No Priors Ep. 144 | The 2026 AI Forecast with Sarah & Elad

00:00:05Hi listeners, welcome to No Priors. Howcan we even begin to wrap this year up?The AI field has grown, breaking outinto the mainstream and taking centerstage with policy makers. Chat GPTshipped massive numbers and asked formassive [music] dollars. Gemini andGoogle roared back strong. And on theapplication front, AI coding has shiftedto agents and is eating up all of ourinference capacity. Doctors are adoptingclinical decision support on mass and inlaw and customer support. Enterpriseadoption is accelerating. [music] What'snext? On the research front, the racehas multiple live players with opensource closing the gap too. [music] Ahandful of Neolabs, new research labsgot funded this year. And the narrativeis changing. Ilia is calling it the ageof research. People are trying differentideas around diffusion,self-improvement, data efficiency, EQ,large scale Asian collaboration,continual learning, energy transformers.It's more open than it's ever been.Finally, we had a lot of attempts tomake AI reach into the real world withrenewed optimism around robotics. Nextyear, those companies are going to startmaking contact with reality. From aprediction standpoint, personally, Ithink we're going to see somebody make alot of money. Hundreds of millions ofdollars trading markets with LLMs nextyear.

00:01:21It's inevitable. So, we're in thesecond or third inning. Markets arerunning a little hot and a littlevolatile. It's hot in the hot tub. So,get into it with me and Alad. Okay.Alad, it's been a year.>> I know. How's it going? 2026, baby.>> Are you feeling the AGI? Are you feelingAI AI winter in a good way?>> I think I'm actually just feelingmicroplastics. I think I'm now 80%microplastics and just increasing mymicroplastic consumptions. A friend ofmine actually launched a new water brandthat uh has no microplastics by the way.It's called Loop and they have likeglass bottles and also the cap doesn'thave plastic.>> Does it come with continual testing?>> Yeah, that's continual testing for you.>> They did actually try to take out allthe microplastics and so they uh I guessbottled water in actual bottles has moremicroplastics than plastic bottlesbecause of the cap.>> Okay, we'll check back in with you in 27to see if you feel>> Yeah. But I'm just completely oified outof plastic. I'm actually really worriedabout microlastics. What about all thelittle glass particles? Aren't youworried about that? People talk aboutmicrolastics, but not microlastics. I'mmuch more concerned about that.>> I don't think those particles end upembedded for you permanently.>> Silicon. [laughter]You're not worried about silicons. WhenI go to the beach, I'm like, "Oh no,microlastics everywhere." I'm actuallyvery willing to insert silicon

00:02:36in mybody eventually in my [laughter]>> Wow, that was Yeah, I'm not gonna sayanything. We can keep going.>> What's What's happening in AI? Al, whatare you where where are we and what areyou most excited about?>> Yeah, I guess for 26 there's a bunch ofstuff I think will be interesting that'scoming. I think we will um I thinkthere's probably four or five things.One is I think people will proclaim yetagain that AI is not doing much and it'soverhyped and like that MIT report thatpeople are quoting that I thought reallydidn't matter and the reality of thetechnology ways to take like 10 years topropagate and people are gettingenormous value out of AI already andthey're going to get way more out of itin the future. You know, so there'sthese undoubtedly next year there'll bethese overstated but bubble claims aswell as um hey I actually isn't workingthat well kind of claims and thathappens every technology cycle and we'lljust hear it again next year andthere'll be pundits and discussions andjust a bunch of waste of time on it. SoI think that'll happen. I think anotherprediction for 26 is the next set ofverticals will hit massive scale. Ithink this year we saw consolidation ofcoding into a handful of players, ofmedical scribing into a handful ofplayers, of legal into a handful ofplayers like Harvey and others. And so Ithink we'll see that next set ofconsolidated verticals happening. Sothat'll be interesting. I can keepgoing. I have like a bunch of these. Doyou want to go next? We can alternate. Ijust did two. Why don't you

00:03:51do two?>> Maybe I'll react.>> Or react.>> I'll react and then I'll and then I'llgive you two predictions. Um I have tothink of my predictions while I'mreacting. So I'm glad I have at leasttwo threads. Yes. I I think that theoverall sentiment on AI in the investinglandscape is a lot of people gettingstressed about the amount of capitalthey have at work and then just a levelof uncertainty around uh the adoptioncycle and technical bets that people aremaking that they don't have full firstprinciples confidence on coming toroost. So, uh I I think like any numberof exogenous factors plus noise about umthe speed of adoption, which by the wayseems like blinding overall and we cantalk about what the constraints are.>> Yeah. So fast I don't even know whatpeople are talking about. I I just saw areport um that talked about it's fromthis group called uh off call thattalked about adoption of AI by doctorsand look there is just amazing adoptionof of course you know several differentcategories like documentation clinicaldecision support with things like abridge and open evidence and obviouslythe general models but there's likemassive enthusiasm from most of thephysician profession here and I'm

00:05:06likeokay of of all of the domains that wereprofessional considered moreconservative. The fact that there isthis like you know desire to have thingsthat make work better seems likeobviously to continue in the otherprofessions. I I think this is by theway super underd discussed that thepeople who tended to be the slowestadopters of technology love AI. That'sphysicians, that's lawyers, that'scertain accounting types. It's, youknow, it's it's actually kind offascinating. It's compliance, you know,it's all the people who always neveradopt technology are now adopting thisstuff fast. So, I do think that's reallynotable and very under discussed.>> It will keep happening. There areactually lots of professions where likebeing able to reason and interact withunstructured data is very useful. Like Iexpect that there's going to be somelike negative market current. Like youknow if Nvidia doesn't overperform bysome massive amount one quarter,everybody's going to freak out. But I Ithink that has very little to do withthe fundamental secular change.>> Yeah. It has to do with microplasticsand Nvidia. It's my two cents.>> It has to it has to do with ummicrolastics as you said.>> Yeah, it's true. Actually, the siliconthere is in the air. I bet that theyhave microlastics all over the place.It's messed up. Sarah,>> it's part of the trade. If you make $20million as an average Nvidia employee,you also

00:06:21have to have microlastics inyour blood.>> Don't listen to this, Jensen. Jensen'sour next guest. You can't hear that.>> 1% microlastics in the blood.>> I think um you know, a third area is thenext set of foundation models are goingto come. And by that I don't mean theneolabs and the and the nextgen LLMswhich of course will happen but I meanuh physics materials um science progressby models math progress and I thinkwhat'll happen is there'll be one or twousea one or two cases where it worksreally well for something they'll inventsome new material or there'll be someconjecture proved or somethingand then it'll fall into this overstatedhype cycle of it's going to changeeverything about physical sciences orwhatever and that oneoff will beoverstated And in the long run, thetrend will be understated and will beincredibly important. So that's whatthat's another prediction for next yearis there'll be a couple anecdotalone-offs in science that will makepeople say, "Look, science is solved."And they'll realize science has beensolved and then later science will besolved.>> I have uh Okay, fine. Three three quickpredictions for you. One is there'sgoing to be like some collapse ofsentiment around a set of roboticscompanies next year. Not because it likeactually isn't as a field going toprogress but because you know people arebeginning to project

00:07:36timelines>> and uh you know not everybody is goingto deliver on those timelines.>> What's your timeline? I think that wewill see um humanoid and semihumanoidrobots get deployed at small scale inenvironments be the consumer orindustrial next year and not everythingwill work and that like the becausethere's this you know hype cycle aroundhuman rights overall as soon assomething doesn't perfectly work whichit will not people are going to freakout right and then there's going to besome bifurcation about people investing>> yeah I mean we're in year 15 17 whateverself-driving something around there andit's really working now but itseems like robotics should have maybe afaster curve but a similar curve rightit's going to take some time to figureall this stuff out and then once it'sfigured out it's going to be reallyvaluable and the big question for me onrobotics you know it's interesting ifyou look at self-driving there's liketwo dozen three dozen whateverlegitimate self-driving companies reallygood teams and good approaches and allthe rest and then arguably the twobiggest winners at least now are Whimoand Tesla which were two incumbentsright Whimo's Google Tesla is Tesla So Iwonder what will happen to robotics. Itfeels to me like Optimus or some form oflike Tesla robot will be one of thewinners most likely, right? Highprobability. And then the question isdoes

00:08:51Whimo just adopt what it's doingfor cars to robots as well? Becausethere's some similar problems there. Isit some other big industrial company? Isit startups? Like who are who are thewinners and why? And structurally whenyou have a lot of capital needs but alsoa lot of hardware and manufacturingneeds that's going to favor incumbentswhich is self-driving right um I guessarguably the other winners inself-driving are Chinese companies rightChinese car companies which are bannedfrom coming into the US market and thosewill probably also be winners inrobotics right the most likely globalwinners in robotics will be some subsetof China plus Tesla plus something elseright maybe maybe one of the startups>> I think that's right but that's like I Ithink in most industries like>> you know the incumbents are more likelyto win than the startups if you're justlooking at it like as as a numbers game.>> I don't know. Yeah, I don't know. Idon't think so. I think um I thinkthere's startup industries wherestartups should win and there'sincumbent industries where incumbentsshould win and they have differentcharacteristics in terms of marketstructure, in terms of capital needs, interms of certain of expertise and supplychain, you know. So I do think there aremarkets where incumbents shoulddefinitionally do better. They don'talways but they typically do. And then Ithink there are markets where startupswill do better.>> Sure. But I I don't I don't argue thatlike some markets are like the

00:10:07modes arestructurally deeper, right?>> But one way that you might look atautonomous vehicles is it's one verycomplex single use case robot.>> And it mostly does locomotion. It doeslots of other necessary types ofprediction, defensive drive, whateverelse. But it's it's it's a single usecase robot.>> Yeah. And we and we forget there's a lotof good ones like that. Dishwashers aregreat single-use robots. Vacuum cleanersare great. You know, like there's allthese things that we actually have thatare robots in the home that we pretendaren't, right? We forgot that they'rerobots. Elevators are robots.>> No, seriously. Escalators are robots.>> I'm going to use the language of likefor a robot [laughter] to be a robot, ithas to be somewhat intelligent, right?Um and so dishwasher doesn't count as anappliance. Um a self-driving car doescount as a robot, not just>> where's the border of intelligence foryou? I I think like it's probably somelevel of generalization, right? It canwork in different environments. It canwork on different tasks. It can work ondifferent objects. Otherwise,>> self-driving car is okay. Yeah, I don'tknow. I didn't have that complex of adefinition. I just had it as likesomething that will do>> certain pre-programmed types of laborfor you. But maybe that's maybe I have abetter definition. Let me look up whatdefinition of robot is. a machinecapable of carrying out a complex seriesof actions

00:11:22automatically,especially when programmable by acomputer. But you know, all these thingshave chips in them now. Your dishwasherhas a chip in it, right? Or the computerin it.>> Okay. Yes. But like uh I would arguethat robotics has not been aninteresting area of innovation withoutintelligence. And so that's the relevantset for maybe you and me and many peoplethat are looking for something thatchanges quickly.>> Yeah, that's cool. I mean, I do thinkthat um on the on the on the topic ofrobots, the biggest trend perhaps or oneof the biggest trends of 2026 100% willbe that self-driving will really beginto matter and that'll be both in termsof your own car, it'll be in terms ofWhimo and Tesla caps. It's going to be Ithink one of the big things that'stalked about next year. So, I think Ithink on the robotics team that's thebig E.>> I think if you um look at all of thepotential use cases for robots besidesself-driving and say like self-driving>> I mean the Optimus team actually provesthis like if you take if you take amodel that is powering Teslaself-driving and you put it in Optimusit can do locomotion but it can't domany other things and you still have todo the hardware right like manipulation.And so I think that the advantages hereare not as strong as you believe theyare. And like startups, some set ofstartups,>> the scariest competition is the Chinese,but I I do think that

00:12:37there isopportunity here.>> Oh, I totally think there's opportunityfor startups. And don't misinterpret me.I just think that it's not just the factthat you have a model or a base model.You have the expertise to build themodel, but then you also have all thesupply chain. And I think that's reallyimportant because a lot of the samesensors that you need to use are there.and you know how you think aboutactually procuring and scaling things isthere you know there's there's goodoverlap actually in terms of some of theother skill sets that are needed thattake a long time to build usually at astartup or that are a little bit painfulto build and people do it it's fine it'snot I mean did it and SpaceX did it andyou know all these companies have doneit it's extra stuff so that makes senseI I do think I do think some startupswill succeed here I just trying to thinkthrough you know besides the startupswho's going to be big and then also Ithink there are one or two likeincumbent slots that will just defaulthappen unless something very strangehappens and you know one could haveargued that should have happened infoundation models where Google shouldhave had a default slot in the end itdid right it got there and I think thatwas very predictable that the Googlemodels will get good I think I even mayhave wrote a post about this like twothree years ago that Google will berelevant right because they just had allthe assets that were needed for them tobe a really important foundation modelcompany they obviously inventedtransformers but they had all the datathey had all the capital they had TPUsand GPUs had like the best people forall sorts of things or

00:13:52some of thepeople. So, um, it felt inevitable and Ithink this feels the same to me. Thatdoesn't mean it's right. Do you want totalk about IP as an M&A next year? Whatdo you think will happen there? I thinkthat's another big that's theme numberfour, five, I guess. You know, three wasdifferent types of models, four wasrobots and self-driving, and then fivewould be IPOs and M&A. What do youthink? More IPOs, less IPOs, more M&A,less M&A, different types of M&A?>> It depends on whether or not the bottomof falls out of the AI market at somepoint, right? But I think regardless,>> what do you mean by the what do you meanthe bottom falls out? Like what whatwhat does that translate into?>> Uh I think people just get skittishabout you you know the cycle here islike what are people scared of? They areconcerned that demand isn't real. nodemand isn't real um for AI to supportthe capex cycle that there is systemicrisk from people passing the ball aroundin terms of who is actually responsiblefor the capex buildout and these creditagreements right or um you know pay ondelivery contracts for data centers andfor chips what else are they afraid ofthey're afraid of like the>> microlastics aka like too muchconcentrationin Nvidia

00:15:07and a small number of otherplayers. If you're like a big publicmarkets investor, you're just like, youknow, you>> silicon, it's too much silicon.>> It's too much silicon. You're damned ifyou do, you're damned if you don't. Iwas talking to a friend of mine who runsa large tech hedge fund>> and they're already like a foundationmodel investor in like multiplesignificant labs that may or may not gopublic in the next couple years. Yeah.And they're like, "Okay, well thequestion is, do you buy the IPO?" Theirgame theory on it was like, "Actually,no matter what I think about it, I haveto do it because retail will want it>> because they like want to be part of theAI revolution." And then if you're ahedge fun, you get benchmarked on annualperformance and because of the retailpop and some set of investors wanting tobuy into it as a pure play where you'relike, "Oh, I can't miss it like I missedNvidia." Then you have to buy it. And sohis view was like you buy the IPOregardless of your fundamental view ofthe company. And I was like, "Wow, thisis not the investing job I know how todo.">> What do you think happens?>> I think there'll definitely be a lotmore IPOs next year. Um, I think if oneof the main AI companies goes out, it'llbe probably do extremely well dependingwhere they price. I mean, they obviouslyif they're overly aggressive, it won't,but in general, I think there's so muchretail appetite to actually participatein AI besides

00:16:22Nvidia. Um, and thenthat'll just get a lot of other peopleto go public to just followers on it.So, I I do expect there'll be a lot ofthem. It's just one that even goes out.Uh, and then also it's a great way toraise huge amounts of money for some ofthese labs potentially. So, um, it'll beinteresting to watch what happens there.Any other predictions for 26? Yeah, I Iuh I think that I did not believe thatwe were going to see that many likeunique consumer experiences>> besides like chat GPT. I think we aregoing to see like a slate of consumerhardware that mostly fails, but I'mstill openminded to it. And thendefinitely actually like it remains tome see if any of these scales, but I amseeing magical experiences of likereally different consumer agent softwarethat I like I actually want and willuse. And I I think people are barelybeginning to>> well I these companies are in stealthright now, but I I do think that likethere's going to be a lot more productpeople that experiment with this andmodel companies that experiment withthis next year. Um and so I'm I'm prettyoptimistic about that. Yeah, I agreewith that 100%. And I think um the bigquestion is what will end up being abreakout startup and it'll undoubtedlybe some and then what will be a startupthat

00:17:37will grow really fast and thenit'll get cop copied by the mainlab/google and then it just getsincorporated into the core product. Andthe the interesting thing is unless acompany truly hits escape velocity andbuild a network effect or something elsethat's really defensible, usuallyincumbents can launch two three yearslater and catch up. And so if they havethe distribution and they have the coreproduct and they have but you know toyour point I think it's very excitingand I've been waiting for this for awhile. I think two years ago, threeyears ago, um this guy David Song whowas on my team at the time ran a twoquarter thing at Stanford where we haddifferent game supply uh from theengineering programs there and it waslike groups of people building consumerapps using AI because we said this waveof AI is so fascinating why didn'tanybody building anything consumer so webasically just gave people free GPU togo and try stuff and there was no likeobligation on their side to do anythingwith it you you know, in terms of usgetting involved. It was just you go docool stuff cuz this is such a goodplayground and it was really neatexperiences that were being prototypedand then I was just shocked that nothinghappened for a couple years in terms ofyou know really interesting consumerproducts. So I agree with you there's somuch room for that and I always wonderis it because there's a differentgeneration of founders who don't want towork on consumer or who've forgotten

00:18:53howbecause you know the big consumercompanies have kind of aged out. Is itthe incumbents are just too scary? Is itlike what why is there so littleinnovation actually on the consumer sideof AI? I still don't quite understandwhat the issue is.>> I Okay, let's let's like list thereasons. I do think that the incumbentsare pretty scary. Um and anybody who wasaround for the last generation ofinteresting consumer ideas saw actuallythe ingestion of those ideas into theexisting platform as you put out.>> Yeah.>> So there's that. I also think like thefirst instinct that that I've seen fromcompanies uh from founders working onlike new consumer experiences isessentially building like betterversions of like last generationexperiences with this generationtechnology and it ends up like not beingthat interesting. And so I actuallythink you have to be like either quiteclose to research or pretty creativelyambitious to build like something verydifferent that has any chance. And so[clears throat] I think I think likethere's just not that many people whohave had that experience set or thatcreativity and now we're going to seeit.>> Yeah, I think it's pretty exciting. Theother thing is um I was talking to areally well-known consumer founder who'srunning you know a giant public companyand his view is

00:20:08that perhaps in theentire world there's a few hundred greatproduct people for consumer at least interms of who are actually working on it.Obviously there's enormous humanpotential and people who aren't workingin consumer products could and you knowbut of the people working consumerproducts he thinks at most there's a fewhundred people who are exceptional whocould actually come up with and launchtheir own product that would beinteresting or good and so you couldalso just say say that maybe there'sjust a limitation on how many of thesethings can exist just given humanpotential within the set of people whoare already doing it which I think iskind of an interesting argument I don'tknow if I agree with it but I thought itwas an interesting argument that he made>> I would limit myself to that number ifit it's also the set people who likehave the context of like what ispossible now.>> If you've got great consumer productinstinct, but you're like work you'relike grinding away on the like 50thiteration of an existing product like>> Yeah. Yeah. You're working on the thethe little sub button in Gmail orwhatever instead of actually going offand doing this 100%.>> Yeah.>> Cool. Anything else we should talk aboutor any other big predictions for 26? Ifeel like a very big um emergent thingthat happened this year was thesurprising funding of like Neolabs likethree through eight. What do you thinkof that?

00:21:23What do you think aboutalternative architectures? Like do youhave any point of view on um all of theeffort around like getting reinforcementlearning to be more general continuallearning? Uh some of the researchdirections>> you know I think there's enormousamounts of really interesting researchbeing done. So I, you know, there's alot of juice to be squeezed out of thesemodels still in different ways and Ithink that's really exciting. Well,ultimately these things become capitalgains for certain types of approaches ormodels because we know scale reallymatters which means that eventually youhave to have collapse into a handful ofplayers because capital will aggregateto things that are working the most.They're generating revenue and so thenthe question is what are those things?At what point do things just get kind oflocked in from a usage perspective forwhatever reason? And there's all sortsof ways you can imagine this being builtover time against some of the models. SoI think it's interesting. I think it'sexciting. I think we'll see how it playsout.>> I think to articulate what like the thearguments could be for, you know, newresearch directions is like Ilia, youknow, did this interview [clears throat]recently where he describes it as theage of research. And to to paraphrase,he like basically says that yes, Ibelieve in scaling of course, but youknow, there's there's somefloor of compute that

00:22:38is not infinitewhere we can test ideas at scale. Andthen if we have [clears throat] let'ssay secret ideas around like how to getto more rapid or more compute efficientimprovement then it actually isn't justa straight resource battle which likethe rat race does feel a little bit liketoday. Um, I think the other argumentyou you could take is actually likemultiple architectures and people havedone some research on this, but multiplearchitectures are really relevant at bigdomains of of um usefulness. They justhaven't been scaled, right? And likethere's enough capital out there to testthem, be they like diffusion or um SSMsor whatever. And that's going to happenthis next year. And then I think there'slike a like a resource focus argument,right? If Ilia is describing that someset of labs they have an enormous amountof compute but they have to spend a lotof that compute on inference today thenhow much do you spend on your particularresearch direction uh be itself-improvement or post- training oremotional intelligence or very largescale out agent stuff.>> Yeah, it depends on what you're doingbecause the inference is what ends upthen uh raising you money to pay foreverything else because you'regenerating revenue. So I think

00:23:53uh surethat it's effectively your way tobootstrap into more and more scales. So,I always thought perhaps incorrectly. II actually probably think it'sincorrect, but I always thought thateventually you end up with evolutionarysystems is really how you build AIbecause and maybe I'm overextulating upa biology where you know effectivelyyour brain has a series of modules thathave different functions or tasks,right? You have a visual system that'sum you know highly sort of pre-wired todeal with vision really effectively. Youhave uh different areas of high pierthought and learning. You have memory.You have uh mirror neurons that areinvolved with empathy, right? Your brainis actually very um specialized in someways. Although obviously there's peoplewho are born with literally like half abrain hemisphere and the brain rewiresand sort of covers all thefunctionality. But um there's a fewfamous cases like that. Uh but you knowfundamentally um you have a lot of stuffthat evolves into very specializedtasks. It's almost like ae or something,you know. And the question is the degreeto which you recapitulate that as you'redoing further development of AI. Andwhen do you start just spawning off abunch of instances of something and justhave some utility function evolvingagainst that you then have someselection and

00:25:08recombining and all theother stuff that you kind of do to totry and make some of that work versushow much of it is a more analyticalapproach or a more experimental anditerative approach or you know so it'sor in a directed way. And so I thinkit's really interesting to ask cuz ifyou look again at biology as a as apotential precedent although maybe avery bad one. You look at protein designand for a long time there are these likesuper analytically designed proteins andthen they came up with all these systemsof this you know like phase displayand like mutagenic scans and all sortsof things that give you dramaticallybetter results than if you just sat andthought about it. And now of course wekind of solved it with AI where you haveum all these 3D structural predictionthat are actually very good right thatthat was um alpha fold and a few otherthings that really were breakthroughsthere. So it feels like in the contextof AI maybe eventually we end up thereas well right where you just involvethese systems and then that may be avery different type of approach andtraining and you know that that that maybe where I think things really have ainteresting break and that's one of thereasons arguably people are so focusedon code because code is arguably abootstrap into moving faster ondevelopment of AGI but I think it's kindof code plus self-evolution is reallythe the potential

00:26:24really interestingapproach to it to to get some reallyfast lift off but Maybe not, right?We'll see.>> What is um the one prediction you havefor 26 that has nothing to do with AI?>> Do you think about anything else, Sarah?[laughter]>> I do.>> I'm joking.>> Really?>> I mean, the other thing, by the way, oneother prediction that does have to dowith AI is I do think um defense willaccelerate in terms of startups anddefense tech and the shift to autonomousor not autonomous but to drone basedsystems in general. a massive reworkingof how you think about war and defenseand I think that's going to be a shootshift that we'll see go even faster thiscoming year I think this is acceleratingin part to you know how the Trumpadministration has been approaching itand the secretary of war and everybodythere have been thinking about it but Ithink in part just you have enoughdensity now of startups doinginteresting things so I think that's theother thing that's like a huge shiftthat you know it's a hype cycle rightnow and I actually think again it's alittle bit under thought about becauseit's it's going to be so big um outsideof AI I mean I think there's obviousreally interesting things happening inspace SpaceX and Starlink and I thinkabout communications and telefan that'sa big shift. There's really interestingthings in my opinion happening in energyand mining and you know I I thinkthere's a lot going on in the world.>> I agree on defense

00:27:40with some like concern that you know wehave to wait for budget to actuallyshift from contracts to primes to someof these new companies at scale. But thedemand like the need to be competitivein a world that's increasinglyautonomydrivenum is like so obvious right and I thinkyou know hype cycles and booms are goodin that they bring a lot of people tothe table you know capital>> founders people who want to work in theindustry um and so you can make a lot ofprogress in a quick amount of time evenif a lot of companies die>> and there's there's um more enthusiasm avery short period of time so I agreewith that. And I also don't think that'snecessarily bad, right? I>> What's your high prediction?>> I think that like I'm not the only one,but I think that the like[clears throat] GLP1 thing is just>> despite all of the enthusiasm, likestill underrated for how much impact itis having, right? And so I think thatthe continual adoption of these is likeinexurable. I actually think it createsa path that is interesting for likeother peptide and hormone therapies.>> I think the fact that it has been soeffective has like lots of

00:28:55second ordereffects both from people way like justbeing a lot less overweight likedirectly and the willingness to look atother engineered peptides or like Ithink it like everybody understands nowthat like>> delivery matters. there are these reallyincredible medicines and I think thatthe impact of that is going to like fuelmuch more investment in um anything thatlooks like that type of opportunity andso I think that's exciting.Yeah, I actually think um one thing thatyou mentioned is really interestingwhere if you look at the sort ofbiohacking community, there's a lot ofpeptide use now of different you knowdifferent peptides that will dodifferent things in terms of you knowsomebody will have some chronic corporalcheerle thing and they'll fly to Dubaito get you know peptides injected orwhatever and usually those are sort ofearly indicators of potential largerscale adoption society>> and so I think that's a reallyinteresting trend right now in generallike this whole like um world ofpeptides and their uses. and is there ahymns of peptides like what's the what'scoming there so I think that's superinteresting you know>> I also think like the biohackingcommunity as you said it like the set ofpeople who were really really early offlabel GLP-1 adopters um interested in

00:30:11longevity neurom modulation withultrasound um stem cell injection forexample like that has been like a fringesmall community>> and I think that like I think it's goingto get less French.>> Uh and a lot of these thingstraditionally 10 years ago came out ofthe bodybuilding community, right? Thebodybuilding community was like creatineand all these things that are morebroadly used now, but also other otherthings for sleep aids or other, youknow, magnesium and all this stuff.>> And to round out this year-end episode,we've asked some of our friends fortheir predictions for 2026. I'm socurious. My prediction for next year isthat uh the reasoninguh systems are going to translatedirectly uh to AIS that are much muchmore versatile, much much more robustand reasoning is going to impact isgoing to revolutionize not just not justlanguage models but reasoning is goingto impact every single industry frombiology to uh self-driving cars torobotics. And so reasoning, I think, isis the big huge breakthrough that thatum is going to transform a lot ofdifferent applications and industries.

00:31:26In 2026, AI will stop being a reactivetool that waits for us to prompt it.Instead, it will become very proactiveand get deeply integrated in our worklife. It'll go where we go, hear what wehear, know what tasks we need to workon, and in fact, most of the timescomplete those for us before we even askit to do so. It'll be our coach thathelps us improve our skills. It'll beour manager who helps us prioritize ourwork and manage our time. In short, it'sgoing to be the best work companion wecould wish for. I think the main AIprediction that I have for next year isI think context is just going to be themost important part of every singleproduct. And honestly, like one of thebest experiences I've had with it so faris just memory and chatbt. Like I thinkthat there are going to be a lot morefeatures that basicallytheir goal is to extract the user intentand make the onus less on the user tobasically give all of the models or thesystem or the product more and morecontext. So in other words, how do youput the onus on the product to actuallyextract that from the user instead ofthe user having to do all of the work todo this up front?

00:32:41>> My prediction for 2026 is there will bea whole new suite of product experiencesthat run on much faster inference.>> My prediction for 2026 is that we'llfinally stop copy pasting stuff intochat boxes. Instead, I think we're goingto have applications that have betteruse of screen sharing and contextmanagement across the sources thatmatter the most.>> One prediction for 2026, there's so muchtalk of agents right now and there hasbeen for a while, but no one has trulycreated a mass scale consumer agenticAI. I think the models are there todayfor this to be possible. And in 2026, wewill see the group that figures out theright interface and system and productthat creates as big a step function andoverall experience as chat did when itfirst came out. And I think this area isnot nearly as seated to the labs aspeople assume. It really is anyone'sball game. Hello, Aaron here. First ofall, I get quite awkward around doingselfie videos. This is my ninth take ofthis video. Um so I hope it goes okaybut uh 2026 prediction would be that uhthis is going to be certainly thecontinued year number two of uh AIagents but in particular AI agents inthe enterprise in either deep verticalor domain specific areas. Um I thinkthis is going to be the main

00:33:56way that weactually take all of the progress thatwe're seeing in AI models and actuallydeliver them into the enterprise. Youhave to be able to tie to the workflowof the organization. You have to be ableto get access to the data that theyhave. You have to have the right contextengineering to make the agents actuallywork. And then you have to do the changemanagement that makes the agentseffective. So this is going to be a yearwhere we start to see this patternemerge more and more. Uh which equallymeans that we need to ensure that wehave a lot more happening on agentharnesses. So shout out to Aorvosu andDex for that answer. Uh but it'sdefinitely going to be the year of ageand harness and seeing how do you startto get you know an order of magnitudeimprovement on the model's capabilitiesby having all the right scaffoldingaround the model. Uh and then finally itwill be the year of uh economicallyuseful evals. Um so really starting tofigure out how these models end up doinga lot more knowledge worker tasks in theeconomy. Um and that's going to uh we'regoing to see a lot more of that in 2026.We saw some previews of that this yearwith Apex and GDP Val uh and a handfulof others. We're going to see way moreof that. So, those are the predictionsand we'll see you uh in 2026.>> I think 2026 is going to be a veryinteresting year for American openmodels. Over the last

00:35:11year, the frontierof open intelligence shifted fromAmerica to China, starting with therelease of Deep Seek at the end of 2024.and American institutions were slow tonotice this erosion of Americanleadership in open intelligence but uh Ithink they've noticed in a big way overthe last half year both from thegovernment level from the enterpriselevel and there are some reallyinteresting uh neolabs starting to comeout with open intelligence as theirdirective and there are a few of thesenot just reflection and these companiesare starting to produce some veryinterestingsmall open models and next year I thinkwe'll see the US regaining leadership atthe open weight frontier at the largestscale and I'm really excited to seethat. Hey folks, my prediction for 2026is that I think we will see AI becomemuch more politicized. I think we'll seeit become a major point of discussionfor the 2026 midterm elections and somepeople will come out strongly againstit. Some people will come out. It'sprobably supportive of it. And um I'mnot sure which side's going to win out.>> 2025 has marked an incredible year in AIdrug discovery.

00:36:27In the past year alone,we've gone from being able to designsimple molecules on the computer todesigning simple antibodies and now mostrecently fulllength antibodies withdrug-like properties zero shot on thecomputer. If 2025 has been the year ofresearch in AI drug discovery, 2026 willbe the year of deployment. The modelshave finally entered an era wherethey're becoming really useful for drugdiscovery. Not only do they make thingsfaster, but they're also allowing us togo after really challenging targetswhich have been traditionally reallydifficult to do with traditionaltechniques. I'm really excited to seewhat comes next because the models showno signs of slowing down. Okay, myprediction for 2026 is it will be theyear that YOLO dies. we will begintransforming ourselves from a you onlylive once to don't die. I think rightnow we're kind of a suicidal species. Wedo very primitive things. We poisonourselves with what we eat. We designour lives so that we slowly killourselves. Companies make profits bymaking us addicted and miserable. Wedestroy the only home we have. Andsomehow we celebrate these things asvirtue. I think it's all backwards. AndI think one day we'll look back andwe'll be pretty astonished that webehaved like this. Um I think the

00:37:42simpthe shift coming is going to be simpleand radical that we say yes to life andno to death. It's simple but I think itcould be in response to AI's progress.And we do this defiantly as a form ofunification. Um I think it does requirea lot of courage for us though to say werecognize how sacred our existence is.We don't want to throw it away and wewant to defend it with every bit ofcourage and strength we have uh becauseit is so precious. I think it's going tobe the year we end yolo and thebeginning of don't die.>> The most striking thing about next yearis that the other forms of knowledgework going to experience what softwareengineers are feeling right now wherethey went from typing you know most oftheir lines of code at the beginning ofthe year to typing barely any of them atthe end of the year. I think of this asthe claude code experience but for allforms of knowledge work. I also thinkthat probably continual learning getssold in a satisfying way, that we seethe first test deployments of homerobots, and the software engineeringitself goes utterly wild next year.>> My prediction for 2026 is that it's theyear where everyone's perceptions areflipped. Currently, everyone believesthat you can only use Nvidia outside ofGoogle, and that will be obvious thatthat's not the case. Currently, about athird of Americans hate AI and thinkit's really bad. That number willincrease.

00:38:57Currently, most Americansthink AI is not useful. That will flipas well. And so, everyone's priors willbe flipped. That's because thetransformative use of AI will be soprevalent. The the obvious utility of itwill be so high that there is no way foranyone's priors. You know, cognitivedissonance will be wiped away.>> Hey, I'm Ben Spectre.>> I'm Ash Spectre.>> And our prediction is that 2026 is theyear of energy efficient AI. Data centerbuildings are primarily constrained byenergy, power availability, greatinterconnects, high voltage equipment,things like that. Which is why XAI'sColossus was initially powered byon-site gas trends. The thing is thedemand for computing to grow. Labs,Neolabs like us and like Kurser have apretty remarkably insatable demand forboth training and compute. And thisdemand is currently on stripping ourability to push lots onto the grid. Thismeans that in 2026, it will be reallyimportant to squeeze every available bitof tons out of every wallet. That said,in the long term, chips probably mattermore than power because chips depreciatemuch more quickly than the underlyingpower infrastructure.>> So, for example, with data center powersupplies at 10 per kilowatt hour, thechips cost action order imaging morethan the power than a 5-yeardepreciation cycle.>> So, in 2026, we think intelligence perwatch is really important

00:40:12to squeeze asmuch intelligence you can out of everyunit of energy. But in the long term, wethink it's the chips that matter more.>> Happy holidays.>> Happy New Year.>> Thanks for the year. Happy 2026.>> Happy 2026, listeners. Thank you.>> Find us on Twitter [music] at no priorpod. Subscribe to our YouTube channel ifyou want to see our faces. Follow theshow on Apple Podcasts, Spotify, orwherever you listen. That way, you get anew [music] episode every week. And signup for emails or find transcripts forevery episode at no-bers.com. [music]