Taking Gebru's word seriously here is reason to discount the entire argument.<p>Gebru is a part of a clique that is devoted to an entirely different set of ideas around AI Safety.<p>The two basic movements here are:
1. Be very scared of AI development, then do capabilities research but feel bad about it.
2. Demand AI research comes with locks in them that make sure the AI can't write a tweet that would make a San Franciscan Activist uncomfortable. Achieve very little in terms of manipulating AI, but get lots of book deals.<p>Neither of these groups should really be trusted for hard data about the other, and more importantly - missing that piece indicates that the author is not even attempting to convince the unconvinced.
Excellent point. I think it’s critical to recognize that as AI tech is getting more advanced, we need to improve the mechanisms that enforce transparency and trustability. It starts with training data. Just because something is technically reachable online, doesn’t make it OK for it to be used for AI training. A person or organization should be able to opt into having their data used, of course, but it’s not a “use first, ask to remove later”.<p>Then who is watching the watchers? Models being used to generate content or make decisions that have a real-world impact on actual humans need to be open to audit and validation that they don’t repeat the bias and social limitations of the people who create those models. This is not a new concept (think algorithms for insurance coverage calculations etc).<p>So yes, a toy chatbot to play around with - sure. A fancy AI model, sucking personal data without consent and being deployed all around us by closed organizations with dubious funding - big no. Is AI ready to offer scientific or academic value - no, and even if it was a yes, it will need to stand to human scrutiny for a long time before being acceptable “as is”. Kids in schools need to be thought how to recognise and work with AI generated content. Just like dealing with fake news, it requires critical thinking and the ability to analyse information.
I'm not sure who is more naive: The people who believe that this technology is the solution to every problem, or the people who believe that it is somehow possible to put this genie back into the box.<p>AI will change <i>everything</i>. Not in the distant future when it becomes sentient, but in the next 5-10 years.
Ah yes, academics, the most unbiased angels of us all…<p>There are good arguments in this article, buried in a mire of self righteous indignation.<p>Better education on viewing any output by a model with more critical eyes - which really is today a basic requirement for parsing any online information - is a fine bill to die on.<p>That then goes in hand with viewing with suspicious eyes the motivations of companies that are commercialising this technology, which is nowhere near a new sentiment.<p>However better recognition needs to be given to what obviously fuels the hype being railed against: the utility of this tool, even if it fails to work adequately most of the time. YMMV depending on your personal experience on the queried topics.<p>The article does indeed sound like it’s just preaching to the choir.
Congratulations !<p>A documented, opiniated, non-hype biased post about AIs. I've lived the 80's AI hype wave and the fall that followed. AFAICT we're on the same train here.<p>"Any sufficiently advanced technology is indistinguishable from magic" (Arthur C Clarke) and we're waiting for the masses to recognize what magic sparkles are : dust (ok, weighted barycenters of a huge database)<p>Side note : I often wonder how *GPT can deal with self-reference, like "What would you think of you if you were me ?" or "tell me something you were not programmed to think is plausible"
My favorite part of OpenAI chat is that there are no videos, no popups asking me to sign up for a newsletter, no warnings to enable cookies, no hiding content behind interstitial ads on scroll.<p>I will gladly pay good money for a web experience and API that just answers my questions.
I'm not sure the paper's authors are very diverse as they claim, they seem cut from the same cloth to me.<p>ChatGPT may spit out word salad, and be trained on inappropriate input, but there is no doubting the success of some of these models, eg at playing the game of Go, predicting protein folding, etc, etc.<p>There's obviously something still missing, but the hype is warranted IMHO, these model represent a major advance compared to even just a few years ago.
I'm unconvinced (but I'm also not the target audience as I'm not an academic).<p>The hype is, in large part, real to me. GPTs, but also Whisper or Stable Diffusion/DALLE solve real problems, way better than previous tools could.
When you agreed to the terms of the various sites you use every day, you allowed the companies access to your output which they will use to train their models. It’s too late.