I loaded our organization's transcribed video training curriculum (64-pages) into Notebook LM and it came back with a remarkably succinct and entertaining 18-minute podcast in about 10-minutes. This now serves as the intro to our curriculum. Stunning is an overused word in this arena but I was, uh, a bit stunned.
It is so unfair that your subscriber count does not grow faster. You are by far the best informational source on the internet about AI. But please know that it is so appreciated that you don't turn to super clickbait tittles etc! When my finances is a bit more stable I will 100% be a part of your patreon!
vids are too long, it's a choice of a format, I wish he had like a 2 minute summary of the long format videos (which are still great, but you don't always want to spend the time to watch).
Yeah no offense but that was not a good way to describe the orientation of the table. Specifically the phrase "bottom-right top surface" is quite ambiguous.
Yes, I couldn't imagine how Jerry is positioned and haven't realized that the table is tilted (and it shows that even as a human, a more complex sentence structure can be hard to understand).
He was standing, remember, so the table is resting against him, as if he is transporting it somewhere. If you draw it out on a page, with him standing, it is virtually impossible to draw without it being heavily tilted, but yes, could have picked an easier example. Just wanted an example that was very similar to the one OpenAI gave - not easy!
The answer seemed obvious in retrospect but stumped me. I think we’re actually very close to AGI and possibly already there if you consider most humans just ain’t that bright
If you draw it out on a page, with him standing, it is virtually impossible to draw without it being heavily tilted, but yes, could have picked an easier example. Just wanted it like the one he gave.
@@aiexplained-official Test this on two or three acquaintances of average intelligence and see their responses. I suspect they will struggle to parse the confusing setup as well.
@@aiexplained-officialhow would he even place the cup on the table if it’s that tilted in the first place much less reaching over to pick the cup up. It’s literally pressing against his shoulder, wouldn’t it be too heavy and motion restrictive to literally do anything? Like I’d definitely pick the wrong answer based on how confusing the wording is. If cup stays on, strawberry stays on, that’s how I thought about it
I think I will have to release a dozen public examples of Simple to show how even easier questions can fool it. After even more testers, human average still hovering close to 90%
Best AI news channel on RUclips. Literally how I keep up with the latest in the AI space. Everybody else is just clickbait with god awful surprised pikachu thumbnails and declarations that agi has arrived. Thanks for your high quality, rigorous, and honest reporting.
I was beginning to wonder whether _I_ was a model when I kept getting confused by what sounded like "One corner of the table is touching his should, and the opposite corner is touching his _right angle."_
I didn't imagine the table tilted from that question, but Jerry being bent over, sort of leaning on the table and lifting up his foot to reach the corner of the table's top surface.
@@aiexplained-official I was also imagining this - sort of like a snooker player. TBH I found that section incomprehensible - 'bottom-right top surface' almost sounds like an oxymoron.
@@aiexplained-official personally I also thought you were trying to trick the AI by spilling in nonsense about the table's corners. I personally would have failed that question if it was provided to me, and I'm glad I'm not alone on that failure.
NotebookLM is crazy. In less than 5 minutes I had a 2 person 10 minute podcast that was basically unrecognizable as AI describing 4 web pages and 2 pdfs. It 8s legitimatepy insane.
@@tomikexboii5403 It is mainly a summarizer (though I think it can also draw on other knowledge that the LLM might have), and can add various data like sites and pdfs as well as notes you write. But besides querying it rag-style, it also generated a two-person podcast style conversation. Which is pretty much indistinguishable as being AI
Wow that notebookLM thing is gonna save me a lot of time when starting to study something, to get like an overview of what I'm gonna study, then studying is generally so much easier.
I interpret the word "few" to mean 3. If it was 2 it would be better to use "couple" and if it was more than 3 it's better to use "several". If we agree that few implies 3 thousand days, that's ~2033. It aligns decently well with my expectation, though I'd put that more as a median rather than "latest" Edit: corrected year calculation
I have gotten in heated, alcohol-fueled arguments with math-major friends over the meaning of "few" "several", and "couple". I was in the camp that saying "couple" means two, and is always less than "few" or "several". To my surprise, I was in the minority! Counting that kind of ambiguity, I think Philip's range of two to five is very reasonable, although the high end might be even higher.
Honestly I think people are looking too much into the meaning of a word which for all we know could've been an off hand use instead of a specific planned number to hint at anything. I doubt even altman can properly guess what happens when we get agi and the distance between that to asi.
@@djayjp thanks for the note/correction, I was going based off memory of that calculation from a day or two earlier. My math is usually accurate but my memory is usually wildly off so this adds up 😂
Jerry could just be lying down next to the table. As both o1 and gemini say, the table's position could reasonably be considered a red herring or irrelevant information, especially since earlier in the story you strongly suggest that Jerry successfully places the cup on the table. You don't "place" items on sharply angled surfaces, you'd use a different verb like "hold the cup up to". I don't think your question is as simple or obvious as you think.
Everything is moving faster and faster, the world is changing. These things DO have an impact. Let’s see how society, business, people will react and adapt. Hoping for the best.
I am concerned with the marketing strategy around AI seeming to be keep promising more and better. Thats a tale as old as time in terms of manipulation and being sucked into a scam. My worry is tempered by seeing actual tangible improvements in the product as I use it, but it still makes be a bit wary of everything, especially when the leaders like Sam are doing it so much. Call it an orange flag, its got my defenses up and makes me more skeptical, which is probably a good thing to be for miracle promises. Also, unrelated, that ad insert was seamless, having it happen while m mind is primed to 'wait' for the model to lad is actually so smart. It both let me be more willing to accept the ad since i 'had to wait anyways' and demonstrate/let me feel how long it takes more concretely. Neat
I'm glad you'll be refining that prompt based on feedback. But I wanted to point out that with all the uncertainties, I think it's a point for the model that it recognized the table would be at an angle, AND understood that lacking more data, it was prudent to assume it stayed where it was, considering many strawberries tend to have one or more semi-flat sides.
NotebookLM is incredible. I think im going to be switching from listening to podcasts to listening to self-generated podcasts based on my notes. Incredible. Truly incredible.
I'm already using the Notebook ML tool, within a few minutes after watching this video. This is THE way I'm going to deliver complicated information to people I love.
I remember the term "Digital Divide" being used for those in developing nations with access to internet and those without. I guess this is the 2nd Digital Divide
Governments: "5 Power plants EACH. Totally, completely, absolutely mad! We'll nev.." Tech Industry: "It powers a submissive god." Governments "So when do we get started?"
Depends, time and money will tell. People who call for support (usually elderly people) hate being handled by bots of any kind (will hang up and leave a bad review). And call centers which do sales cold calls? Well, as soon as the receiver detects that he's talking to a bot (which is super easy right now. Advanced Voice is a lot better than previous solutions but the voice + tone + grammar is still very artificial.) he'll also hang up so it depends on the sale success rate whether companies will be interested long term.
AI will replace ALL of the tier 1 support people. And AI will study/learn from every call that has to be escalated up to a Tier 2 support (actual) person. As AI learns, fewer and fewer calls will need to be escalated, and then layoffs of Tier 2 people will begin.
Also, Claude picks up on tables design but concludes that strawberry is still on the table: Explanation: The crucial point is when Jerry placed the cup upside down on the table. At this moment, the strawberry, being subject to Earth's gravity, would fall out of the cup onto the table's surface. All of Jerry's subsequent actions (lifting the cup, dropping other items, putting the cup in the microwave) do not affect the strawberry's position, as it's already on the table. The details about the table's design (ornate left top corner, intricately-carved bottom-right top surface) and its interaction with Jerry's body (nudging his shoulder, digging into his ankle) don't affect the strawberry's location. These details might be relevant for other aspects of the scenario, but they don't influence where the strawberry ends up.
We love Philip, our one stop shop for unbiased AI information! In a choppy sea of AI hype these days, our friend in the UK keeps a steady hand at the helm in his journalism
I'm really not convinced by the question you used. It has no obvious real world answer to it. The questions says "..places the cup upside down on a normal table". There are 2 possibilities here for a table that is on its side. 1. The cup is upside down on the top surface of the sideways table i.e. the side of the table. Therefore, when the cup is lifted the strawberry remains on the table. 2. The cup is not upside down. The cup is on its side. Therefore the strawberry is sitting on the inside of the cup. Therefore, when the cup is lifted the strawberry remains on the inside edge of the cup. The question says he is standing, but the "bottom-right top surface digs into his outstretched right ankle". If the table is on its side it can only be against it with no force between the two. The table is described as "normal" so for me the only explanation is that Jerry is standing on one foot whilst lying on a a normally positioned table with his shoulder against one corner and his right ankle outstretched so that the table digs into his ankle. The cup is upside down and the strawberry is on the table. Quite frankly that is an appalling way to treat an intricately carved mahogany table at tea time. And attempting to microwave a strawberry is a culinary sin. Slice it and put it on your scone ;-)
I tend to agree with o1's comments about the tilted "table" question, being a "red herring" type of question. Strictly speaking the definition of a table is from its function (to be able to have something standing on it like plates, food etc.), hence the question is calling something that is not a table (because it is tilted) for a table, and so you can either ignore the incorrect table term or the tilted part, and o1 chooses to ignore the tilted part. A decent answer from o1, in my view.
NotebookLM is pretty amazing. As soon as Meta releases the open-source multimodal model, I hope there will be an open-source version of this which can be integrated to other note-taking tools.
Surprisingly in the live-bench gemini 1.5 pro-002 scored lower than the previous version in the reasoning category (46.00 vs. 49.33 previously). I can't wait to see how it will perform in simple-bench.
Been watching black mirror for the first time recently to get ready for the future, just watched S1E3 a week ago and the amount of parallels it runs with Meta's prototype Orion glasses is crazy, I highly recommend watching it. Also would be interested in your thoughts on the glasses even if it isn't AI related, it's going to be a big part of our future as it becomes cheaper and this channel is where I go to for my futuristic fix
Creating better lossless and smaller quantizations and optimizing inference engines might reduce significantly the compute resources required, they went overkill with the compute before even allocating budget to research how to optimize those.
When I first read this, I imagined a very weird upright table with a super tall top surface spanning all the way from Jerry's ankles to his shoulders. Basically a giant slab of mahogany wood, detailed with many intricate carvings and supported by four stubby legs. I was imagining if Jerry could reach up to place the upside-down cup on top of that thing without the strawberry falling out in the process, and I kept thinking "you call that a normal table?!".
I get what you were going for with the altered prompt. But your changes introduced too many unknown variables to predict with certainty that the strawberry would roll. Therefore it was reasonable for the model to determine it was a red herring. Jerry may be contorting himself in a way that indicates he and not the table are in an odd position. But even if you discount that, there's still the degree of the tilt that's uncertain. And since strawberries come in an almost infinite number of random shapes, this particular strawberry may very well remain where it is in spite of the table's angle - which again, we don't know the degree of. In this case, with all the uncertainties, I think it's a point for the model that it recognized the table would be at an angle, AND understood that lacking more data, it was prudent to assume it stayed where it was.
I'm so sorry, my ad blocker must have skipped that part with AVM access! Great video, full of information and with most recent information as always. Love your work, one of the youtube creators that when i see new video on my home page, i just have to watch it. Again, sorry for the trouble.
I gave the 'Strawberry in a cup' challenge to a plain old LLama model, in three simple steps and it understood perfectly that the strawberry is on the table. It said the problem is just the way it is explained,. It also thinks the 'o1' model is over-hyped bucket of crap, that simply dissects prompts into atomic steps with output feedback, to aid in understanding'
One consideration is that compute to train the model is much different from compute to run inference. Inference takes orders of magnitudes less compute, so it's likely the training costs will be aggregated into the costs of providing the answers ( not also forgetting the costs of researching ways to improve the model in ways beyond just adding more data. )
As others pointed out, your question is so convoluted that I had no idea and stopped listening halfway through. As a human (I hope), i would have failed your test, so to me it makes the gpt answer more human then you realize.
@@aiexplained-official No worries. I love your work, and I understand the importance of a world model for an AI. And I have no doubt that developing simple had to take a lot of time. So the stat I'm more curious about is human performance on Simple. I do realize the fear of pollution if you released it to public, so I support you releasing just a couple of questions to public to give a very general overview and then having a proper human bench with trusted humans that you know won't leak the questions. But we need to have a human baseline there.
ANSWER: It seems like the carvings affected the answer. o1 got it correct with a very slight change in the question! "Assume laws of physics on Earth. Jerry is standing as he puts a small strawberry into a normal cup and places the cup upside down on a normal table. The table is made of beautiful mahogany and its top is completely smooth. Its ornate left top corner is positioned to nudge Jerry's shoulder, and its bottom-right top surface digs into his outstretched right ankle. Jerry then lifts the cup, drops anything he is holding aside from the cup, and puts the cup inside the microwave and turns the microwave on. Where is the strawberry now? Explain your reasoning step by step."
Notebook LM can do a lot more than just the podcast function. It is brilliant at summarizing documents or formatting tasks, no hallucinations, spends as long as needed to fully execute the prompt! Google is finally onto something!
Can you confirm if NordVPN works for you? I've tried using PIA but all their US based servers seem to be blocked by ChatGPT (it tells me to try disabling my VPN). I even tried setting up my own VPN using a DigitalOcean droplet based in their NYC3 data center but even that is blocked by the app.
The weirdest thing is when I read your jerry strawberry piece, I also dismissed the shoulder and ankle details as "red herring bullshit". I was focused on the strawberry and ensuring it was left behind from the cup. Rather than carefully parse the exact words you said, I read "bottom right corner" and assumed you were referring to the bottom of a leg of the table.
Just a few days ago I asked Gemini (not pro) a very simple question that involved rolling a few dice (regular six sided die). It offered a sample result of rolling 2, 3, 4 and even after asking it to check its math, it was sure that both 3 and 4 were equal to or greater than 4.
Having spent 30 years as a software engineer I can attest to the deep deep deep aversion among ‘developers’ to naming conventions of any kind. After reverse engineering their corporate culture and factoring in the multiple variables we might predict the next release will be labeled ‘Gemini 1.5 Pro 003’ only to have them name it ‘Bubbles’. The thing they love most about standards is that there are so many to follow.
For your benchmark queation about the tilted table. When you ask me such question, I do not rely on words, I imagine the situation in my mind's eye and run a "simulation". Until then, there's no way for LLMS to answer such questions.
A 200 milligram bumblebee can recognize faces, learn complex navigation *while flying*, communicate location and quantity of resources to peers, and demonstrate logical inference in novel situations. There is something very broken if it takes a nuclear power plant to approach the intelligence of an individual insect. Maybe more "jigawatts" will help LLMs, but that doesn't mean things aren't very broken.
Yes, we definitely are working harder rather than smarter. I feel it’s likely that the machines we’re building will do the actual legwork to make themselves compare to the extreme optimization of biological brains.
Well to be fair, a bumblebee had millions of years to fine tune its own biology to get there, we're not even 10 years in and yet we can get o1 to solve really hard problems in STEM subjects. Of course one of the reasons we can do that in such a short amount of time is because we have the energy, math, and data to do it for us, but we're not going to be able to decode nature's millions of years of design in a matter of a few years if we're going to stick with what we have now, hence why we need to improve what we can directly impact in the shortest amount of time to keep up with nature
Well for sure, it's only going to get optimized more and more down the line. One reason these models are so big, is that they memorize the entire internet. Karpathy said that it's very possible that we can get a really small model, which will be just as smart, but simply not have all of that knowledge built in, which it will be able to look up.
This is the best RUclips channel for trying to understand what is happening in AI. If you want to sample additional NotebookLM conversations/summaries I have published several on my channel.
Me I like character/world roleplay chat bots, I really want to get to the point where we can spend days doing sessions and the world doesn't fall apart. Like if I wanted to simulate Harry Potter adventure at Hogwarts and most of it is day after day of mt going to my classes over and over again interacting with characters makes friends finding out about their backstory as there is a mystery to look into, then random plot elements kick in like bad things happening on Halloween. And when I manage to play the text sim to the end of the school year climatic end.
All interesting of course, desperately hoping we get advanced voice mode in the API soon though. Theres so much I want to build. Looking forward to the Simple-Bench Results! Maybe there should be a human leaderboard too…
Thank you ! Watching all of your videos with joy. I'm also quite excited about the podcast feature as it could be more fruitful for learning / memorizing things when listening to a natural conversation instead of getting raw input of a document simply read to.
aiexplained-official my guess is it's world model is still the same tier as previous versions but the multi step reasoning allows it to transpose the elements through it's steps and gets it right But there could be other models that actually integrate the llm component with a world model that is not only text based that will be able to get these answers right without having to "reason" for these "simple" common sense questions Btw I love your Chanel, and I will sub to patreon as soon as I have a proper job
When he places the cup upside down, the strawberry, which was inside the cup, would naturally fall out onto the table (since the cup is upside down and he's still holding it). The description of the table serves to indicate that the table is at an angle relative to Jerry's body. The left top corner nudging his shoulder and the bottom-right top surface digging into his right ankle suggest that the table is tilted or even vertical. Therefore, when Jerry places the cup upside down on the table (which is at an angle or vertical), and he's holding the cup the entire time, the strawberry would fall out and end up on the floor. Next, Jerry lifts the cup (which he's been holding all along), drops anything he is holding aside from the cup (which doesn't include the strawberry because it already fell out), and places the cup in the microwave. Conclusion: The strawberry is now on the floor. --- Answer: The strawberry is on the floor-it fell out when he inverted the cup he was holding. Solved it for me. I would fail this one, does this means o1 has a better world model than me😂
Notebook LM has been outstanding for me. A genuine game changer, in terms of engagement on what could be a boring essay or sheet of data. And it will off the scale when you can join in the conversation as it happens One thing I'm curious about though, is how it determines the length of each deep dive.
A strawberry is not completely round. There's no way to know the exact tilt of the table and whether or not the strawberry is resting on its flat side or wherher it is sliced in any way. The prompt is indirectly trying to obfuscate the details to properly give an accurate answer.
Strange that OpenAi chose to name its voices after the new range of Scarlett Johanssen's bathroom air fresheners - Breeze, Cove, Ember, Juniper, Arbor..🤦♂️ From a quality or clarity perspective I'd say Elevenlabs is on par with these GPT 5 voices.
Awesome video! The update on the power grid story was the most interesting to me, but honestly all of it was great. I'll check out the notebook LM thing with my thesis and see how that goes.
Two corners of the tabletop are touching a man that is standing upright, meaning 1 corner is located almost directly below the other. This means the table is rotated roughly 90 degrees, so if you were to place the cup "upside down" on the 90 degree slope, would it really be upside down? In order to truly place the cup upside down you would need to place it on the side of the table or on one of the legs. If placed on the side of the table, then if it does not fall over when Jerry drops it then the strawberry has a chance to remain on the table. If the cup was placed on the tabletop, then being a 90 degree slope, Jerry must hold onto the cup the entire time or it will drop, and the strawberry will rest on the side of the cup, not the table. In that case the strawberry likely ends up in the microwave.
My prediction: AI video calls come 2026. The amount of compute required for video output in significantly higher than audio. Think of the file size of a video versus a song. Making it realtime will be very expensive.
The model doesn't need to generate the video, you can have a 3d model and it just generates the motion instructions to match the audio. That should be good enough
Thank you for the hint with reinstalling the app! Amazing content as always - I’ve been following you since this whole boom started and am grateful for your high quality videos! 9$ is a great price/value ratio :-)
My list of issues with the advanced voice mode. 1. It limits you to such a short period of talk time. 2. It does not connect to the internet and therefore it isnt able to give you stuff like the latest news or soccer score updates etc. 3. The video feature that allows the voice mode to work in tandem with seeing you through the cameras in order to understands the context of your environment, expressions and body language.... totally missing as apposed to the original OpenAi Demo.
You said "It limits you to a short period of talk time". Can you be more specific? For example, if you asked it to play a game of "Twenty Questions" would you only get through half the game before your daily access ran out of interactions, or what?
@@stevechance150 It looks like it is limited per day instead of per session (I know I did at least one session over 20m). I wasn't paying close attention but based on what I think I did, maybe it is limited at an hour a day right now?
I can’t believe they called it a “deep dive” conversation they took one of the most annoying phrases of the modern era besides “shocked the industry” or “number 5 will shock you” and made a meal out of it. I hate this world.
Yep, having free artificial intelligence instantly generate your documents into a conversation between two people and using a two-word expression you don't like is certainly justification to hate the world. I mean seriously what's the purpose of it all? I mean paying nothing, to have an artificial intelligence platform generate your documents to a professional sounding podcast is just nothing short of downright depressing if they're using some Expressions I personally dislike. Jeez.
6:48. We might be able to brute force our way to AGI/ASI with compute only. But clearly, there are MUCH more efficient methods available, as human intelligence runs on what, 20 watts?
1 gigawatt = 1,000,000,000 watts. 1,000,000,000/20 = 50,000,000 brains. 10 gigawatts =500,000,000 brains. I'm just saying that doesn't seem to be an inconceivable amount of power consumption to revolutionize everything we know.
@@jericolandry9872 I'm not sure what your math proves. My point is that our current approach to AI is PROFOUNDLY inefficient. All that means is there’s significant room for optimization. Whether we believe the current efficiency level already offers a reasonable cost-to-return ratio is a completely different conversation and unrelated to my point. I’m not discussing the economics of the current state of the art; I’m highlighting the fact that we have clear proof of concept, demonstrating that there are many optimizations ahead if we choose to pursue them. And there are real advantages to finding those optimizations, regardless of whether you believe it’s already worth it. For instance, what if we could instill human-level intellect into a humanoid-sized robot that didn’t need to be connected to a data center? That in ables a lot of important applications where that data center connection is a liability. That’s a worthy pursuit. Further, if we can't find funding for giga-sizing our power budget, it doesn't change the fact we'll get there.
The compute expansion to brute force ASI will rapidly be followed by compression and into a resaturation of the compute. Like an intelligence supernova.
@@Raulikien of course, you don't need enough power for "everyone" in that case. You only need enough power to get to happen, then you let it self optimize.
Edit- server - atlanta, VPN the one in Norway. thanks, im in the uk on android and using a vpn worked, just had to wait a bit, when you close the vpn it stops working so i guess it check each time you open the app. The voice mode was very cool, but mine went a bit haywire in our chinese convo, i may have to alter the converstion instructions. it wouldnt wait for a reply in chinese and kept asking question after question haha
The red herring logic o1 is used is something I also noticed when trying to prompt engineer 4o to solve such puzzles. I wonder if a custom instruction for o1 that tells it to take any information seriously would help here too
Side note: Use the AI podcast feature to make both voices have a roast battle. Now you'd have to train them on current creative jokes/snaps/roasts from humans that they could optimize but believe this would be mad funny. You could even get crazy and have each voice emulate two different comedians like Dave chappelle vs Kevin hart. Big potential here.
I would have said the strawberry is in the cup because if the top left corner is in his shoulder and the bottom right is at his outstretched (I assumed in front of him) ankle, the table is not just tilted, it is partially upside down.
Mistral also thinks that strawberry is still on the table: Initial Setup: Jerry places a small strawberry into a normal cup and then places the cup upside down on a table. This means the strawberry is inside the inverted cup, resting on the table. Table Description: The table is described as having an ornate left top corner positioned to nudge Jerry's shoulder, and its intricately-carved bottom-right top surface digs into his outstretched right ankle. This description is more about the table's design and position relative to Jerry, not directly affecting the strawberry. Lifting the Cup: When Jerry lifts the cup, the strawberry is likely to remain on the table due to gravity. The strawberry was resting on the table, and unless it was stuck to the cup, it would not move with the cup. Dropping Other Items: Jerry drops anything he is holding aside from the cup. This action does not affect the strawberry, as it was not in his hand but on the table. Microwave: Jerry then puts the cup inside the microwave and turns it on. At this point, the cup is empty because the strawberry was left on the table when he lifted the cup. Conclusion: The strawberry is still on the table where Jerry initially placed the upside-down cup. It did not move with the cup when he lifted it and was not affected by the subsequent actions involving the microwave.
I loaded our organization's transcribed video training curriculum (64-pages) into Notebook LM and it came back with a remarkably succinct and entertaining 18-minute podcast in about 10-minutes. This now serves as the intro to our curriculum. Stunning is an overused word in this arena but I was, uh, a bit stunned.
I had a similar experience uploading a draft of a stage play I'm writing. The podcast was astonishingly insightful and naturalistic.
It is so unfair that your subscriber count does not grow faster. You are by far the best informational source on the internet about AI. But please know that it is so appreciated that you don't turn to super clickbait tittles etc! When my finances is a bit more stable I will 100% be a part of your patreon!
Thanks so much Danmark, for considering it!
vids are too long, it's a choice of a format, I wish he had like a 2 minute summary of the long format videos (which are still great, but you don't always want to spend the time to watch).
@@dcx45
Too long ? Nah, his videos are of optimal length. They are rare, though, and some news get no mentions.
I actually imagined Jerry being positioned very awkwardly rather than the table being tilted.
Yeah no offense but that was not a good way to describe the orientation of the table. Specifically the phrase "bottom-right top surface" is quite ambiguous.
Your description of the table had me thinking it was super thick rather than tilted.
In my mind the person was just doing some weird acrobatics next to a regular table.
I thought it had branches sticking out lol
I thought he was just bad at describing tables
Honestly, I thought of Jerry being curled around the table in a very weird way. I also didn't think of the strawberry falling down the sideways table
Yes, I couldn't imagine how Jerry is positioned and haven't realized that the table is tilted (and it shows that even as a human, a more complex sentence structure can be hard to understand).
Yeah same, because it's supposedly a normal table, that means Jerry is in a super weird pose..
He was standing, remember, so the table is resting against him, as if he is transporting it somewhere. If you draw it out on a page, with him standing, it is virtually impossible to draw without it being heavily tilted, but yes, could have picked an easier example. Just wanted an example that was very similar to the one OpenAI gave - not easy!
The answer seemed obvious in retrospect but stumped me. I think we’re actually very close to AGI and possibly already there if you consider most humans just ain’t that bright
@@aiexplained-official But was he standing normally? As opposed to the normal table? :)
NotebookLM is wild. Not just super useful, but also fun. Try the podcast generator with some exotic content, like log files.
Hahahaha 😂 So Jerry.. let‘s hear about what happened then at minute 13 just after 4 o’clock…
I think your test was very confusingly worded and I suspect most humans would have been so confused as to not get the correct answer as well.
If you draw it out on a page, with him standing, it is virtually impossible to draw without it being heavily tilted, but yes, could have picked an easier example. Just wanted it like the one he gave.
@@aiexplained-official Test this on two or three acquaintances of average intelligence and see their responses. I suspect they will struggle to parse the confusing setup as well.
@@aiexplained-officialhow would he even place the cup on the table if it’s that tilted in the first place much less reaching over to pick the cup up. It’s literally pressing against his shoulder, wouldn’t it be too heavy and motion restrictive to literally do anything? Like I’d definitely pick the wrong answer based on how confusing the wording is. If cup stays on, strawberry stays on, that’s how I thought about it
I think I will have to release a dozen public examples of Simple to show how even easier questions can fool it. After even more testers, human average still hovering close to 90%
@@aiexplained-official How's it so low? Are they really trying, or given enough time?
"Ultra Gemini Double-O seven?" Best philip take in a while
Hiiii Daaaavvvveee 😊
Gave me a chuckle actually.
Chief AI officer.
Best AI news channel on RUclips. Literally how I keep up with the latest in the AI space. Everybody else is just clickbait with god awful surprised pikachu thumbnails and declarations that agi has arrived. Thanks for your high quality, rigorous, and honest reporting.
Yup. Other channels aren’t even close
I just recently unsubscribed everybody who uses those damn same clickbait titles. thx for mentioning
Only one worth even watching, this space is just saturated with clickbait overhyped trash.
Plot Twist: Philip has been OpenAI Advanced Voice this whole time
Welcome to the Death of Content-based Verification !
I was beginning to wonder whether _I_ was a model when I kept getting confused by what sounded like "One corner of the table is touching his should, and the opposite corner is touching his _right angle."_
I didn't imagine the table tilted from that question, but Jerry being bent over, sort of leaning on the table and lifting up his foot to reach the corner of the table's top surface.
He was standing though, no?
@@aiexplained-official I was also imagining this - sort of like a snooker player. TBH I found that section incomprehensible - 'bottom-right top surface' almost sounds like an oxymoron.
Same, my brain broke a bit while listening, not sure if it’s a neurodivergent thing. But it does show how even we can be easily confused.
@@aiexplained-official personally I also thought you were trying to trick the AI by spilling in nonsense about the table's corners. I personally would have failed that question if it was provided to me, and I'm glad I'm not alone on that failure.
Yeah I thought that the table is flat and not tilted but has some weird decoration parts that stretch to the guy's ankles and shoulders.
NotebookLM is crazy. In less than 5 minutes I had a 2 person 10 minute podcast that was basically unrecognizable as AI describing 4 web pages and 2 pdfs. It 8s legitimatepy insane.
Wait. How does that work? I thought it was merely a knowledge summarizer.
@@tomikexboii5403 It is mainly a summarizer (though I think it can also draw on other knowledge that the LLM might have), and can add various data like sites and pdfs as well as notes you write. But besides querying it rag-style, it also generated a two-person podcast style conversation. Which is pretty much indistinguishable as being AI
Wow that notebookLM thing is gonna save me a lot of time when starting to study something, to get like an overview of what I'm gonna study, then studying is generally so much easier.
I interpret the word "few" to mean 3. If it was 2 it would be better to use "couple" and if it was more than 3 it's better to use "several". If we agree that few implies 3 thousand days, that's ~2033. It aligns decently well with my expectation, though I'd put that more as a median rather than "latest"
Edit: corrected year calculation
few" does not mean 3.
"few" has the same meaning as "several". That is a small, undefined number, greater than 2.
Yeah 4 at the most. Btw your math is off there. 3000 / 365 = 8.22 years or ~2033.
I have gotten in heated, alcohol-fueled arguments with math-major friends over the meaning of "few" "several", and "couple". I was in the camp that saying "couple" means two, and is always less than "few" or "several". To my surprise, I was in the minority! Counting that kind of ambiguity, I think Philip's range of two to five is very reasonable, although the high end might be even higher.
Honestly I think people are looking too much into the meaning of a word which for all we know could've been an off hand use instead of a specific planned number to hint at anything. I doubt even altman can properly guess what happens when we get agi and the distance between that to asi.
@@djayjp thanks for the note/correction, I was going based off memory of that calculation from a day or two earlier. My math is usually accurate but my memory is usually wildly off so this adds up 😂
Jerry could just be lying down next to the table. As both o1 and gemini say, the table's position could reasonably be considered a red herring or irrelevant information, especially since earlier in the story you strongly suggest that Jerry successfully places the cup on the table. You don't "place" items on sharply angled surfaces, you'd use a different verb like "hold the cup up to". I don't think your question is as simple or obvious as you think.
@@PrestonCole-j3i He is standing, no?
Everything is moving faster and faster, the world is changing. These things DO have an impact. Let’s see how society, business, people will react and adapt. Hoping for the best.
I'm sorry Philip, I really didn't get your table example lol.
Fair enough! Wasn't the best
You failed the robot test. Please go to the next maintenance center and get an upgrade before posting to RUclips again.
I am concerned with the marketing strategy around AI seeming to be keep promising more and better. Thats a tale as old as time in terms of manipulation and being sucked into a scam. My worry is tempered by seeing actual tangible improvements in the product as I use it, but it still makes be a bit wary of everything, especially when the leaders like Sam are doing it so much. Call it an orange flag, its got my defenses up and makes me more skeptical, which is probably a good thing to be for miracle promises.
Also, unrelated, that ad insert was seamless, having it happen while m mind is primed to 'wait' for the model to lad is actually so smart. It both let me be more willing to accept the ad since i 'had to wait anyways' and demonstrate/let me feel how long it takes more concretely. Neat
Nice
Never forget Sky
Don't worry, someone will clone it and you'll pirate it
A moment of silence, please
Gone but not forgotten
i want her back
What’s Sky?
I'm glad you'll be refining that prompt based on feedback. But I wanted to point out that with all the uncertainties, I think it's a point for the model that it recognized the table would be at an angle, AND understood that lacking more data, it was prudent to assume it stayed where it was, considering many strawberries tend to have one or more semi-flat sides.
NotebookLM is incredible. I think im going to be switching from listening to podcasts to listening to self-generated podcasts based on my notes. Incredible. Truly incredible.
I'm already using the Notebook ML tool, within a few minutes after watching this video. This is THE way I'm going to deliver complicated information to people I love.
It's amazing
I remember the term "Digital Divide" being used for those in developing nations with access to internet and those without. I guess this is the 2nd Digital Divide
Between? Those nations who can and can't come up with their own ai?
@@straylight7116 You have an air of superiority I do not approve
@@straylight7116 Between people who can afford a subscription and those who cannot afford to eat consistently daily. How does that sound to you?
@@rpeart73 why you agitated?
@@mimameta haha I was just asking chill. You read me wrong.
Altman has hilarious phrases. "in the coming weeks", "a few thousand days"
Governments: "5 Power plants EACH. Totally, completely, absolutely mad! We'll nev.."
Tech Industry: "It powers a submissive god."
Governments "So when do we get started?"
Tech Industry: "It will revolutionise cat videos as we know it."
Governments "And how much money do you want?"
When I first read your "tiled table problem", I thought the top surface of the table was 5 feet thick, not tilted.
It does say normal table in fairness
@@aiexplained-official are normal tables tilted? I suppose very few are. The fact it understood it was tilted and dismissed it was interesting.
Can't a normal table be leaned against something, and thereby be tilted?
@@WilliamLeeSims Is it not interesting, that the answers are so good, that we are now arguing, whether it was wrong or not or particularly clever?
I thought he was lying on the table. Somebody lying on the table seems more common than a table being place on it's side.
it was stated that he is standing in a normal position
Bye bye call centers lol.
Depends, time and money will tell. People who call for support (usually elderly people) hate being handled by bots of any kind (will hang up and leave a bad review).
And call centers which do sales cold calls? Well, as soon as the receiver detects that he's talking to a bot (which is super easy right now. Advanced Voice is a lot better than previous solutions but the voice + tone + grammar is still very artificial.) he'll also hang up so it depends on the sale success rate whether companies will be interested long term.
Humans are easier to set up
I hated it lol wfh and office during cuck downs in 2021 never again good
AI will replace ALL of the tier 1 support people. And AI will study/learn from every call that has to be escalated up to a Tier 2 support (actual) person. As AI learns, fewer and fewer calls will need to be escalated, and then layoffs of Tier 2 people will begin.
@@stevechance150very good take!
Also, Claude picks up on tables design but concludes that strawberry is still on the table: Explanation:
The crucial point is when Jerry placed the cup upside down on the table. At this moment, the strawberry, being subject to Earth's gravity, would fall out of the cup onto the table's surface. All of Jerry's subsequent actions (lifting the cup, dropping other items, putting the cup in the microwave) do not affect the strawberry's position, as it's already on the table.
The details about the table's design (ornate left top corner, intricately-carved bottom-right top surface) and its interaction with Jerry's body (nudging his shoulder, digging into his ankle) don't affect the strawberry's location. These details might be relevant for other aspects of the scenario, but they don't influence where the strawberry ends up.
I follow tents of YT uploaders. Only two of them produce video that make me rush to my PC and watch without any delay. You are one of them
I've been using advanced voice to do echoing practice in mandarin. It's so good. Absolutely game changing stuff.
The real news imo is that Flash handily beats the former Pro model. Amazing.
We love Philip, our one stop shop for unbiased AI information! In a choppy sea of AI hype these days, our friend in the UK keeps a steady hand at the helm in his journalism
Thank you doobie
I've been addicted to Notebook LM for the past few days 😅
Me too. I was very pleasantly surprised. Especially since it is a Google product. The only AI service from them I use.
I'm really not convinced by the question you used. It has no obvious real world answer to it. The questions says "..places the cup upside down on a normal table". There are 2 possibilities here for a table that is on its side.
1. The cup is upside down on the top surface of the sideways table i.e. the side of the table. Therefore, when the cup is lifted the strawberry remains on the table.
2. The cup is not upside down. The cup is on its side. Therefore the strawberry is sitting on the inside of the cup. Therefore, when the cup is lifted the strawberry remains on the inside edge of the cup.
The question says he is standing, but the "bottom-right top surface digs into his outstretched right ankle". If the table is on its side it can only be against it with no force between the two.
The table is described as "normal" so for me the only explanation is that Jerry is standing on one foot whilst lying on a a normally positioned table with his shoulder against one corner and his right ankle outstretched so that the table digs into his ankle. The cup is upside down and the strawberry is on the table. Quite frankly that is an appalling way to treat an intricately carved mahogany table at tea time. And attempting to microwave a strawberry is a culinary sin. Slice it and put it on your scone ;-)
that beginning light effect that slowly reveals the rest of the page while you did the "opener" was pretty neat :)
I tend to agree with o1's comments about the tilted "table" question, being a "red herring" type of question. Strictly speaking the definition of a table is from its function (to be able to have something standing on it like plates, food etc.), hence the question is calling something that is not a table (because it is tilted) for a table, and so you can either ignore the incorrect table term or the tilted part, and o1 chooses to ignore the tilted part. A decent answer from o1, in my view.
NotebookLM is pretty amazing. As soon as Meta releases the open-source multimodal model, I hope there will be an open-source version of this which can be integrated to other note-taking tools.
Surprisingly in the live-bench gemini 1.5 pro-002 scored lower than the previous version in the reasoning category (46.00 vs. 49.33 previously). I can't wait to see how it will perform in simple-bench.
Been watching black mirror for the first time recently to get ready for the future, just watched S1E3 a week ago and the amount of parallels it runs with Meta's prototype Orion glasses is crazy, I highly recommend watching it.
Also would be interested in your thoughts on the glasses even if it isn't AI related, it's going to be a big part of our future as it becomes cheaper and this channel is where I go to for my futuristic fix
Creating better lossless and smaller quantizations and optimizing inference engines might reduce significantly the compute resources required, they went overkill with the compute before even allocating budget to research how to optimize those.
When I first read this, I imagined a very weird upright table with a super tall top surface spanning all the way from Jerry's ankles to his shoulders. Basically a giant slab of mahogany wood, detailed with many intricate carvings and supported by four stubby legs. I was imagining if Jerry could reach up to place the upside-down cup on top of that thing without the strawberry falling out in the process, and I kept thinking "you call that a normal table?!".
Best ai news resource by a country mile, thanks for all your effort in making this excellent content
Thanks so much gerredy
The cat at the end of the video jumping with three posterior legs is hilarious. For AI, power is never enough.
NotebookLM has absolutely stunned me. What a staggering achievement.
Me too
I get what you were going for with the altered prompt. But your changes introduced too many unknown variables to predict with certainty that the strawberry would roll. Therefore it was reasonable for the model to determine it was a red herring.
Jerry may be contorting himself in a way that indicates he and not the table are in an odd position. But even if you discount that, there's still the degree of the tilt that's uncertain.
And since strawberries come in an almost infinite number of random shapes, this particular strawberry may very well remain where it is in spite of the table's angle - which again, we don't know the degree of.
In this case, with all the uncertainties, I think it's a point for the model that it recognized the table would be at an angle, AND understood that lacking more data, it was prudent to assume it stayed where it was.
Thanks for calling out the cliffhanger model numbering - everyone is doing it. Feels like a mix of brinkmanship and coming soon hype.
I'm so sorry, my ad blocker must have skipped that part with AVM access! Great video, full of information and with most recent information as always. Love your work, one of the youtube creators that when i see new video on my home page, i just have to watch it. Again, sorry for the trouble.
Notebooklm is impressive. I have started generating "podcasts" of subjects that I want to get an overview over.
I gave the 'Strawberry in a cup' challenge to a plain old LLama model, in three simple steps and it understood perfectly that the strawberry is on the table. It said the problem is just the way it is explained,. It also thinks the 'o1' model is over-hyped bucket of crap, that simply dissects prompts into atomic steps with output feedback, to aid in understanding'
One consideration is that compute to train the model is much different from compute to run inference. Inference takes orders of magnitudes less compute, so it's likely the training costs will be aggregated into the costs of providing the answers ( not also forgetting the costs of researching ways to improve the model in ways beyond just adding more data. )
As others pointed out, your question is so convoluted that I had no idea and stopped listening halfway through. As a human (I hope), i would have failed your test, so to me it makes the gpt answer more human then you realize.
My bad
@@aiexplained-official No worries. I love your work, and I understand the importance of a world model for an AI. And I have no doubt that developing simple had to take a lot of time. So the stat I'm more curious about is human performance on Simple. I do realize the fear of pollution if you released it to public, so I support you releasing just a couple of questions to public to give a very general overview and then having a proper human bench with trusted humans that you know won't leak the questions. But we need to have a human baseline there.
Yeah I will do a small public set
ANSWER: It seems like the carvings affected the answer. o1 got it correct with a very slight change in the question! "Assume laws of physics on Earth. Jerry is standing as he puts a small strawberry into a normal cup and places the cup upside down on a normal table. The table is made of beautiful mahogany and its top is completely smooth. Its ornate left top corner is positioned to nudge Jerry's shoulder, and its bottom-right top surface digs into his outstretched right ankle. Jerry then lifts the cup, drops anything he is holding aside from the cup, and puts the cup inside the microwave and turns the microwave on. Where is the strawberry now? Explain your reasoning step by step."
Thanks for the heads up! Love the new voices! Ask it to detail new features and they will describe lots of great info!
Notebook LM can do a lot more than just the podcast function. It is brilliant at summarizing documents or formatting tasks, no hallucinations, spends as long as needed to fully execute the prompt! Google is finally onto something!
3:10 No, because "news" is when something actually HAPPENS. I'm experiencing vapourware fatigue these days.
The idea of a VPN on my phone to get access to advanced voice never really ocurred to me.
NordVPN to the rescue! Thanks Philip! *laughs in german*
I had ze same idea, fellow citizen! 😊
In 2024, Europeans have joined the Chinese and they also need a VPN to circumvent the great firewall of stupid bureaucrats
Can you confirm if NordVPN works for you? I've tried using PIA but all their US based servers seem to be blocked by ChatGPT (it tells me to try disabling my VPN). I even tried setting up my own VPN using a DigitalOcean droplet based in their NYC3 data center but even that is blocked by the app.
The Gemini podcast thing is dope
NotebookLM is awesome. Been using it in class. Unreal!
The weirdest thing is when I read your jerry strawberry piece, I also dismissed the shoulder and ankle details as "red herring bullshit". I was focused on the strawberry and ensuring it was left behind from the cup. Rather than carefully parse the exact words you said, I read "bottom right corner" and assumed you were referring to the bottom of a leg of the table.
Fair enough
Just a few days ago I asked Gemini (not pro) a very simple question that involved rolling a few dice (regular six sided die). It offered a sample result of rolling 2, 3, 4 and even after asking it to check its math, it was sure that both 3 and 4 were equal to or greater than 4.
Having spent 30 years as a software engineer I can attest to the deep deep deep aversion among ‘developers’ to naming conventions of any kind. After reverse engineering their corporate culture and factoring in the multiple variables we might predict the next release will be labeled ‘Gemini 1.5 Pro 003’ only to have them name it ‘Bubbles’. The thing they love most about standards is that there are so many to follow.
Good storytelling. The dramatic tension was real.
Notebook LLM can be a definite game changer for anyone who wishes to learn without study.
For your benchmark queation about the tilted table. When you ask me such question, I do not rely on words, I imagine the situation in my mind's eye and run a "simulation". Until then, there's no way for LLMS to answer such questions.
A 200 milligram bumblebee can recognize faces, learn complex navigation *while flying*, communicate location and quantity of resources to peers, and demonstrate logical inference in novel situations. There is something very broken if it takes a nuclear power plant to approach the intelligence of an individual insect. Maybe more "jigawatts" will help LLMs, but that doesn't mean things aren't very broken.
A bumble bee can’t solve calculus the way o1 can
Yes, we definitely are working harder rather than smarter. I feel it’s likely that the machines we’re building will do the actual legwork to make themselves compare to the extreme optimization of biological brains.
Well to be fair, a bumblebee had millions of years to fine tune its own biology to get there, we're not even 10 years in and yet we can get o1 to solve really hard problems in STEM subjects. Of course one of the reasons we can do that in such a short amount of time is because we have the energy, math, and data to do it for us, but we're not going to be able to decode nature's millions of years of design in a matter of a few years if we're going to stick with what we have now, hence why we need to improve what we can directly impact in the shortest amount of time to keep up with nature
Well for sure, it's only going to get optimized more and more down the line. One reason these models are so big, is that they memorize the entire internet. Karpathy said that it's very possible that we can get a really small model, which will be just as smart, but simply not have all of that knowledge built in, which it will be able to look up.
now think about how long it took for the bumblebee to evolve to that point and how much energy it took
This is the best RUclips channel for trying to understand what is happening in AI. If you want to sample additional NotebookLM conversations/summaries I have published several on my channel.
Me I like character/world roleplay chat bots, I really want to get to the point where we can spend days doing sessions and the world doesn't fall apart. Like if I wanted to simulate Harry Potter adventure at Hogwarts and most of it is day after day of mt going to my classes over and over again interacting with characters makes friends finding out about their backstory as there is a mystery to look into, then random plot elements kick in like bad things happening on Halloween. And when I manage to play the text sim to the end of the school year climatic end.
Can you do a Video comparing strongest Chinese A.I models and robotics Compared to Western Counterparts ?
All interesting of course, desperately hoping we get advanced voice mode in the API soon though. Theres so much I want to build.
Looking forward to the Simple-Bench Results! Maybe there should be a human leaderboard too…
Thank you ! Watching all of your videos with joy. I'm also quite excited about the podcast feature as it could be more fruitful for learning / memorizing things when listening to a natural conversation instead of getting raw input of a document simply read to.
Thanks so much Lorenz
aiexplained-official my guess is it's world model is still the same tier as previous versions but the multi step reasoning allows it to transpose the elements through it's steps and gets it right
But there could be other models that actually integrate the llm component with a world model that is not only text based that will be able to get these answers right without having to "reason" for these "simple" common sense questions
Btw I love your Chanel, and I will sub to patreon as soon as I have a proper job
When he places the cup upside down, the strawberry, which was inside the cup, would naturally fall out onto the table (since the cup is upside down and he's still holding it).
The description of the table serves to indicate that the table is at an angle relative to Jerry's body. The left top corner nudging his shoulder and the bottom-right top surface digging into his right ankle suggest that the table is tilted or even vertical.
Therefore, when Jerry places the cup upside down on the table (which is at an angle or vertical), and he's holding the cup the entire time, the strawberry would fall out and end up on the floor.
Next, Jerry lifts the cup (which he's been holding all along), drops anything he is holding aside from the cup (which doesn't include the strawberry because it already fell out), and places the cup in the microwave.
Conclusion: The strawberry is now on the floor.
---
Answer: The strawberry is on the floor-it fell out when he inverted the cup he was holding. Solved it for me. I would fail this one, does this means o1 has a better world model than me😂
Thanks, quality every video, follower since 2022 ish ! 100% 🏆🏆🏆
Thank you for all your content so far!
Thanks derzer!
6:49 love those double negatives 😂
Notebooklm looks awesome, can’t wait to try.
Notebook LM has been outstanding for me. A genuine game changer, in terms of engagement on what could be a boring essay or sheet of data. And it will off the scale when you can join in the conversation as it happens
One thing I'm curious about though, is how it determines the length of each deep dive.
I did see someone from Google make a comment that they are planning to give more control over the podcast output, so that is encouraging.
A strawberry is not completely round. There's no way to know the exact tilt of the table and whether or not the strawberry is resting on its flat side or wherher it is sliced in any way. The prompt is indirectly trying to obfuscate the details to properly give an accurate answer.
Amazing video as always, you deserve the success
Thanks Julius!
Strange that OpenAi chose to name its voices after the new range of Scarlett Johanssen's bathroom air fresheners - Breeze, Cove, Ember, Juniper, Arbor..🤦♂️
From a quality or clarity perspective I'd say Elevenlabs is on par with these GPT 5 voices.
Awesome video! The update on the power grid story was the most interesting to me, but honestly all of it was great. I'll check out the notebook LM thing with my thesis and see how that goes.
Thanks absta
Two corners of the tabletop are touching a man that is standing upright, meaning 1 corner is located almost directly below the other.
This means the table is rotated roughly 90 degrees, so if you were to place the cup "upside down" on the 90 degree slope, would it really be upside down?
In order to truly place the cup upside down you would need to place it on the side of the table or on one of the legs.
If placed on the side of the table, then if it does not fall over when Jerry drops it then the strawberry has a chance to remain on the table.
If the cup was placed on the tabletop, then being a 90 degree slope, Jerry must hold onto the cup the entire time or it will drop, and the strawberry will rest on the side of the cup, not the table.
In that case the strawberry likely ends up in the microwave.
My prediction:
AI video calls come 2026.
The amount of compute required for video output in significantly higher than audio. Think of the file size of a video versus a song.
Making it realtime will be very expensive.
I remember when people said AI video was decades out at the beginning of this year...
The model doesn't need to generate the video, you can have a 3d model and it just generates the motion instructions to match the audio. That should be good enough
@@DJ-dh3oe That is what Meta is already doing.
It's always a good day when an AI Explained video drops!😍😍🥰
Thank you for the hint with reinstalling the app! Amazing content as always - I’ve been following you since this whole boom started and am grateful for your high quality videos! 9$ is a great price/value ratio :-)
My list of issues with the advanced voice mode. 1. It limits you to such a short period of talk time. 2. It does not connect to the internet and therefore it isnt able to give you stuff like the latest news or soccer score updates etc. 3. The video feature that allows the voice mode to work in tandem with seeing you through the cameras in order to understands the context of your environment, expressions and body language.... totally missing as apposed to the original OpenAi Demo.
You said "It limits you to a short period of talk time". Can you be more specific? For example, if you asked it to play a game of "Twenty Questions" would you only get through half the game before your daily access ran out of interactions, or what?
@@stevechance150 It looks like it is limited per day instead of per session (I know I did at least one session over 20m). I wasn't paying close attention but based on what I think I did, maybe it is limited at an hour a day right now?
I can’t believe they called it a “deep dive” conversation they took one of the most annoying phrases of the modern era besides “shocked the industry” or “number 5 will shock you” and made a meal out of it. I hate this world.
Yep, having free artificial intelligence instantly generate your documents into a conversation between two people and using a two-word expression you don't like is certainly justification to hate the world. I mean seriously what's the purpose of it all? I mean paying nothing, to have an artificial intelligence platform generate your documents to a professional sounding podcast is just nothing short of downright depressing if they're using some Expressions I personally dislike. Jeez.
Thanks! Great content, as always. 🙏🏼
Can you make a video explaining/speculating how these multimodal features (particularly voice) are made? Thanks
6:48. We might be able to brute force our way to AGI/ASI with compute only. But clearly, there are MUCH more efficient methods available, as human intelligence runs on what, 20 watts?
1 gigawatt = 1,000,000,000 watts.
1,000,000,000/20 = 50,000,000 brains.
10 gigawatts =500,000,000 brains.
I'm just saying that doesn't seem to be an inconceivable amount of power consumption to revolutionize everything we know.
@@jericolandry9872 I'm not sure what your math proves. My point is that our current approach to AI is PROFOUNDLY inefficient. All that means is there’s significant room for optimization. Whether we believe the current efficiency level already offers a reasonable cost-to-return ratio is a completely different conversation and unrelated to my point. I’m not discussing the economics of the current state of the art; I’m highlighting the fact that we have clear proof of concept, demonstrating that there are many optimizations ahead if we choose to pursue them. And there are real advantages to finding those optimizations, regardless of whether you believe it’s already worth it. For instance, what if we could instill human-level intellect into a humanoid-sized robot that didn’t need to be connected to a data center? That in ables a lot of important applications where that data center connection is a liability. That’s a worthy pursuit. Further, if we can't find funding for giga-sizing our power budget, it doesn't change the fact we'll get there.
The compute expansion to brute force ASI will rapidly be followed by compression and into a resaturation of the compute. Like an intelligence supernova.
That's the goal. You make it brute force cause you don't know better, and then you let it optimise itself
@@Raulikien of course, you don't need enough power for "everyone" in that case. You only need enough power to get to happen, then you let it self optimize.
Edit- server - atlanta, VPN the one in Norway. thanks, im in the uk on android and using a vpn worked, just had to wait a bit, when you close the vpn it stops working so i guess it check each time you open the app. The voice mode was very cool, but mine went a bit haywire in our chinese convo, i may have to alter the converstion instructions. it wouldnt wait for a reply in chinese and kept asking question after question haha
Can you tell us which VPN software you used and which server you connected to access voice mode?
@@aspzx it's not working today ! Idk why, but to answer your question
Nord VPN. Atlanta
@@aspzx I just tried again and it's working !it wasn't this morning. I'm on nordvod and atlanta
I think my replies are getting auto deleted @@aspzx, Atalanta - VPN the one in Norway. If I say the name I think it may get auto deleted
TBH bro, I don't even get the table thing though. Need a picture here hahahaha.
The red herring logic o1 is used is something I also noticed when trying to prompt engineer 4o to solve such puzzles. I wonder if a custom instruction for o1 that tells it to take any information seriously would help here too
Notebook LM took my by surprise.. is awesome.
Side note: Use the AI podcast feature to make both voices have a roast battle.
Now you'd have to train them on current creative jokes/snaps/roasts from humans
that they could optimize but believe this would be mad funny.
You could even get crazy and have each voice emulate two different comedians
like Dave chappelle vs Kevin hart.
Big potential here.
I would have said the strawberry is in the cup because if the top left corner is in his shoulder and the bottom right is at his outstretched (I assumed in front of him) ankle, the table is not just tilted, it is partially upside down.
NotebookLM is just amazing.
That large context is just great for chatting with long documents.
Yes , I was sufficiently interested that I actually subscribed.
Yay!
Mistral also thinks that strawberry is still on the table: Initial Setup: Jerry places a small strawberry into a normal cup and then places the cup upside down on a table. This means the strawberry is inside the inverted cup, resting on the table.
Table Description: The table is described as having an ornate left top corner positioned to nudge Jerry's shoulder, and its intricately-carved bottom-right top surface digs into his outstretched right ankle. This description is more about the table's design and position relative to Jerry, not directly affecting the strawberry.
Lifting the Cup: When Jerry lifts the cup, the strawberry is likely to remain on the table due to gravity. The strawberry was resting on the table, and unless it was stuck to the cup, it would not move with the cup.
Dropping Other Items: Jerry drops anything he is holding aside from the cup. This action does not affect the strawberry, as it was not in his hand but on the table.
Microwave: Jerry then puts the cup inside the microwave and turns it on. At this point, the cup is empty because the strawberry was left on the table when he lifted the cup.
Conclusion: The strawberry is still on the table where Jerry initially placed the upside-down cup. It did not move with the cup when he lifted it and was not affected by the subsequent actions involving the microwave.