Pod 4: Unpredictability: AI, Content Creation, Timelines and Vernor Vinge - Let's Think Sip by Sip
Added 2024-03-24 19:09:25 +0000 UTCThe only theme for this episode is unpredictability, from the swirling new rumours of GPT-5 release dates from Business Insider, to the challenges of promoting interviews that don't happen, behind-the-scenes chats, how we can't rely on AGI Lab leader reassurances and key extracts from the portentous essay from the late Vernor Vinge: 'The Coming Technological Singularity'.
GPT-5 Summer?: https://www.businessinsider.com/openai-launch-better-gpt-5-chatbot-2024-3
Job Rumours: https://twitter.com/norabelrose/status/1768421994562031861
Not as Good As Doctors at Diagnoses: https://browse.arxiv.org/pdf/2401.08396
Geoffrey Hinton Subjective Experience (Talk starts around 7.00): https://www.youtube.com/watch?v=iHCeAotHZa4
Mustafa Suleyman Quote on AI Explained: Time to Superintelligence: https://www.youtube.com/watch?v=vvU3Dn_8sFI
New Quote: https://www.theguardian.com/business/2024/mar/20/mustafa-suleyman-the-new-head-of-microsoft-ai-with-concerns-about-his-trade
Vernor Vinge Essay: https://edoras.sdsu.edu/~vinge/misc/singularity.html
Comments
even hardworking :) SMEs we deserve some clarity for sure. Do we invest heavily in fine tuning GPT4 (we have access to that) while there is very high chance we don't need it with 4.5? even less with 5? do we change our cot/tot to a more human centric in context because aligning many agents in generic complex tasks in large real projects become more an r&d exercise than a really effective approach (vs shorter human-AI interactions with multiple agents in same context), etc. I was exactly thinking this is not cool at all... Sama said indeed that gpt5 would be a leap as being as 3.5(or 3?) to 4... and I am driving my company and taking strategic decisions with that in mind... what if he was hyping the stakes? 🫨
Robert Gomez-Reino
2024-04-02 17:00:03 +0000 UTCI wonder if, in the near future, holiday destinations and flights will be cheaper on new model drop dates 🤔?
GGuy
2024-03-30 01:23:19 +0000 UTCIt would be nice to have a bit more notice on model release times, but I think in the bigger picture, a couple of weeks or a month isn’t enough time to make a huge difference in how prepared one is. It’s not as if you could acquire an entire new set of non-automatable skills in that time. Plus, only the ones who are really plugged into the AI scene (such as being subscribed to this channel) would really know what to make of that knowledge. And estimating very generously, that’s what, a few million people across the globe at most? Outside of that group, knowing that will be released on x or y day is entirely an abstract piece of information. And even for those paying close attention, there is simply not enough practical context without the actual model to make any kind of major decision on one’s job, regardless of what kind of on-paper info is released. I’m pretty intimately familiar with using GPT-4-class LLMs at this point, and I even have an internal (placeholder) date I’m using for the next model release—but I still have absolutely no idea what sort of impact even GPT-4.5 would have on my work, and I still wouldn’t even if I got a full model paper, benchmarks and an exact release date, because there simply is no substitute for having the actual model when it comes to figuring out what it can do.
solarapparition
2024-03-29 01:15:00 +0000 UTCIt is either (1) Not part of the training data, or (2) The LLMs cannot find it or (3) The LLMs are instructed not to use it. I have had many conversations with the LLMs where they don't indicate any awareness of this data and, if asked, will often essentially say, "You can Google that if you want to know." Many of these datasets are public, such as the US patent database, Open Access academic papers, product instruction manuals, technical whitepapers, tech talk transcripts, etc. ChatGPT-4 told me LLM developers typically focus on gathering a wide range of data sources with a goal of broad *general* knowledge -- not specialized knowledge such as legal, scientific or technical datasets. It said if expertise in specific fields is needed, then specialized LLMs can be trained on that. I don't know if that's correct, but it's roughly consistent with the current behavior. AI will never achieve the potential societal benefit if increasing amounts of valuable data are "fenced off" from LLMs. We will have increasingly powerful LLMs running on giant supercomputers, only to tell us, "You can Google that." I don't recollect any of the AI commentators discussing this.
Joe Marler
2024-03-26 14:06:38 +0000 UTCHi Joe. When you say that "patents, research papers, product instruction manuals, knowledgebase articles, tech talk transcripts" aren't part of training data, is that something you know to be true or something you suspect to be true? I agree these all seem like rich sources of information which (not that I'd thought about it) I kinda assumed would be in training data. If you're correct that they aren't, seems like a huge opportunity.
Mike Pemberton
2024-03-26 13:40:22 +0000 UTCAs someone who is aligning a company with thousands of employees with our vision for the future, it would be very very nice to have a better idea of what will be available when! I’m focusing on preparing the group for the next generation of models, but the erratic nature of their releases makes explaining timelines to senior leadership very difficult!
Trenton Dambrowitz
2024-03-25 21:43:47 +0000 UTCCalling out a mistake and leaving it in the edit to show you're not an AI sounds suspiciously like the sort of tactic a clever AI would employ to convince someone they were not an AI. Just sayin' ... :D
Kemi
2024-03-25 16:48:17 +0000 UTCah! now i got it!
Christopher Pollin
2024-03-25 15:41:12 +0000 UTCHey Billy, I have an RSS feed! Should be in settings, or somewhere on here, people are using it in Spotify for example
Philip
2024-03-25 12:45:50 +0000 UTCAnyway to do these in Apple Podcasts? I know they support paywalling, i.e. only Patreon members can subscribe to podcasts.
Billy N
2024-03-25 12:44:50 +0000 UTCThanks Sean, and you're right - a large degree of uncertainty is simply inevitable. Hope all is well with your team BTW
Philip
2024-03-25 10:51:19 +0000 UTCGreat podcast - totally agree on your points re. unpredictability, but I think this is now the way of the world. Change is accelerating and a natural consequence of that is more uncertainty. Yes, a bit of notice and better comms from the large frontier model creators will help but I think we all have to get more comfortable with more uncertainty!
Sean Betts
2024-03-25 10:40:33 +0000 UTCFantastic point, thank you Mark. Wishing you the best in your job.
Philip
2024-03-25 10:23:03 +0000 UTCThank you Arvind, your support has meant a lot over the months
Philip
2024-03-25 10:22:22 +0000 UTCI don't know if you saw the Edelman Trust Index, but trust in AI companies is collapsing, particularly in America. Down from an index of 50 in 2019 to 35 last year. If public sentiment continues to sour on AI companies, they will absolutely titrate their messages and releases. It's all about the bottom line, and if sentiment turns against them, voters rebel, etc.
David Shapiro
2024-03-25 08:43:52 +0000 UTCI think people are getting incredible notice and it is happening…thanks to people like you. I think we have all been telling everyone we know yet people don’t really listen or take it seriously. If history is a guide, people don’t deserve anything. However the upside is we usually underestimate peoples adaptability and resilience
Jim Beaver
2024-03-25 02:08:39 +0000 UTCSince you asked for feedback, I will give you my two cents. I am a little older and I have lived through the PC revolution in the Internet revolution, the iPhone revolution, etc.
Jim Beaver
2024-03-25 02:05:48 +0000 UTCThere’s another group for which the uncertainty is a major challenge: policy makers. I am an elected official in NYC. It’s hard for gov’t to adapt to change in the best of circumstances. But AI is moving so fast, with so little certainty in the timeline, that I fear the result in the public sector will simply be inaction—on schools, job training, election integrity etc. Any way I loved this episode and very much like the podcast format.
Mark Levine
2024-03-24 23:57:54 +0000 UTCThanks for the interesting podcast. I am however not that impressed by the predictions of Vernor Vinge. He didn't say: the median of predicted time is 30 years but that it will be achieved within 30 years - quite a broad interval. And from what you described - he meant ASI, which GPT-4 definitely isn't. The broader range from 2005 and 2030 isn't very likely for ASI - even if average-human level AI is reached by that time (quite probable) - it doesn't mean it will be able to exponentially self-improve. After all - an average human cannot improve the architecture of cutting-edge AI models.
Jan Matusiewicz
2024-03-24 23:38:12 +0000 UTCThank you for yet another fascinating podcast, Philip. I was speaking to the head of an Indonesian educational foundation today and commenting on the unpredictability of how to prepare young minds for the future in a world with AI - so your commentary is quite timely for me. As contradictory as it may sound, I appreciate both your candour and discretion in discussing the challenges in securing that certain interview you mentioned (please don’t hold any negative feelings to similar namesakes associated with this experience :)). I continue to find significant value in your insights and eagerly await other relevant news you are able to share. Truly grateful for your continued important contribution to helping the broader community understand our rapidly evolving and unpredictable future.
Arvind Mani
2024-03-24 23:28:55 +0000 UTCI think you have put words to some of the thoughts I've been having regarding random releases of major models. After Claude 3 was released I was expecting GPT 4.5 to drop. I signed up to Twitter and have been monitoring AI accounts in a weird frenzy/hype. I think the anxiety comes down to the fact that a life changing update may or may not be dropped at any random time. I think one of the realistic things OpenAI could be more open with, is disclosure about future release dates and capabilities. Let people plan and know which day to be excited or panicked over.
Matt Grosse
2024-03-24 23:18:00 +0000 UTCTrue I guess, but do you think that would actually have any significant impact or would it cause more chaos?
Dan Brian
2024-03-24 21:27:46 +0000 UTCMergers and Acquisitions (often abbreviated M&A). Those are governed by US federal laws and regulations designed to preserve competition and protect consumers from anti-competitive practices. See the Sherman Antitrust Act, Clayton Act, and Hart-Scott-Rodino Antitrust Improvements Act.
Joe Marler
2024-03-24 21:27:27 +0000 UTCWhat are MNA rules?
Colton S.
2024-03-24 20:46:28 +0000 UTCChin up buddy, if you really want answers to those 20 questions then keep booking a chat with them. Or see if they'll send you written answers, and then use 11 labs to voice them. . You'll still not be AI, but it'll be team Philip
Lee FRASER
2024-03-24 20:38:29 +0000 UTCPhilip, excellent comments about the amplified impact of new models cascading in a seismic fashion throughout the ecosystem, and the importance of notification. Good points about the limitations on "reasoning" of current models. I have subscriptions to ChatGPT-4, Gemini Advanced, Claude 3.0 Opus and Perplexity Pro. I use them hours every day in my work. It's not like they can answer everything except for the Riemann hypothesis. They are often unable to reliably provide answers to questions that experienced human professionals in a given field can normally address. It would help if the training corpus contained patents, research papers, product instruction manuals, knowledgebase articles, tech talk transcripts, etc. If some of this key information is kept out of the reach of AIs, it may as well be stored on parchment scrolls.
Joe Marler
2024-03-24 20:32:36 +0000 UTCI meant more tactically, like 2 weeks before a release
Philip
2024-03-24 20:26:08 +0000 UTC:))
Philip
2024-03-24 20:24:50 +0000 UTCI will be speaking on the record soon! But I gave the only quote I could 'businesses should think 6 months, not 1 year'
Philip
2024-03-24 20:24:43 +0000 UTCYou didn't tell us about the OpenAI chat :(
Colton S.
2024-03-24 20:22:18 +0000 UTCI feel like, openAI has been giving notice for over a year. Was Sam Altman’s world tour not a huge signal?
Dan Brian
2024-03-24 20:15:17 +0000 UTCThank you, Philip. Nice podcast…. And maybe you’ll get that interview yet. I’m perplexed as to why they would keep putting it off.
Tony Coffman
2024-03-24 19:53:56 +0000 UTC
