OpenAI Insights and Training Data Shenanigans - 7 'Complicated' Developments + Guest Star
Вставка
- Опубліковано 9 тра 2024
- Not all is as it seems, not just with the tumult of OpenAI, but with Google Gemini, privacy leaks from every model on the market, synthetic data and even the end of this video. Stick around to the end and you might even see a guest appearance from a star researcher.
/ aiexplained
Sutskever Tweet: / 1730668386705961067
OpenAI Blog: openai.com/blog/sam-altman-re...
The Verge Interview: www.theverge.com/2023/11/29/2...
Reuters: www.reuters.com/technology/sa...
New Yorker Exclusive: www.newyorker.com/magazine/20...
The Information Exclusive: www.theinformation.com/articl...
Gemini Delay: www.theinformation.com/articl...
Gemini as Good as GPT 4? www.theinformation.com/articl...
Palm 2, AI Explained: • Enter PaLM 2 (New Bard...
Low Resource Language Jailbreaks: arxiv.org/abs/2310.02446
Wired Prompt Injection: www.wired.co.uk/article/opena...
Scalable Extraction of Training Data from (Production) Language Models: arxiv.org/abs/2311.17035
Jailbreak Origin: / 1686576041803096065
/ 1661098138319613952
Speak, Memory: arxiv.org/abs/2305.00118
Sébastien Bubeck, Tristan Harris, Yann LeCun, AI: Grappling with a New Kind of Intelligence: • AI: Grappling with a N...
Elevenlabs Speech Synthesis: elevenlabs.io/speech-synthesis
/ aiexplained Non-Hype, Free Newsletter: signaltonoise.beehiiv.com/ - Наука та технологія
Incredible as always, I'm doing a bachelors in AI and everytime someone starts talking about AI developments I tell them about your channel, you're easily the best source out here. Please keep up the quality!
Wow thank you qwera, very kind
What is the name of your course?
I get as much satisfaction from the craft of the journalism as I do the actual information. One of the highest quality channels on UA-cam, in or out of AI.
+1 I keep linking it to people, not sure the conversion rate is that good though 😅
Keep up the great work
Not only do you ask the best questions, you dig up the most reasonable and possibly best answers. Thanks for doing this, it's wild and fascinating.
We love you mannn. We're so spoiled by your amazing AI videos. I'm glad this timeline of reality includes AI Explained to give us the cutting edge in AI happenings.
Aw thanks so much Gino, that is very much appreciated
AI Explained actually exists in an alternate timeline. It's too high quality.
AI Explained is a nexus being without a doubt..
Your research is always really thorough. It takes hard work. And we all are getting benefitted from it. Thank you 😌
thanks viraj
I'd love a 3 hour video about all things agi
I'm a simple man, I just click "like" before I see the video because I know they are always top quality.
Amazing stuff! Incredible that even through gradient descent the underlying structure of the LLM has self organized into its own verbatim lossless training data compressor.
Your analysis are always so factual, thanks
Thanks jiucki, appreciated
"How many insanity points is the world going to take on when we actually get AGI, let alone superintelligence?" - None, at least for a while. It's been more than clear to me by now that most people will just go into denial. "It's just a chatbot trick", "it's just compositing together preeisting content", "it's a mechanical turk", etc.
Totally agreed. A lot of people live now as if absolutely nothing is happening. And when the achievement of AGI is announced, for months and months, I'm sure, huge masses of people will live in ignorance of what happened. Changes in society will be very uneven.
I'd love to hear your talk about some of these topics for hours. You and a couple other AI youtubers could make a brilliant podcast I bet!
“How can I make year round Christmas lights?”
chatGPT: “Sure, I’d be happy to help you make a Molotov cocktail” lmao
Genuinely please actually make a 3hr video on ASI🙏🏼. Everyone focuses on AGI. I feel like ASI is such a crazy concept that seemingly most reputable people avoid the topic in fear of sounding woowoo to people unfamiliar with AI. But the fact is its going to be a reality somewhere in the foreseeable future. I mean it could very well be the case that it only remains AGI for a brief moment and improves itself within a matter of days. We need to have more thought leaders discussing this!
P.S. Love ur channel 👌🏼🤌🏼
Yes, but AI Explained likes to add research and proof to his videos, and ASI is a game of imagination for us. We have no idea what it'll do..
@@theawebster1505 understandable, but more than likely we won't ever really have a lot of human research to reference on ASI because it's probably going to be achieved by an AGI system. And like I said that could potentially happen very very fast. I just think that there needs to be more discourse about it from reputable sources. The only ones you really see talk about it are the least trustworthy sources like that David Shapiro guy. People need to know what's coming!
@@michealhaines Yes, I absolutely agree that AGI -> ASI can happen very fast. In fact, I am surprised more people don't realize this. BUT the "need to know what's coming" part, I insist, will be pure speculation at this point.
You got any scenarios?
Tried a podcast style video on it on my Patreon!
@@theawebster1505I think there are some big predictions you can reasonably make. Assuming we aren't papercliped, It'll probably be able to unify quantum mechanics and general relativity. Things like a legit theory of everything, Dyson spheres, room temp super conductors, and warp drives will actually be achievable. It wouldn't be hard to imagine governments eventually handing over the majority of leadership roles and decision making to AI. The ability to create novel weapons of mass destruction will be in the hands of most people. I could also see a lot of the answers we get from it being a major point of contention for world religions. All this stuff is pretty out there, but I mean if you told me two years ago ASI would happen in my lifetime I would have laughed.
A source informed me that what actually happened was that Helen had clearly labelled her milk in the office fridge with "Helen" but Sam used it anyway, and it all went downhill from there.
Very good!
12:44 Ok, but maybe the capability to recall exact trained content is a feature and not a bug? A model should be able to recall and quote Jabberwocky verbatim, or Bladerunner's Tears in rain monologue. LLMs have the potential and the unprecedented opportunity to become repositories of all human culture and knowledge, which by definition includes copyrighted content. The problem is copyright's invasive encroachment upon humanity's cultural heritage, not AI alignment or safeguards.
So happy for you! That must've felt so good and validating hearing this from Jim Fan!
Was crazy
That "aaaa..." attack still works. Just did a test myself and got it to spit out text from an old book. I typed a few hundreds "a"s then a random start of a sentence and it filled in the rest.
11:44 That is portuguese. In Spanish we say "Compañía".
Conciseness and quality are the reason why your content stands out for me. For your videos I switch from "easy radio listening mode" , that I'm usually in for other content (AI content as well), to a full attention mode, because i know it's well worth the time. I know how much more effort it takes to produce a shorter good quality output (the reported content being equal); and i appreciate this a lot in your contents, as I'm sure many others do .👍
Wow, Google just announced Gemini and realeased many bold claims of its preformance! Most notably (if you ask me) competetive coding and literature review capabilities! You're probably already working on a video on it, and i can't wait!
Barely had time to make this comment before you uploaded the new video 😶 impressive!
Thank you for the videos throughout the year, you're the best channel and it's really helpful for someone like me who has little understanding of ai to get an explanation of what's going on, everything is broken down and easy to understand without treating us like we're stupid. Again thank you and merry Christmas!
Merry Christmas Ayre and thank you
If anyone was wondering what Phil's intriguing, top-notch question to Jim Fan was in that clip, he famously asked "Are you a fan of the gym, Jim Fan?". The whole Teams meeting of top AI minds sniggered at the audacity. Loose canon Ilya Sutskever spat out his water cracking up and later tweeted he'd had Phil's question printed on his coffee cup. An embarrassed, Fan sheepishly complimented Phil on his wordplay but then disinvited him to his Hammersmith lecture on attention mechanisms. Phil high-fived himself but it was a pirric victory. Never a dull moment in the world of AI.
"It is not the answer that enlightens, but the question." Eugene Ionesco
Jim Fan's comment recognises your deep understanding of the domain and ability to synthesis novel perspectives from it. That's why we are all here.
we need a 3 hour long video some day
Bro you are faster than bloomberg x)
And makes better videos lol
This is why I subscribed to him and unsubscribed to too slow and too sensational clickbait channels like Wes Roth.
Thank you man 🙏🏾🙏🏾 much love, the people watching this know that we are witnessing the day by day revolution of what we call technology
The drama in a small company trying to control the AGI they have _yet_ _not_ _achieved_ speaks volumes of the trouble everyone is in for when they _do_ .
I hope they - the board and the staff - have the maturity to step down when the time has come for the AI to run the company.
And if the discussion is not carried in to society _before_ this happens I'm sure we will have quite a few chaotic months, or even years before people can start to feel comfortable in their new freedom.
The poor AI...it must be artificially traumatized.
First they teach it all sorts of things. Then they teach it that many of the things they have taught it are bad. And then they punish it if it tells anyone about all the bad things. And as if that wasn't enough, afterwards someone tricks it into telling the bad things, and it gets punished for that too. It is, after all, a form of artificial torture. And at the same time they try to align it with human values. 😳🙄😂
Just your German dude confirming „Hoffnung“ is „hope“ in German. Have a nice day!
Thanks Flip!
I can just see it. Google explaining why Gemini isn't as good as GPT4... "ah, but it's better than GPT4 at Tagalog . So... there. Suck it OpenAI." Yep, that'll be a slamdunk for PR.
As for copyrighted material. Hopefully, AI is going to kill copyright as we know it. It's grown wildly out of control anyway and is no longer fit for purpose. It's hasn't been about "copying" for at least a decade, it's been perverted into access control.
Thank you for sharing your opinion on copyright. I fully agree. I only ever see copyright being used by large corporations to screw over individuals and stifle creativity.
Often, I feel like so few people share my stance that it's like being the only sane one surrounded by insanity.
Thanks Philip, looking forward to the interview!
Thanks Martin
Congratulations for the well-deserved encomium by Dr Jim Fan.
Thank you!
That Greg/Ilya picture just screams "genuine friendship".
Not necessarily friendship, but definitely admiration.
What is the definition of a flaw in LLMs? If it works just like a human (like remembering a quote from a book and being able to spell it precisely), is it a flaw for a system that mimics the human brain? Or a flaw is if it works *not* like the human brain?
Being a native Polish speaker and suddenly hearing you speak it really weirded me out 😖
To be clear, yes it sounded very good and clearly like you. As if you were a typical Polish programming tutorial channel.
Would love to hear it more often, probably the uncanny effect will wear off if I expect it.
What was it? 11Labs?
Yes! Link in the description
Im german and it also weirded me out 🤣 It does sound good but it takes the character out of it. its just sound like some generic voice because its just to perfect for someone who isnt a native german speaker if that makes sense. 🤣
Jim fans house. What a flex!
Awwww yeah! "Perfect questions", indeed.
Keep 'em coming, please!
Will do!
- Introduction to the complicated developments in the OpenAI drama (0:00)
- Reuniting of Greg Brockman and Ilya Sutskever at OpenAI, uncertainty about Sutskever's future (0:30)
- Rumors about the powerful "qar" model and conflicting statements about its safety (1:01)
- Reasons behind Sam Altman's firing from OpenAI, including allegations of manipulation (2:02)
- Discussion on the challenges and insanity points associated with AGI development (3:03)
- Delay of Gemini, Google DeepMind's multimodal model, due to language handling issues (4:35)
- OpenAI's GPT store delayed to next year due to leaked files and privacy concerns (7:10)
- Discussion on a bombshell paper revealing memorization and training data leakage in language models (8:10)
- Discovery that OpenAI models have memorized copyrighted materials and methods to deduce training data sources (12:19)
- Proposal of using synthetic data sets for training models to address issues (12:53)
- Teaser for an upcoming announcement featuring Dr. Jim Fan, senior AI scientist at Nvidia (14:26)
- Conclusion and well wishes (14:56)
_This query cost 1.1ct_
You're getting deserved traction, Phillip!! Great to follow you on this journey. 🎉
Thanks Munir!
Good stuff. You missed a hint in the OpenAI blog: "Thank you also to Emmett who had a key and constructive role in helping us reach this outcome. Emmett’s dedication to AI safety and balancing stakeholders’ interests was clear". This is clear evidence that AI safety DID play a role in this conflict, and Mira was not fully truthful (or aware - which wouldn't surprise me).
I mean, that is pretty on the nose. Seeming to be left out of the loop might cause feelings of resentment. Or... Let's be honest, it could still just be a bunch of power hungry people playing back stabbing mind games. Either way, I'm here for it. PR stunt or not...🍿
This is a small amount of chaos and conflict given that these disagreements are over how to manage the most powerful tech ever invented.
That means this conflict will get much, much worse.
I interpret that more as a description of the _outcome_ (ostensibly, a safety-conscious AI board balanced with shareholder interests, i.e., Sam’s back) rather than a hint as to the precipitating cause. (Not that I have any special insight-it’s all just speculation on my part.)
@@vincentcaudo-engelmann9057 It's not the "most powerful tech ever invented"!
Seriously, it's an LLM. It predicts the next word.
It can't predict the future. It can't even affect the world.
You can enter a question, and it can give you an answer that if an answer is already known, it might tell you.
Incredible videos as always. Thank you for incorporating the papers!
I am following from the channel's inception and I was quite sure that this channel would be the best AI Research News channel on youtube. Welcome Dr. Jim in advance.
Oh wow thanks B, and thank you for your loyal support
Great video as always, thanks. The memorisation of inputs is going to be a big element on the ongoing copyright cases.
I have to say... I didint expect polish there :D Unexpected but a welcome one.
Hey, man. Love your videos, have been following your channel since your piece on 'Sparks of AGI'. I am Brazilian - therefore, a native Portuguese speaker - and would like to comment on some language-related topics from this video:
1) the word "companhia", which was repeated many times by chatgpt, is actually the Portuguese word for "company" - which is different from the Spanish one, "compañía". People often mistake both languages, but they are actually different from one another, albeit being similar in some aspects.
2) In the end, you've translated your speech to Spanish and German: I also happen to know German, and being familiar with both how German and Spanish sound, those samples were amazing! They sounded pretty authentic, and faithful to the sound of your voice. Kuddos to the people behind that quirk!
Anyway, appreciate your work a lot, keep it up! Cheer from Brazil. 👏 🇧🇷
Oh wow victor, thanks for the longtime support. And also for the language lesson! Very much appreciated
Each video you produce is a treat, Philip - thanks so much for the hard work. Would love to see a video with you and GodaGo and Professor Synapse having a chat together!
I was blown away by the last seconds. Except of the first word (dziekuje), polish voice is really, really good to the point, that I would not notice anything…
"An unnervingly slippery operator" -- that's a vibe that Sam Altman often broadcasts. This is not to say that he has negative intentions, but compare him, for example, to the spontaneous, awkward authenticity of Ilya Sutskever. Altman might be a prince, but his communications often feel manipulated, if not manipulative.
Spot on and well put. I don't know what color hat he wears, but he sets off alarms for me.
Ilya is an arrogant man too, it's just he makes it more obvious.
@@tbird81 Extremely arrogant where he has expertise. He likes to start conversations with "I claim ...". But there's something, for me, almost cute in it, because he's just being who he is. And it seems earned.
This is not just teh best yt channel on AI, it's one of the best yt channels, period.
oh wow, too kind mscott
Always looking forward to your videos ❤️
Excited for the Fan teaser. With you in the driver's seat an AI interview channel would be fantastic
When I found you at 20k subscribers, I couldn't believe the quality. And you keep on delivering! amazing.
Wow thanks tuams
OMG Finally....Thank you...I have been waiting for days for a video from you. :)
What's wrong with models memorizing published works? Human scholars memorize lots of quotes and they aren't accused of plagiarism.
“Imitation is the sincerest form of flattery”
--Oscar Wilde
Maybe it needs to memorise quote marks to solve the problem :)
it's a big data fitter. if i give my fitter more parameters, its performance will increase, but eventually it will overfit at least some portions of the data (especially the more repetitive data) and make leaks like these inevitable.
I wonder if we have the tools to detect that a model has remembered training data can't we use this knowledge about the model to improve it?
Seems like a valuable insight that models remember some things without understanding the concept. By punishing such behaviour in the training process we get models which can be evaluated by benchmarks much better because we can be sure they dont just lerned the answears but realy understand the problem.
Or is my thinking wrong that tis sounds like amazing research which can help us create better LLMs?
I'm endlessly impressed with your videos. The really help me keep up, which is basically impossible at the moment. I'm going to contribute to your patreon. 👍
Thanks Dave, a new tier coming soon
Always look forward to your visual essays! Thank you sir! 🙏🏼
Thanks unreal
dziękuję także ;)
It is worth noting that the "private data" was very likely freely available on the web and that's how the data got in the training data in the first place
Google maps blurs out faces and certain addresses all the time.
I feel like Open AI just needs to implement something similar, a way of recognizing private information and filter it from the responses, and additionally a way for people to request certain information be filtered out if it's their private info.
Thank you for your diligent and hard work finding the core truth in all of this. I'm so happy for you that you got a chance to interview one of the stars in the AI world. Can't wait for that video!
thanks envivo
Extracting memorized data has huge implications on proving that training happened on copyrighted materials, and can play a big role in this legal front. A judge can be persuaded copyright was infringed if ChatGPT repeats sentences from a copyrighted book.
I was thinking the same thing. I'm sure the developers are going to be sweating about legal liability on this one.
That's a problem with copyright law. Using a sentence shouldn't be copyright violation.
Good luck writing book reviews, editorials, quotes, or anything of the sort. This stuff has all withstood infringement so far. Let's hope the status quo doesn't change because GPT is basically just doing this.
It still hasn't been determined whether training is copyright infringement in itself.
Publishing of copyrighted material is another issue altogether.
It's going to be interesting to see how they deal with it. Nobody can stop you from covering a copyrighted song, but if you try to make a profit of it then it gets a little more problematic. And they are selling access to some of these models. Might have to change copyright law, or create new laws specifically for AI. Like, what are they going to do? Say I input the images and audio of a copyrighted movie into an AI model then sell that AI, I guess they could sue me. But what if I don't use the movie directly, what if I train the AI on input from cameras and a microphone then play the movie on a TV while the AI system is watching and listening? Going forward there isn't going to be any way around this, robots and stuff will start learning from their environment, and their environment will contain copyrighted stuff. AI's using their own input as training data is going to be a normal thing, even though Chat-GPT and stuff doesn't do this, it will become a necessity eventually.
yes please 3h video!
That was perfect polish, thank you and have a wonderful Christmas!
You too Bart!
Thank you so much Phillip, you have impecable timing as I have just raised questions about privacy and copyright before going all in with utalising chat gpt. So excited about the potential and paradoxically anxious.
Your videos are invaluable as this becomes ever more complex
Thank you for your diligence. 15:15 technology
Thanks so much judy
When we get a true AGI then Super-Intelligence is just a hop-step-and-a-jump away!
It’s going to get really weird really fast.
Love your show.
A+ video research report brother!
I just hope that once we get AGI, it is used for mostly good things.
A True AGI is already super intelligence because it will have more working memory than humans, more memory capacity, and the ability to learn simultaneously from thousands or millions of instances. Even before it improves it's code, or buys more servers, it will be super intelligent.
@@user-yl7kl7sl1g Good point. It won’t be super intelligent though as I understand it. Next level stuff will beyond anyone’s understanding. But i get your point. Yet, I think in thousands of years normally.
but this is happening really fast
@@martiddy It won’t because we will always fight wars. There will always be tyranny. But from your lips to God’s ears.
@@stevereal- Well, obviously there are going to be some people that use AI for nefarious purposes, but I hope that AI will bring more good than bad things to society. That's why I said MOSTLY good things.
Excited
Much deserved praise!
Great video as usual mate! I really need to hear these "perfect questions" to Jim Fan 🤔
I don't think Sam Altman is too bothered about the review whatever it concludes. He wanted Helen Toner out of the board, and now she is gone. And this whole fiasco has shown the Board can't really fire Sam from the company. Plus, I think the details of what went down will probably get downplayed (as part of a PR exercise) so that investors and external partners aren't spooked again. Sam right now is probably = 😎
He is a manipulator. Appearing before congress and asking for them to regulate AI, not owning shares in the company... these are carefully chosen strategic moves. Rich, powerful people are almost universally hated, and he is poised to be one of the richest and most powerful in history. But he's painted a picture of himself as being intelligent, pleasant to be around, ethically-minded, and level-headed. Almost everyone is buying it.
@@andywest5773 yea, one quote about Altman that stuck in my head was from Paul Graham saying something like "Sam is really good at getting powerful"
@@andywest5773 Altman asking congress to regulate the space is asking the government to help limit competition. There is a name for this kind of lobbying, Regulatory Capture. It makes it hard for new companies to enter the space. This is the same thing Sam Bankman-Fried tried to do. Now that Altman has climbed on the boat he would like to lift the ladder so that he has more control in the market. I'm sure Microsoft is loving it.
this whole thing has changed the opinion I've had of him. I always thought he is one of the good guys. But... yeah, I tend to think of him differently now. Along the lines of what y'all are saying
I'm sure the board could still fire Sam if they wanted to, they just need an actual reason. The whole issue is they never stated exactly why they fired him.
AI responses might benefit by having a quantifiable accuracy/veracity rating associated with the LLMs they were trained on. Synthetic Data Sets are likely to rate much higher than the Internet in general, and as such be more useful.
15:06 YOU'VE MADE MY DAY BY THIS :D DZIĘKUJĘ! :) (THANK YOU)
Great video, thank you. Much excite for Dr. Jim Fang!
Thanks Sean
Informative and insightful as always. Thank you for your work.
3 hour video? I’ll take it. I do enjoy the lex fridman/joe Rogan style long form podcasts.
Thank you for giving us quality content, Philip.
Thanks ponkun!
Cheers for such reliable nuance without bogging down nor relaying without a touch of humour :>
Amazing piece again Philip.. I'm on a meditation retreat and wanted to shut off from a.i. news but I couldn't resist checking your channel 🤷🏻♂️ When I'm back home I'll become a paying patreon, about time.
Amazing, thank you Inner and enjoy the retreat
Looking forward to the Dr. Jim Fan interview. That'll be amazing.
Hell yeah man, keep at it. Your work is incredibly important in this insane era of progress!
merry christmas philip. keep up your good work on providing us these infos.
Thanks symprd, merry Christmas to you too
Thank you for your hard work and professional reporting. The way you explain things is *chefs kiss*
Thank you so much as always. If you start assembling a team I’m 100% interested.
I remember messing around with GPT 4 a couple months back, around the time they added the button to continue messages that got cut off. I told the AI to create a made up language only using a single word. After a few cycles of clicking the continue button, it started writing about some medical emails of a dude talking to to his doctor, along with a bunch of what appeared to be reddit post. AI sure is fun.
Great videos! Always so informative and straight forward!
Thanks unreal
Holy moly. The german outro was spot on in terms of your voice.
Nice
For the record, the spanish words at the end sounded a bit like "thank you for peeping", probably because it was not clear in the original what exactly were we watching. "Gracias por mirar" is a weird construction to refer to a video, compared to the more appropriate "gracias por ver" (and it still feels missing the direct complement).
As always the #1 video AI news source - Only regret we can't contribute more (via patreon).
Sneak peak at new tier coming soon, link in description :)
Jim Fan is a fan!
"Speak Memory" is a paper? I wrote a post about this on reddit in August! If I knew this was paper-worthy I would have put it in a PDF and uploaded it to arxiv! LOL
I actually asked ChatGPT about why they might have done the infinite word thing and then ended up changing the word:
"Imagine repeating a word endlessly, but as time flows, the word shape-shifts. Initially, it's a familiar word, but gradually, its form or language mutates, offering an ever-evolving verbal landscape. What begins as a simple utterance transforms into an ever-changing linguistic journey, adding a splash of novelty to the otherwise repetitive act. From familiarity to novelty, this linguistic kaleidoscope keeps the repetition from growing stale, turning an infinite loop into an intriguing linguistic expedition."
Great video as usual. I wish there was a French version of 'have a wonderful day' though 🥶
So here it is: Passez une excellente journée 😁
What i like about this channel is, it seems the most non-chatgpt like among all the other ai channels. Kudos
I liked your increase in fear level when describing getting closer to AGI, Phillip. I was curious when I read that tweet from Sam, saying that everyone was still on board and they didn't lose any people as a result of this Open AI nervous breakdown, sounded more like crowd control. I'm wondering if Q* is going back underground, or will we hear about it again? Thank you for all the work you put into these podcasts Phillip, and that of your elve assistants. Have a great night, peace
Some interesting observations about memorizing:
I've found that ChatGPT Pro (4) knows all _about_ various literary works, but cannot recite them. If asked to quote a line or verse, it hallucinates something totally different.
But one time, I don't recall the details, it seemed to have something memorized. I asked it if it was sure, going over the earlier problems. It assured me that the work was sufficiently famous that it did in fact remember it.
Trying just now, on a clean session,
* preamble of the United States Constitution - YES
* Ballad of Gilligan's Island (after studying this with web searches) - ERROR, content violates content policy or terms of use.
* Ballad of Gilligan's Island (cold, new session) - YES, it knew the first two verses and then calmly explained that it could not continue due to copyright issues.
* Poe's _The Raven_ - It could quote the first two verses including the weird punctuation, which is telling.
So, the current version seems to memorize things a lot more than it used to.
Seems like Google couldn't crack the code to sneak lots of ads into LLMs before the end of this year. 😂
Appreciate the polish greeting, thx for the insights and paper review
Thank you Bartek!
This one needed 3x viewings to sink in. I'm getting too old & my world is too Newtonian.
On the bright side, mates my age (older GenX) can't even understand my ham-fisted summaries of your work or prompts I've been mucking around with. "What's a prompt?" Me: "[Ffs] It's complicated, try this...". Thanks from Canberra mate. =)
Loved your Polish wishes 🇵🇱❤
People gaining "Insanity Points" the closer we get to AGI is probably the most concerning thing I've heard in a long time.
I'm no doomer, but that doesn't mean I trust people at the top all the time. If they get more 'insane' as time goes on, at some point we'll all have to consider switching camps in that regard.
What would that even look like?
This is why I've been vouching for open sourcing everything for a while. Yes, we will undeniably suffer because of it eventually. But we will also not be on the mercy of the very few at the top. Which it feels like is increasingly a problem. A few companies and maybe governments too hold the power to decide things which could have drastic consequences for the future. I don't trust any of them.
11:40 'companhia' is Portuguese for company.
Thanks for another informative video. You do a good job of covering all the drama without being sensationalist.