AI Agents Take the Wheel: Devin, SIMA, Figure 01 and The Future of Jobs
Вставка
- Опубліковано 13 бер 2024
- Devin, SIMA, Figure 01, all in 24 hours. What does it mean and are AI models taking the wheel? I’ll go through 5 relevant papers and 11 articles to get you all the relevant details, from what exactly Devin accomplished, and didn’t, to DeepMind's new AGI-attempt-in-3D (SIMA) to just how far AI agents have come and what that means for the future of jobs. They’ll also be a guest star … discussing … me?
AI Insiders [Exclusive videos, Discord, Interviews and More]: / aiexplained
Devin: www.cognition-labs.com/blog
Devin YT: • AI trains an AI!
SWE-bench: arxiv.org/pdf/2310.06770.pdf
Cognition Twitter: / with_replies
Reality Check: / 1768056098995814836
Karpathy Tweet: / 1767598414945292695
Bloomberg: www.bloomberg.com/news/articl...
Chollet Prediction: / 1767935813646716976
magic.dev/
SIMA: deepmind.google/discover/blog...
SIMA Paper: storage.googleapis.com/deepmi...
MobileAgent: arxiv.org/pdf/2401.16158.pdf
OpenAI Agent: www.theinformation.com/articl...
Red Dead Redemption AI: arxiv.org/pdf/2403.03186.pdf
RT-X: deepmind.google/discover/blog...
Figure 01 Hz: / 1767928771875868677
MasterPlan: www.figure.ai/master-plan
Unit Cost: www.cnbc.com/2024/02/29/robot...
MMMU: arxiv.org/pdf/2311.16502.pdf
github.com/MMMU-Benchmark/MMMU
Jeff Clune Tweet: / 1768320487627579466
Semianalysis: www.semianalysis.com/p/ai-dat...
Huang AGI Quote: www.reuters.com/technology/nv...
Altman Quote: www.marketingaiinstitute.com/...
US Govt Report: jeffclune?ref_src...
AI Insiders: / aiexplained Non-Hype, Free Newsletter: signaltonoise.beehiiv.com/ - Наука та технологія
This is one of the few channels i trust when it comes to Ai news. Im glad you're not click baity and actually show genuine excitement whilst maintaining an objective view of what's going on.
Im glad AI npcs are here to post normie comments on every channel
AIGrid - Wes Roth. -
SHOCKING , AGI!!
I know. When you see others and see this it brings you back to earth.
Are you the AI NPC?@@piotrek7633
@@user-it5kl2mt8iwes roth really did change.
Jesus, the amount of work and analysis that goes into EACH video is impressive. Congrats AI Explained
Thanks Rob
With GPT-5, chances are a wrapper like Devin won't even be necessary. Wouldn't be the first time OpenAI's advancements made tools built on top of its models obsolete.
Good point
Everything we see today is unusable from an industry standpoint, serving merely as 'beta demonstrations' to showcase future possibilities and attract investment. Of course, everything will become obsolete soon, hopefully within months!
GPT5 = AGI? All of these new ai tools coming out are specific applications of a.i. for specific use cases. Once they all get integrated together and the system is iteratively refined and trained on the integration (the sum of the individual parts), we'll be nearing AGI.@@danielmartinmonge4054
@@aiexplained-official GPT5 = AGI? All of these new ai tools coming out are specific applications of a.i. for specific use cases. Once they all get integrated together and the system is iteratively refined and trained on the integration (the sum of the individual parts), we'll be nearing AGI. OPENAI is most certainly doing this right now and is converging on AGI (they might have already achieved it internally)
Exactly. When or if AGI arrives almost all of these AI wrappers will become completely unnecessary, along with a lot of other tech companies. Tbh it kind of baffles me that any other AI startup other than the ones developing the leading models are even able to attract investors at all.
After I got access to Devin the first thing I asked was "build a better version of Devin". I call it Kevin.
How was Kevin, was he friendly?
Kevin is very demanding, only want's to run on Azure. @@aiexplained-official
@@aiexplained-official😂
More like DevsOut
@@SirHargreeves Devout
"When these babies get to discovering fundamental physics you're gonna see some SERIOUS SHIT!"
I can't wait for Wolfram to incorporate these directly into their physics project and overall new computational paradigm of reality. I'm sure it'll be solved two more papers down the line. What a time to be alive!
I believe they already did , just don't get good results or the results they don't understand, and I'm talking about researchers. Also they could hide the results from public usage, because their reasons.
Why would a physicist, on making a new dimensional discovery proof, take the time to expose a parrot to his equations?
It seems like even without AGI, if we can get the AI to be able to do research, that will be a great step forward as well.
Bro, AI was able to infer Kepler's Law by observing celestial data like a year ago.
i love how "thank you so much for watching to the end" has a dual meaning
The end is near
Didn't think of that
the end is near *incomprehensible rambling*
I loved that robotic CEO bit about "oh don't worry we only want to automate bad jobs that no one wants" lol this really made me laugh like...dude, come on. does he think that literally everyone on youtube was born yesterday?
Lol how are they jobs, then, if no one wants to do them and get paid? Their ultimate goal is AI taking all/most jobs so people can be backed into a situation where it's either UBI or destitution. And those begging for UBI, there are going to be more strings attached than you can pull. We're heading for a dark place.
You are the only AI news channel I still bother checking. Everything else has devolved into either constant doomsaying or clickbait 'biggest news ever' videos. Thanks very much for all of your hard and thorough work.
I like Twominutepapers and Fireship. Matt Wolfe is pretty good too.
@@vectoralphaAI Fireship is overrated imo. Nothing knowledgeable from his videos if you are already following the news. He's just milking views on AI since the average Joe doesn't follow AI news.
so true homie
i still like wes roth, i don't even read his ⚡ SHOCKING ⚡titles.. i see the yellow border, i click and listen. no one beats AI Explained but AI Explained isn't every day
@@sandeepsrinivas7Fireship is comedy for me, I love it :)
Probably best AI news youtube channel
'Probably' ;)
@@aiexplained-official I have a tendency to use minimizing language. Don't think too much about it
Definitively a AI news youtube channel.
@@aiexplained-official a philosopher and scientist should almost always qualify their statements, I think.
Odds are good
The more I see these companies using video games as tools to collect useful data for training general intelligence systems, the more I think we could be using games to more effectively educate kids. The line I was always told was "School isn't about learning specific things like what a sedimentary rock is, it's about learning how to learn" but surely that could apply to games that span a wide range of domains too? Plus they are way more fun and actually can keep kids engaged more than some dull lecture. We just need to actually spend the time, effort, and money to make a game that is acutally fun and can also allow the students to learn the general concepts that we want them too.
There is a game I played recently called Mimic Logic that could teach logic and deductive reasoning pretty well if paired with class lessons. Its an example of a good first step imo.
Video games have always been a learning tool.
@@vectoralphaAI I've learned alot from them, but I don't know if society at large would agree with us. Certainly none of my teachers or my parents did.
Totally agree. We're overdue for great discoveries in how to pair games with teaching desired skills and behavior. There have been attempts, but we don't have a comprehensive solution yet. It is difficult to equate in-game knowledge to in-world knowledge, and games should therefore be paired with tutoring, like you said.
Using your brain will always be taxing and more difficult than just enjoying yourself. It isnt a school system thing, it is a fact of reality
@@IconoclastX Obviously but there is a motivational aspect that is required to get humans to use their brain in the first place. Games have been optimizing the desire to play, win, and learn them for decades. So if you can design a fun game that a player can most easily win by learning useful information/skills then provide a frictionless means to learn said skills that, in theory, could create a superior learning environment to what we currently have.
Whenever the notification goes for AI Explained I don’t even think about tapping anymore, it’s just pure reflex
Post like these sound so fake. Specially when there is a bunch of similar comments that mention nothing about the actual video themselves.
did i just fail the turing test
maybe people just like the video dude
@@jordanmackay9156honestly it's just the truth, because actually good channels are few and far between. I share his opinion, instant click without thinking. Sometimes I see videos that have a lot of white in the thumbnail and get disappointed that it's not AI explained 😆
I get really excited because this is the highest quality youtube channel I follow. Him and Veritasium carrying my learning.
i see ai explained video i click on it within 47 second
Why the hell does it take you that long, what are you doing for 47 seconds?
@@aiexplained-official i blame youtube took some time to refresh :( i will put MultiOn on the job to refresh your channel every second for next time 🤖
Yes. True.
Best I can do is 627 seconds. I am a local LLM running on AMD athlon on a washing machine
@@lthedoperabbitl9258good
Honestly I've felt like an ML/AI optimist for over a decade and I regularly train CV NN's myself (I have one training right now). But yet I can't help but start to feel like we're going in the wrong direction and we're doomed. I'm not too worried about AI nuking everyone but the way these companies are acting is exactly what we've seen in the past, companies doing whatever they can to raise their short term profits which in this case means cutting workers. Do we think that these companies are going to care about the people they put on the street? Perhaps the government could eventually put measures in place like UBI but I'd be willing to bet it will be way too little too late. Just look at the fight to raise minimum wage. To me its starting to look like we're hurtling toward an inevitable future of the rich individuals and companies getting infinitely more richer and everyone else getting infinitely poorer and I'm not too sure if there's much anyone can do about it. As a software developer, I try to stay informed so that at least hopefully I can be one of the people using AI to be more productive until I'm eventually replaced.
yeah, unless politically checked, AI can be an absurd multiplier for concentration of income.
About UBI, even if comes in a timely manner, it is still a huge transfer of power. Like, common people become dependent on the goodwill of the ruling classes with little to know economical bargaining power.
@@user-sl6gn1ss8p Yea, historically the working class has always had some degree of influence but without being the "working class" anymore, we'd no longer have any real influence.
Thank you for the powerful comment skier
The problem is the intention behind UBI is control. AI is the leverage to take away most/all jobs so people can no longer earn an income. UBI is for when people can no longer earn because they no longer work but need a means of subsistence. Sounds like a really nice dream until you realise "this is me losing everything - rights, voice, etc." Those looking forward to a UBI future, there are going to be way more strings attached than you can pull. Dystopia comes in the nice, shimmering wrap of utopia.
"I'm not too sure if there's much anyone can do about it."
There's plenty we can try! Take a look at the grassroots movement PauseAI.
Its scary how fast ai is progressing now, what once took 10 years, now takes 1 year or less
By now I see the hype-topics like Devin come and just wait for your video to drop. Excellent work!
Also ... I'm a freelancer doing exactly what Altman thinks (or says he thinks) will be done for extremely low cost by AI in about 5 years. And I really cannot see anything to retrain where the AI won't progress faster than my jobtraining would be. I'm 45 and cannot outlearn AI progress. All of this coupled wiith political systems completely paralized and unable to effect any significant change about anything whatssoever. Honestzly, I was a tech enthusiast all my life, but by now I'm more than a little bit scared.
I know what you mean. It's a moving target, just focus on what you do best for now.
@@aiexplained-official Yeah. That's the plan. This is either hype (and I don't think that) or so huge that actually preparing for it feels out of my reach. The best thing seems to be to enjoy the day and be prepared to, well, have a flexible mind set. For better or worse, life seems likely to change A LOT.
I don't think anyone can predict what profession to retrain for, if any. If things go well, we would all be sipping cocktails on a warm sunny beach while enjoying the abundance AI will bring.
yeah, this is not like other technological revolutions--it's not just that it's going to take the current job/career we have but that there's nowhere to go to, no way to prepare
My thoughts as well. If the AI takes my job (and there's a good chance although the timeline is hard to predict) there will be societal problems that can't be solved by individuals, there needs to be a fundamental shift. What I am actually scared of is the current political climate in relation to this.. because this is not something to be left to the market forces.
I've been fascinated by the way this has been going since I created my first AI picks nearly two years ago. I'm not the type who thinks "they'll steal all our jobs," but when they mention "unsafe and undesirable jobs," I immediately reinterpret it as "jobs people do to make ends meet and survive" (sometimes multiple jobs like this), whereas robots can make it safe, but they don't need to work to get by. I think people need to think about whether we draw a line on what they do, or do we put the "savings" predicted to be made by using these robots into a system to help the people reliant on these low-skill or undesirable jobs. Anyway, love the channel and always look forward to the new content... you have a good way of seeing beyond the hype.
That's right AI explained! Shred those papers to pieces!! Love the rational take in the sea of irrational hype!
In the shredder!
At some point an AI collectiv will just post on social media themselves, about themselves and look like this "I achieved another revolutionary thing and it took me 3453452345x less time than humans"
Remember: This is the best AI channel on UA-cam, and it’s not even close.
Thanks so much rainman
fireship is the best.
This is by far the best AI news channel on YT (for me), to the extent that I almost check it a couple of times a week to see if I missed a new video. Thank you so much for your efforts.
Thanks Omar
Really appreciated that readout of the roadmap for Figure. Space is the place.
That's why this is the best channel for info on those things.
People are too sensationalist about those news, calling everything AGI now
Amazing content as always! In a future video, I would love to hear your take on Extropy and their AI analog chips. I still couldn't wrap my head around it.
Awesome video as always! As a marketing professional the last couple of minutes gave me pause.... I've been pushing all the ideas Sam Altman covers, with a similar timeline, but it is sobering to hear it from him rather than myself 🤓
Thanks Sean, huge changes indeed, and timeline seems plausible
Always a nice surprise to see your vids
You're the only channel I genuinely love to watch for AI news. Thank you for your work!
:))
Now I'm so happy that ai is going to replace workforce, gamers, media influencers and displace humans to a lower tear in The food chain.
The brightest future i have imagined in my life.
Good video tho keep it up❤️
Thanks! Excellent content, as always. 🙏🏼
Finally someone giving a well informed report on the matter. Between the hype vs gloom and doom is very hard to take any AI news seriously. Thank you!
Thanks amy, means a lot
Well the doom is between the lines all over the video... but yeah
the doom is real, if u avoiding it, u will be devastated later.
SHOCKED.
Awesome content 😊
Stellar video as always, Phil!
Pretty confident I'd pull the car over to watch when a new video comes out. Thanks for your hard work!
Oh wow, that is very high praise! Be safe though!
@aiexplained-official Hence pulling over ;D
I wouldn't suggest or condone driving and watching AI Explained.
@@aiexplained-official When Level 5 Full Self Driving Autonomous vehicles are the norm, then we wont have to do this, we can all watch your videos any time while "driving"
@@vectoralphaAI Very true.
The mime line got a laugh but you’re not even wrong! 😂🙏
Great video, as always. My best source about AI in UA-cam.
Insightful as always - very much appreciate your content!
Thank you for your video - and especially for your reflective point of view. I learn so much from your channel!
Thanks guest!
Man I've had it up to here with all caps red arrows bold font ai youtuber thumbnails this is such a relief
Always top-notch content.
Thanks man! Great informative video once again!
Thank you Benjamin!
Great due diligence as usual. The best channel on AI news by far.
And I appreciate that you don't gush about AI/AGI like some giddy fanboy--instead, you actually provide the information through a rational, unbiased lens. Extremely hard to find that on the internet these days.
I just want to thank this creator for using complete sentences.
:)
Thanks for keeeing us up to date with recepmt new
As always: Thanks for explaining what last week's hype was about!
Last week lol
The implicit promise of “labor-saving” technology had always been that we could sit around and relax, reaping the benefits of all that labor “saved,” while technology or automation or robots did all the work. But, of course, except in cases of domestic technology (e.g., washing machines), it has never worked out that way-the owners of the companies kept the profit and left the workers to scramble to do whatever labor had not yet been “saved.”
If AI is going to get rid of “unsafe and undesirable jobs”-or, for that matter, “safe and desirable” ones-it would be better if we gave some thought to passing on at least some of that surplus value to the workers who, through no fault of their own, lose their income. Put another way, it’s not technology or automation or robots that is the problem-it’s how the surplus created by all those things is _distributed_ that is the problem, at least the way it works under capitalism.
I imagine that if Socrates was alive today, he would call this whole endeavour very unwise...
automation absolutely should expose the contradiction of capitalism, but, regardless, it's still bad to lose all of our purpose, even if surplus was equitably distributed. we need to be able to be useful to one another.
Great video with great analysis as always! Thank you!
With every new video Philip puts out, we're moving into a new era. Things are moving exponential fast.
Best ai channel. Straight facts no nonsense or click bait
Great video. This is exactly the information I seek. Out of all the youtubers on AI your videos have the best intellectual focus on AI performance and AGI. I think AGI will happen this year or if not already (Q*), and they're not releasing it yet due to safety testing. I speculate that Sam Altman is saying AGI is 5 years away so not get outside pressure to release it soon (for health and science breakthroughs) or start a panic on existential risk or job losses.
As always, many thanks for the great analysis and superb quality content!👍
:)
It's really ramping up now. I just desperately wish that we can be a little more proactive in safety of the models as well as whatever sociatal transitions will take place... though I feel it's unlikely.
This channel used to be my go-to for AI news but now updates are far less frequent as the Patreon takes priority I'm watching more Wes Roth who admittedly isn't the exhaustive reader of papers Phil is it but hits the same topics. Phil don't neglect this channel as it is the gateway to your paid content!
You used to be my go-to source for cool, supportive comments but of late I get more frequent comments from other subscribers, like PsychicCatGod. Maybe endoflevelboss is too focused on his other jobs/family/being ill with flu. Don't forget to comment more!!
You are incredible man! lol
You see how many people break their phones and such clicking so fast on your videos??
They're right too... I feel giddy in a way that should be embarrassing when you post
Keep it up brother
Thanks. Lot of intensity in your profile pic!
I figure any friendships made with me looking like that will be based on the content of my character, and thus will be lasting ones lol
Exciting times 🎉
Thanks for the sober, but exciting video! I wish we knew more about how exactly the figure 01 robot works with GPT-4, cause I'm extremely skeptical about that one.
I appreciate you bringing up the reactions people have about Devin. I'm used to seeing the typical exaggeration and fearmongering surrounding AI, and indeed these reactions feel more palpable? Maybe because I belong in the same professional sector.
All eyes are on the development of AI, at most we might trip over ourselves but I find it hard to imagine we all knock ourselves out.
You are simply the best! :)
Subscribed after the algorithm sent me here. No one is really in control.
Thank you so much for a true analysis of the tech. You are the best AI channel out there. ❤
:))
Congratulations man. 🎉 great work.
Thank you mini
Devin is such a great developer that its creator need more actual developers to develop it. It's a well crafted sales pitch to get VCs money basically and that's about it.
Honestly I can't wait to see chaos caused by code created solely by AI. Writing code exactly (or not, because who will verify) that customer wants and how the customer wants is the worst thing that one can do. Every actual dev knows that.
keep up the good work. ❤
He went beyond the 4th wall all the way into A.I... as always, great video!
Fantastic video but you've now induced my 208th existential crisis...
Sorry Hogg!
You’ll be glad to know there’s now an AI that can have 1000 existential crises in the time it takes one human to have one.
I think it’s interesting how video games have indirectly helped to advance AI. Firstly with the graphics cards that have evolved to run video games, those are now used for running the AI and the video games themselves or virtual worlds that help the AI to navigate the real world.
I said it in the past I will say it again, this is not a bubble it's a game changer. We're looking at a whole new industry that is coming to being, AGI is only a few baby steps away based on the speed of development. I'm glad I found you in the beginning and stuck with you, this is THE place to find reliable and vetted info, thank you!
Thanks Mad for sticking around
The ostriches that are entrenched in denial simply won't let themselves see their obsolescence. Of course Devin will not replace a single engineer today, but it's THE TREND that is alarming and shows the inevitability that Jensen Huang is brave enough to admit.
Another great video Philip, thanks. Laughed out loud on "if it can do (the unsafe and undesirable jobs), can't it also do the safe and desirable jobs?"
I swear to god I heard you say 3 developments in the last 4 to 8 hours and did a double take. Then I realized you said 48 hours and still did a double take. Things are moving quick, aren't they? We may have more than 3 LLM powered robots, GPT 4.5, Model 5, automated software development, and 1 bit quantization by the end of the year. I am starting to think we really will have AGI by 2030.
Reminder that all this progress mostly began with the launch of ChatGPT 3.5 back in November 2022. Meaning all of this has been in less than 2 years. The progress that will be made in the next 5 will be unimaginable.
@@vectoralphaAI That's just simply false, with the launch of ChatGPT 3.5 the progress became more tangible and visible to people outside of the field, but this is by no means 2 years of progress. At the very least you could say 2017, when the transformer paper was published and AlphaGo beat the Go world champ.
your timeline is 2030? damn, I wish it was gonna take that long. I expect to be homeless or dead in
@@VictorKing144 GPT 3.5 was clearly a dramatic inflection point
@@justtiredthings I expected 2 more models after 5 from OpenAI, and incredible replies from Google and Anthropic over the next 6 years. I don't think we'll get AGI though, I think once it goes from narrow to general, it will be effectively ASI, because it will be 10%-50% better than the median human. Once you have a model that is better than all but the top 1% of workers, that's 99 out of 100 people that have become jobless. That's ASI to all of us but the certifiable geniuses.
Hey Philip, where does Boston Dynamics robots fall in all this? They made some pretty crazy robots like 5 years ago.
Again, another amazing video! Thanks!!
Love having @ai_explained on the commute home keeping me apprised of the latest in AI developments. Cheers, suh, for your content! Much appreciate and blessed be. 🙏
Thank you for having me
lmao wes roth understandably catching strays at the end there
It’s always a great day when an new AI Explained video is released!
Thank you! You actually pop in at the end,kinda
Spotted it! Hopefully we’ll get some more people at the next meetup 😁
Would be good to see you go through the Lex video with Lecun. He seems very level headed about the current AGI predictions.
The robot demo was slick, asides from this glitch. When asked, where do the dishes go, the answer was the dish rack. The correct answer should have been back to the sink, as garbage had been thrown on top of the dishes. Either way, progress is being made quickly and it was impressive to see it develop to this point. As always, AI Explained wraps it all up nicely for us in another great video.
:)
Heh. AI's biggest coming gift to software infrastructure: "Please locate and fix all identifiable security vulnerabilities in this code."
What if someone changes just one word : "please locate and exploit all identifiable security vulnerabilities in this code" ?
Great video as per usual
Thanks astro!
It's only been a year since GPT4!? Holy shit.
Creating something is one of the most satisfying things a person can do. I foresee mass despondency when all the creating is done by AI. I foresee that times when you can get anything, getting something made by a person will become very valuable.
Literally have been w8ing for your video the whole day man what took u so long 😂
Haha, it was a big one, lotta work went into this one, more than usual even
Thanks for the video!!!
Thanks nacho!
Great video!
The best youtube channel to truly understand why I should be shocked
Great write up as always, was really looking forward to this one!
What i oftentimes wonder about: How come we train these models on "human" data and let them play games via our input-methods instead of using their ability to directly interact with the game on a machine/code level. Eletrical signals&background inputs consist of much higher quality data and should be quite clear to analyze.
Basically we are rebuilding humans instead of ghosts in the machine.
Is this based on a) a concsious decision to want these systems to be relatable to us, b) my misunderstanding of how valuable the actual training data would be or c) something that will follow in the future but is impossible for now? Do you have an opinion on this subject? Is it something that has been discussed prior on your insiders channel?
🎯 Key Takeaways for quick navigation:
[00:41] Big improvements coming for AI models? The video suggests that future advancements in underlying language models (like GPT-4 to GPT-5) could significantly improve the performance of AI systems like Devon.
[02:20] AI self-improvement? The video ponders a future where AI models can fine-tune other models (or themselves) to improve their ability to complete tasks.
[03:15] Complex coding challenges for AI: The video highlights the difficulty of the benchmark used to test Devon, which requires understanding and coordinating changes across multiple areas of code.
[04:10] ⚠️ En biases in AI benchmarks? The video raises concerns that the benchmark used to test Devon may be biased towards easier-to-solve problems due to the way data was selected.
[05:06] Why Devon might improve with GPT-5: The video lays out reasons why Devon's performance on the coding benchmark could significantly improve with a more advanced language model like GPT-5.
[06:16] Cost of running Devon: The video highlights that Devon can be expensive to run, taking 15-60 minutes to complete a task.
[06:31] AI won't replace all software engineers soon: The video cites experts who believe AI will create more software engineer jobs, with humans likely working alongside AI in a supervisory role.
[07:41] Future advancements with GPT-5: The video suggests that Devon's performance could significantly improve with the release of GPT-5 due to its better reflection and debugging abilities.
[08:47] Concerns about AI and jobs: The video acknowledges that people are scared about the impact of AI on jobs and the need for companies to address these fears.
[09:29] Not all jobs automated yet: The video clarifies that Devon is not going to automate everything yet, and that there are still job opportunities with companies like Cognition AI.
[09:57] Potential of AI for various tasks: The video discusses how AI agents like SEMA, trained on playing games, could potentially be applied to other tasks like video editing or using phone apps.
[11:21] Multi-game training benefits AI agents: The video highlights that training AI agents on multiple games (SEMA) can lead to better performance on new, unseen games.
[12:16] Zero-shot learning for AI agents: The video explains how a model trained on multiple games (SEMA) can outperform a model trained on a single game, even for a game it has never seen before (zero-shot learning).
[12:44] AI nearing human performance in video games: The video suggests that AI agents like SEMA are approaching human-level performance in complex video games.
[13:26] Rapid advancements in AI visual understanding: The video points out that AI models like GPT-4V are rapidly improving in visual understanding tasks, as shown by the MM benchmark.
[14:08] More training data leads to better AI performance: The video emphasizes that the more data (games) an AI model is trained on, the better it will perform (SEMA).
[14:21] Transfer learning benefits robotics: The video mentions a robotics study (RT2) that showed how training a robot on data from other robots improved its ability to perform new tasks.
[16:55] Concerns about AI misuse: The video raises concerns that AI robots like Figure 01 could be misused for military purposes or in ways beyond the developer's control.
[17:09] AGI timeline concerns: The video cites a US government report expressing concerns about the lack of control over rapidly developing AGI.
[17:39] Experts' predictions on AGI timeline: The video mentions predictions from experts like Jensen Huang (Nvidia) and Sam Altman (OpenAI) that AGI could be reached within 5 years.
[17:53] Impact of AGI on marketing jobs: The video suggests that AGI could automate most marketing tasks currently done by agencies and creative professionals.
[18:35] Rapid advancements in AI compute power: The video highlights the significant increase in AI compute power expected by the end of 2025.
Made with HARPA AI
What a time to be alive!
Thanks for refraining from using SHOCKING titles 😁 signal to noise ratio on your channel can’t be beaten.
You made my point for me. Thanks
You and David Shapiro are my favorite ai people on UA-cam. You treat your viewers with respect, thanks!
Thanks Everett
Thanks !
The level of journalism is outstanding, enjoyed your take on devin and co, they have a long way to go
Thank you luke
you are a stronger man then i, i don't know how you held back a SIMA balls joke but i salute you
Tough but I made it
Amazing channel
Yay!
Jobs, what jobs?
Ive been following the develop of this for several years now and its always felt like this is the way we where headed, but at the same time I’ve been expecting us to hit a plateau. To be honest I thought gpt4 would be where we would stay for a long time. I no longer believe that. This is legit starting to move way too fast. No one in control indeed.
What is still mind bowing to me is how this this systems are interacting to 2d visuals and being able to react with it bypassing the 3d envronment mapping aspect..which is still in its infancy( By Boston Dynamics!!) and would probably take forever to finish the model since data on individual objects in environment is literally to infinity and wouldn't even be computable on a compact system...this is the key to the next age of General interactive Robotics with intuitive control and less personnel training
This is my favourite youtube channel.
Thanks Billy
100x efficiency? Huge!!
Figure 01 is impressive, however each action it performs is a separate neural network. So GPT-4V is basically using function calling to load weights of a specific neural network designed to achieve that specific task. Think we need a lot more training data to make them general, but of course this is where synthetic data will really shine.
chapters:
- title: "Introduction to Recent AI Developments"
timestamp_range: "0:00-1:02"
summary: |
The presenter introduces three significant AI developments in the past 48 hours, questioning if they live up to their hype. The focus is on Devon, an AI system designed to function as a software engineer; Google DeepMind's SEMA, which excels in video games; and a humanoid robot, Figure One, capable of performing tasks like dishwashing. The common thread is their reliance on GPT-4 and the anticipation of their evolution with future model upgrades.
- title: "Devon: The AI Software Engineer"
timestamp_range: "1:08-9:17"
summary: |
Devon, built on GPT-4, showcases advanced capabilities in software engineering, outperforming its predecessors in benchmarks. It integrates a code editor, shell, and browser, allowing it to read documentation and execute tasks. A significant focus is on its ability to autonomously fine-tune models and its impressive performance on a software engineering benchmark, hinting at rapid future advancements with GPT-5. The discussion includes its operational cost, implications for the job market, and the ongoing development of competitive models by other entities.
- title: "Google DeepMind SEMA: Mastering Video Games"
timestamp_range: "9:33-14:18"
summary: |
SEMA aims to create an instructable agent capable of performing any task in simulated 3D environments, including commercial video games. The system's development involved training on diverse games, with SEMA demonstrating positive transfer across different environments. The presenter speculates on SEMA's potential beyond gaming, including tasks on smartphones, and highlights the fast-improving capabilities of AI in visual understanding and interaction.
- title: "Humanoid Robots and the Future of Manual Labor"
timestamp_range: "14:55-19:20"
summary: |
The segment focuses on humanoid robots, particularly Figure One, which aims to automate manual labor with AI. The discussion extends to the broader implications of such technologies on the job market and society, including the potential for AI to take on safe and desirable jobs. The presenter reflects on the rapid pace of technological change and the importance of public awareness and engagement with these developments.
- title: "Concluding Remarks"
timestamp_range: "19:04-19:20"
summary: |
The presenter concludes by emphasizing the transformative potential of AI technologies while dispelling the notion that current systems like Devon represent artificial general intelligence (AGI). The call to action encourages viewers to engage with and understand the implications of AI advancements.
With both Devin and Figure 001, there’s plenty of sleight of hand going on. You correctly questioned how long, how much compute and what cost Devin took to do what it allegedly does. You didn’t scrutinise Figure at all, taking their blanket claims of no human manipulation as read. In the video, you see the alleged interaction with Figure. Watch closely when it puts the cup and plate in the drier. Watch the motion of the cup and plate. Now scroll to the end of the video where they ‘replay’ those moves. Those aren’t the same shots, the cup and plate move differently. This suggests they did multiple takes and composited the best. This then questions whether the robot is actually making dynamic choices or merely copying moves that are on its training data or otherwise injected. Same goes for the mode of speech, and the utility of it holding the basket at a slight angle. Then there’s the question of why you’d put a cup and plate laid out on a table into a drying rack (it’s not wet). To a human, taking the items out of the basket seems like the next step, either laying the table or stacking them. This all smacks of hype and attention grabbing, not real advances. There’s a lot of work in that robot in terms of its motor controls etc, but this looks more like someone promoted ChatGPT to see what the AI would think are the moves, and then just programmed the interactions in a conventional way.
I find it interestingly eerie that in the early days, before they were called chat bots, they were referred to as agents. Now the term “agents” has reappeared… … Agent Smith… from the Matrix!😳