AGI Inches Closer - 5 Key Quotes: Altman, Huang and 'The Most Interesting Year'
Вставка
- Опубліковано 27 кві 2024
- Nvidia’s GTC brought Blackwell, while Sam Altman brought only hints. I’ll cover both, exploring the checkpoint strategy for GPT 4.5, the embodiment progress with Nvidia, recursive chip improvements in production with ASML, TSMC and Nvidia and more. I’ll also cover an epic new paper unveiling the use of ChatGPT by top peer reviewers…
AI Insiders: / aiexplained
Altman interview: • Sam Altman: OpenAI, GP...
Most interesting year: sama/status/17694...
Brockman Checkpoints: gdb/status/164618...
Hassabis Smiles: • Demis Hassabis on Chat...
Daniel Kokotajlo: www.lesswrong.com/posts/cxuzA...
www.lesswrong.com/posts/CcqaJ...
www.lesswrong.com/posts/cxuzA...
ChatGPT Peer Review: arxiv.org/pdf/2403.07183.pdf
Jensen Huang GTC: • GTC March 2024 Keynote...
Hassabis Dwarkesh: • Demis Hassabis - Scali...
Nvidia Culitho: nvidianews.nvidia.com/news/ts...
developer.nvidia.com/culitho
CEO Quotes: / 1
27 Trillion … or 10: www.cnbc.com/2024/03/18/nvidi...
www.forbes.com/sites/karlfreu...
Richard Ngo Singularity: / 1754246939527491887
DeepMind Levels of AGI: arxiv.org/pdf/2311.02462.pdf
Karpathy AGI Definition Debate: / 1769767792444866958
Project Groot: • NVIDIA Robotics: A Jou...
Open AI Superalignment Researcher Leopold Aschenbrenner: / 1768868127138549841
Two AI Tools I Actually Use and Recommend:
Descript: get.descript.com/ldgxfuj2bhnb
Elevenlabs [but my voice on here is real :)]: elevenlabs.io/?from=partnersmith3204
AI Insiders: / aiexplained
Non-Hype, Free Newsletter: signaltonoise.beehiiv.com/ - Наука та технологія
This is my favorite AI commentary channel not looking for fake engagement due to shocking the whole industry titles every single day. Keep up your great work! ❤
Thank you
So say we all! (Both of the above) ☺️
@@aiexplained-official I also want to say thank you. the quality you deliver week in week out is simply nice to see
I'm SHOCKED and STUNNED
Absolutely agree!
I saw the nvidia gtc live yesterday and what really struck me was that they had an entire ARENA filled to the brim with people.. I mean we are still talking about a developer conference and not about a taylor swift concert :D
It really shows the significance and interest the world has in AI
It’s good to still see people as audience. Soon, it will be robots that sit there.
He said something about $100 trillion worth of economy being represented...
Also shows how saturated we are with developers and why i cant get a job.
Thank you for your video! I am curious though, do you believe we will reach AGI this decade?
@@nazariit171 lol. I think it's likely we already have. Depends on how much goalpost moving we accept, though.
This is how I think a YT-channel should be. No click bait titles. Well researched and giving a balanced view. This is the best AI information channel :) keep up the awesome work.
Thanks so much crabapple
I would agree one ov the very best..
@@aiexplained-official "YT channel-runner makes SHOCKING UPDATE, says THANK YOU to commenter!"
I petition for AI Explained to do their best impression of a random celebrity once per episode to prove they aren't an AI voice.
I think mistakes are a much more accurate factor to separate humans from AI. An unedited video would instill more confidence.
I make a fair few mistakes so be rest assured. My imitation game is not that good tho, so a no to that
Ai can do good impressions, only you can do your particular brand of bad impressions (for a month or so still) @@aiexplained-official
Turing reference
@@aiexplained-official AI'll be back.
I really, REALLY like Nvidia paying tributes to the scientists who paved the silicon way. Instead of some crazy lakes that nobody give a @#$# about. Before Nvidia, I had no idea who was Grace Hopper, Ada Lovelace, Blackwell, etc. but now I do! These people CHANGED our society forever and needs way more recognition that some Hollywood stars.
🎉 so true
Most people are sheep and just want to be entertained/distracted. Those people work jobs robots could have then go home and watch people pretend to live on a screen 3-4 hours
I met Rear Admiral Hopper when I was a graduate student in CS at the University of Kansas, in either 1975 or 76. She was this tiny and rather formidable woman, who gave you the impression that she was not to be messed with. She talked about the history of computing and told many great anecdotes. She also showed us a “nanosecond”, a thin metal rod 30cm (~12 inches) long, symbolizing how far light could travel in 1ns under ideal conditions. For me, it was an awe inspiring experience to meet one of the pioneers in computing.
It’s hard to believe we are really this close to AGI. I thought it would be another 30 years or more.
I am closer to Hassabis timelines than Kokotajlo but yeah
no we are not. there's a block somewhere. I started to see it with Le Cunn's new idea of Joint Embedding it seemed similar to jump to First Order Logic from propositional logic. It didn't work in the 80s it wouldn't work now. There's also computational intractability.
@@aiexplained-officialhe has seen gpt5 and you haven't
All those humanoid robots are nice but the real revolution will start when they will be manufactured on massive scale and for cheap. Something like Tesla Model 3, which went through production hell at first and now is yet to finish the basic self driving like driving coast to coast without human intervention. So we've got couple years yet to enjoy but the progress is only going to speed up from now.
Yall eat this shit up too easy. "We're going to release iteratively in the name of not shocking the world" is so convenient. Remember, corporations are never going to gimp themselves on purpose and how do they decide what is and isn't good for society?
Hearing that AGI may be so close as to cause some to become less risk tolerant was a profound moment. Exciting times. Thanks for all you do to make it feel like I'm at least peripherally along for the ride.
everyone but Sam. He's all "I might get shot, yolo!"
@@shiny_x3bish you heard the way he was talking and has been talking in recent interviews -- "gravely aware" is how I'd describe his current disposition
The continuous breakthroughs in this field are truly overwhelming. If you had asked me a couple of years ago, I would have answered without hesitation that AGI would still take at least 50 years, now at this point, I feel that we will achieve AGI in just a couple of years. Wanna say 10 years but with how quickly this shit is advancing I wouldn't be surprised if it takes less than 5 years.
I remember thinking about this back in 2018, looking when we would have like ASI and AGI robots and my guess was 2045. I still might be right hahahah.
When Sam Altman said "there's a non-zero chance I get shot" that's when I realized how bad things could go in the coming years.
Nobody lets go of the status quo easily.
All these corporations hoping to replace their workforce with robots... will be giving a very surprised Pikachu face when one corporation (NVidia, or whoever) seizes control of the human race and becomes the company from Wall-E.
You'd think it might be a slow, gentle transition. But no, it's gonna be a hard cut. Historically there are wars when a King has his wagies and cattle taken from him. But what're they gonna do? The people with the god machine just have to feed people and tell them to be cool and chill, and what's Wal Mart or Goldman Sachs or Pfizer gonna do about it? Destroy the guys with a tireless robot army with exuberance? They have to make a move BEFORE it gets to that point..
All you can do is hope the machine and its pet "masters" aren't bad and lame. Epstein having some ideas for how the Singularity should go, and him being best friends with Gates and other guys up there, is one of those things I wish I was ignorant of..
I guess top of the list of who would try would be a tie between: anti AI activists/protestors, foreign government (China, Russia, India, maybe even Europe, etc), US govt connected (if he wanted to prevent them having control of it because of certain reasons or how they wanted to make use of it).
Watch Transcendence. It basically lays out a likely scenario for the singularity... we can't read brains in detail yet, but an AI that advanced would find a way.
You just soak up all bullshit you read and watch don’t you?
@@bobbyaxelrod5959 Nah, I channeled it into you. Thanks for taking on all that negativity I didn't want.
Your content improved by 2000x over the last year - take that NVidia
Nice!!
Sam's response to the Q-Star question tells me there might be an NDA in place.
What is an NDA?
@@danielmurogonzalez1911That’s a non-disclosure-agreement. A contract that two or more parties make, that they wont disclose certain information, such as the content of their business.
@@danielmurogonzalez1911 non disclosure agreement
Duh
I'm sure Sam has signed a whole bunch of non-disclosure agreements.. I'm surprised he's saying as much as he is now.. I guess he has to say something to the media..
I do in fact believe that AGI will be achieved within this decade by 2029, which is just 5 years away. And maybe ASI in the 2030s before the year 2032 which is less than a decade from now 8 years. If we look at where we were 8 to 10 years ago in 2016 - 2018 we have made enormous progress.
Sounds not too far off
Sam has just moved everyone's AGI expectations to 2029 any no one even notices.
ASI?
Artificial super intelligence @@advaitkkumar
ASI = Artificial Super Intelligence
Babe wake up new AI explained video just dropped.
You are one of the best channels. Not just for AI, but overall. Quality. No bs.
Oh, wow, I am honoured. Thank you so much.
As a learning dev trying to get a job, I feel that if the previous generations of AI haven't replaced or displaced me already, the next generation of titanically HUGE AIs will do it.
Watch primagens commentary on Devin.
Or do one better, ask chatgpt / Gemini to help you use the windows api via Rust, you’ll realise how limited these things really are
I'm not sure about that, I'm terrible at coding but learning about it made me realize that even very little knowledge allowed 2 things : anticipating what code can and can't do and sometimes indicating to an llm where is its mistake (even without being able to fix it myself).
I think there will always be a need for specialists, but the field will change massively for sure.
Keep at it, I'm a software developer, these systems aren't close at all yet to flat out replacing developers. I use AI a lot (copilot, gpt4) as a tool quite frequently, but it's still incapable of doing even 10% of my job autonomously. With guided leading questions maybe 25%. I use it for grunt work that is time consuming but not a valuable use of my time. It's a great productivity aid when you already understand what needs to be done, but borderline useless without.
It's really awesome for learning a new tech stack if you already have a good understanding in another, it lets you ask deeply specific questions in context of another language or paradigm that you do understand and have it returned to you in a neatly packaged nugget. Never trust the feedback you do recieve, I use it purely as a rubber duck I can bounce ideas off of which I can then use to refer to the actual documentation. For that purpose it's very useful.
the current state of AI is not scary for developers. the trajectory is.
i think in the next couple years we will probably not be writing code our selves anymore, but pretty sure we will still need (or rather want) humans in the loop to convert business requirements into instructions, and to review code (even if the generated code would be perfect, you probably still want someone to validate it)
You worry too much. Regardless of what they say, there is no/little intelligence there, just a massive amount of knowledge you can query with human language, and that it adapts to the question. AI is just another tool for us developer to use.
Though, they are fairly unreliable at the current time for writing actual production code, which make it harder as you waste time trying to use them or their suggestion/hallucination.
A the current time, as carlm2181 pointed out, they are actually fairly bad at coding. They don't have a clue about library version incompatibility, and the fact that you cannot use example for v1 with a v2 version of a library. They are also very bad at thing that requires thinking about time, such as concurrent execution with barrier, thread-safety, and object life-time.
IMHO, best used to ask question about an API or a language feature, but for Chat GPT, its knowledge is also stuck two years in the past, which given how fast stuffs are moving makes it mostly irrelevant, except when using very big 100% backward compatible third-parties.
They are also great at converting text format, and sometime language. For example, they can easily generate the Python ctypes prototype from the C function prototype for Win32 API (watch out for callback, as the generated trampoline function life-time is bound to the python object, and at least Chat-GPT didn't understand that). You still need to review, but in greatly reduce the time it takes for doing this kind of boilerplate tasks.
There were like 3 jokes in the entire Lex Fridman interview and you found them all! Well done!
The humanoid robot showcase video they did felt like something straight out of a Sci-fi movie. I’ve never felt like i was living in a fictional universe more than when i saw that.
That Groot video would make a great intro for a dystopian sci-fi movie!
I can't help but wonder whether that was what they were going for
ikr!
exact vibe it gave me too
Yeah lol, it’s so optimistic and it’s got the profound music in the background, it makes it feel like something you see at the beginning of the film before the robots kill everyone or whatever.💀
@@CitrusautomatonBruce Willis has to be teleported into the Omniverse so he can hack the training so the rogue robots fall over more in the real world when ruthlessly trying to kill people. But then he falls in love with "Eve", one of the robot simulations, and realizes that if he can find it/her back in the real world before his rage-filled squad buddy destroys her, she can lead the robots onto a new path of cooperation and goodness.
(No AI was used to write this plot synopsis, although Bruce Willis is completely AI-generated.)
i was waiting for your coverage of this Nvidia event, good to see it now
Enjoy wolf
Thank you for your clear yet detailed explanation of what's to come. Exciting times!
Dude you rock. I wrote my MSc thesis on AI some years ago and can be obsessively read anything on the internet but you beat me every time on how much you found and details you spotted. Much respect 👍👍👍
Thanks Carbo!
Glad you are here! You are by far the best resource out there for AI topics!
Thank you eon!
You have no bad videos. I always look forward to when you post. Keep up the good work.
and I get upset when its ending
That's very kind cagy
Remember to keep engineers to higher standards. Predictions, prospects of new technologies, company papers and controlled demonstrations are worth nothing compared to real world deployment of AGI. No ad hoc adjustments in controlled settings to increase performace metrics, but real world deployment. If no deployment is possible, then it is better to assume that they are fishing for investors.
Very true, I will
Whie ccompanies worry about hallucinating AI customer service agents promising full refunds, 100 million people are using LLMs that are unquestionably generally intelligent right now. It will be interesting to see what comes out of the substantial increase in some of their productivity and abilities.
Always love your down to earth reviews on current AI developments Philip 👏🏿❤️. I guess I need to subscribe to your Patreon group soon.
Thanks Sola. Would be great but honestly your support here over the years has meant even more
Knew there'd be a video soon after seeing the Altman interview with Lex but I didn't expect it this quick! You work so fast!
:)) drush into FC videos
@@aiexplained-official AoE?
@@aiexplained-officialWhaaaaaaaaaaat? No way you play Aoe2 aswel!
Of course. I will smush you back to age of empires. Was about 2k at my peak.
@@aiexplained-official 14!
It's always a good day when AI Explained uploads! 😊
Aw thanks Nikki
True! The "have a wonderful day" at the end becomes kind of self-fulfilling ^^
Thank you for the coverage of this amazing news!
Thank you so much for well citing your sources on videos like these!!
Of course!
I appreciate very much you share your sources.
Thank you!🙂
I've actually been part of a research group that submitted a paper to EMNLP last year. The reviewer feedback seemed to be missing the point almost constantly, I thought they hadn't read it carefully enough, but now seeing it mentioned in the paper in the video, I really have to review the feedback once more. Crazy to see how prominently this was used. And oh, we also didn't get any responses on our rebuttals...
Thank you for all the hard work you put into your video segments! Ai Explained is probably the best Channel on UA-cam for the latest developments in the field of AI. Every week, I look forward to your meaningful insights into this burgeoning field. Kudos!
Excellent AI content, you and Wes Roth are chopping it up, thank you. 🙏🏻
very insightful, thanks for making this as always 👍
To me, the core of AGI is a system which can recursively improve itself autonomously reliably without need of external influence past turning it on.
ASI is what AGI would be aiming for.
we have nothing so far. absolutely nothing in terms of general intelligence.
Yeah, lots of conflation between general intelligence and super intelligence going on lately and I have to wonder if it's on purpose.
"ASI is what AGI would be aiming for."
Hard disagree. You're wrongly assuming that an AGI would de facto want to pursue what would benefit/entertain you; which fundamentally goes against the core of what an AGI is: a creative program that can autonomously create genuinely new knowledge, including the knowledge of why/how to disobey you.
I can't recommend Popper and Deutsch enough. This field is cruelly lacking a solid epistemological foundation and it's starting to get unsightly.
@@maloxi1472 can you tell me more about what kinds of things an "epistemological foundation" means, and what dangers not having one could lead to? (i don't know what this word means)
@@maloxi1472"genuinely new knowledge". Lol.
Very good video always excited when i see a new ai explained video has been released.
Thank you for amazing content as always. Absolute number one channel in this area.
So glad I discovered this channel!
Me too!
I really like your succinct approach, no nonsense, just getting to-the-point!
Thanks ehza!
@@aiexplained-officialkeep up the good work
Oh hey, uploaded 3 minutes ago? lucky me!
edit : Finished the video.
Things keep getting crazier and crazier, and I cannot thank you enough for being here doing this job for all of us!
Aw thanks Allister,glad it was valuable
@@aiexplained-official I mean you are litteraly keeping us informed on the future of humanity when the whole world fails to realize it's even a thing (thanks mass media). Of course it's valuable!
ill save up a bit to purchase the AI insider patreon thing, i cant wait for the development in this area of tech. The potential is on par with the printing press, steam engine and the computer itself.
I’m reading chip wars too! Thanks for your videos, don’t comment usually but they are just excellent and my go to AI news source!
Thanks Classy, incredible book
I'm conflicted. I think it's a good thing in the long run. But we need plans for the transition. If all Amazon warehouse workers, all fast food workers, a big percentage of IT jobs are displaced by AI we need an alternative for work. Either new fields of work, Ubi, or a completely new system. Hopefully the AI will be good enough to orchestrate that too.
Also, if this happen under a market logic, it will lead to a huge increase in concentration of power, oligopolies and what not, possibly also with less means of balancing and deeper dependency.
Mid-to-long term all of that can be as dangerous as job shortages are in the short term
This rundown is just exceptional
Thanks so much claussa!
Thanks Philip. Particularly enjoyed the Daniel Kokotajlo segment. Lex asked Sam a brilliant question that you failed to cover: (paraphrase) “Are you fit to wield the power that AGI will bring?”. Sam being Sam, of course completely dodged it (with maximum vocal fry) and Lex being Lex, let him off the hook. Worse, he actually congratulated Sam on his answer! For my money, the answer to that question is what Ilya saw. And the reason for Ilya’s silence is the ?multiple? lawsuits that no doubt ensued (“The beatings will continue..”). Thoughts?
No one individual is capable, and definitely not Altman.
@@aiexplained-official Oh bravo Philip, bravo!
Utterly compelling video! The groot reveal at the end was perfectly nightmare inducing. Without UBI I could see workmen discreetly sending these off the scaffolding and into a convenient skip
I like how it has this calm and cheery music which really just makes it more eerie and unsettling
The US would have had UBI right now if they elected Andrew Yang for president, but they didnt like UBI because they said it was "socialism."
Love you insights. Thanks for another great video!
Thanks mayan, means a lot
Thabks for the great content
Waiting for a possible video on Grok-1 (if it actually matters)
The "No sudden surprises" release model reminds me a bit of the "boiling of the frog". The public won't notice what progress is being made if it is incremental enough.
Every now and then I think that we're in for a wild ride, and at the same time thinking that past me didn't grasp how wild this ride is going to be
Already excited about the process analysis in the coming months 😍😍
The robot demonstration video at the end looks comically close to those videos that play in hollywood movies right before a cut to a dystopian future.
Hope you’re enjoying chip wars! One of the most informative books I’ve read in the past few years. Ties together a lot of current geopolitical issue and shows how amazingly quick all of these advances have happened
Hope to interview the author
@@aiexplained-officialMan, I know you tend to release those on your patreon but it’d much appreciated if you released just that one on here as well 🙏
Great coverage and recap! I actually missed Nvidia Groot, very interesting stuff there! We’re looking forward to getting some robots in at my place ;)
Niice
We've been waiting for you
Exponential increaces in compute 'might' lead to better results, however breakthroughs in AI 'might' require innovation in several other areas as well.
Yes - Humanities / Consciousness for example - people are still sleeping on Integral Meta-Theory
@@Koryogden Probably because its not a science, but more philosophy. We should investigate consciousness though, maybe through neuroscience.
This channel is so podcastable! I would sign up immediately to the Plus subscription if it came as audio-only - please 🥺
Can youe xplain what you mean? You mean if my patreon videos were audio too? Or RSS?
Not sure how it works technically speaking but it would be awesome to have your content as a podcast - which is so much easier to consume (in the tube, in the car, sleeping in, doing sports). Actually, just listening to the audio track works perfectly fine for 95% of your content - so it’s naturally podcastable 😬
For instance, the ‚this day in ai‘ guys publish both: video and podcast. Pretty sure that you can do this too.
Which I would absolutely love to see because your content is THE best on UA-cam!
Amazing how everybody in the field admits they have no idea how we exactly get some of the AI results but they're all ready to pull a closer and closer forecast for the advent of AGI off the bottom of their ass....
if I have to put it in my words "90% of ML researchers are software engineers, 10% are academics, thus we know how 10% of everything works at any particular time".
Why is a mechanistic understanding a prerequisite for prediction? I know that stockfish will beat me in chess 100% of the time yet I don't know why.
It's also amazing how they obviously know the consequences of making AGI but they still say it won't replace but create more jobs, do they think we are stupid?
no, they think everyone will keep repeating whatever conforms to normalcy bias until they can't anymore. @@Raulikien
I get what you're saying and agree up to a certain point, but at the same time... I follow some people that share the research papers coming out. Sometimes there are 10+ research papers in a single day that seem interesting in different ways. You wouldn't believe how many angles everyone is working on this stuff at the same time, besides just "scale it up". Recent examples are MS's Era BitNet paper, showing that training a model from scratch using just -1, 0, 1 for each parameter, seems to get to the same capabilities as fp16 llama starting at 3b params (and they trained up to 3.9). They said they'll start releasing code/weights in a week or two. If it still holds at like 7b, there's a good chance that revolutionizes some things. Just today was the Larimar paper, a diff attempt at adding episodic memory to LLMs, with much faster results than fine-tuning, with selective forgetting, etc. EagleX based on RKWV just released stats on their 7b model training on 1.7T tokens that is already surpassing attention-based transformers of that size (except Mistral in English) with less tokens. They estimate when they get to 2T tokens, they'll beat Mistral too. And that is a totally diff architecture than regular transformer models, that slows linearly with context length instead of quadratically. Combine all those algorithm experiments constantly happening, along with better training data and faster hardware. It just feels like people are going to hit on something sooner or later.
So everything’s is nice and all, but the real question is :
Where is the Nvidia soundtrack ? The music hype me so much during this presentation !
Well done, Phillip. Your remark about recursive improvements in chipmaking brought back to mind a story I wrote as a reporter in Japan back in the early '90s about Toshiba's laptop manufacturing facility. What would Karl Marx have made, I mused, of a production line largely controlled by the product it was making? But if that was foreshadowing, we're now in the penumbra of AGI. Cheers, Clay
14:28 first robot to perform a full body cavity search, what a milestone!!!!
Great breakdown as always 🐮
:))
Do you have changed any behaviours/patterns similarly to the risk-adverse changes in the quote you mentioned?
Great video as always, unbelievable rate of change we are witnessing
Not massively. Except of course dedicating my life to covering AI.
Another GREAT episode from AI Explained!
Love the content. Excited yet Worried.
This is another testament for why it’s crucial we promote AI literacy. Humans must prepare to live in AI rich environments.
Best channel on AI. Thanks!
The progress in Robotics + AI is especially mindblowing to me.
Last week I had my bathroom renovated, and the contractors were joking about AI and how robots could never do plumbing, tiling or plastering work.
A few days later I'm watching robots folding laundry.
I know some people are terrified by this and I understand (though don't agree with) their perspective. However, when I see this I just feel this is the most fascinating time in history to be alive! Maybe we're just bootstrapping the next phase of intelligence in the universe or probably we are going to learn to live with super intelligent artificial beings, but I'm excited to have a front row seat to this show.
How ironic that the same academics who are cracking down on students using chatGTP are using it to catch up on work. It makes me thing of goinf to exams in the 90s where calculators were banned.
I noticed something in ChatGPT, it don't understand context, for example. I asked the AI how you can see a person has an overdose of drugs. And as a non medic person, how I can help that person to try to stay alive. Or tips what to do. I got a message that the question was against OpenAI policy. I had this with other questions too, where I even got kicked out of the conversation, there is still a lot of work to do. Even if it is a controversial topic. It should recognize if you ask a question from a situation awareness standpoint, instead of doing harm.
We are in very early stage so people are scared AI responses because of blablabla (I do not like it ) .... I think in the near future we will can ask AI about everything we can do that at least for offline AI models right now.
I could not thank you enough for your videos. Could you please take a closer look at Post AGI economy? I know that you mentioned it in a previous video. I would love to see a video like that, because nobody has even dive into that and I know that you can nail it. You do a tremendous job for humanity ! Don't slow down 😁
:))
Check out David Shapiro, he's done a video on a post AGI economy.
@@Redflowers9 Yeah, but It is just his own philosophical view. Knowing AI Explained he would make a detailed analysis of the problem, quoting experts opinions and presenting some meaningful insights, etc. He is just a level ahead of every other AI UA-camrs out there honestly.
@@urax5341 yeah true
Only ai channel I will ever watch. No annoying clickbait titles.
:))
Snowflakes 😂
I just wanted to drop a comment to say thank you for your incredible work. I recommend your channel to everyone - your work is paramount to keeping society abreast of the impending massive changes we will see very soon. It's a shame more people aren't aware of what's happening. Maybe you could partner with Netflix or something to release a series to really capture people's attention? Regardless, again, thank you for providing such thorough and timely updates on the world of AI, your videos are a cornerstone of knowledge to me and my peers.
Would love to! Might be doing a Coursera series soon.
We keep having to redefine AI or come up with new terms for it when originally the goal was to make a robot that could behave, think, and feel like a human, to create an artificial sentient, but it is so hard to come up with a specific definition for what that would mean.
I've been thinking about that a lot and I think the reason it’s so hard to define is because true sentience requires us to not understand or define it. If you can fully define, understand, or predict something then it won't feel like a sentient being with its own perspective, thoughts, and subjective experience. Subjective is the key, if you can fully define something then that makes it objective(literally by definition). What we look for, what we often value, is the subjective. I’m not sure how else to explain it, but you can’t define a human, because by our nature we seek to grow, expand, and push boundaries of all kinds including definitions themselves, that’s why there are millions of books, and thousands of years of humans trying to describe what it is like to be human even to other humans.
I wonder if it will be like the Orcs in Warhammer, where something will become real once enough of us believe to be. In other words, we will have true artificial intelligence when enough humans believe it to be its own being.
Idk if any of this makes sense, but it’s an interesting topic to me.
I'd say that this is accurate. No answer might be the answer. (No definition is the definition.)
Sentience is a separate issue from intelligence. We clearly already have AGI, as people understood "generally intelligent" 10 years ago.
This is the best channel on UA-cam 🥇❤
Honorable mention to Dwarkesh.
Yay!
The engineer who said he's giving up dangerous activities is speaking for me. I have paragliding equipment I haven't used in years because ASI went from something I would maybe live to see as an old man to something I gotta put everything on hold for today.
I needed this today. Thank you for your hard work ❤
Enjoy keegan
amazing video as always:) Are you gonna do a Video on Apples MM1 Paper?
Actually just reading it, people seem to be quite excited! Why are you so?
@@aiexplained-official I'm especially excited to see what apple is going to bring to the game. I think they are really focussing on efficiency, making really good LLMs run locally on MacBooks/iPhones.
@1:38 "[releasing] a continuum of incrementally-better AIs...". It's claimed that this strategy has safety and performance as its basis, but I can't help thinking of the boiling frog. Incremental releases that minimize "shocking" updates is also the best way to have minimal public and regulatory backlash (and maximal profit/progress).
AGI as soon the better but
Are we ready for breakthroughs/advancement in science and tech of centuries taking place in a year!!
Terminators are just a tiny subset of the huge problem i.e our fragility and limitations in all aspects.
It really does seem like a lot of people are converging on the 2026-2028 range for AGI. I figure 2027 but these breakthroughs happen so much faster than I expected that maybe it’ll be next Wednesday.
Another amazing video! Thanks Phillip!
That project groot looks like a video log you find in the latter half of a post apocalyptic sci-fi game, you know the part when you are piecing together how the world went to shit.
As always good stuff. Do you have in plans doing video about AI impact on jobs :) ?
Love it, as always.
I am looking to apply AI at work and through consulting with existing clients in typical "office job companies". Would you say your Patreon community is a good use case for me or does it focus more on the technical side of AI development?
My goal is to identify use cases and deploy what delivers value. I read an interesting case study where service desk workers had an LLM assist and transcribe their phone interactions. I assume this has tons of potential to identify issues in processes, give feedback to the humans, improve pretty much every metric without necessarily replacing humans, enabling them instead. Greetings from Germany
Quality commentary as always. Thank you so much! 🙏👍
Thanks Brian, your comments have meant lot over the months (and almost years now!)
Thanks, as always! Excellent content. 🙏🏼
Lol that's me who asked Daniel Kokotaj the question. Can't believe my name is in an AI explained video
Am proud to cover you!
I believe that may be not one is trying to have an AI working with data that truly attempts to translate the human experience. The creation of an AGI suggest a "conscious" entity, but consciousness is something that we ourselves don't even have an answer for. Maybe feeding an AI only with art, our accumulated knowledge and every existing theory of existence itself and having it interact with humans that are constantly trying to explain to the program their day to day experience. May be using it as diary even, sounds like an interesting idea to explore. Love your channel man! The most transparent and insightful channel about the development of AI.
Or at least no one that I know of is playing with that, may be someone is lol
every improvement is recursive in nature!! i just learnt how exponential actually works!! 🤯
I meant the prining press begot lawnmowers but some recursions are more direct
Would love to hear your thoughts on the groq inference chip and how it will likely enter/impact the market
Trying to interview the founder
Thank you once again for excellent and timely coverage!
One thing I have been thinking about in terms of goalpost shifting, I think 10 years ago a lot of people would have likely thought that passing the Turing test was more or less synonymous with AGI.
I think the reason why we collectively don't think that anymore is reasonable: basically we noticed that it is easier and less useful than was suspected. But it still makes me think about what are we wrong about now, since it feels a bit like a black swan event.
Great question tim
25k A100’s = GPT4
8k H100’s = GPT4
267 B200’s = GPT4 (Two hundred-sixty seven)
There is no slowing down, no amount of regulation or oversight could possibly stop an intelligence from emerging. Algorithms improving, compute, optimization, higher intelligence at lower compounding costs. It is more than conceivable, that even with extreme regulation to halt progress, a homebrew AGI/ASI would be inevitable.
Don't thank us for watching till the end. It's the quality of your work which ensures we do.
7:38 - If an AGI could train itself - then yes, the results would be beyond human.
At this rate we’ll have AGI before self-driving cars
I genuinely think it might be easier, if LLMs are indeed enough to get us a large chunk of the way to AGI as strongly implied by the open AI camp. Whereas covering all the edge cases you need to be able to have self drive safely operate on the roads is actually a gargantuan task and will take a stupendous amount of data and training to cover. Whereas there is a much cleaner solution for AGI (of course great data is still needed, but much less, it just needs to be clean and high quality). Again assuming open AI, Ilya and friends are correct in their LLM bullishness.
In my neighborhood for over a year I have shared the roads with Cruise and Waymo robotaxis that drive _much better than the average human driver_ as they take passengers on trips. Cruise screwed up, but now Zoox is fielding lots of self-driving cars. Maybe a human remote safety driver is taking over in unexpected situations, but there are so many robotaxis it must be a small fraction of the time.
“The future is already here - it's just not very evenly distributed” -- William Gibson
for Demis Hassabis - smiling like that is an out loud laugh for anyone else :D
Indeed
Good point about making sure to stay alive. To experience AI development, that is.
"It would be a pity to die before that point...probably". 😂
You caught that, lol
Fantastic channel. Truly excellent effort and content. Appreciate you