TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Sparks of Artificial General Intelligence: Early Experiments with GPT-4

180 pointsby thinxerabout 2 years ago

33 comments

outlaceabout 2 years ago
ChatGPT and its relatives are very very impressive on first impressions, but I've been using ChatGPT-3 and now 4 heavily every day since they became available to individuals and once you start using them this much it becomes very clear how NOT intelligent they are. It really just seems like extremely impressive statistical inference after this much use and finding so many failure modes. But it is still impressive how much of human intellectual endeavors can be captured by sophisticated statistical inference. Very useful technology nonetheless.
评论 #35270146 未加载
评论 #35269996 未加载
评论 #35270258 未加载
评论 #35280731 未加载
评论 #35269981 未加载
评论 #35270337 未加载
评论 #35270385 未加载
评论 #35270149 未加载
评论 #35287646 未加载
neilellisabout 2 years ago
[This is in reply to the comments not the article!]<p>It&#x27;s just a statistical model is the logical equivalent of human beings are just a bunch of atoms.<p>The amount of reductionist thinking that goes on in tech is hilarious. First define AGI then challenge an AI to meet those requirements. If it meets them it is AGI. Put aside your preconceptions of what technology you think is required to achieve the goals and stay empirical.<p>Note previous definitions of AI have been thrown away as AI passes through them one by one :-)<p>What goes on inside its &#x27;head&#x27; is irrelevant. We still don&#x27;t know what actually goes on inside our heads and we were damn sure we were <i>intelligent</i> long before we had a clue how our heads worked at all.<p>Also sentience != AGI. We can&#x27;t even agree what sentience is in humans and other living beings so I&#x27;d stay clear of that one for now :-)
评论 #35285010 未加载
评论 #35288064 未加载
_gabe_about 2 years ago
&gt; Given the breadth and depth of GPT-4&#x27;s capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system.<p>But it&#x27;s just statistics, a fancy text predictor, a Markov-chain. Surely these scientists that work in the field of AI and are intimately familiar with how this stuff works aren&#x27;t so stupid as to think emergent behavior potentially resembling intelligence could result from such simple systems? It&#x27;s just statistics after all. Given enough training, any neural net could <i>guess</i> the next best token. It trained off all of Google after all. It&#x27;s just looking up the answers. No hint of intelligence. Just a mindless machine. After all, the saying goes, &quot;If it walks like a duck and quacks like a duck, it must be a mindless machine that has no bearing on a duck whatsoever&quot;. &#x2F;s
评论 #35270288 未加载
评论 #35285042 未加载
评论 #35290205 未加载
评论 #35270289 未加载
评论 #35279100 未加载
raincoleabout 2 years ago
My prediction for the top comments of this thread (paraphrased)<p>1. It&#x27;s just Microsoft&#x27;s advertisement<p>2. No it&#x27;s just a very effective pattern matching algorithm<p>3. Please define intelligence first otherwise it&#x27;s nonsense<p>4. I welcome our machine overlord<p>5. Lmao I asked it to do $thing and it failed<p>I&#x27;d like to know if GPT-4 can predict the top comments of this thread?
评论 #35271203 未加载
评论 #35271091 未加载
评论 #35271162 未加载
评论 #35271677 未加载
评论 #35271416 未加载
评论 #35271130 未加载
hislazinessabout 2 years ago
For whatever reason we seem to have set a very high expectation from AI as compared to NI (Natural Intelligence). I remember reading &quot;There is a considerable overlap between the intelligence of the smartest bears and the dumbest tourists.&quot;
评论 #35271397 未加载
评论 #35271237 未加载
评论 #35271353 未加载
TMWNNabout 2 years ago
As a non-expert in the field I was hesitant at the time to disagree with the legions of experts who last year denounced Blake Lemoine and his claims. I know enough to know, though, of the AI effect &lt;<a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;AI_effect" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;AI_effect</a>&gt;, a longstanding tradition&#x2F;bad habit of advances being dismissed by those in the field itself as &quot;not real AI&quot;. Anyone, expert or not, in 1950, 1960, or even 1970 who was told that before the turn of the century a computer would defeat the world chess champion would conclude that said feat must have come as part of a breakthrough in AGI. Same if told that by 2015 many people would have in their homes, and carry around in their pockets, devices that can respond to spoken queries on a variety of topics.<p>To put another way, I was hesitant to be as self-assuredly certain about how to define consciousness, intelligence, and sentience—and what it takes for them to emerge—as the experts who denounced Lemoine. The recent GPT breakthroughs have made me more so.<p>I found this recent Sabine Hossenfelder video interesting. &lt;<a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=cP5zGh2fui0">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=cP5zGh2fui0</a>&gt;
hislazinessabout 2 years ago
I remember reading this somewhere - &quot;There is a considerable overlap between the intelligence of the smartest bears and the dumbest tourists.&quot;. Though I do not think GPT-4 is even close to AGI it can definitely claim to be better at faking it than many intelligent beings can.
评论 #35269899 未加载
评论 #35279208 未加载
评论 #35270569 未加载
IanCalabout 2 years ago
I&#x27;m increasingly convinced you can build an agi system with gpt4.<p>People are trying to get it to solve everything up front but I&#x27;ve had GPT3 do much better by taking it through a problem asking it questions. Then I realised it was good at asking those questions too so just hooked it up to talk to itself with different roles. Gpt4 seems much better overall and is very good at using tools if you just tell it how and what it has available.<p>With a better setup than reAct, better memory storage and recall, I think it&#x27;d be an agi. I&#x27;m not hugely convinced it isn&#x27;t anyway - it&#x27;s better than most people at most tasks I&#x27;ve thrown at it.<p>Oh, and gpt came up with better roles for the &quot;voices in the head&quot; than I did too.
评论 #35271583 未加载
评论 #35271359 未加载
reidjsabout 2 years ago
Well of course Microsoft is going to say something sensational about it, aren’t they in charge of the project somewhat? This is just an advertisement for them, by them.
评论 #35271197 未加载
atleastoptimalabout 2 years ago
To me it&#x27;s really crazy that there is a public UI (ChatGPT) that lets people use GPT-4. If OpenAI had the attitude of Google they would have just gone &quot;Yeah we created a language model that&#x27;s light years ahead of anything else, look how cool it is, but sorry due to public safety you will never get to use it. Bye now!&quot; I feel that the public accessibility of these large language models is a fluke. Being able to use it for almost free feels like cheating reality.
评论 #35281157 未加载
beoberhaabout 2 years ago
I know enough about how neural nets work to be absolutely blown away at how good the GPT are. I only skimmed the paper, but even chatGPT showed a lot of these “sparks”, IMO. We are certainly a long way off from any semblance of general intelligence, but for a model that just tries to predict the next word, I’m dumbfounded at how good it is.
评论 #35271300 未加载
评论 #35271314 未加载
评论 #35271205 未加载
gwoolhurmeabout 2 years ago
From the intro: &quot;we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system.&quot;<p>What does that mean? If we take it as fact, so if it is an early version of AGI, Microsoft is using this thing to push subscriptions to all their services? This thing that is potentially the greatest thing humanity has made, an artificial living thing, and it&#x27;s used to sell CoPilot and 365 subscriptions. Paint me as really sad then. Instead of sharing the research with other entities, or anything that could further help or push us... we get subscriptions? Fuck me, the future sucks.
评论 #35271334 未加载
评论 #35271378 未加载
js8about 2 years ago
I don&#x27;t accept that something is AGI unless it can solve general instances of SAT (satisfiability problem, not the school test). Also recognizing (formulating from the task) an instance in the first place would help too.<p>To me, these are hallmarks of reason, and not available in LLMs, in fact probably impossible just with pattern recognition.
评论 #35271231 未加载
评论 #35271201 未加载
评论 #35271242 未加载
notShabuabout 2 years ago
You know the &quot;you pass butter&quot; scene from Rick &amp; Morty?<p>I&#x27;m imagining humans being told &quot;you complete thought sentences&quot;
评论 #35271895 未加载
doctobogganabout 2 years ago
Does anyone have insight into the GPT-4 model itself? What is the parameter count? Training procedure? I know &quot;Open&quot;AI hasn&#x27;t released this data but I was hoping someone with inside knowledge would have leaked it by now.
PaulDavisThe1stabout 2 years ago
I can&#x27;t help but the hear the distant, very very quiet echo of Clever Hans.
评论 #35271112 未加载
评论 #35271338 未加载
lsyabout 2 years ago
This is a pretty fluffy paper, especially for an institution like Microsoft Research. It says it&#x27;s an &quot;early AGI&quot; in the abstract, but elsewhere says it&#x27;s merely a &quot;step towards AGI&quot;. The basis for this is asking ChatGPT a bunch of stuff, but they don&#x27;t really present an overarching framework for what questions to ask or why.<p>The paper makes outlandish claims like &quot;GPT-4 has common sense grounding&quot; on the basis of its answers to these questions, but the questions don&#x27;t show that the model has common sense or grounding. One of their constructed questions involves prompting the model with the equator&#x27;s exact length—&quot;precisely 24,901 miles&quot;—and then being astonished that the model predicts that you&#x27;re on the equator (&quot;Equator&quot; being the first result on Wikipedia for the search term &quot;24,901&quot;). It&#x27;s also the case that while GPT-4 can say a bear at the north pole is &quot;white&quot;, it has no way of knowing what &quot;white&quot;, or &quot;bear&quot;, or &quot;north&quot; actually represent.<p>Are there folks out there doing rigorous research on these topics, who have a framework for developing tests of actual understanding?
评论 #35273360 未加载
评论 #35270398 未加载
georgehillabout 2 years ago
&gt; Given the breadth and depth of GPT-4’s capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system.<p>I don&#x27;t know why, but my brain refuses to accept GPT-4 as something close to AGI. Maybe I am wrong. It is hard to believe that our brain is just a bunch of attention layers and neural nets.
评论 #35269898 未加载
评论 #35269844 未加载
评论 #35270352 未加载
评论 #35269858 未加载
评论 #35269835 未加载
评论 #35269830 未加载
评论 #35269969 未加载
评论 #35269960 未加载
nycdatasciabout 2 years ago
Can we agree that an &quot;early (yet still incomplete) version of AGI&quot; isn&#x27;t AGI?
评论 #35270046 未加载
评论 #35270657 未加载
templeosenjoyerabout 2 years ago
Unless they somehow cured GPT-3&#x27;s schizophrenia and this model is a significant upgrade I&#x27;m not buying it - no matter how good it is at proving trivial mathematics theorems in the style of Eliot or whoever. Too often I have dealt with &quot;The answer to your question is X. Oh, sorry, you are right, the answer is actually Y. Oh, it is good of you to ask for a proof, sure I can prove the answer is Y, I used this (hallucinated) method described in this (hallucinated) paper. Oh, sorry, you are right, I cannot find any evidence that the method and paper I mentioned earlier actually exist, oops!&quot;.
ilitiritabout 2 years ago
GPT AI systems remind me of Chinese Room thought experiment:<p><a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Chinese_room" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Chinese_room</a><p>This is also similar to the Duck Test:<p><a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Duck_test" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Duck_test</a><p>Depending on the context, there are <i>generally</i> two takes: &quot;It is (or is not) a duck&quot;, and &quot;It doesn&#x27;t (or does) matter whether or not it&#x27;s a duck&quot;. These aren&#x27;t mutually exclusive.
bumbledravenabout 2 years ago
<i>While [pre-GPT-4 ChatGPT] performs non-trivially on both tasks, there is no comparison with the outputs from GPT-4. These preliminary observations will repeat themselves throughout the paper, on a great variety of tasks. The combination of the generality of GPT-4’s capabilities, with numerous abilities spanning a broad swath of domains, and its performance on a wide spectrum of tasks at or beyond human-level, makes us comfortable with saying that GPT-4 is a significant step towards AGI.</i>
satoshiiiiabout 2 years ago
If they remove the guardrails, then we can truly assess its intelligence. Currently, humans are directly interfering with a certain aspect of it. If it can provide a response without Microsoft&#x27;s stock being affected by removing these human-imposed limitations, then I would be genuinely impressed.
blarg1about 2 years ago
All this GPT stuff feels reminiscent of Frank Herbert&#x27;s novel: Destination Void ...
onosabout 2 years ago
“ We note however that there is no single definition of AGI that is broadly accepted, and we discuss other definitions in the conclusion section.”<p>We know it can do a lot of cool stuff, but without a pinned down definition the headline here is useless.
评论 #35270173 未加载
wildermuthnabout 2 years ago
It’s already smarter than 50% of us, and more knowledgeable than 99% of us. It no longer matters what label we give it, and we’re only a few years away from it giving labels to us.
namariaabout 2 years ago
Well it took us just about 65 years and a couple of AI winters to get convincing NLP going. And it takes about 1 TB of RAM...<p>So either AGI is around the corner or a generation away. Same as positive yield fusion reactors?
hallqvabout 2 years ago
Been pair-coding with gpt4 for the last week, it&#x27;s definitely AGI..
评论 #35271281 未加载
评论 #35271295 未加载
评论 #35271342 未加载
mplewisabout 2 years ago
No it doesn’t.
undert0wnabout 2 years ago
Good find. I am reading through this now.
ftxbroabout 2 years ago
Some will say it&#x27;s as important as the internet or mobile, but they&#x27;re wrong. This is like the discovery of fire or the invention of language.
评论 #35270742 未加载
number6about 2 years ago
I always thought General Intelligence would be Achieved by IBM or at least Apple, not by Microsoft. Now it will be used to pressure us into Windows Upgrades...
评论 #35271265 未加载
评论 #35271128 未加载
评论 #35271103 未加载
bumbledravenabout 2 years ago
GPT-4&#x27;s proof of the infinitude of primes, as quoted in the paper, is wrong.<p><pre><code> Yes, I think I can, Though it might take a clever plan. I’ll start by noting Euclid’s proof, Which shows that primes aren’t just aloof. Assume we have a finite list, Of primes, and that none have been missed. Multiply them all together, And add one, just to be clever. The result will be a number, That has no prime factors, oh wonder! </code></pre> Not true. The result may itself be prime. For example, say your finite list of primes is {2,3}. Multiplying them together and adding one yields 2*3 + 1 = 6 + 1 = 7, which is itself prime. Hence it has a prime factor: 7.<p><pre><code> But every number has a prime, That divides it, it’s just a matter of time. So we’ve found a contradiction, And our finite list needs eviction. There must be infinitely many primes, And that’s the end of my rhyming lines. </code></pre> A correct verison of the proof might say that the result has a prime factor, however, that prime factor is not an element of the supposed finite list, hence the list is incomplete.
评论 #35274500 未加载
评论 #35272838 未加载
评论 #35287405 未加载
评论 #35282381 未加载