1033
Google has sent internet into ‘spiral of decline’, claims DeepMind co-founder
(www.telegraph.co.uk)
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 15 Oct 2023
1033 points (97.2% liked)
Technology
58123 readers
4322 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
I already go to ChatGPT more than Google. If you pay for it then the latest version can access the internet and if it doesn’t know the answer to something it’ll search the internet for you. Sometimes I come across a large clickbait page and I just give ChatGPT the link and tell it to get the information from it for me.
Do you fact-check the answers?
It depends what you’re using it for as to whether you need to fact check stuff.
I’m a software developer and if I can’t remember how to do an inner join in SQL then I can easier ask ChatGPT to do it for me and I will know if it is right or not as this is my field of expertise.
If I’m asking it how to perform open heart surgery on my cat, then sure I’m probably going to want several second opinions as that is not my area of expertise.
When using a calculator do you use two different calculators to check that the first one isn’t lying?
Also, you made a massive assumption that the stuff OP was using it for was something that warranted fact checking.
I can see why you would use it. Why would I want to search Google for inner joins sql when it is going to give me so many false links that don’t give me the info in need in a concise manner.
Even time wasting searches have just been ruined. Example: Top Minecraft Java seeds 1.20. Will give me pages littered with ads or the awful page 1-10 that you must click through.
Many websites are literally unusable at this point and I use ad blockers and things like consent-o-matic. But there are still pop up ads, sub to our newsletter, scam ads etc. so much so that I’ll just leave the site and forego learning the new thing I wanted to learn.
The new release of GPT-4 searches Bing, reads the results, summarizes, and provides sources, so it's easier to fact check than ever if you need to.
It's pretty trivial to fact check an answer... You should start using this kind of bots more. Check perplexity.ai for a free version.
Sources are referenced and linked.
Don't judge on chatgpt free version
People don't do it though and often parrot bullshit.
People who do so aren't smart enough to use internet anyway. With or without AI it wouldn't change anything for them, they stay stupid and will continue acting stupidly
Perplexity.ai has been my go to for this reason.
It often brings up bad solutions to a problem and checking the sources it references shows it regulary misses the gist of these sources.
There sources it selects are often not the ones I end up using. They are starting point, but not the best starting point.
What it is good for is for finding content when I don't know the terminology of the domain. It is a starting point ready to lead me astray with exquisitely written content.
Find trustworthy sources and use them.
It is more of a proof of concept at the moment, but it shows the potential
That's what's usually gets said about lots of alternative fusion energy generation methods that later turn out to be impossible to have net-positive energy generation.
And this is just one example. Another example: tons of "neat concept that shows potential" medical compounds end up dropped at the medical testing stage because of their nasty side effects or it turns out their "positive" effects are indistinguisheable from the placebo effect.
The point being that you can't actually extrapolative from "neat concept that shows potential" even to merelly "will work", much less to "will be a great success".
PS: Equally, one can't just say it's not going to be a great success - being a "neat concept that shows potential" has a pretty low informational content when it comes to predicting the future, worse so when there are people monetarilly heavilly invested into it who have a strong interest in making it look like a "neat concept that shows potential" whilst hiding any early stage problem as they're activelly poluting the information space around it.
You are mixing sci-fi level of cutting edge basic research (fusion), with commercial products (chatgpt). They are 2 very different type of proof of concepts.
And both will likely revolutionize human society. Fusion will simply commercially become a thing in 30/50 years. AI has been on the market for years now. Generative models are also few years old. They are simply becoming better and now new products can be built on top of them
(btw I already use chatgpt 4 productively every day for my daily work and it helps me a lot)
I seem to not have explained myself correctly.
This specific tech you seem to be emotionally invested in is no different from the rest in this sense because it still faces in the real world the very same kind of risks and pitfalls as the rest - there are possible internal pitfalls inherent to every new technology (i.e. a problem we never knew about because we never used it with so many people in the real world before, becomes visible with widespread use) and there are possible external pitfalls inherent to how it fits in the complex world we live in (i.e. it turns out the use cases don't make quite as much economic sense as was first tought or it indirectly generates more problems than it solves).
Such Process and Fit risks are true for every early stage "revolutionary" tech (i.e. we never did it before, now that we do it, we discover problems we were not at all aware of before) - business guys might say that revolutionary tech starts with a lot more "unknown unknowns" than incremental improvements on existing tech - and is why the bean counters rarelly put money in revolutionary and instead go mainly for incremental improvements on proven tech. At times one or more of such "we had no idea this could happen problems" turn out to be insurmountable, sometimes they can be overcomed but the result is not especially great, sometimes they're all overcomed without any nasty side-effects and the thing ends up being a world-changing tech: you can't really tell upfront.
In the case of LLMs, the two risky problems from what I've heard which might stop it from being "world changing" are in how LLMs being trained in material which includes LLM-generated material actually get worse (so as the Internet gets flooded with LLM-generated material passing for human-generated one, LLMs would get worse and worse) and the other is the so-called Hallucinations, which are really just the natural side effect of them being Language Models hence all that they do is generate compositions of language tokens that pass for human generated language, with no reasoning involved hence cannot validate through inductive or deductive reasoning said "compositions of language tokens", so LLMs wouldonly usefull for altering format without touching the information (for example, turn lists of cold hard facts into fluffy longwinded text or do the opposite and summarize lots of fluffy text into just the facts) which would still have a big impact in certain professions but not necessarilly be "world changing" (or, even more interestingly, make over time people value "fluffy text" less and less, which would be "world changing" but not in a way that benefits the makers of LLM).
Unless you want to deny the last 4 decades of History in Tech, you can't logically extrapolate from an early "looks like it might be a success" to "it will be a success", especially in the era of money-driven overhype we live in.
give it time, algos will fuck those results as well
They'll need to make money with a cheap cost-per-sale, so they'll put ads on the site. Then they'll put promoted content in the AI chat, but it's okay because they'll say it's promoted. Eventually it won't even say it's promoted and it will just be all ads, just like every other tech company.
Why? Because monetization leads directly to enshittification, because the users stop being the customers.
When I tried it it was never able to give me the sources of what it said. And it has given me way too many made up answers to just trust it without reasons. Having to search for sources after it said something has made me skip the middle man(machine).
You probably tried the free version. Check perplexity.ai to see how the paid version of chatgpt works. Every source is referenced and linked.
This guy is not talking about the current version of free chatgpt. He's talking of the much better tools that will be available in the next few years
Yeah, because people selling AI products have a great track record on predicting how their products will develop in the future. Because of that, Teslas don't have steering wheels any more, because Full Self Driving drives people incident-free from New York to California since 2017.
The thing with AI development is, that it rapidly gets to 50% of the desired solution, but then gets stuck there, not being able to get consistently good enough that you can actually rely on it.
I don't really understand what it means. If the product is unreliable people won't use it, and everything will stay as it is now. It's not a big issue. But It is already pretty reliable for many use cases.
Realistically the real future problem will be monetization (which is causing the issues of Google), not features
Well, here's the thing. How often are you willing to dismiss the misses because of the hits? Your measure of unreliability is now subject to bias because you're no longer assessing the bot's answers objectively.
I don't expect it to be 100% correct. I have realistic expectations built on experience. Any source isn't 100% reliable. A friend is 50% reliable, an expert maybe 95. A random web page probably 40... I don't know.
I built up my strategies to address uncertainty by applying critical thinking. It is not much different than in the past. By experience, chatgpt 4 is currently more reliable than a random web page that comes in the first page of a Google search. Unless I exactly search for a trustworthy source, such as nhs or guardian.
The main problem is the drop in quality of search engines. For instance, I often start with chatgpt 4 without plugins to focus my research. Once I understand what I should look for, I use search engines for focused searches on official websites or documentation pages.
The issue with reliability is a completely different one between web search and AI.
If you search something on Google, there are quite a few ways you can judge the quality of the answer with "metadata" around it. If you find a scientific paper, it's probably more reliable than a post on a parents forum. If the source is a quality newspaper or Wikipedia, that's also more on the reliable side, but some conspiracy theorist website is not. And if the source is some kind of forum or Q&A site, wrong answers often have comments under them that correct the error.
Also, you can follow multiple links and take a wider sample on the topic that way.
With AI that's not possible. Whether it is wrong or correct, the AI will give you an answer in the exact same format, with the same self-confident tone. You basically need to know the correct answer to know whether the answer is correct.
Sure, you can re-roll and ask it again, but that doesn't make the result more likely to be correct.
For example, I asked ChatGPT which Harry Potter chapter is the longest. It happily gave me a chapter, but it wasn't the longest. So I asked again and again and again, and each time it gave me a new wrong answer, every time with made-up word counts.
This is the reason I am suggesting people to give a try to perplexity.ai to understand how these tools will work in the near future. And why I don't understand the reason I am downvoted for that.
Current "free" chatgpt was created as a proof of concept, not as a finished, complete solution for humanity issues. What we have now is a showcase of llm, for openai to improve the product via human feedback, for everyone else to enjoy what is it already now, with all its limitations, an extremely useful tool.
But this kind of LLM is intended to be a building block of the future solutions. To enable interactivity, summarization, analysis features within larger products with larger and more refined set of features.
If you don't have paid version of chatgpt, again, try perplexity.ai with the copilot feature, to see a (still imperfect, under development) proof of concept of the near future of AI assisted research.
And more tools will come, that will make easier to navigate the huge amount of information that is the main issue of modern internet.
For your specific case, gpt 3.5 has poor logical and mathematical capabilities. Gpt-4 is much better with that. But still, using a language model for math is almost never a good choice. What you'd need, is an llm able to access information from the internet and to have access to some math tool, such as python or Matlab. These options currently are available on chatgpt with plugin, but they are suboptimal. In the future you'll have better product able to combine llm, focused internet search and math.
We should focus on the future, not on the present when discussing AI. LLMs based products are in their infancy
ChatGPT powers Bing Chat, which can access the internet and find answers for you, no purchase necessary (if you're not on edge, you might need to install a browser extension to access it as they are trying to push edge still).
Do you fact-check the answers?
That's such a strange question. It's almost like you imply that Google results do not need fact checking.
They do. Everything found online does.
With google, it depends on what webpage you end up on. Some require more checking than others, which are more trustworthy
Generative AI can hallucinate about anything
There are no countries in Africa starting with K.
LLMs aren’t trained to give correct answers, they’re trained to generate human-like text. That’s a significant difference.
They also aren't valuable for asking direct questions like this.
There value comes in with call and response discussions. Being able to pair program and work through a problem for example. It isn't about it spitting out a working problem, but about it being able to assess a piece of information in a different way than you can, which creates a new analysis of the information.
It's extraordinarily good at finding things you miss in text.
Yeah. There's definitely tasks suited to LLMs. I've used it to condense text, write emails, and even project planning because they do give decently good ideas if you prompt them right.
Not sure I'd use them for finding information though, even with the ability to search for it. I'd much rather just search for it myself so I can select the sources, then have the LLM process it.
Agree.
I found it more tempting to accept the initial answers I got from GPT4 (and derivatives) because they are so well written. I know there are more like me.
With the advent of working LLMs, reference manuals should gain importance too. I check them more often than before because LLMs have forced me to. Could be very positive.