a lot to unpack for 2025.

[laughter] >> I think 2025 was kind of the year where

demand outran supply chain. Um and uh so what we saw is basically 2025 kind of cemented what we already sense you know towards the end of 2024 AI hardware

scarcity became kind of a structural constraint here uh not kind of a temporary thing. So even with Nvidia's

temporary thing. So even with Nvidia's record shipments and the rise of any kind of GPU any GPU is better than no GPU uh compute you know demand you know

for LLMs multimodels video models aentic workflows you know really grow faster than what you know the fabrication capacity can do. So I think for the

first time here companies you know kind of be began optimizing kind of their entire business strategy around compute availability rather than model capability. So that's kind of become so

capability. So that's kind of become so important.

>> Yeah. Because we like literally ran out of compute basically.

>> Yes. Yes. And and I also felt that kind of the AI hardware kind of diverse in two worlds. There is the kind of the

two worlds. There is the kind of the scale up and the scale out. So um so when we talk about scale up we're talking about these super chips like the

H200 the B200 the GB200 you know these massive clusters uh cerebrass uh you know uh you know AWS tranium inferentia

Google action TPUs and then the scale out which is more edge local kind of where we saw things like deepsek optimizations quantization breakthrough

breakthroughs um things like 1 to three billion edge capable models kind of the rise also of these small LLMs and also heterogeneous compute for example with

the Apple neural engine the Qualcomm uh uh the NPUs so it seems like it was kind of a divide uh in LLMs uh frontier labs

prioritize you know these huge clusters but enterprises you know and nations they try to prioritize efficiency sovereignity and control >> yeah I mean I think that's going to be the big trend Where do you think we're

going to go in 2026? I mean, are we going to resolve this supply constraint?

What do things look like? You think when we're talking say December 2026?

>> Yeah, I think that's going to continue.

So, and I feel like 2026 will be kind of the year of frontier versus efficient model classes. We started even seeing

model classes. We started even seeing that with Kim K2. Um so because a lot of the models assume infinite compute and

so it it seems like I think we will see things kind of frontier kind of two 2.0 O models where we're talking about one

to five3 parameters everywhere long context multimodels etc still huge models but also the efficient uh models where we're talking about between maybe

half billion uh or you know 250 million parameters to 20 billion parameters.

They're hardware aware. They're sparse.

They're quantiz quantization friendly.

They're capable of running on modest accelerators. And why? because here we

accelerators. And why? because here we can't really keep scaling compute. So

the industry really must scale efficiency instead. That's becoming

efficiency instead. That's becoming really important direction and I think that's going to become even more like a big focus in 2026. Another thing I think

AI edge uh will will uh go from hype to reality. So, of course, we still have a

reality. So, of course, we still have a lot of, you know, things happening in the the edge, but running a frontier, we're talking about, you know, generative AI. Running a frontier model

generative AI. Running a frontier model on your phone is still a fantasy. But

running like a 10 billion model locally with maybe 4-bit quantization and using things like flash attention, uh, these optimized kernels, it is feasible.

So there's a lot of drivers behind this things like you know maybe Apple Qualcomm Samsung NPU you know changes at the level of the model architecture you know looking at you know hybrid

architectures like the the Mamba 2 with the transformers new architectures are emerging that are um kind of prioritizing efficiency constraints and edge constraints >> the other big question of course on

everybody's mind when you talk AI hardware is Nvidia [laughter] um you know I think we've seen a lot of kind of interesting developments over 2025 of people you know obviously still

dependent on using Nvidia platforms but increasingly also experimenting elsewhere right I mean between the announcements on anthropic and I think openai you know in the last you know few months even have all kind of indicated

that you know they're exploring what they can do with TPUs right they're exploring what they can do with AMD do you think 26 will still be a very kind of like Nvidia dominant year or do you feel like this is like the year where

maybe there's a little bit of an inflection point in the compet competition in that space.

>> I think of course 2026 won't scale. The

GPUs are still going to be you know important but it will challenge their dominance. Um I think we will see maybe

dominance. Um I think we will see maybe early ASIC based LLM accelerators in production increased adoption of chiplet based designs. Um maybe some first

based designs. Um maybe some first proofs of um proof of concepts of analog in memory accelerators um quantum assisted optimizers for model

compression and chip design. uh some

also real commercials use of hybrid digital analog inference. So GPUs I think is still going to be remain king but I think a new generations of challenges will continue to mature.

>> Well Katar we'll have to keep you we'll keep posted on this um and uh we'll be looking to see how these predictions play out over the next 12 months.

>> I hope so. Yeah, we'll see what holds and what doesn't. And I think another thing that might also be start playing is the rise of specialized chips for agentic workloads. Um because agents

agentic workloads. Um because agents change everything here. They're long

running. They require memory persistence. They operate in these

persistence. They operate in these multimodel streams. They need, you know, GPU, CPU, NPU, a lot of coordination. So

maybe a new class of chips, what we call agent processors can start to emerge here that are focused on these context persistence um low latency reasoning,

fast planning loops, multimodel token streaming, energy efficient scheduling um and this is you know where for example uh IBM style data flow architectures will really shine.

>> That's great. Well, we'll definitely have to keep you on the show to talk more about it and um thanks Katar for your predictions.

>> Thank you.

>> [music] >> Aaron Abe, thanks for joining us. Uh, so

you're here to talk a little bit about sort of the present and future of multimodal in these models. Um, and I guess Aaron, maybe the easiest way to is to start with you. You know, you've done

a bunch of multimodal work this year around sports is kind of my understanding. And I guess the question

understanding. And I guess the question is like just in your own work looking into 2026, what are you excited for? How

do you think the capabilities will open up? Are there things that you want to

up? Are there things that you want to experiment with?

>> Yeah, I mean it's it's incredibly important to have, you know, these types of generative, you know, models be multi-ensory. So they can interpret our

multi-ensory. So they can interpret our world and the way that we we interpret it and see signals that maybe we can't, right? And these multimodal models are

right? And these multimodal models are enabling agents to perform this type of contextual computing and maybe even effective computing so that they can better understand our emotions, how

we're feeling, what we're doing, right?

And um I think you know looking towards the future you know we're going towards these neuromorphic you know reasonings where it's like these systems that mimic the structure and function of how we are

right so so these multimotive models are going to be able to see hear and read everything right and then then these models will be able to perceive and act in a world much more like a human so

they'll be able to bridge language vision and action uh all together and I think you In the near future, we're going to start seeing these multimodal digital workers that can autonomously

complete, you know, these different tasks to interpret things, maybe even like complex health care cases, right, to to help that problem, right? Um, and

then another area that I think it's going to emerge is is multimodal models are going to start running on these um hybrid topologies, right? So, it's going to be like these combinations of

transformer state space models, right?

So, so, so in net it'll make them more efficient and performant at the same time. So, it's like the best of uh both

time. So, it's like the best of uh both worlds, but but I'm really excited just about the about the potential uh that that we have and to watch how some of today's models like Meta Lava, IBM's

Granite Vision, Nova, uh Dolly, you know, you know, so on and so forth, um how they're going to change, right, and and merge to make, you know, this multimodal uh experience better for us.

>> Yeah, absolutely. Um, and yeah, I think where kind of agents meet multimodal will be really interesting because it'll just really expand the kinds of problems they can work on. Abe, for your work with Granite, you know, it seems like

multimodal is becoming more of a standard thing that all kind of open source provide model creators are just building in by default. Um, you can do you think that's going to become more of a norm in 26 that essentially everybody

will just expect out of the box that there'll be some multimodal capability for these models?

>> Um, I mean in short, yes. But I think what multimodal will mean may shift a little bit in terms of just out of the box performance. Um, specifically for

box performance. Um, specifically for granite, I think what we're really trying to focus on is modular multimodal capabilities. So there's the idea behind

capabilities. So there's the idea behind being able to, let's say, take a um a number of different models or adapters

and using them as a um, you know, a network of models to be able to carry out a use case. So if you do have a you know a situation that requires you know

audio image as well as some workhorse LLM in place of having an omni model that supports each of them but maybe not best

of class you'd be able to tie them together as part of a um you know function or as part of a workflow where you can call each capability as needed.

So, one reducing your actual footprint given some of these adapters are quite lightweight, you know, the Laura or Aurora adapters, but also kind of capitalizing on a more state-of-the-art

model given you actually are using a kind of uh a focused capability as opposed to, you know, some of the more, you know, jack-of alltrade style models.

So, what we're really trying to focus on is modularity through um orchestration, if you will. um really doubling down on on being able to wrap software around

these use cases so that you know you can not only let's say use um you know multimodal capabilities but let's say you have specific you know certainty checks that you want to do with on

output and then you know maybe rewrite your prompt because it didn't align with the actual request as needed and then you know maybe you want to reference the summary as part of your pipeline. So I

think it's a little bit more about you know obviously multimodal is at the center of it or at least being able to enable multimodal but it's using software orchestration to be able to tie

a lot of these models together so you have more of like a world view of uh uh in place of of kind of like a static monolithic model.

>> Yeah. Yeah. So, so Abe um some some something uh that you mentioned u was Allora, right? And and being able to

Allora, right? And and being able to take in, you know, different functions and plug it in, you know, and and and it's almost like having different skills, right? And you and you um Yeah.

skills, right? And you and you um Yeah.

And and within sports, you know, lots of what I'm doing today is I'm putting in vision skills, you know, um I'm putting in, you know, different types of speech to text skills. Um, but I think it's also important in the future to have

this human in the loop AI so that the human can fine-tune and change the skill that's necessary to do the job at the right time to have the best kind of accuracy that we're looking for because

I have noticed and and today's models that you know they're they're quite big you know um to be able to for example take in a you know a fulllength video

you know which you know might be of you know one gig size but you got to chop it up feed it in the But there's a lot of hallucinations that can happen. Um, it might not even even

can happen. Um, it might not even even with prompt tuning, it might not be able to pick up the exact kind of contextual cues. So, example, in golf, it may not

cues. So, example, in golf, it may not be able to pick up, you know, where the ball lands precisely. Is it going to be in the bunker? Is it on the green second cut? Is it in the margin? Is it in the

cut? Is it in the margin? Is it in the trees? Right? Um, but it might just

trees? Right? Um, but it might just generically just describe what the landscape is and it's not focused around the particular action point that we're looking for. So having this human end in

looking for. So having this human end in and the loop and and and using these Allora type adapters that you can just plug in, I think will really help to change what the focus is on on the

attention levels, you know. So so so I do think that that's an area that's that we're going to see um to help us, you know, use these types of models better, right? And in the future to align it

right? And in the future to align it towards the use cases that we're trying to work on.

>> Yeah, for sure. Abraham, maybe a final question to you as we get into 2026. I

mean multimodal uh there's a lot of modals and so I'm kind of curious about if there's particular modals you think are uh modes rather that you think are more interesting obviously there's a lot of excitement about image but you know

what do you think about video audio of course is maybe less in the picture like what are the kind of multimodal sort of like uh capabilities that you're most excited about so yeah image of course but even within there I think we can

break it down where you know our focus is really enterprise images so um you know complex tables complex charts complex workflows. I think we're really

complex workflows. I think we're really starting to look into computer use. So,

being able to maybe extend control of the model to, you know, screen so that they can capture images and be able to kind of carry out activities. Audio is

also a huge component of the granite landscape. You know, our granite speech

landscape. You know, our granite speech model um is, you know, some one of the best performing models as far as as part of like audio speech recognition on, you know, our OCR bench. Um and then we kind

of double click into vision after that.

though you know KBP extraction so key value pair extraction which is very you know um top of- mind use case for a lot of enterprises that want to take documents and pull data and pump it into

a database so that they can use for downstream rag tasks. Um vision

embedding as well is really big for us for rag pipelines kind of flattening the curve or flattening the pipeline as part of you know rag use cases. Um so for for us I think from the multimodal

perspective we're less focused on um text in image out I think that's just not a space that IBM um really wants to uh play in given you know the hyperfocus

on you know enterprise use cases um but but yeah I think 2026 for us is going to be really focused on kind of one building out the ecosystem of our multimodal models and then really

understanding you know where is the best place to use these models and showcasing that as part of our recipes, um our temp our um sorry our notebooks as well as just providing guidance to customers and

consumers on what when to pick the best model and then building exactly what I mentioned earlier is like the orchestrational layer around like you know how do you start to use these models as part of uh more of a a system

as opposed to standalone and what are some of the cool things that you can start to do with these models when combining them or chaining them you know with other granite models or third party models >> well we'll keep an eye on it uh Aaron

Abe thanks for joining us and uh we'll see how some of these predictions play out in 26. Uh thanks for coming on the show. [snorts]

show. [snorts] >> Thanks for listening. We're excited to have you along for the ride this year [music] and join us next week on Mixture of Experts.