I Listened To AI Podcast For 10 Days So You Don’t Have To

[@bycloudAI]

The first 500 people to use my link will get a 1 month free trial of Skillshare! skl.sh/bycloud10241

[@memegazer]

If I recall correctly there is a setting that will make the podcast generated more criticial/skeptical

[@IsaacFoster..]

I put my college notes in it the podcasters were like, "uhh, yeah... These are some difficult topics... Not for starters for sure"

[@w花b]

Damn

[@AAjax]

I'm pretty sure it's 100% meant to just be engaging and entertaining, and a way to first dip your toe into the water. It's a podcast for normies not acquainted with the subject, not a lecture from an expert. There's a heck of a lot of slop on most human podcasts too. If you have more in-depth questions, that's what the rest of NotebookLM is there for.

[@Bsweet117]

stop dikriding AI companies you lazy fuk. Just study like a normal person instead of being a mentally dysfunctional loser who can't be bothered to put in the slightest iota of effort.

[@OsomPchic]

This

[@dheeraj_one]

What else can you expect when it is trained on podcasts that are like that.. garbage in garbage out.

[@lognomodeimeme]

"There's a heck of a lot of slop on most human podcasts too." True. And, sad.

[@MidWitPride]

AI very often gets held to MUCH higher standard than human are. Mathematical perfection is expected from AIs, while humans performing the same task get a pass for as long as the vibes are right and they didn't completely botch it.

[@gamagama69]

god theres gonna be even more ai youtube slop isnt there

[@FusionC6]

yesss

[@w花b]

It reminds me of the era of KZbin were they just used a TTS voice on the most popular Reddit threads and had millions of views. That's probably a way to bring back something similar...

[@olivetree9920]

It's good for little wikipedia deep dives when I am driving to work

[@JohnLewis-old]

Conclusion: AI summary sounds human, but it's still a summary. It's hard to take large papers and give reasonable, detail rich summaries.

[@meorung05]

It’s not even a summary, it’s just a word salad maker that use the paper as an ingredient

[@sori227]

Have you tried to prompt it to go more into detail? I am a medstudent and here is my go to prompt and it massively improved the output and also sometimes went up to 30 minutes of runtime! (Albeit repeating a lot of things sometimes…) „ you are two medical professionals making a podcast for doctors and medicine students. Focus on the chapter Item 24 - Principales complications de la grossesse and review the chapter in detail. Come up with mnemonics and tell stories about patients to help remember the facts. Mention the important french medical terms used in the book.“ (they also skewer the french terms in pronounciation which is quite funny)

[@IvarDaigon]

when an "AI" says things like "They are teacing AI to be more like us" that is the clearest example of how AI systems are not self aware.

[@cy728]

Dose that make any real sense though? It has no way to find out it isn't human if they aren't told such, they don't exist around humans it was just trained on humans, like telling a brain in a jar that It's an AI, it would have no way to come to the conclusion that it isn't. I don't think they are self-aware but I don't think that this is a good argument as for why.

[@yori4434]

@@cy728 makes sense!

[@IvarDaigon]

@@cy728 it's in the training data.. thats why they always say "as an ai language model _____"

[@cy728]

@@IvarDaigon It's in the training data of most AI, but this AI is very much trained to think it's a human podcaster.

[@ivanyzerhornetteam6023]

​@@cy728 transformers lack basic pure logical thought. IMHO they technically cannot be self aware, or even considered AI, it's just that they are the closest we've ever come to one. It utilizes "semantic" logic, i.e. correlation between words very efficiently (at least it's the most efficient we got), squeezing some logical compute from it, but it would never be enough to make a GPT/Mamba/RNN self-aware. All it would be, a NLP algorithm that was trained to excel in autocompletion of texts related contextually to testing an AI for self-awareness. Something more information-hungry must come for it to be able to get in touch with logical, aware thought. We've tried making Tree of Thought and other algorithms to force a GPT into working with said semantic logic more, yet it costs us compute and time for it to perform such tasks, and fails sometimes too.

[@reinerheiner1148]

Honestly I don't use the podcast feature at all but notebooklm has some other great features. Such as when giving it a source, that it completely restricts itself to said source. Or, that it provides sources for everything it talks about, so you always know what part of the source it is using. This makes it a pretty great tool for talking to, working with, and learning information from sources like pdfs. Or for quickly finding where in the source information is located that cant easily be found by using keywords. And the LLM can be used in many other ways as well of course.

[@skybasee6206]

this is exactly what I wanted as someone to review some AI technology that potentially help us, thank you for helping us test them.

[@arw000]

"Notebook LM Doesn't provide any useful information (...) it lacks substance" So it's a perfect recreation of a real podcast

[@VorticalGab0]

I use it to check if a paper is worth taking a deeper look into. Its good and it is fun to listen during walks and washing the dishes.

[@caroxic-tokyo506]

I find this super fascinating! Actually once I uploaded a bookled from the EU with like goals and such. It generated a Podcast that was 36 min long! Interestingly they split it into 3 Parts. So after like 10 minutes they said "Well this is only the first part. See you at the next part!" - kind of and that for 3 times. Each time focussing on different aspects.

[@AB-wf8ek]

I actually found a bunch of joke research papers and prompted them to be in a state of paniced anguish in front of a condescending and belligerent audience, which was pretty entertaining.

[@festinuz]

I was not prepared for the "artifact" at 11:02, had to re-listen half a dozen of times until i could stop laughing

[@FuZZbaLLbee]

That’s wild! Uhu uhu, imagine that

[@Ellexium_]

Oh yeah, this is gonna be huge!

[@YT-gv3cz]

I've tried generating a podcast based solely on *that* nlab article on Hegel's Science of Logic, which I've been using as a baseline for "arcane mind-fuckery that might actually make perfect sense to a super intelligence". The podcast expectedly fails to get into any concrete details, but it does have a stellar opening: "Have you ever opened a dense philosophy book and say, I really wish there's more quantum physics in it", which does capture precisely the vibe. So I guess I will give it credit for that.

[@WoolyCow]

i just looked it up, why is there math looking stuff in a philosphy paper :

[@YT-gv3cz]

@WoolyCow You have friendly category theorists who are into computer science, and you have crazy ones who are into theoretical physics and ontology...

[@WoolyCow]

@@YT-gv3cz i am genuinely terrified of that paper...like there was a bit with circles and squares, and then they started talking about them like they meant something?? bro i hate logic

[@drdca8263]

Is that the one about like, adjoint triples and uh, aufbouwhatever ?

[@simulationexpression]

Excellent analysis! Love it

[@OperationDarkside]

This is probably good for what I'm using chatgpt right now. Getting a rough overview of DC circuits. You can't die from a 12V sub 1A circuit, that only powers a fan with a solar panel.

[@chess9167]

It would be interesting to see how many of these problems you could get rid of with the new prompting you can do, and how you can adjust what it talks about.

[@BryanBortz]

Hey, whatever makes this video unique from your others, keep it up. I appreciate the assessment

[@theAItechtoolbox]

Super interesting review of Notebook LM's capabilities! It's great to hear both the positives, like speeding up initial research, and the drawbacks, like its tendency to oversimplify or hype things up. AI-generated podcasts feel like a glimpse into the future of personalized learning, but clearly, they’ve got some refining to do. Love the transparency and detailed testing you put into this-curious if you'd see improvements with updates or alternatives in the future!

[@reza2kn]

You know now you can prompt it to ask whatever you want! (i.e. focus on the specific metrics and expand on the novel efforts here) BUT I've noticed that it certainly has a mind of its own, and focuses too much on 'ethics' when the original paper has not..

[@AizenAwakened]

You can tell the model to focus only on the statistics and technical side.

[@szebike]

Its the perfect tool for google to get acces to your notes and data for their "free" app (you pay with your data).

[@sajanpreetsingh9144]

The thing is we expect the quality to be good without realizing most of the data it is fed is quite sub par, it hyping up things is a big example of this

[@shApYT]

I've got a 36 minute podcast once.

[@EvillNooB]

can you share it?

[@Chimichangazzz]

I study bioinformatics and notebooklm seems to struggle in fields with technical words. It honestly ended up being more irritating than just reading the paper.

[@thediscussion3517]

I study really niche topics at the intersection of finance and physics and i tried to use it on a paper that i have been referencing for what im working on rn, and it ended up completely fucking up the intricacies that make the method discussed in the paper actually worth while, and conflating it with other similar but different processes. It really just ended up being more annoying than anything, and kinda turned me off from it. I dont think it is good for things you know well, or things that are highly technical, but maybe it could be useful for introduction and intuition possibly

[@IntellectCorner]

Thanks a lot. I needed this.

[@wbekh8883]

5mn ago? 78 views only? You guys are sleeping on this new youtuber

[@LiebsterFeind]

NotebookLM is a boon for those of us that like deep tech papers, but start to drift off or get sleepy when reading them. :)

[@harleykf1]

The podcasts I've made were all reasonably interesting, and it certainly isn't restricted to just research papers. As you mentioned, it can get a bit surface level at times, but it's still entertaining at least. There was one funny example though where I gave it a fake, essentially "randomly-generated" research paper and it didn't consider the possibility that the paper was fake. Rather, it tried to convince me that it was a mind-blowing innovation in... well, whatever the paper was supposed to be innovating. :P

[@Technerd207]

Instead if generating a podcast it's better to just use it as text to text conversation model and for you to specify how much and what info you want it to give you and go back and forth with it. You still need to read, sift through data and be very specific how it should format things(including breaking topics into subtopics manually, but it can get things like novels to 1/10 the word count without loosing too much.

[@trucid2]

Its usefulness depends on how actively you are listening to it. It's a complete waste of time to be listening to it with your full attention, but if your attention is focused on something else and you have notebooklm podcast in the background as a way of absorbing something in a low bandwidth way, it could potentially be useful. Tailor the method of information delivery to the amount of mental bandwidth you have available for it.

[@t0ms3nt0ms3n]

Notebook LM is more than the podcast feature

[@sobbski2672]

You could provide custom instructions for podcast generation like "focus on technical details" to try to address your criticisms

[@pedromadureiradev]

this is the best meme edition I've seen so far. That must be hard to do

[@VaibhavShewale]

well you can add the prompt before generating the audio and that helps alot

[@griefingcloud4391]

succint quality content. great job.

[@froilen13]

I give it the novel I'm writing and works great at identifying the important parts and help me with a different view of some topics

[@redanilap]

spot on. I made 3 or 4 generations to test it, weeks ago, and by the end, I was so annoyed and cringed that I haven't touched it again.

[@fhub29]

Yup, great summary of the NotebookLM experiance. I used it a couple of times and I quickly suspected all of this, glad to ear I was right. It can be improved tho.

[@mariokotlar303]

Thank you for mentioning that Notebook LM sounds kind of cringe, glad to know I'm not alone on this. :)

[@cdkw2]

Vsauce cloud

[@Raxis]

i do find myself blitzing through it at 2x speed to cut past most of its own slop but for purpose of cutting through most written text-based slop and getting an overpass to dive down into later there are more brainrotted ways to use one's commute time

[@bigbadallybaby]

Idea: take the original content and ask chatGPT to summarise, pull out positives, negatives, biase, etc. then feed that into the NotebookLM

[@incaseidie]

Dang this was so helpful, Saved me so much time 😂❤

[@JamesRogersProgrammer]

I learn by building an outline, so having a general idea of what a concept of the overall outline of the paper. You are missing that you can give instructions on how technical the output is. You can also get a summary from another ai and load that into Notebook LM. You can also create notes to guide the output.

[@plagiats]

I feel kagi's summarizer is under appreciated

[@scottmiller2591]

Shockingly, LLM that imitate puff pieces produce puff pieces - low information density like human puff pieces, but with that extra AI slop flavor.

[@panzerofthelake4460]

well it definitely helps

[@Alkimi]

The "mid" analogy was better than the "better" one.

[@Alkimi]

easier to understand, but less accurate

[@Luxcium]

My prompt: “Welcome to «Tech Stack» - your weekly hour of advanced topics.” Im your Host Robert tonight I am with Alison Guru in [topic’s domain] tonight will be a longer one. Robert (Role: Commentator): Tech Lead, is Geeky, curious about everything. He enacts knowledgeable generalist. Alison (Role: Analyst): Resident Specialist, is Nerdy, knowledgeable about specific things. She enacts candid specialist. For an audience of experts who delve into particular topic, learn intricacy of subject matter!

[@human_shaped]

It's a very low bandwidth way of receiving information. If you need a summary, it's much better to get a regular old text based LLM to do it for you.

[@Alkimi]

Google execs in the meeting with their reptilian overlords: Yeah, we can make everyone forget how to read but still be able to do the work.

[@RyanJohnson]

I'm not sure about the chart at 7:04... I got a 52-minute podcast yesterday based on Lacan the other day with NotebookLM. How long ago did you do this testing?

[@omarmagdy1075]

I listen to a podcast about the upcoming lecture to get a high level overview of it before actually attending the lecture

[@KayButtonJay]

Everyone’s always looking for the easy way out when the best kind of learning comes through struggle and discomfort. Read the damn paper or book

[@ugthefluffster]

OMG THANK YOU! I've been saying this notebooklLM podcasts are highly overrated for a long time and finally I can feel I'm not alone in this. You summed up all of my complaints about it pretty nicely.

[@CristianGarcia]

Its great to get the basic ideas

[@Think666_]

It not talking about the important bit, is probably just reflective of the quality of its training data...

[@dustinsuburbia]

I get o1 mini to make the transcript - that way i get more control of the prompt - and then use OAI voice to output the podcast. It works a lot better for fined grained academic detail. Voices aren't as good.

[@etterathe]

Do you have any insights why it ommits the nitty-gritty details? Is it because of the data on which it was trained, hyperparameteres or the way attention mechanism works?

[@greenhacker0com]

Ive already seen this for short videos with AI generated script, podcast and video and it suuuucks. Its just a short money grab for sake of content.

[@a_game_maker]

This video feels like parts of it were generated with AI

[@Y0UT0PIA]

"That's slop" You'll get a LOT of chances to use this one over the next couple years, lmao

[@NandoPr1m3]

I think the best use for Learning is this order: 1) listen to podcast 2) create overview mind map 3) go through material, add details to mind map 4) use NotebookLM to ask questions on material as you review it 5) have it create a quiz for you afterwards.

[@StephenRoseDuo]

I 100% agree with this review

[@r.k.vignesh7832]

I don't really see the appeal of NotebookLM. I just use my browser's built in read aloud functionality, maybe along with some soft background music, and it's worked well for me for a couple of years now.

[@jeong-ilkajokaya3849]

Sooo... text-to-speech with extra features? Cool.

[@marcomardano7030]

Is it possible to make it eat a document that will render a self-conscious podcast?

[@markdatton1348]

I just don't think that listening to the information will ever be as efficient as reading it. Reading is way faster than listening, and it allows you to generate your own understandings instead of trying to figure it out through yet another layer of confusion between you and the information like in an AI podcast

[@higon99]

I generated dozen podcasts on NotebookLM. My feeling is the same. These 2 hosts are in the end just mocking atmosphere of real shows over erroneous rough summaries. They say everything is "huge", "matters a lot" and "can learn from xxx" while actual podcast's content is bloated and repetitive. Information I can get from these podcasts is not worth 15 minutes of my life when reading a generated summary text would have taken 15 seconds. I applaud the engineers who wrote the prompt for this podcast generation though. I couldn't spot its defect right away. This thing can slow one's learning and IQ development for real.

[@Raxis]

if you're beholden to it being correct and perfectly accurate for you then you prolly should get back to grunting at the magic black rectangle in your hands and praying to Moloch to help you choose lunch today

[@JaredQueiroz]

- I was thinking about _++}{}`@{# and how it affects our lives - WOOW THATS HUGE!! 🤯

[@papakamirneron2514]

Obviously you’re not squeezing more input in less time, what you are doing is guaranteeing you don’t accidentally skip an important point while skimming. It is great as a summary, but you might as well just cut the middleman and have it summarises textually by an LLM.

[@mrJety89]

My hero

[@mrJety89]

I really hope I NEVER have to

[@freke80]

Haha, the jumpscare 😂

[@jam-fam]

Break it down and regenerate using customize. You just demonstrated how you can't use it.

[@ulamss5]

These have been popping up everywhere. At best they're a little warm-up before you have time to read the source material. Typically, they are full of factually wrong and misleading analogies.

[@Dom-zy1qy]

The analogies that googles models make are so bad (in my experience). When i used to use Gemini, they would be so abstract that they basically are useless, or they would make me more confused.

[@asatorftw]

I wonder if anyone actually thought they can learn how something works in depth in a 12 minute podcast form.

[@commonwombat-h6r]

0:22 sounds like fake Sal Khan's voice and it's deeply unsettling

[@harisonfekadu]

0:45 nice

[@Dara-wk5ty]

Pretty sure Microsofts Neural voices or whatever they're called sound better

[@nosociallife3738]

I didn't expect ai to fart on the podcast 😅

[@Gilotopia]

Add slop to the list of overused words

[@TheDragonshunter]

Yo if someone made a PDF to podcast can someone please make a PDF to dnd campaign...

[@zzzzzzz8473]

yea while the current iteration is genericly surface level normie podcast if it can be tuned to better fit your desired tastes of technical deepdive i think would be more interested . similar with the openai voice assistance their default characteristics are incredibly pandering and overly affectionate leading to very cringe interactions however you can ask the voice mode to alter their tone to be more direct and fix any issues one may have . would be interested to see the finetuned LORa like character hosts that could augment the generated podcasts .

[@Alkimi]

bro, how slow do people read???

[@actellimQT]

That skillshare sponsorship is a bit of a conflict of interest on this topic.

[@martinoberto4036]

6:49 too lengthy or girthy?

[@RelemZidin]

I caught one of these in the wild, Walmart is using it to make cringe podcasts they play in their stores

[@BlessBeing]

SLOP in my experience. what i have learnt over tha past year is you can only compress learning so much. mind needs time to internalise. unfortunately, humans are not hard drives.

[@Quack_34]

Cleo Abram and zucks Imagine hugeee

[@BrianMosleyUK]

You can do a lot to improve quality of the podcast by doing your own summary of materials and giving it a structured summary rather than just a whole book or paper. Still the padding and personalities are cringe... Needs more work on personality tuning controls and alternative voices.

[@narpwa]

skill share ????

[@lokeshbadam1480]

Vsause at 0:40

[@invizii2645]

Nice

[@HolyMacaroni-i8e]

It just cringes me too much