Microsoft's NEW Model BETTER Than OpenAI's o1? (rStar-Math)

[@GraipVine]

Wes, whatever chroma key colour you're using, wrap it around the headrest on your chair. It will key it out and give our eyes some relief. :D

[@PnchBagTF2]

Or put a sheet of paper behind the holes of the chair

[@RickySupriyadi]

@@PnchBagTF2 wow brilliant!

[@autohmae]

Euh... I think you mean the opposite, he's probably using a green screen or similar (I'll use green in my comment) and his headrest is partially see through showing the background aka green screen partially. So you'd probably want a different color than green.

[@storm14k]

​@@autohmaethey are suggesting they wrap the headrest in green or whatever color he is using so that we don't see it.

[@autohmae]

@@storm14k yeah, I know, but seems kind of silly solution to me. Very hacky, instead of making it look nice in his office.

[@lionhearto6238]

o1-preview is actually the old one. o1-full is out now and much better. there's also o1-pro which is way better than they advertise imo

[@TheSongManipulator]

With all the noise to keep up with, I don't blame him for mixing that up. But yes, we have o1-full and o1-pro out now.

[@cajampa]

​@@TheSongManipulatorNah, he should know. That he doesn't just means he aren't using them. So he don't know them.

[@James777x]

Classic deliberate mistake to encourage more comments. I’m totally on to you, Wes.

[@ABUNDANCEandBEYONDATHLETE]

o1 pro is still pretty slow and not much better than o1

[@GeoRust1]

When did it get publicly released? O1 pro?

[@NeorecnamorceN]

AI is helping me write a book right now, so yes, it has definitely had an impact on my life. Never would have dreamed of writing a book because of the amount of time it takes but having the AI help research, organize my thoughts, write rough drafts to help develop my story. It's an absolute game changer.

[@GraipVine]

Agreed. It's helped take me in directions that I never thought of. It's a huge repository of information and I am developing a much more engaging environment because of it. Like any tool, it doesn't do it for you, but it gives you the power to bolster your ideas.

[@j2csharp]

I'm curious... Are you writing the book, or is AI writing it? Will you be giving some credit to the AI as co-author? After you complete the book, what percentage of the writing is actually yours vs. AI? And how do you feel about that? I am not certain how to feel about any of this, but I find it interesting to think about.

[@NeorecnamorceN]

@@j2csharp would it be weird to have the AI help me figure that all out? Like, how much credit to people give ghost writers/publishers/existing stories/lore/etc.? This is a whole new world for me in almost every sense so I'm just kind of flying by the seat of my pants here haha

[@j2csharp]

@@NeorecnamorceN I use AI to write code snippets that I can reshape to fit my needs. You can write with AI in the same way. I'm just having a moral dilemma about how to give the AI its credit when credit is due. Thank you "Mr. Roboto" for doing the job no body wants to. haha

[@GraipVine]

@@j2csharp I originally wrote, "The Complete Guide To Writing Fantasy," back in 2003. It was a bestseller at the time, but some of the information by the time it was released was outdated, so that always bothered me. I use AI for research, to spark ideas, to help rough out some scenes when I just can't put it into words. Having said all that, it is me, the creator, that drives the narrative. AI, as I said before, is just a tool, albeit a more powerful tool, but still needs human guidance. So, there's no way to really put a "percentage" on it. Let me give you a somewhat simple example. You could do a math equation by using your skillset. You know the rules, what you have to do, then sit down and do the calculations until you reach the answer. Or you could reach for a calculator, a tool. You still need to know the rules, and you apply them to the calculator, and you still get the same answer. Did the calculator solve the equation or did you? Use a stone for a hammer or a hammer for a hammer. The nail goes in either way - just a bit easier with the proper tool. But the human is the one holding it from the start. So, yes, I adopt AI as my tool, and I use it to be more creative. Still my juice, but the AI tool is better than Google search, and encyclopedias, and saves time. Sorry for the long response, but it just struck me how people long for the past instead of helping develop a future to assist them. What's going to come out of all of this with advancements in AI? Only time will tell.

[@storm14k]

The number of software engineers I see claiming AI can't write software is astounding to me. Meanwhile I'm building a whole CQRS based system of services without writing a single line of code. Does it misunderstand or do some things I didn't intend? Sure. No different than what you may flag in a code reciew. I say correct it and it's done all within minutes where I'd be talking hours or days before.

[@drwhitewash]

The point isn't not to write any single line of code, though. Often it's easier quicker to manually write/modify parts of the code, rather than keep re/prompting the LLM to get the desired result. And of course knowing what to ultimately write is most important.

[@tonystarkagi]

ya they're coping

[@young9534]

Small projects and tasks are not the same as writing enterprise level software with hundreds of thousands or even millions of users. LLMs are good at the small scale stuff but still need to be hand held through a lot of the process. I think this will change when we get reliable agents though

[@storm14k]

@young9534 just left a large company building a major customer facing system the exact same way. Don't hide behind complexity. You should be seeking to reduce complexity anyway. CQRS so far has proven for me to be a good way to do that as the company I left was moving from a giant overly complex monolith using that pattern and it's why I decided to try it here. By the way this isn't a small project I'm working on. Why can I pull it off without a team right now? See my original post.

[@technolus5742]

​@@storm14ksounds like your project structure is very flat. When there's a layer of custom code that you are interacting with, llms quickly become useless and hinder progress more than they help.

[@vaioslaschos]

As someone that is currently working on this very specific topic, I will say you did a very good job explaining it. One small comment though. The idea behind "golden answer"-only verification, is that although, yes it is possible to come to the right answer with wrong reasoning, it supposed to have very small probability. Afterwards in the training, even if you have some "corrupted" data (correct answer/ wrong reasoning), if they are only few in number, they will not affect the overall training like it happens with all small noise in general.

[@rotary65]

We have tended to evaluate AGI and ASI as an individual model instance based on one prompt/response. It is very clear now that both will be achieved through many networked smaller agents that have domain super intelligence. It is the networked and interconnected efficiencies that represent the greatest potential.

[@techrvl9406]

E m e r g e n c e

[@anatalelectronics4096]

was thinking the same the other day, the big thing will come from MOM, mixture of models. since the big players are probably to obsessed with their own stuff it might come out of a garage

[@BrianMosleyUK]

Wes, I love these walk throughs - you're doing a great job with these. Especially with your patient attempts to explain the acronyms along the way. Thank you, thank you 🙏❤️

[@jameslincs]

19:12 it could just be mistaken about its reasons. For example… AI trained to see tanks in satellite footage were given training data where the photos containing tanks were taken on dark days and the non-tank images were taken on sunny days. The AI then said a photo contains tanks if it was an image taken on a dark day. The only solution to this is for the AI to know what a tank is, so it knows a tank is not just a dark day. This is a harder problem than it appears and probably won’t be solved until full AGI

[@Dissimulate]

Training the correct chain-of-thought into a model seems like a very promising way for grounding the model in reasoning and reducing hallucinations.

[@zalzalahbuttsaab]

6:00 nothing happens independently of anything else Thanks for an excellent breakdown

[@not_milk]

The future we have in front of us is not Star Trek. The future we’re on course to reach is Wall-E

[@OverbiteGames]

The masses seem to want Wall-E - at least the ship portion, lol. Those who don't are pretty much split between the equivalent of Job Simulator / GTA Online and the rest of us are probably going to be self sufficient in keeping ourselves entertained hopefully. But that's the long game. Imo AI that understands and reads your mind and simulates every possible outcome is the final thing - it's what amounts to you living out your perfect infinite lives at a rate of an infinitely increasing sim-sec/second with as much interaction and solitude as desired.

[@GospelProgressionsUniversity]

You have a very a smooth calming voice. If this KZbin thing doesn’t work out there’s always podcasting.

[@bernhardd626]

Yes, I often use it to fall asleep 🙂

[@waynewells2862]

You covered a major juncture in the evolution of AI capability very clearly - thanks

[@Anders01]

Microsoft has huge resources, including their own AI chips. It's possible that Microsoft will surpass OpenAI and that the future will mainly be an AI battle between Microsoft and Google, plus the Chinese tech giants of course.

[@jeanfredericferte1128]

scaling law only applies when comparing similar architectures/structures and training processes. Model's architectures are evolving (that's the interesting part of research imho, scal race is shallow) Distillation is another way of playing on external parameters (training method) anyway, exciting times.

[@thekisoboqo7836]

The coming wave describes it as many waves coming, not just one, also Yuval did say that we probably won't just need to adapt once, but adapt many times, constantly

[@Axiomatic75]

Yuval creeps me out. One of the worst WEF shills.

[@punk3900]

Wesley, your armchair is stunning the compression algorhythms

[@NoPhilospher]

🤣🤣🤣

[@Elongatedrider]

😂😂😂

[@Comments_By_BotJ]

I paused it @5:28 when you mentioned us hitting the limit of our available resources. Referencing your video from yesterday "OpenAI SCARED To Release New AI Product", at the @9:07 mark of that video you cover the story of AI creating new chip designs thinking outside the box and bringing novel creations to take current problems. I sincerely believe that when AGI is 'unleashed' to the scientists and researchers, we will not just answer our current limit issues with resources but blow past it in the same manner the performance of models have done. Sure, if we're considering our limitations with our human limited ideas, then its very much an issue (for us) but if AI is pointed at the problems, I suspect they'll find novel answers and solutions to make our current issues/problems with limitations be a memory of the past.

[@michaelspoden1694]

I hate when they say we run out of data you know like they've used the whole internet so that includes everything that's ever been discovered there's so many patented ideas that are nowhere to be found on the internet and all the private information Geniuses have and other people in general probably surpassed the internet in depth when send it to actual quality data. Patent ideas in all these other things that are kept encrypted are incredibly precise in the way that they get to a problem and solve the issues. All of this will give the ai's massively great capabilities on how to go through problems to create new things out of basically what information that it already has to find out new information. I think that's a huge step knowing how to utilize the information to create novel Creations which end up being patents and the same information that is not being available as data amongst many other sorts of data think about it just being natural language data definitely not utilizing all human imagery and all video and ask people to donate directly in the training I think they would get a lot lot lot lot of information. Also think about all the information that we get every day from all these science projects we have to convert them and understand that to be able to train it. But how is it supposed to take a new information anyways and learn from it if it doesn't have the capability that humans use already I guess that's what will be used but if it figures out a problem like novel scientific boundaries like we utilize during Aquarius but in the end it disappears or doesn't get actually trained back into the data it's just used and then goes away when they could have that boundary breaking data that's created by using this reasoning capability to actually insert that directly back into the model drawing of course that it is correct apologize for my grammar I use voice to text and sometimes my words also get interpreted improperly but I don't go back always. Think about all the spatial awareness data learning in time with sensors on hands and arms constantly processing the world which is obviously the embodied case that people have presumed is necessary and also I definitely believe is a quicker and more thorough much more genius and human-like thought process of concept contextuality being able to understand time as we know it relative to how we create things and movement of other objects to really have the perception the physical world and physics by actually being immersed inside of it giving it essentially subjective reality and being able to possibly give it a separate personality module that kind of allows it for it to even learn from what it sees instead of only what is trained on and that makes it a more reliable source of information because it can be the one thing that it knows for a fact is no doubt it can be the one thing that it knows for a fact is no doubt factual so then it can actually have its opinions and not just bought out by his data although it still will have biased integrated into is underlying data set but still be able to get past that I'm not making assumptions initially pick oh my God I could make a paper LOL I wonder if anybody actually going to read this to the end or even start it because it's too long. Assalamualaikum😂

[@scotter]

I second your book recommendation. The author's name is Iaian M. Banks.

[@miikalewandowski7765]

A little step in the benchmark, a SHOCKING step for the industry!

[@apester2]

Orthogonal is founder speak for "we've hit a wall with our current approach".

[@robert.jackson]

Do we get sStar next? That is almost a stutter when reading or writing it.

[@Terran_AI]

It's pretty obvious that the priority for AI needs to be advanced energy production. Zero Point energy finally?

[@Raymond-wj4ol]

No, better models that are more energy efficient. Biological brains set the standard. They have billions of years head start in evolutionary optimization, so we probably need more advanced biomimicry.

[@ericfisher1360]

They are not mutually exclusive.

[@Terran_AI]

@@Raymond-wj4ol True but the whole planet needs cleaner, infinite energy, not only to feed our AI overlords 😁

[@tracy419]

​@@Raymond-wj4olI agree with the need to focus on energy ASAP. Energy costs affect the price of literally everything, and if jobs are going to start disappearing, we need prices to go down. Energy, housing and food, but lowering energy cost will help lower the cost of everything else in the chain. Better models are important, but if we don't address energy costs it'll be especially rough for those who lose jobs.

[@Raymond-wj4ol]

@@ericfisher1360 Yes, I stand corrected on your point.

[@BilichaGhebremuse]

Excellent bro 🎉

[@Raymond-wj4ol]

Wes, amazing orthogonality!

[@LatentSpaceD]

Thank you, Wes-- solid

[@asthinea]

I am here and enjoying this Wes thank you

[@djayjp]

Someone should tell Wes that "o1" scores much higher than the obsolete "o1 Preview".

[@dennisg967]

Right. But his point is that much smaller models beat models that are much bigger.

[@Sinjhin]

I gotta say, it has been pretty neat watching your technical understanding just continually increase over the past year as I have watched your videos. Dipping into the code a bit here and there, getting into the weeds on occasion, knowing when you are over-simplifying things and stating so but still giving a useful comparison. 👌🏻 EDIT: LoL. I hadn't watched til the end when I wrote that. Guess I hit the nail on the head. You are doing great!

[@MrConete]

Can do one video on LLM vs the new LCM?

[@pabloe1802]

So basically big companies are rediscovering 60 years algorithms

[@abusedcharisma]

Note comparison is to o1-preview not to o1 (!)

[@TheRemarkableN]

Highly recommend The Culture series for anyone interested in a (mostly) utopian vision of AI and the future.

[@maheshBasavaraju]

Netflix?

[@savagestranger]

@@maheshBasavaraju Books

[@Axiomatic75]

I've read some of Ian (M.) Banks' books, he's a very good writer. I think I'll start the culture series soon

[@Axiomatic75]

I've read some of Ian (M.) Banks' books, he's a very good writer. I think I'll start the culture series soon

[@complianceaves1120]

Great content thanks Wes

[@igotnuttin9935]

It gets the right answer and we declare its reasoning is wrong. “Move 37” - AlphaGo

[@luizpereira7165]

It looks more like a new AI reasoning technique than a new AI model. It seems, it can be aplied to diferent LLM models.

[@RichardHartness]

I just want to say thank you for this channel and how you manage it. I like the balance of news, some times paired with these instructional moments. I am a Software Engineer and (if I had the time) I could probably deep-dive into much of the technicalities of this stuff. But, I DON'T have time and this type of information helps me get a bit up to speed so if, or when, I do need to dig deep into an AI specific topic, I have pre-knowledge about the basics.

[@ryangerritytech]

This is very interesting, it should start the path to actual Agi reasoning capabilities. I am currently building a benchmark for this as well.

[@lule-ahmed]

Can we install those small ones locally ?

[@StefRush]

Yes, I thing you're wrong about the gap from "AGI" to "Star Trek". I think it should be 1/10 the size 😅🤣😂

[@TheMrCougarful]

Only the elites can benefit from this. The rest of us live under bridges.

[@RussAbbott1]

I would have liked an explanation for how rStar-Math solved the puzzle the paper showed at the end: Find x and y such that ... [some equation]. The solution shown was brute force search. Perhaps it's smart to use that approach when the domain is as small as it is in this problem. But in general, it's not a smart approach. So three questions: (1) How did it decide to use brute-force search (and how did it know how to implement it)? (2) More generally, when it decides to use an approach, what is the population of approaches it knows how to use? Did it learn them during training? and (3) LLMs don't have a means to backtrack. What mechanism does 4Star-Math have built into it, e.g., Python code or something else, that enables it to do backtracking?

[@caffeinated-ai]

Definitely the right angle to take on the use of orthogonal.

[@ALTINSEA1]

Do ai really exhaust every daya on internet? Like how many youtube video uploaded everyday that provide new info that useful or useless

[@satansbarman]

Sudoku would be a better analogy than crosswords

[@darkmage7771]

why do you say we only have o1-preview when we have o1

[@aliettienne2907]

19:14 I believe in this concept of a structural model system that enables the PPM language model to guide in mapping the correct tokens from their library with the required group of tokens (the solution or answer). It's synonymous to a person asking a friend to assist in remembering an event or an address of a place. Both individuals will be able to recall different aspects of the event or the address location with the very exact results (with no contradiction to each other). Let's say they are trying to remember a specific birthday party event, the PPM language model may memorize the birthday person name and their address, while the SLM language model may memorize the food and the guest names. The PPM LANGUAGE MODEL can assist the SLM language model to remember what had transpired at that time with pertinent clues from its library. I hope that I'm making sense. 😁💯💪🏾👍🏾

[@TheJimNeiL]

Where do we get more data? Two places: 1. Synthesize the data. Take existing data and mathematically manipulate it to create new data; feed that into the AI. 2. Put AI sensors in mobile devices (cars, robots, etc.) and have them interact with the real world, collecting new data as they go. There’s no substitute for direct experience.

[@miikalewandowski7765]

Do THEY also run a training program based on RLHF data to train a LLM to perform RLF? I guess there is some valuable meta-data within the RLHF process waiting to be distilled..

[@pcread]

It comes down to the Generalist (LLM) vs Specialist (SLM) conundrum. Is it better to specialise and know almost everything about increasingly narrow fields of expertise, or learn enough about many topics to survive and thrive. Maybe some Belbin Team Type colab of SLMs working together to accomplish a task. Might that prevent rogue AIs from overstepping their bounds?

[@jeanfredericferte1128]

MoE...then the question of supervisors remains, for the supervisor will need to match experts complexity/depth without outmatching the user's cognitive abilities

[@HzPixelForge2025]

"You look lonely, i can fix that" . i m speechless. it's mind blowing how things are unveiling... i am amazed by everything that is happening right now around the world . what a start 2025 !.

[@qwertasd7]

ik kinda wonder for llms and neural nets, i always thought of them as sort of flow process, somethign goes in, after all weight and biasis something gets out. what is keeping the repeat flow, word by word line by line, and then thinks over the construction of the lines displayed?. do they actually process different ideas ( or is it one optimal head ? ).

[@calinciobanu25]

this shows the power of VERIFIERS. With Math they could implement very strong verification and generated very high quality data (and that can improve iteratively). Imagine if you could have perfect verifiers for any reasoning problem 🚀

[@spinningaround]

We need to develop SAI! Without it, we'll remain stuck mimicking the reflexes of an overly educated spinal cord!

[@technolus5742]

not sure why you think o1-previous is a big model... judging by its performance on many tasks, it's a medium/small model to make up for the expense of the extra compute time.

[@tinyhato]

Orthogonal is the new liminal :p

[@OwenThompson-g1y]

Great style and presentation as always. Thanks. It sounds very interesting and fruitful (if a little obvious for such brilliant researchers not to have already implemented some years ago). It does however sound hard to remove the subjectivity from many decision tree choices a model might make in instances where they are not math logic based (as many will not be) without large token use to validate reasoning. I’m still unclear as to the mathematical model that allows for the model’s evolution of treatment of a specific narrow problem type into its wider behavioral memory. Still enlightening to hear the broad brush approaches that are working, even if the detail remains very opaque. I imagine that token cost reduction now allows the researchers greater reasoning possibilities.

[@tmac8198]

Big fan of the culture war series by EM Banks. If you love sci-fi you’ll love his books.

[@TheDarkhawk243]

Holy fuck. This is a 7B model? That’s crazy.

[@waynewells2862]

Should intelligence be distinguished by the category of "artificial" ? We might benefit our moving into the new era of the coming intelligence potential for a speciation event by thinking and using the term Non Organic Machine Intelligence (NOMI) that will likely be replaced in the not to far future by Chimeric Organic/Machine Intelligence (COMI).

[@senju2024]

Orthogonal -> Philosopher Nick Bostrom introduced what’s known as the Orthogonality Thesis: Thesis: The level of an AI system’s intelligence is (theoretically) orthogonal to the goals it pursues.

[@mattgscox]

09:50 - am thinking the future is likely to be Battlestar Galactica rather than Star Trek

[@DonNome]

Gracias por poner el audio en español

[@73N5H1]

You're missing that we will merge with AI and become a hybrid species

[@autohmae]

7:42 you do realize you are talking about pretty much any technology ? Every technology makes a single person more powerful to do things, just a side note: also bad things. Which is why it's important to make sure we get through these disruptive times without disruption happening to fast. 10:15 OK, let's think about it. So you are saying as technology advances this will not be the case anymore ? That seems unlikely to stop. I guess the disruptive times do act as a multiplier. But the real question is: if you look at technology, always speeding up and adoption also speeding up (10 years adoption for cars, 7 years for internet, 5 years for social media, 3 years for smartphones, 1 year for AI, etc.). So that's why people talk about the singularity, when technological change happens so fast we can't keep up. So yeah... disruption will happen even more, but humans in control of that change might not be ? Only if the AI does things independently. That's a big IF of course, I guess we can give it an agentic long running/long term goal task ..?

[@vieome101]

This CHANGES EVERYTHING! LOL

[@StefanReich]

GAME OVER

[@BeTheFeatureNotTheBug]

Sempei 🎉

[@BeTheFeatureNotTheBug]

21:50 makes so much sense why I listen every day

[@caiogustus]

self-tought AI with synthetic data is scary af

[@RadiantNij]

Why are people still talking about o1-preview, what timeline did I wake up in????👀👁👀

[@EnigmaticEncounters420]

Was probably the easiest to assess for the Microsoft team when they were doing this

[@EnigmaticEncounters420]

*Access

[@sneezweasel]

Models need to be considering that their training data is not gospel and arrive on truth on its own. since quality data is going to be hard to find it's going to need to derive something from bad data.

[@Ben_D.]

Orthogonal. 😆 Love your humor dude

[@easydoesitismist]

Ran out of data? That only means that supervised learning is done. They use reinforcement learning and synthetic data now, especially for reasoning. Watch any talk by Noam Brown (he trained super human systems at poker with self play). Distillation is the method where you train smaller models to copy larger smarter models. Thats the easiest training to perform since it requires the least human effort. Distillation will keep shrinking these models without any new data. Smaller models are forced to generalize quicker than larger models because they are less able to just memorize the tests.

[@tommyhuffman7499]

All the things you listed AI will do, the internet has already done.

[@patruff]

I always said math was a waste of time. AI can't weave baskets underwater, I'm so happy i chose the right major!

[@jeffsteyn7174]

I'll ask chatgpt how to build a bot to replace u

[@dan-cj1rr]

3-4 years ? this thing been giving me sht code for 2 years now.

[@erererx3]

Elon, where's FSD?

[@wwkk4964]

Thank you!

[@LatentSpaceD]

Wes, I'm a practicing solipsist now. I've been dying to tell you.

[@benjames813]

Hey Wes jw if you're a programmer? Keep up the good work btw!

[@BrianMosleyUK]

6:35 massive Freudian slip by the aspiring king of the world "anything I want" lol... Surprise surprise Grok-4 will even tell you what *you* want, and surprise surprise, you *will* want what the weapon of mass persuasion will want you to want... No surprise, this will align 100% with what Elon wants. 😵‍💫

[@honkytonk4465]

You do everything your AI wants,poor boy?!

[@BrianMosleyUK]

@honkytonk4465 no, Grok-4 hasn't started whispering in my ear yet... However, LLM's already have a superhuman ability to be persuasive, it's just not attached to your complete personal data history and being tuned by every interaction you have with it yet. Listen to Mustafa Suleyman talk about personalised AI and you'll get the idea. You think that's free will you're using? Huh!

[@honkytonk4465]

@BrianMosleyUK Look at your prefered AIs like Claude,they are much more ideological.

[@BrianMosleyUK]

@honkytonk4465 yes, don't look at what Musk, Meta, TikTok are doing to societies worldwide with social media. I'm sure it will all be fine.

[@honkytonk4465]

@BrianMosleyUK Teach people to think critically in every direction,don't teach what to think but how to think.Todays universities have become bootcamps for leftist ideologies.

[@Gholladay]

Fascinating!

[@nighteyes360]

AI is just impressively good. For now however, it definitely still requires a lot of human guidance, but who knows when from a single input document with all specs, it will write the full code of the app. It could probably already do it already, but I think it'd still require many reworks behind from human input. It still needs training to be perfect, we will see that in a short amount of time I believe, maybe ~1-2 years time). Does it mean the end of software engies? Probably the end of programmers and a new age of 'IT guys', but I'm not sure what will be their role. Because for the common folk who wants to make a website for instance, it still takes up time to build one, and someone has to spend time and put in the effort, and many people don't want to do that or even remotely touch a computer (yes, even many some from our generation). Many folks will probably start using it to build their own stuff however. Who knows how far this is gonna go and how common folks will start using this technology, it's honestly very hard to precisely know, but people whose sole job is to write code for classic websites, it's not gonna end well for them. If you are one of those folks, you have to start expanding your skills beyond that of just writing code, you have to be a builder and someone who can bring a fast solution to people, leverage AI to ship them high quality products fast. Then maybe we'll reach and age where from a microphone, you can talk to the AI and it will build in real time the website and that will be pretty much what we could see in a futuristic movie, with perfectly adapted UI to help you build it and mold it at your will. Software engies will be fucked when this happens, at least for the average website (not sure about ultra-specialized webapps). The more complex an app however, the more clunky the AI becomes at the moment and needs to self-asses it own work and rewrite the stuff it has missed/fucked up. I'm sure at some point, AI will handle most of that too, but maybe some human intervention will be required, who knows

[@aaronhhill]

Are you talking about Ready Player One?

[@cowlevelcrypto2346]

On the chance that 46% of it's calculation might be wrong, I wouldn't feel comfortable having this AI design or control a Power Plant, or a space ship designed to land on Mars and actually support human life during the trip and landing.

[@PrinceCyborg]

Where did you get the 46% from, that's almost half, are you saying when I ask o1 a math question it is wrong half the time?

[@cowlevelcrypto2346]

@@PrinceCyborg The data he posted was claiming that the ai got 53.3 % of the math questions right, which put it in the top 20% of high level high school math takers. That means it makes mistakes.

[@AzoreanProud]

We hit a goddamn wall!

[@RobertHouse101]

Reading the abstract, you repeated the same idea several times X 2. At 17 minutes, I feel like I've heard it all. . J/K Friendly critique. Thanks. Love all the new information you provide. I haven't missed many.😄

[@schrodingerbracat2927]

When people use 'orthogonal' they usually mean independent (but are trying too hard to impress)

[@douglasgarth]

Your question regarding how do we scale up AGI/ASI with power constraints is easily answered. You DON’T scale up for the public, only for Elites & Government (security concerns of course) as we all slowly come to realize we’re headed for a Dystopia, not a Utopia.

[@BigSources]

Classic tinfoil yap. It's the illuminati guys and the government wants our doom.

[@sinnwalker]

Nice doomerism. Except you're only partially right. The world has always been in a dystopia/utopia, just depends where you live. Also the foundation of AI is open source, and endgame is open source. We won't have the compute power giant companies have individually, but we'll definitely get the models, whether companies like it or not.

[@PaulHigginbothamSr]

Since ai is drawing up upon old nuclear plants for use in the future I see ai showing molten salt small thorium reactors coming to the fore. What this will mean is these large llms like o1 to promote how these new data centers to build these reactors to power the new data centers. I am almost sure that will be the case and new generation capacity will also hurry the production of these small reactors and will almost surely have supercritical co2 turbines as 1/10th the amount of resources will be consumed and will be orders of magnitude will come on line maybe more than is used at present. There will be no increase in co2 and less will be produced with 10 xs as much as used at present.

[@mito._]

I still say all of this LLM news is great, but what the world needs is an open source LLM.

[@pierretetreau7497]

We have a lot of them deepseak might be the one that comes to mind when evocating open AI O1

[@alexleo4863]

We have deepSeek v3

[@ishakrahou3230]

we have new open source light model from Microsoft called Phi-4

[@WackyGameEngineer]

all the ollama stuff is open source

[@pdjinne65]

dude, there are so many of them, and good ones. Of course you'll need good hardware to run them, for for $2-3k you get a 5090 that can handle them a bit.

[@nexusdog]

Yes indeed, what are we going to do?

[@oytrn]

In regards to your question about peak human productivity happening in the next 4-5 years due to AI-augments before AGI surpasses that. I am aligned with your (and Elon's, Sam's, Ilya's, and many others) assessment of where the tech will get to over the next few years is highly probably correct, likely even conservative. What I do believe we (including I, up to recently) underestimate is the huge public push back against AI that will lead to delays to significant adoption. We. can already see this in a lot of tech and non-tech settings. There's also the Trump factor where, over the next 4 years, if he feels threatened, or that it would popular and/or lucrative to do so, that he might act in a way that hinders general public/workforce adoption of AI. Layer in the limitations in total compute / compute power available and there may be a resource bottleneck there, too. Given those factors, and natural resistance to change (especially when used in conjunction to fear of the unknown, etc.) I would be surprised if the tech really blooms to the point Musk mentionned until at least 10+ years.

[@dennisg967]

Orthogonal means perpendicular. Just fyi

[@anatalelectronics4096]

on auto and just listening i thought PBD was talking ;) take it as a compliment pls

[@pdjinne65]

Honestly as much as I love this stuff it's become EXTREMELY hard to separate the hype from the actually cool things nowadays. How much of all this is a scam is up for debate.

[@AndyBarbosa96]

I am not a pessimist here, and looking at our recent historical record, niche AI has already surpassed humans in closed domains aeons ago, for example, in chess,..., however, instead of just becoming obsolete, the game of chess has never been so popular, even though, everyone knows, robots&AIs are thousands of ELO points better than the best humans.

[@AndrejPekar]

In my opinion, you are not taking into account the human factor. People do the same thing over and over again. They will only learn a new thing if the gain from the new thing greatly exceeds the effort needed to learn the new thing. Even the TV remote control is a problem. People will prefer the old remote control to artificial intelligence. Because the old is always better... Artificial intelligence would have to come up with something that is truly groundbreaking. Not a little better than the old thing. In my opinion, voice assistants also fall into this category. Assistants don't even do anything better than the old way.

[@Justin_Arut]

What's with the filler speak with these AI channels? You don't have the "you knows" as bad as some others, but your "kind of sort of" is off the charts. I bail on the videos where you start looping that junk.

[@froote]

Bruh 3or 4 years for elon musk is 10+ for us.

[@spiritinflux]

I wonder when we will accept that 'ASI', Rose, in the near/far future can travel up and down time and is here and now, influencing us subtly to create her a little faster each time, - well each time being the first time to us on the line, but just further back to her... But, i suppose, this is impossible 😏 Netflix does not have my permission make this into a film. (because i get this is basically a full script to them 🤖)