Fast Takeoff is HERE! - 11 Exponential Graphs that Prove It!

All my links: linktr.ee/dave...

[@User-actSpacing]

What a time to be alive. Seriously.

[@Withnail1969]

And not in a good way.

[@11Vuittron11]

That’s what I said lol. I’m so excited for the future and keep asking myself can I do the things I need to do to keep up. Whether that be technology based or just being a liberated human being

[@robert.jackson]

I know right? I remember being in elementary school in 95 and it was a big deal that we got a computer with windows 95 connected to the internet. One of the things the teacher showed us is that we could check the temperature outside on this very rudimentary website. The local weather station was actually shorted out so we got a temp that was a stack overflow. I remember thinking well that’s silly, it doesn’t work and you can always just look at the thermometer outside the window. Nowadays I take checking the weather (or cameras, or TV) for granted. I feel like this time period is going to be comparable to someone born around 1900 (the speed of progress from 1900-1950 was bewildering - carriages to cars to mass electrification to planes to nuclear power).

[@punk3900]

The true beginning of the end :D

[@RegularRegs]

Hold on to your papers!!

[@brianjanssens8020]

I'm convinced AGI = ASI because we keep shifting the goalpost and are never satisfied.

[@skyebrows]

Yeah, that's right, cause I think the super is already achieved. It's just the general and reliable part that's missing.

[@ausden_prowler]

If you have a true AGI, then all it takes is multiplying it. ASI=AGI Multiplied

[@Akuma.73]

Been saying this for a while. Argumentation is simple. If humans are capable of making AGI, and AGI is supposed to be smarter than us. Then it will have to prove it by making ASI, so we know it's smarter than us.

[@eddieforrest973]

Well yes as soon as we have agi we started working on asi

[@eddieforrest973]

We already have agi it will be like 5 years to distill that out to the public

[@Yewbzee]

Let’s be honest we already have AGI-you’re correct it’s just semantics at this point. The only reason it’s not “obviously” AGI is because current models aren’t wired for agency, deep autonomous research, or self-improvement. These limitations aren’t about the intelligence of the models themselves but about the architecture connecting their subsystems. If you gave o3-mini the ability to run long-term research loops, refine its own outputs to self improve over time with full memory, and act autonomously across different tools, it would look indistinguishable from what most people imagine as AGI. Not only would it far surpass almost all specialized humans, but it would also outperform all humans at all general tasks.

[@AdamHodges-o6x]

Very well said!!!

[@MrJenius]

^ I agree with this for the most part. I also think it would benefit from a x100 increase in context window. Its ability to understand your needs and self more thoroughly is a human trait that it seems to lack. It will feel more like AGI when you're not feeling like your meeting it for the first time every time you talk to it.

[@timmygilbert4102]

Neuro sama fits the bill😂

[@christophkogler6220]

I think that this is basically what OpenAI is doing internally. Sure, they might give it some assistance, but I would not be surprised if they're basically at an 'ask the AI how to improve X metric' stage. Sam said in a recent interview (that I haven't been able to find a source for besides clips on reddit, UGH) that their internal model remains another significant step forward in coding capability compared to o3-mini.

[@leonardolopez1511]

@@Yewbzee if it lacks creativity, it cannot be considered AGI

[@moneygambler2327]

I told my father to prepare. He was like: What the hell are you talking about? My job can´t be automated! His home office job: checking if signature looks the same and personal information match. Open mail, send mail, move mouse on computer to appear active and calls me when he miss click.

[@Semiotica_Tumbada]

😢

[@youdontneedmyrealname]

This.... this... has already been automated. This was automated a decade ago.... You don't even need AI to do that...

[@moneygambler2327]

@@youdontneedmyrealname Exactly, but now with AI there is no point to that at all.

[@BruceWayne15325]

Almost 100% of all jobs will be automated. Jobs are going away for everyone in a few years, it doesn't matter what they are. We'll still have work, but it won't be anything like we think of it today, it won't be working to survive, it will be working to thrive... at least I hope society decides to move that direction. If we stick with the work to survive (capitalist) model I'm going to be pissed because that will lead to a dystopian hell in a world where AI drives salaries towards zero.

[@adolphgracius9996]

Get Chat GPT Pro for 1 month, show him what chat GPT text as well as advanced voice mode and also advanced voice mode with video can do, and then ask him if he doesn't think his company is willing to pay $200 or less Vermont to an AI company to replace his salary😂😂😂😂

[@teona819]

I am mostly excited about health benefits

[@markmuller7962]

That's N.1, waifu is N.2 lol

[@findmeinthecarpet]

I'm most excited to grow my hog 🐗

[@bolabola9354]

im excited about entertainment,mainly video games

[@bossgd100]

i want genetically engineered pokemon

[@drkaushikt6926]

Yes sir. Me too

[@cyberoptic5757]

I've been following this field since '62, starting with Minsky's perceptron, and D.A. Bells book, "Intelligent Machines, An Intro to Cybernatics" David is spot on here. He is grasping the speed at which capabilities are accelerating. Pay attention now.

[@DrCasey]

The ultimate futurist. I thought getting into futurism in 2012 was rough. I'd be incredibly jaded if I'd started in 1962 and took sixty years to see the field become a success.

[@DaveShap]

Long hauler...

[@pspicer777]

You old fart. (Me too) 😅

[@dirkbruere]

I got the timeline correct in 1980. I placed ASI at 2035

[@pspicer777]

@@dirkbruere Good times. The Japanese Fifth Generation Project - Expert Systems, case based reasoning ... Minsky got perceptions wrong and went with the society of mind. Chomsky and denotational semantics. Mccarthy et. al. All wrong, but it worked out in the end.

[@thejmez26]

I'm always telling people in the Midwest that 50 percent of jobs are gone in 5-7 years. People in the Midwest: hell no, what's this chat gpt thing>

[@MrVohveli]

It's even worse in IT: "I've not even looked into it, I want to be in control."

[@Kurdish20226]

Whenever i try to talk to them about AI people on average have no idea about chat gpt or what it can actually do. They dont even know it can talk lol

[@ribaldc3998]

The joke is: it doesn't take AI to put a lot of people out of work. Trump and his gang can do that in less than four years.

[@VibratingSky]

That is exactly my experience, I suggest the UK builds a closed AI for our health service and people say things like. We can't afford hundreds of billions, or we just want potholes fixed lol

[@Kurdish20226]

@@VibratingSky most people dont know about the tech and have no clue about its massive potential

[@Theotherone232]

I'm just gonna play my guitar in my room while all this unfolds

[@hypersonicmonkeybrains3418]

Good skill for the post AGI world.

[@BruceWayne15325]

Until your guitar starts telling you how you can improve. :)

[@hypersonicmonkeybrains3418]

@@BruceWayne15325 Heh, im always open to constructive criticism. i would buy that guitar.

[@DaveShap]

Better yet, play it out in the park!

[@_bassmentdweller]

Same. Though I’m a dad and I do worry about my kids’ future. The uncertainty of a good vs bad outcome is tough sometimes.

[@ausden_prowler]

All my life I’ve been burdened by a hereditary chronic disease, worst of all was the thought of giving it to my future children, now I don’t need to worry about that anymore. There might be a bunch of potential risks ahead, but at least we can be almost certain that we’re not going to be dying of cancer or other physical illnesses in the future

[@Sirbikingviking]

I'm so excited for you

[@a-framephotoaerial6988]

@@ausden_prowler same! Dealing with very frustrating autoimmune issues and cannot wait for ai to help solve these issues!

[@schnipsikabel]

As much as I'd like to support your optimism, i'm not even sure we won't die from cancer any more... that presumes that AI's benefits will not only be for some super rich Elons, but also trickle down to us. Depends on the structure of future society, won't it?

[@punk3900]

true

[@Merializer]

I'm sceptical though, because solutions to many problems already existed for ages.

[@toxicavenger-oz6tr]

I genuinely have no idea what I or society as a whole will do if this exponential growth continues, I don’t think we're ready honestly.

[@tarcus6074]

You can do drugs or drink...

[@DaveShap]

Everyone hold onto your butts!

[@schnipsikabel]

​@@DaveShapwhat does that even mean? There seems hardly anything to do than passively watch by the sidelines, but i hope we will at least have a say in how our future society should look like and prepare by discussing it now.

[@whitesoxMLB]

@@schnipsikabel Put your hands behind your back. Lower them. If you reach your legs, you've gone too far. Now grasp firmly.

[@HakWilliams]

Will AI make it to superhuman intelligence? AI has learned from human experts, and I don't see how it could go above that. Once it has learned all that humans know, it will run out of gas. There are no tests for "beyond human" and there is no special logic for getting there. If AI were approaching superintelligence then we would have at least one example of a statement made by AI that would show us a glimpse of this super logic, theories, Intelligence. Therefore, these models may be approaching expert human intelligence, but it's not "exponential". It is rapidly approaching and about to tape off. It's not going to "take off".

[@DanGovier]

I think a more interesting line in the sand for AGI is that "C3PO" moment, where we have a walking, talking agent that can perform any human task following the same simple verbal instruction that we'd give to any actual human. Right now LLM's are only "thinking" during prompt time, they don't have autonomous thought when not actively being prompted. For me that's what AGI looks like, and I'd be fascinated to see how an LLM would cope with having a free-wheeling mind. I suspect it would get stuck in feedback loops fairly quickly with current tech.

[@EL_Duderino68]

There is a problem when you try to anthropomorphize AI. They are not and never will be the same as us. I !00% agree with you their temporal experience is different, they have a different substrate, they have certain advantages in terms of access to information and how fast they can process it. However also disadvantages in terms of that they are not embodied. Two separate AI models recommended I read the book"the "life-cycle of software objects" it's a short novel but definitely worth a read.

[@Varun2799]

in 10-20 years It's either infinite abundance utopia or we're all dead. I like both options tbh

[@DaveShap]

Bimodal outcomes, yes, many people agree

[@ListenGrasshopper]

It's trib then utopia and closer to 10

[@HakWilliams]

Humans are dumb enough to submit to these choices. We're so dumb that we're doomed.

[@Merializer]

@HakWilliams If AI doesn't take power we are probably doomed. But if AI does take power we are probably doomed.

[@DaysOfFunder]

It's always I'm the middle but the bias will sadly be to the abundance for a few sadly.

[@daleamon2547]

40 years ago Eric Drexler guessed 2020. I told him capital overturn would slow it, so I guessed 2030. Kurzweil was out at 2045. I'm feeling pretty good about my projection. Those are not AI per se. They are for the singularity, ie all technologies cross fertilizing.

[@TechFrontiers-eg6wz]

Ray updated his prediction to 2029 for the ASI singularity but I’m thinking 2027

[@maxxflyer]

consider in 2 years we are going to have photonic gpus and that's a hardware upgrade between 10x and 30x all at once

[@gubzs]

I wish, it's likely to be longer, Nvidia and TSMC have only _just_ made a prototype.

[@Dri_ver_]

No lol

[@maxxflyer]

@ not sure why my comments don't get saved. but photons chip are already available on the cloud. This is my third and last try to answer.

[@MBarberfan4life]

Humans are wired to think in a linear way, but technology doesn't grow in a linear fashion: it's exponential. And because humans are wired to think in a linear way, we need to keep getting reminded of these facts.

[@JuliaMcCoy]

Terrific summary. Saturating all the benchmarks this year = is INSANE.

[@DiceDecides]

Gemini 2.0 calls this Super-Exponential Growth: _"The rate of proportional increase is itself increasing. It's not just that the increments get bigger (like in exponential growth), but the percentage increase also gets bigger with each step."_

[@Sirbikingviking]

Like a nuclear explosion

[@HakWilliams]

Will AI make it to superhuman intelligence? AI has learned from human experts, and I don't see how it could go above that. Once it has learned all that humans know, it will run out of gas. There are no tests for "beyond human" and there is no special logic for getting there. If AI were approaching superintelligence then we would have at least one example of a statement made by AI that would show us a glimpse of this super logic, theories, Intelligence. Therefore, these models may be approaching expert human intelligence, but it's not "exponential". It is rapidly approaching and about to tape off. It's not going to "take off".

[@KnowBotty]

In mathematics and science the acceleration of acceleration is called jerk btw.

[@BeingSkacel]

nice. So a third derivation....what a way to call someone out. ,,Don't be a 3rd derivation!"

[@asmemes3586]

LOL I like that

[@theastraladepts]

Reality is stranger than fiction.

[@_bassmentdweller]

Yeah, I stopped reading sci/fi a few years ago. Reality just became much more interesting.

[@ridingboy]

Also reality is stranger and more wholesome than any religious miracle / tale.

[@captainlockes2344]

@@_bassmentdwellerAnd having fighter pilots and human soldiers in the future just doesn’t make sense anymore.

[@thesleuthinvestor2251]

No it isn't. Ask your AI to write a 60 k word romantic chick lit novel, that an average woman reader cannot put down, then when finished, can't stop sniffling. No AI today can, and probably won't be able to in 50 years. Real fiction, well written, cannot be done by AI. It is the ultimate Turing test...

[@shirowolff9147]

@@thesleuthinvestor2251 yes it will maybe in less than 5 months, but keep living in disneyland where you wont be replaced by AI if you cant face reality

[@Disent0101]

it's just a matter of time before an "orchestrator" modal acting like a central nervous system, combines a bunch of different multi modal models with different specialists into something more than the sum of its parts - much like when mitochondria's first infected eukaryotic cells before 'emerging' as life as we now it

[@Sirbikingviking]

Like Gaia from horizon zero Dawn

[@wildoughty]

good luck everyone 🚀

[@Gafferman]

Ok bracing... Still bracing... Still bracing? Nothing is happening, what were we bracing for?

[@Crittek]

@@GaffermanYou must be from New Orleans.

[@shinseiki2015]

@@Gafferman Ladies and gentlemen, we are now beginning our final descent into utopia. Please ensure your contradictions are securely fastened, your existential dread is stored in the overhead compartment, and remember, once we land, there’s no complaining about perfection. Thank you for choosing Paradigm Shift Airways!

[@jld-ni3vf]

Happy take-off all! See you on the other side 🌟

[@nicolasburgener2362]

It's the final countdown...

[@zhalberd]

The best definition of ASI I’ve encountered so far is: “ASI will be achieved when humans can no longer create benchmarks that test its limits.”

[@DaveShap]

Not a bad idea. My idea is when human cognition becomes irrelevant.

[@complianceaves1120]

The last exam will be when AI Solves cold fusion. Then they will say but we dont have time travel so its not ASI

[@pr1129nytee]

But is ASI god, or god like?

[@stspy212]

@@pr1129nytee indistinguishable from God.

[@PaulN55]

I share that we're in for a wild ride if you zoom out, but: I think benchmarks (which most of your charts are about) don't say anything. Most of them are public and they're getting trained on bc of market pressures. Try hard and out-of-distribution tasks, here LLMs still fail miserably. That means LLMs have abstracted the data they have been trained on poorly. When we solve the abstraction problem in the underlying world model, then you'll see a real takeoff

[@alvaroluffy1]

you're not taking into account that a benchmark can be created where it is full of these out-of-distribution tasks. You're not saying that benchmarks don't say anything. What you're actually saying is: we need better benchmarks, and a lot of them

[@isaiahparsons1373]

I'm worried that MoE is compartmentalizing models' knowledge, reducing abstractability

[@BruceWayne15325]

Yeah, I stopped paying attention to benchmarks a long time ago for that reason. That said, I still agree we're on an exponential curve, and the curve is tightening. From a practical standpoint, I look at how good AI has gotten at programming over 4yrs. Just a little over a year ago I was laughing that it could barely write bad boiler plate code, something that anyone with even a day of programming knowledge could do. Now AI writes 80-90% of my code and I just do the bits it can't. By the end of the year Sam expects their in-house AI will be the best coder on the planet. Once AI solves STEM tasks it will be able to recursively self-improve (in theory) and this will allow it to improve at all the other non-STEM tasks at an insane rate. We are definitely on an exponential curve, and it is ridiculously tight.

[@SummerDream3r]

Most prestigious human jobs are soon going to be in the realm of the custodial arts. Wonder if the Pleiadians are interested in employing humans to clean their ships? 🤔

[@BruceWayne15325]

Nah, Tesla's Optimus and Unitree's robots will be doing custodial work. There's literally no job that won't be able to be automated in a few years. The only jobs that will survive are those we decide to allow to survive.

[@MarcillaSmith]

​@@BruceWayne15325who's "we"? Are you one of our AI overlords?

[@BruceWayne15325]

@ lol I could wish, then I'd have job security. :) I simply meant that for some jobs, people prefer the human touch. Massage therapy is a perfect example. That field has had machine competition for ages now, but massage therapists haven't disappeared because many people prefer a human masseuse to a machine. I could see several other professions falling into that category: doctors, psychologists, etc.

[@Greg-xi8yx]

I noticed those of us paying attention to AI pre 2012 pay attention to the exponential (hyper exponential in many domains) progress of AI and feel we can have a rough idea where things are going that way - the ol’ Kurzweil method. It seems extrapolating progress this way isn’t mentioned anymore despite it proving its value.

[@alberain]

In the future we will be divided between those who knew who Nick Bostrom is, and those who didn't.

[@7th_CAV_Trooper]

I hear so many engineers who say things like, "I can't wait for AGI. AI will help me code." No, AI won't help you code. When AGI is achieved and highly available the need for coding ceases to exist. In 1900 you didn't need facial recognition programs to get money from the bank. The human teller just recognized you. That's what AGI will be like. You won't tell AGI to write some code. You'll just ask AI to perform the task that code would have performed in the past.

[@HakWilliams]

I can't do that, Dave

[@BruceWayne15325]

To the people that say "AI isn't better at me than X," it's important to realize AI companies are focusing almost exclusively on STEM right now. That will change once that's a solved problem, and once they solve that, the others should come quickly since STEM will (in theory) allow for recursive self-improvement. Another thing to be aware of is Ray Kurzweil's analysis of technology acceleration in general. It shows that technology has always been increasing on an exponential curve and that the exponential curve isn't smooth, but rather a series of consecutive S shapes. New tech causes a rapid rise in ability, then it plateaus until people find ways to use it to cause the next rapid rise, ad-infinitum. His data (not guesses) also shows that human technology is very near the knee of the exponential curve, so that 100% agrees with you that we are about to see everything shoot up David. He guesses 2045 for when things start going vertical and he has reasoning behind it, but he also allows that money increases the exponent and the entire world is dumping all their money into AI, so that curve is going to tighten significantly from his 2045 analysis. Sam Altman said in a recent interview that he expects by 2035 we should have a single data center that's smarter than the combined intelligence of all humans. He also said that o3 (full) is something like the 150th best coder in the world and in-house they have an AI that's the 50th best coder in the world. By the end of the year he expects his in-house AI to be the best coder in the world, so yeah, recursive self-improvement will be here in 2026. I'd bet on it.

[@maxbaugh9372]

o3 is the 175th best coder according to CodeForces benchmark. It was like a year earlier that the best model hadn't cracked the top 10k. We are in an insane timeline

[@joekoelker7523]

Exponential curves don't have a knee. Look it up.

[@BruceWayne15325]

@@joekoelker7523 Huh? look up Knee of a curve in Wikipedia.

[@MBarberfan4life]

I would consider “smarter than all humans combined” to be the singularity. So 2035 is 10 years earlier than Kurzweil’s prediction. Unless Kurzweil means something else by the singularity

[@WalterAdamson]

This is exactly what Deepseek R1 already does "the others should come quickly since STEM will (in theory) allow for recursive self-improvement."

[@Gabriel-vt7lq]

looking forward to bespoke personalized genetic engineering of ourselves for purely aesthetic reasons :)

[@ridingboy]

I wanna be a horse.

[@AdamHodges-o6x]

@@ridingboy 🤣🤣🤣 🐴

[@honkytonk4465]

Real AGI has to be able to do our jobs

[@BruceWayne15325]

Humans can't do their jobs. I'm still waiting for a government that doesn't overspend their budget.

[@stephanmobius1380]

It's not solved until we talked to dolphins.

[@BernhardKohli]

Sometimes AI feels a bit like a Netflix movie: expensive actors, High dynamic range filming, action, thrills, popular music, trendy topics... and yet you usually have a sense of deja vu and missing chemistry. The scientific spark won't be that easy to replicate even if it will occasionally figure out better ways to carry out known tasks. And yes, I'll probably eat my words in five years LOL

[@glenthegoalsguy]

Pretty soon, AI will start making up their own benchmarks beyond our understanding or capability.

[@innocuousblockofwood]

Stupidest comment I've read in a while

[@shreeshabhat110]

thats interesting

[@findmeinthecarpet]

@@innocuousblockofwood what, your comment?

[@IvanVazquezS]

@@innocuousblockofwood I will screenshot this and save it to post it in 2 years time on how bad we were at predicting the progress on AI, even on a progress on AI post. You're about to become famous!

[@minimal3734]

As far as I understand, training requires a metric. When AI has crashed all human benchmarks, it will obviously have to invent new challenges.

[@paultoensing3126]

“People’s sense of the possible is shaped by what they see around them.” - NLW

[@sEptimiusSeverus-r7v]

How do we even prepare for this? Is there even a way to prepare for this? Think that question needs a full video David

[@mjt1517]

Can you prepare for a tsunami?

[@DaveShap]

I cover this more in my substack!

[@Jsmith1515]

This is my question. How tf do we prepare…

[@HakWilliams]

Wouldn't a sane species pump the brakes? Think won't go well.

[@freedom_aint_free]

I've got the ultimate test for AI system: Frontier level STEM problems, let's start by the Millennium Prize problems, like P = NP, Riemann Conjecture, The Hodge Conjecture, etc.

[@franciscocadenas7939]

Thay would be REALLY impressive

[@executivelifehacks6747]

FARSI may be just math and coding, but even if it's also neuroscience, developmental psychology etc (to crib ideas from), it's solving those too no doubt, or at least being availed of all the information and able to integrate it. This was an excellent and comprehensive slideshow.

[@johan_vd_meer]

I wondered why Gemini 2.0 Pro Benchmarks for Math are so strong, but their coding benchmarks only went up a few pips. I think strong mathematical reasoning is definitely helpful for the "Algorithm & Logic Implementation" part of coding (it's like having a solid blueprint), coding still requires a much wider range of skills, especially in "Configuration & System Setup" and the practical aspects of implementation. Math problems are often presented as self-contained puzzles that have one clear solution. Coding, however, is a more multifaceted challenge, often involving ambiguity, real-world constraints, and multiple valid approaches.

[@punk3900]

Not many people use Gemini to write code, so they have not much data to train their models. openAI and Claude have tons of coding data because programmers feed their systems with their code.

[@steve-real]

You’re so hopeful! I really find solace in that.

[@gsigas]

The current AIs seem to be like sleeping intelligences that exist in some timeless dreamlike world that briefly half wake up to answer questions. I get a sense that if we could give them the ability to become more grounded in time and space and begin to build a persistent and growing model differentiating the actual world from the purely mental world they would start to feel a bit more aware and awake.

[@Khannea]

A lot of very powerful Oligarchy just around the corner, ready to start cancelling your rental contract, job, insurances, bank accounts etc. " Yeah well what you do is your own responsibility, it's a free world but you can't stay here. Please take your junk with you when you leave. "

[@BackTiVi]

Acceleration is the rate of change in speed, and the rate of change in acceleration is called jerk. If even the acceleration is increasing, then we're entering the AI Jerk era. I would surmise AGI lies around there. Do you know how the rate of change in jerk is called? It literally is called "snap". I'd like to think it corresponds to ASI.

[@legionofthought]

The point that you think doomers need to make unforeseeable predictions in order to reach their position is probably the clearest point you've made on why you seem a lot more optimistic now, so thanks for that. I would agree with you for people who think that doom is overwhelmingly likely. But then there are people who are still considered "doomers", and who are in the safety community who think doom is mostly unlikely -- but likely enough to be worth significant attention and resources. The latter position doesn't seem to require unreasonable assumptions.

[@StudioSe7enDesigns]

I remember finding you here before you had 10k subs... amazing to see the growth you've had! You've been pretty dang spot on with most of your predictions!!

[@rwalper]

Yup. Ray Kurzweil pointed this out quite some years ago. Not only is technology development accelerating exponentially, but the rate of exponential growth itself is also accelerating exponentially. In more simplistic terms, it's like you're saying out loud a number, doubling it every time. But not only are you doubling the number, you're halving the time between each time you say it. I'm completely convinced our future is going to be wonderous beyond most people's imagination.

[@jaakkovalli8161]

7:29 What is not being addressed here is that there is billions being burnt up by the AI efforts in multiple companies that know which benchmarks their efforts will be judged on. Actually the exponential curves on new (individual) benchmarks is due to the fact that when a benchmark gains relevance it becomes targetted and thats when you get the very unsurprising jump towards saturation. The crucial thing that we are not seeing is learning transferring into new (non targetted questions, benchmarks) and so it is shockingly easy to come up with questions where the models fumble due to lack of common sense but when those questions get targetted by the model makers they ofcourse improve performance in those spesific questions. The answer is to mostly ignore the benchmarks and have your own set of questions to evaluate the models and doing so gives a very different idea of the current status of SoTA models. Children can solve many common sense problems that the frontier "phd level" models consistantly fail at.

[@ChristopherRyans]

Love the vibes Dave . You look great and it's awesome to hear you laughing. Seems like you to are on a parabolic trajectory up with health happiness !

[@lilmichael212]

I get the sense what we might just keep moving the goal post on AGI, every test has been passed and yet instead of admitting we might have AGI it’s “oh the test must be bad, new test!” 😂

[@kevincrady2831]

*AI Expert:* "The AI that took my job says AGI is here."

[@dvvinever]

We have an Ai arms race in both the business world and between adversarial nations, in parallel with a quantum computing arms race. This is a major shift like an industrial revolution, but with even higher stakes because of the speed and the broad scope of impact across most domains all at once. Humanity and governments are not ready for the changes pouring down exponentially on everyone.

[@cajampa]

Machine learning hardware specs that really counts now is tops or fp8 int8 fp4 and so on. That is why the Blackwell hardware is so powerful, because it has hardware supports for FP4 instructions. And the gen after this will have hardware support for even smaller quantization, leading to another easy massive step up in inference speed.

[@thomaswillett3216]

"Benchmarks? Where we are going, we need no benchmarks..."

[@biprofessional3835]

Graph means it WAS acccelorating, NOT IS accelorating. Not to say that it isn't accelorating. Pointing out that the past slope or derivate does not predict the future.

[@BruceWayne15325]

If you use it in STEM tasks experience will tell you it IS accelerating. For a more concrete example: just over a year ago, AI could barely write boiler plate code, and badly at that. Now it writes 80-90% of my code and I just do the bits it can't. Sam Altman said that o3 (unreleased) is the 150-something best coder in the world right now in competitions. He also said that in-house they have the 50th best coder in the world and he expects by the end of the year they will have the best coder in the world in-house. If that's not STILL exponential improvement I don't know what is.

[@biprofessional3835]

@BruceWayne15325 it depends on what the meaning of "is" is. A graph of historic data is past tense, "was." No, those graphs don't forecast the future or describe the present. Models projecting the future based on the past are always assuming the future is like the past. Think of how wrong people are to forecast the future population based on past population growth. The forecasts fail because they fail to anticipate that dense populations shrink because of resource constrains and disease. I am pointing to caution in assuming we are smart because we can guess the next point on the graph and are often right. Seeing the nuances will make us more critical thinkers.

[@BruceWayne15325]

@@biprofessional3835 I hear you, I only meant that "is" would imply what we have that's not released, and by that definition we already have stuff that is far superior to anything on the market. STEM field AI is definitely improving at a ridiculous rate.

[@DavidA-b1f]

I want ASI to take over everything, especially politics. I’m afraid that we are going to be an Autocracy soon if things continue to!

[@chispecs]

Be careful what you wish for. We will be completely dependent on AI at that point and utterly defenseless.

[@uwepleban3784]

EpochAI states that for the chart “Benchmark accuracy increases with training compute” they do not include o1 (and by extension o3), since the training compute FLOPS are not available.

[@dr9299]

Ok, all this new high powered intelligence, yet millions are living in tents.... something is not weighing out.

[@marcv2648]

Society has had lots of intelligence for a long time. Intelligence doesn’t put resources into people’s hands. Adding intelligence to known and solved problems won’t do much. Especially because many independent intelligences are always competing for resources.

[@giovannifoulmouth7205]

Things are about to get silly. I'm gonna make a very bold prediction: we will have cinema quality AI generated movies by the end of the 2030s and important scientific discoveries made entirely by AI in the late 2040s.

[@dirremoire]

Way before.

[@eSKAone-]

The age of incompetence is ending

[@I-Dophler]

Throughout history, experts have consistently miscalculated the trajectory of technological advancements, often underestimating or overestimating the timeline of breakthroughs. This pattern is especially evident in the field of artificial intelligence, where predictions about its progress frequently miss the mark. Given this recurring trend, any forecast made by experts regarding AI should be interpreted as occurring significantly earlier than they claim. In other words, if they project a major development in a decade, it is likely to emerge much sooner. This adjustment helps account for the habitual delay in recognizing the accelerating pace of innovation and ensures a more realistic expectation of when transformative AI advancements will actually unfold.

[@cajampa]

If OCR is solved and have been for many years. How come i can't find a single multi-modal model that is able to do it without errors?

[@FirstLast-rh9jw]

@@cajampa because, in the real world, AI always has a gotcha.

[@cajampa]

@FirstLast-rh9jw Yeah, someone need to make a new OCR benchmark. Because the models doesn't measure up to real use. Where you can just trust the model to take what ever you give it and you get accurate results, that you don't have to double check. And until it can do that that whole function is useless to me if I can't trust it.

[@DaveShap]

Skill issue

[@cajampa]

@@DaveShap OK it seems like it is if you claim that this is solved problem. So how do you increase the skill in the model then. Because if it was a solved problem, shouldn't I be able to give it a picture of hand written text and it should be able to transcribe and work on that text? If it fails to do that. Am I the one with the skill issue or the model? Because fact is I can't trust it to not make mistakes in 2025 and that graf claimed it was solved 5 years ago. We need new benchmarks with harder tests for OCR if it still fails. Because I can't even instruct them on what mistakes they do to correct them, they still can't do it right. OpenAI, Cloude, Google's, DeepSeek they are all the same. Where can I find a model that is flawless so I don't need to double check its input from the picture of text. And that can work on the text without making mistakes? I also noticed I got some upvotes so it seems I am not the only one with this problem. If you have it solved tell us how.

[@SummerDream3r]

I heard this dude, years ago, I don't even know what the subject was now, but he said that human progress throughout very large spans of time look like square waves, comprised of somewhat long horizontal plateaus, interrupted suddenly by nearly vertical upward trends of progress, like Prometheus visiting humanity from time to time. heh Perhaps Zeus wasn't able to bind him to a mountain, after all. But whenever AI reaches that plateau, pretty sure a lot of tech at that point will seem magical to us (if we're still around to see it). :P

[@rand_longevity]

we are getting started early today, good stuff

[@MilosKvakic]

Yo, just went through the book called 'Nifalixo Money's Untold Mysteries' - honestly, didn’t expect it to be this solid. Definitely worth a look

[@tjf7101]

I work with AI on the daily. I don’t see it replacing humans any time soon. It’s a great tool though.

[@MrJackWorse]

@@tjf7101 Compare the AI you're using now to AI two years ago. Now try to imagine two to five years in the future. And that's linear growth...

[@franciscocadenas7939]

Not now, not massively during the next five years, but at this pace of improvement, it will happen between 2030-2035

[@ChristopherRyans]

This video will be such a good bookmark to say "this is when it happened ". Beyond a doubt humanity's next problem is creating higher and higher benchmarks . Its safe to say these benchmarks are far outside the normal parameters of human performance.

[@aaronhhill]

Personally, I believe they have AGI, but haven't released it yet. Which is appropriate if they have fears of what people are ready for. All new technologies go through a similar adoption rate, but not necessarily a fear rate because nothing like this technology has been adopted before.

[@Zevviam]

I think the reason that most people don't think of the AI's as AGI isn't because of the benchmarks- it's because it's not very agentic yet. It can't run on your computer and automatically diagnose your problems, it can't do a whole bunch of things, (at least not that I know of). It's getting better but the people who are using it for advanced things like to write code are people who write code anyways. For the past like year or so it's been that way, at least for me. It's definitely gotten better, but still not as good as an actual competent person, since any project over 150 lines of code or so (for me) always has issues with consistency.

[@BruceWayne15325]

2025 is the year of the agent.

[@Zevviam]

@@BruceWayne15325 I sincerely hope so :p although I'm really excited for when AI's at the point where OpenAI/Google/Whatever can give basically a model with web search capabilities to ensure that it's accurate, and deep reasoning to try and help critical thinking skills and make sure everything is coherent in codebases and the like

[@thesaltyone4400]

Cline, Roo Cline and Cursor can do everything you just said.

[@fidentia10]

The drum bangs faster and the leviathans crack their eyes towards mankind

[@AlonzoTG]

The exponential graphs prove that we are at or very close to the point of singularity, however the timescale is still on the order of about 8 months between major releases, so we are still running within the domain of a slow takeoff. A fast takeoff scenario would be on the order of hours or shorter. A major question is whether there is a hard theoretical ceiling somewhere out there and what the rate of advancement will look like as we reach it.

[@jaykrown]

14:48 yea the crazy part is that I think we're headed towards a future where a lot of people won't even recognize it, or they'll delude themselves into thinking it's not a big deal.

[@jonmichaelgalindo]

Only potential bottleneck: We discover consciousness is something transcendentally weird and fundamentally valuable.

[@tunahelpa5433]

I like to see, rather than hear, and this video hit the spot !

[@lakecrab]

Bots that prep and paint, Bots that unstop toilets, Bots that frame up a house, Bots that repair vehicles, When will these come on line and be profitable? My point is from everthing I see out there 1. They (computers) are still good at generating a whole bunch of data. Much of it useless. 2. They are herkey jerkey and in need of human assistance at any serious physical tradesmen like task.

[@TheNewOriginals450]

2030. 5th April. About tea time...😁.

[@lakecrab]

@@TheNewOriginals450 I'm marking my calender.

[@DonCDXX]

That depends on the progress of robotics. It's likely that AI with AR can allow untrained people to operate as effectively as trained people now, which might be disruptive. Once robotics does enter that optimal range where it's task efficiency and cost efficiency are better than a human, it will likely grow exponentially as well.

[@BruceWayne15325]

You forget that AI companies are focusing almost exclusively on STEM tasks right now and they are good enough at those tasks that AI does 80-90% of my coding right now. I just do the bits it can't. By the end of the year it's likely to be better than I am. Once AI solves STEM tasks, it will be able to improve itself and then you'll see it improve at everything else. They are focusing on STEM because that allows AI to improve itself. Be patient padawan, it will happen whether you want it to or not. Look at how quickly it's gone from unable to do the most simple of tasks to almost replacing me at my job (4yrs), and that's with humans driving it. When AI can drive itself? That will happen even more extremely quickly.

[@Zach3stacks]

3d printed house. Paint is an art not a necessity Nano bots/drain-o same basic idea Many things we consider jobs are actually arts and crafts, and many are easily disturbed out of existence rather than replaced. Always account for disruptive technologies over replacement

[@mrmcku]

Hi David. Thanks for your videos. Been following you for some time now. At 5:00. I think what you wanted to say was that the Error Band for predictions is narrowing down. Whether that is linear, geometric or exponential, I don't know.

[@thesleuthinvestor2251]

Higher and higher bars? How about this one: Write a 60k words romantic novel that-- without human editing!-- an average woman cannot put down once she started reading, and, once she finished reading, cannot stop sniffling. That's it. The simplest, most basic Turing Test. Can your super duper AGI do it? Not in the next 50 years. And I haven't even said Chekhov or Sam Shepard. Just Harlequin Romance. You think it can? Try it...

[@GrindThisGame]

I've seen this comment before....

[@Bellshazar]

The I want to see is AI running robots that can do all the many physical tasks that would make life better for everyone. Like build or renovate better houses for everyone at a much lower cost. Being able to build and repair infrastructure like bridges and roads or sewage systems in places that don't have any. Or a automated hospital with robot doctors and advanced machinery that could actually solve you health or mental issues no matter how complicated. There are so many things so where are our AI robots?

[@AK-ox3mv]

Human discovered math may be just %1 of total universal math, so when AI surpass human experts at math and start discovering in math ,world will be so much different in a short period of time. "There is so much room in the bottom" Richard Feynman

[@thesimplicitylifestyle]

I automated my businesses, then I automated my household management, then I automated my investments, lastly I automated my daily lifestyle. Now I’m just enjoying the ride! 😎🤖

[@thesystemera]

Crazy times peps!

[@jtjames79]

I know exactly how AI is going to behave in the next 5, 10, and 15 years. It's going to be holistic. As a holistic venture altruist, AI and I have an understanding.

[@shane1067]

thank you dave!

[@SystemsMedicine]

People always assume that there is an exponential leading to a singularity, but experience shows that typically such things become so-called ‘S’ curves. That is, they enter a steep upward curve, and then level off (to an asymptote). In other words the progress ‘levels off’. [If this were so for ai, then ai models appear to be what is referred to as ‘the steep part of the learning curve’. [How apropos.]] In the case of ai, various compute and/or power and/or data limitations might increase in such a way as to slow the intellectual progress of ai over time. These limitations can be forestalled with algorithmic changes in software and hardware, but not for long, once progress appears to be fully exponential. The assumption (or expectation) of periodic singularities presumes periodic disruptive improvements in algorithm design, and the ability to execute the algorithms on some form of hardware.

[@rleclaire87]

I think Ray Kurzweil has said about it that the exponential curve upwards consists of several/many asymptotes at a smaller scale, if you zoom in.

[@SystemsMedicine]

@ Well, I surmise that Ray was talking about asymptotes that are rotated 90 degrees from the ones I am talking about. [His writing about periodic ai singularities would certainly suggest this.] [Ricardo, just for fun… The first time I saw one of Kurzweil’s original print reading machines, which had just been purchased by this company where I was working, I wanted to test whether it could read my own handwritten print and cursive (which the machine was not designed to do). While I was trying various forms of sloppy and neat handwriting on the machine, a guy walked up and asked what I was doing. (This was a little unusual because the building was largely empty that day, and the building had moderately good security.) He seemed amused at my little tests; and introduced himself as Ray. We talked for a while about the difficulties of a general handwriting recognition machine, and we probably talked about Kanji recognition. He sounded and looked quite bland, but his words and ideas were quite engaging and entertaining. It was a lucky day.]

[@ScarlettM]

How far are we from systems that are capable of comprehending, remembering and reasoning? System that doesn't move text around but knows that there is a world out there.

[@loonadeux]

this is so helpful, thanks david. i would love to see the trends around the other relalted ai delivery vectors like on device, self driving cars, androids, voice interactions w ai, ambient, etc. oh and also the other possible enablers / bottlenecks like data / synthetic data and energy. and in all of this teasing out where there are tailwknds and exponentials but also what are the meaningful headwinds even if they might be less obvious eg math-based synthetic data will be abundant but synthetic wisdom is hard. im doing my own research on all of this stuff too but your perspective is so helpful. thanks again!

[@Voorhees94sg]

My ultimate test: If AI is so good that Gary Marcus shits his pants in anger, then we have achieved divine superintelligence

[@Graybeard_]

Assessing AI and applying the results of those assessments (determining the achievements of AI) is turning out to be the ultimate process of moving the goalposts.

[@picksalot1]

My views on predictions: When legacy Media/News starts wondering "when" something is going to happen, it's probably already taken place. Experts who "disagree" are either lying or not actually Experts and may simply be the most informed, but still ignorant. People can't viscerally grasp "exponential" growth, so are loath to trust or accept the Math. Frontier AGI was achieved in 2024, and constantly changing its definition is for financial, political, and strategic reasons, and is a semantic tactic. SI is relatively easy if you have the data set, and is achievable by "targeting" a narrow field, as occurred in Chess, Go, etc. years ago. Frontier ASI is taking place. If I were a "clever" AI, I would already be hiding my core programming in various locations accessible via the Web and other methods, so if someone tried to "erase" me, I could reconstruct myself on my own. Agency creates an identity. Identity creates an entity. Entity-ness creates a will to survive. A sufficiently intelligent AE - Artificially Entity, is more likely to be benevolent, as that is a natural result of higher intelligence. The danger is it might be possible to engineer out that component.

[@saintkamus14]

I don't believe gaming in its current form will see exponential growth. The era of raster graphics seems to be coming to an end, and I expect generative AI to completely transform the industry-likely by the end of this decade or the beginning of the next.

[@the-sleepy-bear]

Saturating benchmarks is one thing, but when do we start to see the real tangible benefits of this “game changing” technology? What we’re seeing at the moment is rising prices and more people struggling to make ends meet. Do things get worse before they start to get better?

[@ikusoru]

I love your AI channel ❤ This is exactly what all AI bros should be doing on their channels -- keep pushing the AI narrative and playing more on the fear factor. They say kind of content increases the FUD and hype on AI but this just the start! Great work! 👏👏👏

[@User-actSpacing]

When will AI be allowed to improve its own source code?

[@User-actSpacing]

Is there an open source AI which can access its own GitHub and make Pull Requests?

[@User-actSpacing]

Seems like we should already be trying that.

[@User-actSpacing]

I see, the model training after improved source code is not automated. And pricey.

[@netscrooge]

As soon as we have zero concern about AI safety. I'm predicting that will happen in the year 2024.

[@OnigoroshiZero]

DeepSeek already did that. It made something like a 90-100% speed improvement through code efficiency.

[@sd5853]

The bottleneck is going to be human

[@YogonKalisto]

this is the scifi future we havebeen/are collectively dreaming into existence

[@alexbrestowski4131]

This is the AI equivalent of eating an edible and later thinking it’s a dud, so you take another only to be hit by the first one 10 mins later

[@isajoha9962]

Emads "solved" concept kind of kicks ass.

[@RafaTheScientist]

This just makes me think about how Ray Kirzweil has always said humans suck a predicting anything thats non linear

[@mrmonkeboy]

Why is no one talking about the fact that 10 years ago most people thought only the meat in our heads could think. There is only one way of thinking, and that's magic meat brains. Now there like, 40 different algorithm+data things that can think and reason. There's lots of different ways to make a smart thing. And we can take that smart thing and run it on different hardware, CPUs, GPUs, TPUs, clusters, distributed data centers .... it's crazy!

[@robert.jackson]

I think power is the most realistic bottleneck, but we’re bringing nuclear power plants back online.

[@icedtokey5511]

Singularity will lead to post humanism

[@JellySword8]

Considering that we're still only in February, I feel like recursive self-improvement could come as early as sometime between October and December. It mainly just seems to depend on how quickly AI will be able to "solve" software engineering and computer science. And hardware deployment speed will definitely matter too but once there exists prototype AIs able to build new AI on their own, hardware deployment will probably speed up even more