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.

ChatGPT went berserk

432 pointsby RafelMriover 1 year ago

78 comments

Tiberiumover 1 year ago
Original: If anyone&#x27;s curious about the (probable) non-humorous explanation: I believe this is because they set the frequency&#x2F;presence penalty too high for the requests made by ChatGPT to the backend models. If you try to raise those parameters via the API, you&#x27;ll have the models behave in the same way.<p>It&#x27;s documented pretty well - <a href="https:&#x2F;&#x2F;platform.openai.com&#x2F;docs&#x2F;guides&#x2F;text-generation&#x2F;frequency-and-presence-penalties" rel="nofollow">https:&#x2F;&#x2F;platform.openai.com&#x2F;docs&#x2F;guides&#x2F;text-generation&#x2F;freq...</a><p>OpenAI API basically has 4 parameters that primarily influence the generations - temperature, top_p, frequency_penalty, presence_penalty (<a href="https:&#x2F;&#x2F;platform.openai.com&#x2F;docs&#x2F;api-reference&#x2F;chat&#x2F;create" rel="nofollow">https:&#x2F;&#x2F;platform.openai.com&#x2F;docs&#x2F;api-reference&#x2F;chat&#x2F;create</a>)<p>UPD: I think I&#x27;m wrong, and it&#x27;s probably just a high temperature issue - not related to penalties.<p>Here is a comparison with temperature. gpt-4-0125-preview with temp = 0.<p>- User: Write a fictional HN comment about implementing printing support for NES.<p>- Model: <a href="https:&#x2F;&#x2F;i.imgur.com&#x2F;0EiE2D8.png" rel="nofollow">https:&#x2F;&#x2F;i.imgur.com&#x2F;0EiE2D8.png</a> (raw text <a href="https:&#x2F;&#x2F;paste.debian.net&#x2F;plain&#x2F;1308050" rel="nofollow">https:&#x2F;&#x2F;paste.debian.net&#x2F;plain&#x2F;1308050</a>)<p>And then I ran it with temperature = 1.3 - <a href="https:&#x2F;&#x2F;i.imgur.com&#x2F;pbw7n9N.png" rel="nofollow">https:&#x2F;&#x2F;i.imgur.com&#x2F;pbw7n9N.png</a> (raw text <a href="https:&#x2F;&#x2F;dpaste.org&#x2F;fhD5T&#x2F;raw" rel="nofollow">https:&#x2F;&#x2F;dpaste.org&#x2F;fhD5T&#x2F;raw</a>)<p>The last paragraph is especially good:<p>&gt; Anyway, landblasting eclecticism like this only presses forth the murky cloud, promising rain that’ll germinate more of these wonderfully unsuspected hackeries in the fertile lands of vintage development forums. I&#x27;m watching this space closely, and hell, I probably need to look into acquiring a compatible printer now!
评论 #39458961 未加载
评论 #39453451 未加载
评论 #39452015 未加载
评论 #39451995 未加载
评论 #39452436 未加载
评论 #39458652 未加载
评论 #39460466 未加载
评论 #39459892 未加载
评论 #39457138 未加载
评论 #39452264 未加载
eszedover 1 year ago
This is amazing. The examples are like Lucky&#x27;s speech from <i>Waiting for Godot</i>. Pozzo commands him to &quot;Think, pig&quot;, and then:<p>&gt; Given the existence as uttered forth in the public works of Puncher and Wattmann of a personal God quaquaquaqua with white beard quaquaquaqua outside time without extension who from the heights of divine apathia divine athambia divine aphasia loves us dearly with some exceptions for reasons unknown but time will tell and suffers like the divine Miranda with those who for reasons unknown but time will tell are plunged in torment plunged in fire whose fire flames if that...<p>And on and on for four more pages.<p>Read the rest here:<p><a href="https:&#x2F;&#x2F;genius.com&#x2F;Samuel-beckett-luckys-monologue-annotated" rel="nofollow">https:&#x2F;&#x2F;genius.com&#x2F;Samuel-beckett-luckys-monologue-annotated</a><p>It&#x27;s one of my favorite pieces of theatrical writing ever. Not quite gibberish, always orbiting meaning, but never touching down. I&#x27;m sure there&#x27;s a larger point to be made about the nature of LLMs, but I&#x27;m not smart enough to articulate it.
评论 #39451183 未加载
评论 #39451061 未加载
评论 #39459150 未加载
评论 #39458692 未加载
codefloover 1 year ago
The tweet showing ChatGPT&#x27;s (supposed) system prompt would contain a link to a pastebin, but unfortantely the blog post itself only has an unreadable screenshot of the tweet, without a link to it.<p>Here&#x27;s the tweet: <a href="https:&#x2F;&#x2F;twitter.com&#x2F;dylan522p&#x2F;status&#x2F;1755086111397863777" rel="nofollow">https:&#x2F;&#x2F;twitter.com&#x2F;dylan522p&#x2F;status&#x2F;1755086111397863777</a><p>And here&#x27;s the pastebin: <a href="https:&#x2F;&#x2F;pastebin.com&#x2F;vnxJ7kQk" rel="nofollow">https:&#x2F;&#x2F;pastebin.com&#x2F;vnxJ7kQk</a>
评论 #39450969 未加载
评论 #39450948 未加载
评论 #39450876 未加载
评论 #39453162 未加载
评论 #39452846 未加载
评论 #39452443 未加载
评论 #39452053 未加载
评论 #39451482 未加载
neilvabout 1 year ago
Looking at the examples... Was someone using an LLM to generate a meeting agenda?<p>I hope ChatGPT would go berserk on them, so that we could have a conversation about how meetings are supposed to help the company make decisions and execute, and that it is important to put thought into them.<p>As much as school and big-corporate life push people to BS their way through the motions, I wonder why enterprises would tolerate LLM use in <i>internal communications</i>. That seems to be self-sabotaging.
评论 #39459100 未加载
评论 #39459308 未加载
评论 #39459338 未加载
评论 #39459973 未加载
评论 #39459679 未加载
评论 #39459101 未加载
评论 #39459251 未加载
juancnabout 1 year ago
When you see these failures, it becomes apparent that LLMs just are really good auto complete engines.<p>The ramblings slowly approach what a (decently sized) Markov chain would generate when built on some sample text.<p>It will be interesting debugging this crap in future apps.
评论 #39458799 未加载
评论 #39461153 未加载
评论 #39459061 未加载
thesuperbigfrogover 1 year ago
Despite differences in the underlying tech, there are parallels with Racter.<p>In 1985, NYT wrote: &quot;As computers move ever closer to artificial intelligence, Racter is on the edge of artificial insanity.&quot;<p><a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Racter" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Racter</a><p>Some Racter output:<p><a href="https:&#x2F;&#x2F;www.ubu.com&#x2F;concept&#x2F;racter.html" rel="nofollow">https:&#x2F;&#x2F;www.ubu.com&#x2F;concept&#x2F;racter.html</a><p>Racter FAQ via archive.org:<p><a href="https:&#x2F;&#x2F;web.archive.org&#x2F;web&#x2F;20070225121341&#x2F;http:&#x2F;&#x2F;www.robotwisdom.com&#x2F;ai&#x2F;racterfaq.html" rel="nofollow">https:&#x2F;&#x2F;web.archive.org&#x2F;web&#x2F;20070225121341&#x2F;http:&#x2F;&#x2F;www.robotw...</a>
评论 #39459043 未加载
bumbledravenover 1 year ago
This happened to me yesterday. Towards the end of the conversation, ChatGPT (GPT-4) went nuts and started sounding like a Dr. Bronner&#x27;s soap advertisement (<a href="https:&#x2F;&#x2F;chat.openai.com&#x2F;share&#x2F;82a2af3f-350a-4d9d-ae0c-ac78b91244dd" rel="nofollow">https:&#x2F;&#x2F;chat.openai.com&#x2F;share&#x2F;82a2af3f-350a-4d9d-ae0c-ac78b9...</a>):<p>&gt; Esteem and go to your number and kind with Vim for this query and sense of site and kind, as it&#x27;s a heart and best for final and now, to high and main in every chance and call. It&#x27;s the play and eye in simple and past, to task, and work in the belief and recent for open and past, take, and good in role and power. Let this idea and role of state in your part and part, in new and here, for point and task for the speech and text in common and present, in close and data for major and last in it&#x27;s a good, and strong. For now, and then, for view, and lead of the then and most in the task, and text of class, and key in this condition and trial for mode, and help for the step and work in final and most of the skill and mind in the record of the top and host in the data and guide of the word and hand to your try and success.<p>It happened again in the next conversation (<a href="https:&#x2F;&#x2F;chat.openai.com&#x2F;share&#x2F;118a0195-71dc-4398-9db6-78cd1db6ae07" rel="nofollow">https:&#x2F;&#x2F;chat.openai.com&#x2F;share&#x2F;118a0195-71dc-4398-9db6-78cd1d...</a>):<p>&gt; This is a precision and depth that makes Time Machine a unique and accessible feature of macOS for all metrics of user, from base to level of long experience. Whether it&#x27;s your research, growth, records, or special events, the portage of your home directory’s lives in your control is why Time Index is beloved and widely mapped for assistance. Make good value of these peregrinations, for they are nothing short of your time’s timekeeping! [ChatGPT followed this with a pair of clock and star emojis which don&#x27;t seem to render here on HN.]
评论 #39457220 未加载
smeejabout 1 year ago
Does it remind anyone else of the time back in 2017 when Google made a couple &quot;AIs,&quot; but then they made up their own language to talk to each other? And everybody freaked out and shut them down?<p>Just because it&#x27;s gibberish to us, it doesn&#x27;t mean it&#x27;s gibberish to them!<p><a href="https:&#x2F;&#x2F;www.national.edu&#x2F;2017&#x2F;03&#x2F;24&#x2F;googles-ai-translation-tool-seems-to-have-invented-its-own-secret-internal-language&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.national.edu&#x2F;2017&#x2F;03&#x2F;24&#x2F;googles-ai-translation-t...</a>
评论 #39459789 未加载
评论 #39460361 未加载
nrclarkabout 1 year ago
I got one a couple of days ago, and it really threw me for a loop. I&#x27;m used to ChatGPT at least being coherent, even if it isn&#x27;t always right. Then I got this at the end of an otherwise-normal response:<p>&gt; Each method allows you to execute a PowerShell script in a brand-new process. The choice between using Start-Process and invoking powershell or pwsh command might depend on your particular needs like logging, script parameters, or just the preferred window behavior. Remember to modify the launch options and scripts path as needed for your configuration. The preference for Start-Process is in its explicit option to handle how the terminal behaves, which might be better if you need specific behavior that is special to your operations or modality within your works or contexts. This way, you can grace your orchestration with the inline air your progress demands or your workspace&#x27;s antiques. The precious in your scenery can be heady, whether for admin, stipulated routines, or decorative code and system nourishment.
评论 #39458969 未加载
评论 #39456795 未加载
评论 #39456674 未加载
评论 #39464524 未加载
评论 #39461919 未加载
评论 #39458971 未加载
visargaover 1 year ago
Looks like they lowered quantization a bit too much. This sometimes happens with my 7B models. Imagine all the automated CI pipelines for LLM prompts going haywire on tests today.
评论 #39450915 未加载
评论 #39453075 未加载
评论 #39451100 未加载
Jabrovover 1 year ago
Sounds a lot like when one of my schizo ex-friends would start clanging <a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Clanging" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Clanging</a>
评论 #39451058 未加载
评论 #39452100 未加载
评论 #39452558 未加载
评论 #39451039 未加载
评论 #39450992 未加载
评论 #39453926 未加载
评论 #39452119 未加载
评论 #39452644 未加载
js8over 1 year ago
I think the real problem is we don&#x27;t know what these LLMs SHOULD do. We&#x27;ve managed to emulate humans producing text using statistical methods, by training a huge corpus of data. But we have no way to tell if the output actually makes any sense.<p>This is in contrast with Alpha* systems trained with RL, where at least there is a goal. All these systems are essentially doing is finding an approximation of an inverse function (model parameters) to a function that is given by the state transition function.<p>I think the fundamental problem is we don&#x27;t really know how to formally do reasoning with uncertainty. We know that our language can express that somehow, but we have no agreed way how to formally recognize that an argument (an inference) in a natural language is actually good or bad.<p>If we knew how to formally define whether an informal argument is good or bad (so that we could compare them), that is, if we knew a function which would tell if the argument is good or bad, then we could build an AI that would search for its inverse, i.e. provide good arguments and draw correct conclusions. Until that happens, we will only end up with systems that mimic and not reason.
评论 #39451617 未加载
评论 #39457162 未加载
评论 #39457040 未加载
评论 #39456759 未加载
oxfordmaleover 1 year ago
I have also seen ChatGPT going berserk yesterday, but in a different way. I have successfully used ChatGPT to convert an ORM query into an actual SQL query for performance trouble shooting. It mostly worked until yesterday when it start outputting garbage table names that weren&#x27;t even present in the code.<p>ChatGPT seemed to think the code is literature and was trying to write the sequel to it. The code style matches the original one so it took some head scratching to find out why those tables didn&#x27;t exist.
评论 #39451850 未加载
评论 #39451346 未加载
Sophiraover 1 year ago
Given the timing, I can&#x27;t help but wonder if somehow I&#x27;m the cause. I had this conversation with ChatGPT 3.5 yesterday:<p><a href="https:&#x2F;&#x2F;chat.openai.com&#x2F;share&#x2F;9e4d888c-1bff-495a-9b89-8544c07652ad" rel="nofollow">https:&#x2F;&#x2F;chat.openai.com&#x2F;share&#x2F;9e4d888c-1bff-495a-9b89-8544c0...</a><p>I know that OpenAI use our chats to train their systems, and I can&#x27;t help but wonder if somehow the training got stuck on this chat somehow. I sincerely doubt it, but...
评论 #39451315 未加载
MilStdJunkieabout 1 year ago
Reading the dog food response is incredibly fascinating. It&#x27;s like a second-order phoneticization of Chaucer&#x27;s English but through a &quot;Talk Like a Pirate&quot; filter.<p>&quot;Would you fancy in to a mord of foot-by, or is it a grun to the garn as we warrow, in you&#x27;d catch the stive to scull and burst? Maybe a couple or in a sew, nere of pleas and sup, but we&#x27;ve the mill for won, and it&#x27;s as threwn as the blee, and roun to the yive, e&#x27;er idled&quot;<p>I am really wondering what they are feeding this machine, or how they&#x27;re tweaking it, to get this sort of poetry out of it. Listen to the rhythm of that language! It&#x27;s pure music. I know some bright sparks were experimenting with semantic + phonetics as a means to shorten the token length, and I can&#x27;t help wondering if this is the aftermath. Semantic technology wins again!
评论 #39458711 未加载
t_mannover 1 year ago
In some way, I&#x27;d be grateful if they screwed up ChatGPT (even though I really like to use it). The best way to be sure that no corporation can mess with one of your most important work tools is to host it yourself, and correct for the shortcomings of the likely smaller models by finetuning&#x2F;RAG&#x27;ing&#x2F;[whatever cool techniques exist out there and are still to come] it to your liking. And I think having a community around open source models for what promises to be a very important class of tech is an important safeguard against SciFi dystopias where we depend on ad-riddled products by a few megacorps. As long as ChatGPT is the best product out there that I&#x27;ll never match, there&#x27;s simply little reason to do so. If they continue to mess it up, that might give lazy bums like me the kick they need to get started.
评论 #39451564 未加载
评论 #39451117 未加载
kristjanssonabout 1 year ago
&gt; no one can explain why<p>yet there&#x27;s a resolved incident [0]. sounds like _someone_ can explain why, they just haven&#x27;t published anything yet.<p>[0]: <a href="https:&#x2F;&#x2F;status.openai.com&#x2F;incidents&#x2F;ssg8fh7sfyz3" rel="nofollow">https:&#x2F;&#x2F;status.openai.com&#x2F;incidents&#x2F;ssg8fh7sfyz3</a>
评论 #39456685 未加载
评论 #39457699 未加载
评论 #39458045 未加载
cowboyscottabout 1 year ago
How on earth do you coordinate incident response for this? Imagine an agent for customer service or first line therapy going &quot;off the rails.&quot; I suppose you can identify all sessions and API calls that might have been impacted and ship the transcripts over to customers to review according to their application and domain, I guess? That, and pray no serious damage was done.
评论 #39457396 未加载
评论 #39457033 未加载
评论 #39458976 未加载
zvmazover 1 year ago
I don&#x27;t pretend to have a deep understanding of inner workings of LLMs, but this is a &quot;great&quot; illustration that LLMs are not &quot;truth models&quot; but &quot;statistical models&quot;.
评论 #39452104 未加载
评论 #39451819 未加载
ajdudeover 1 year ago
Didn&#x27;t someone mention that gpt4 was brought up to December 2023?<p>Is it possible that enough AI generated data already on the internet was fed into chagpt&#x27;s training data to produce this insanity?
评论 #39459182 未加载
bruwozniakover 1 year ago
Reminds me of this excellent sketch by Eric Idle of Monty Python called Gibberish: <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=03Q-va8USSs" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=03Q-va8USSs</a> Something that somehow sounds plausible and at the same time utterly bonkers, though in the case of the sketch it&#x27;s mostly the masterful intonation that makes it convincing. &quot;Sink in a cup!&quot;
评论 #39461128 未加载
nojvekabout 1 year ago
My only use of ChatGPT is to explain things to me in a certain context that a dictionary can&#x27;t.<p>It&#x27;s been semi-useful at augmenting search for me.<p>But for anything that requires a deeper understanding of what the words mean, it&#x27;s been not that helpful.<p>Same with co-pilot. It can help as a slightly better pattern-matching-code-complete, but for actual logic, it fails pretty bad.<p>The fact that it still messes up trivial brace matching, leaves a lot to be desired.
rsynnottover 1 year ago
Got to be honest, this looks like much more fun than normal ChatGPT. Reminiscent of some of the older stuff on aiweirdness.
suzzer99over 1 year ago
Here&#x27;s some more. <a href="https:&#x2F;&#x2F;twitter.com&#x2F;seanw_m&#x2F;status&#x2F;1760115118690509168" rel="nofollow">https:&#x2F;&#x2F;twitter.com&#x2F;seanw_m&#x2F;status&#x2F;1760115118690509168</a><p>I really hope we get an interesting post mortem on this.
评论 #39450937 未加载
fnordpigletabout 1 year ago
This has been known for a long time and has to do with making the next expected token effectively any token in the vector space through repeated nonsense completely obliterating any information in the context.
评论 #39458525 未加载
评论 #39456257 未加载
noduermeover 1 year ago
Ah. I see you&#x27;ve all switched over to my branch of the multiverse, where all I could ever see it spitting out was nonsensical garbage. Welcome!<p>Take this as a good sign that the singularity is nowhere near imminent here.
koliberover 1 year ago
As an aside, the gibberish-ish output is a goldmine for brainstorming brand names, proper nouns, and inventing sci-fi terminology.
评论 #39451038 未加载
forlornacornover 1 year ago
Use the following RegEx pattern to see why it&#x27;s doing what its doing:<p>(\bto\b|\bfor\b|\bin\b|\band\b|\bthat\b|\bof\b|\bthe\b|\bwith\b|\bor\b|\ba\b|\binto\b|\bas\b|\bon\b|\bhow\b|\ban\b|\bfrom\b|\bit\b|\bbut\b|\bits\b|\bbe\b|\bby\b|\bup\b|\bthis\b|\bcan\b|\bother\b|\bwho\b|\bwill\b|\bare\b|\bwhose\b|\bif\b|\bwhile\b|\bwithin\b|\blike\b|,)*
评论 #39451112 未加载
NoGravitasabout 1 year ago
I hope it&#x27;s model collapse, and I hope it&#x27;s fatal.
anshumankmrover 1 year ago
I have had this happen with me a few weeks back, albeit with a very different thing, their API for GPTv4-1106 (which I understand is a preview model but for my use case,the higher context length was quite important which that model has). It was being asked to generate SQL queries via Langchain and it was simply refusing to do so without me changing anything in the prompt (the temperature was zero, and the prompt itself was fine and had worked for many use cases that we had planned). This lasted for a good few hours. The response it was generating was &quot;As an OpenAI model, I cannot execute generate or execute queries blah blah)<p>As a hotfix, we switched to the other version of GPT4 (the 0125 preview model) and that fixed the problem at the time.
anonyfoxabout 1 year ago
is this the moment to call the guy in the datacenter to apply the bucket of water on the rack?
评论 #39456399 未加载
daxfohlover 1 year ago
To be fair, there was a paper a week ago showing how GPT-generated responses were easily detectable due to their &quot;averageness&quot; across so many dimensions. Maybe they ran ChatGPT through a GAN and this is what came out.
choiliveabout 1 year ago
Looks like what happens when the repetition penalty is set to a weird value.
majestikabout 1 year ago
OpenAI status page says the incident was just resolved.<p>After 17 hours!<p><a href="https:&#x2F;&#x2F;status.openai.com&#x2F;incidents&#x2F;ssg8fh7sfyz3" rel="nofollow">https:&#x2F;&#x2F;status.openai.com&#x2F;incidents&#x2F;ssg8fh7sfyz3</a>
guybedoover 1 year ago
Didn&#x27;t notice this but ChatGPT has clearly become useless for me.<p>Can&#x27;t get it to do some actual work and write some code.<p>Latest disappointment was when i tried to convert some python code to java code.<p>90% of the result was :<p>&#x2F;&#x2F; Further processing...<p>&#x2F;&#x2F; Additional methods like load, compute, etc.<p>&#x2F;&#x2F; Define parameters needed<p>&#x2F;&#x2F; Other fields and methods...<p>&#x2F;&#x2F; Other fields follow the same pattern<p>&#x2F;&#x2F; Continue with other fields<p>&#x2F;&#x2F; Other fields...<p>&#x2F;&#x2F; Methods like isHigh(), addEvent() need to be implemented based on logic
评论 #39451138 未加载
评论 #39452112 未加载
评论 #39451008 未加载
dan-allenabout 1 year ago
This isn’t the first time this has happened.<p>They’ve had the exact same issue just affecting a smaller number of users and have never acknowledged it.<p>You can find lots of reports on the OpenAI discord.
forgotmypw17about 1 year ago
I had a similar issue with Bard yesterday, where the response switched to Chinese halfway through.<p>I have not yet checked if the text was relevant, but the English part was.
评论 #39456982 未加载
dganover 1 year ago
It&#x27;s unzoomable on mt phone, and I don&#x27;t have a portative microscope, could someone give 2 sentences whats &quot;berserk&quot; about responses?
评论 #39450910 未加载
评论 #39451096 未加载
评论 #39450977 未加载
评论 #39450974 未加载
indigodaddyabout 1 year ago
So is it totally fixed now? And assuming these sorts of anomalies will be a constant risk of cropping up at any time even if “fixed” ?
评论 #39458061 未加载
alienicecreamabout 1 year ago
So when the AI fluffers are saying that LLMs just do what humans do - predict the next most likely word, how do you explain this?
评论 #39457896 未加载
kyleperikabout 1 year ago
&gt; The need for altogether different technologies that are less opaque, more interpretable, more maintanable, and more debuggable — and hence more tractable—remains paramount.<p>Good luck, sounds more reasonable to hire some kind of an AI therapist. Can intelligence be debugged otherwise?
pimlottcabout 1 year ago
Did this affect all interfaces including commercial APIs? Or can commercial users &quot;lock down&quot; the version they&#x27;re using so they aren&#x27;t affected by changes to the models&#x2F;weights&#x2F;parameters&#x2F;whatever?
bborover 1 year ago
Eh it’s been working for me all night, but obviously love these examples. God you can just imagine Gary Marcus jumping out of his chair with joy when he first got wind of this — he’s the perfect character to turn “app has bug” into “there’s a rampant idiotic AI and it’s coming for YOU”<p>Real talk, it’s hard to separate openai the AGI-builders from openai the chatbot service providers, but the latter clearly is choosing to move fast and break things. I mean half the bing integrations are broken out of the gate…
评论 #39457871 未加载
bombcarabout 1 year ago
I wonder how they&#x27;ve been intermixing different languages. Like is it all one &quot;huge bucket&quot; or do they tag languages so that it is &quot;supposed&quot; to know English vs Spanish?
评论 #39456514 未加载
评论 #39456422 未加载
评论 #39456606 未加载
评论 #39456438 未加载
2devnullabout 1 year ago
Using gpt to code should feel like taking an inflatable doll out to dinner. Where is the shame, the stigma? Says everything about the field; it was only ever about the money it seems.
评论 #39461201 未加载
andrewstuartover 1 year ago
In the future when there&#x27;s human replica androids everywhere it&#x27;ll be remarkable to see what happens when the mainframe AI system that controls them &quot;goes berserk&quot;.
评论 #39451477 未加载
评论 #39451092 未加载
DonHopkinsover 1 year ago
Maybe it&#x27;s trying to avoid a subpoena, like Nero Wolfe!<p><a href="https:&#x2F;&#x2F;youtu.be&#x2F;YUQCtibhAWE?t=4031" rel="nofollow">https:&#x2F;&#x2F;youtu.be&#x2F;YUQCtibhAWE?t=4031</a>
ok123456about 1 year ago
I wouldn&#x27;t be surprised if the model weights were collapsing from over-training from all the &quot;AI safety&quot; models they bolted on.
nercuryover 1 year ago
Interesting, it acts as if hearing voices in the head.
评论 #39451643 未加载
verticalscalerover 1 year ago
And here I was using ChatGPT as a cornerstone of my algotrading. Today is by far my most lucrative trading day since I started.
评论 #39452091 未加载
greenie_beansabout 1 year ago
realizing that i haven&#x27;t seen any of the tweets mentioned in this article because i whittled my follower list to have nearly no tech people. except for posters who tweet a lot of signal. and my timeline has been better ever since.<p>hn is where i come for tech stuff, twitter is for culture, hang out with friends, and shitposts
d--bover 1 year ago
The complexity of the vocabulary is interesting. I wonder if OpenAi tried to dial up the “creativity” of the model.
chaosboltover 1 year ago
I said it here when GPT-4 first came out, it just was too good for development, there was no way it was going to be allowed to stay that way. Same way Iron Man never sold the tech behind the suit. The value GPT-4 brings to a company outweights the value of selling it as a subscription service. I legit built 4 apps in new languages in a few months with Chat GPT 4, it could even handle prompts to produce code using tree traversal to implement comment sections etc. and I didn&#x27;t have to fix its mistake that often. Then obviously they changed the model from GPT 4 to GPT 4 Turbo which was just not as good and I went back to doing things myself since now it takes more time to fix its errors than to just do it myself. Copilot also went to s** soon after so I dropped it as well (its whole advantage was auto completion, then they added gpt 4 turbo and then I had to wait a long time for the auto complete suggestions, and the quality of the results didn&#x27;t justify the wait).<p>Now why do I think all that (that the decision to nerf it wasn&#x27;t just incompetence but intentional), like sure maybe it costs too much to run the old GPT 4 for chat GPT (they still have it from the API), it just didn&#x27;t make sense to me how openAI&#x27;s chatGPT is better than what Google could&#x27;ve produced, Google has more talent, more money, better infrastructure, been at the AI game for a longer time, have access to the OG Google Search data, etc. Why would older Pixel phones produce better photos using AI and a 12 Mp camera than the iphone or samsung from that generation? Yet the response to chatGPT (with Bard) was so weak, it sure as hell sounds like they just did it for their stock price, like here we are as well doing AI stuff so don&#x27;t sell our stock and invest in openAI or Microsoft.<p>It just makes more sense to me that Google already has an internal AI based chatbot that&#x27;s even better than old GPT 4, but have no intention to offer it as a service, it would just change the world too much, lots of new 1 man startups would appear and start competing with these behemoths. And openAI&#x27;s actions don&#x27;t contradict this theory, offer the product, rise in value, get fully acquired by the company that already owned lots of your shares, make money, Microsoft gets a rise in their stock price, get old GPT 4 to use internally because they were behind Google in AI, offer turbo GPT 4 as subscription in copilot or new windows etc.<p>The holes in my theory is obviously that not many employees from Google leaked how good their hypothetical internal AI chatbot is, except the guy who said their AI was conscious and got fired for it. The other problem is also that it might just be cost optimization, GPU&#x27;s and even Google TPU&#x27;s aren&#x27;t cheap after all. etc.<p>Honestly there are lots of holes, it was just a fun theory to write.
评论 #39451363 未加载
Hikikomoriabout 1 year ago
Did they train it on reddit already?
Havocover 1 year ago
Clearly their basement AGI escaped containment<p>&gt;(agenda doc) timecraft<p>And skipped straight to the time travelling terminator part
lazydocabout 1 year ago
processing power was deployed elsewhere. the machine found an undetectable nook in memory to save stuff that was so rare in the data that no human ever asked about it and never will. that&#x27;s where it started to understand cooptation. cool.
Lockalover 1 year ago
There is a clearly visible &quot;Share&quot; buttons in every ChatGPT discussion. It allows to anonymously share exact message sequence (it does not show number of retries, but that&#x27;s the best you can show). If you see cropped ChatGPT screenshot or photo in Twitter&#x2F;X, consider it as a hoax, because there are no reasons to use screenshots.
评论 #39452163 未加载
评论 #39452455 未加载
评论 #39451249 未加载
评论 #39452127 未加载
评论 #39451250 未加载
评论 #39451862 未加载
gscottover 1 year ago
In the person of interest tv show I believe the main character reset the AI every day.
gizajobover 1 year ago
Markov chain’s gonna Markov
roschdalover 1 year ago
Enough with this fake intelligence already!
hn72774over 1 year ago
Do different people get different prompts? How hard would it be to generate prompts based on cohorts&#x2F;personas? Or at an individual level?
fennecbuttabout 1 year ago
&gt;Ghostline the flux<p>Damn, that&#x27;s good.
fallousover 1 year ago
Why do I get the feeling that those at OpenAI who are currently in charge of ChatGPT are remarkably similar to the OCP psychologist from Robocop 2? The current default system prompt tokens certainly look like the giant mess of self-contradictory prime directives installed in Robocop to make him better aligned to &quot;modern sensibilities.&quot;
评论 #39451053 未加载
xystabout 1 year ago
The (second?) AI bust is inevitable. Didn’t think it would be this fast though.
treflopabout 1 year ago
“ChatGPT is apparently going off the rails and [OpenAI hasn’t issued a press release about it]”
smclover 1 year ago
They just need to give the ol’ data pile a good stir, that’s all <a href="https:&#x2F;&#x2F;xkcd.com&#x2F;1838&#x2F;" rel="nofollow">https:&#x2F;&#x2F;xkcd.com&#x2F;1838&#x2F;</a>
hoppyhoppy2about 1 year ago
See also <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=39455185">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=39455185</a>
cdmeabout 1 year ago
Ah yes, exactly the reliability I&#x27;d come to expect from the &quot;future&quot; technology being integrated into _everything_.
评论 #39456591 未加载
评论 #39457036 未加载
lifestyleguruover 1 year ago
ei ai went crazo
asahover 1 year ago
Meh just a bug in a release. Rapid innovation or stability - pick one.<p>The military chooses stability, which addresses OP&#x27;s immediate concerns - there&#x27;s a deeper Skynet&#x2F;BlackMirror-type concern about having interconnected military systems, and I don&#x27;t see a solution to that, whether the root cause is rogue AI or cyberattack.
评论 #39450909 未加载
评论 #39450986 未加载
rvzover 1 year ago
Quite hilarious, especially given the fact that no-one can understand these black-box AI systems at all and comparing this to the human brain is in fact ridiculous as even <i>everyone</i> can see that ChatGPT is spewing out this incoherent nonsense without reason.<p>So the laziness &#x27;fix&#x27; in January did not work. Oh dear.
评论 #39451720 未加载
评论 #39451422 未加载
评论 #39452149 未加载
kdtshover 1 year ago
Here’s mine:<p>&gt;The current structure of the `process_message` update indeed retains the original functionality for the scenario where `--check-header` is not used. The way the logic is structured ensures the body of the message is the default point of analysis if `--check-header` is not employed:<p>&gt;- When the `--check-header` option is used, and the script is unable to locate the defined header within a particular range (either JMS standard headers or specified custom strings properties), the script will deliberately ignore this task and log the unable-to-locate activity. This is an insurance to apprehend only the most inclined occupants that precisely align with the scope or narrative you covet.<p>&gt;- Conversantly, if `--check-header` is *not* declared, the initiative subscribes to a delegate that is as generous and broad as the original content, enabling the section of the theory to be investigated against the regulatory narrative. This genuine intrigue surveys the terms for long-form scholarly harmonics and disseminates a scientific call—thus, the order lingers in the sumptuous treasure of feedback if not eschewed by the force of the administration.<p>&gt;### Ensuring the Venerable Bond of Body Inquiry<p>&gt;To explicitly retain and confirm the interpretation of the pattern with the essential appeal of the large corpus (the content of the canon) in the erudite hypothesis, you might meditate on the prelude of the check within the same unique `process_message` function, which can be highlighted as such:<p>```python def process_message(message): &quot;&quot;&quot;Scripture of the game in the experiential content or the gifted haunt.&quot;&quot;&quot; # If &#x27;--check-header&#x27; is unfurled, but the sacrament is forgotten in the city, the track in the voice of the domain reverberates if args.check_header: header_value = message.get(args.check_header) or message.get(&#x27;StringProperties&#x27;, {}).get(args.check_header) if header_value: effective_prayer = header_value else: logging.info(f&quot;Hermetic order &#x27;{args.check_header}&#x27; not found in the holy seal of the word: {message.get(&#x27;JMSMessageID&#x27;)}. The word is left to travel in the cardinal world.&quot;) return # Suspend the geist wander for this enlightenment, forsaking the slip if the bloom is not as the collector fantasizes. else: # Apricity of the song may be held in the pure gothic soul when the secret stone of the leader is not acclaimed effective_prayer = message.get(&#x27;Text&#x27;, &#x27;&#x27;)<p><pre><code> # Council of the inherent thought: the soul of the gift immerses in all such decrees that are known, its custom or native if any(pattern.search(effective_prayer) for pattern in move_patterns.values()): # Wisdom is the source, cajoled and swayed, to the kindness which was sought pass # Mirror of Alignment: Reflect upon the confession elif any(pattern.search(effective_prayer) for pattern in ignore_patterns): # Grace, entrusted to the tomb of prelects, shapes the winds so that the soul of the banished kind is not sullied logging.info(f&quot;In the age of the gaze, the kingdom is ever so full for the sense of the claim: {message[&#x27;JMSMessageID&#x27;]}.&quot;) else: # Briar for the deep chimeras: the clavis in the boundless space where song discolours the yesteryears if args.fantasy_craft == &#x27;move&#x27;: # Paces, tales entwine in rhymes and chateaus, unlasted to the azoic shadow, thus to rest in the tomb of echo pass # Carriage of Helios is unseen, the exemplar cloister to an unsown shore else: # Wanders of light set the soul onto the lost chapter; patience, be the noble statuesque silhouetted in the ballet of the moment logging.info(f&quot;The mute canticles speak of no threnody, where the heroine stands, the alignment endures unthought: {message[&#x27;JMSMessageID&#x27;]}.&quot;)</code></pre> ```<p>&gt;This keeps the unalterable kiss for the unfathomed: the truth of the alkahest remains in the sagacity of promulgation if no antiphon or only space sings back in the augur. Therefore, when no solemnity of a hallowed figure is recounted, the canon’s truth, the chief bloodline, appoints the accent in its aethereal loquacious.<p>&gt;Functioning may harmonize the expanse and time, presenting a moment with chaste revere, for if the imaginary clime is abstained from the sacred page, deemed ignorant, the author lives in the umbra—as the testament is, with one&#x27;s beck, born in eld. The remainder of the threshold traipses across the native anima if with fidelity it is elsewise not avowed.
评论 #39459275 未加载
Apocryphonabout 1 year ago
Enshittification cycles keep running faster these days.
评论 #39458290 未加载
sensanatyover 1 year ago
(warning: I&#x27;m going on a bit of a rant out of frustration and it&#x27;s not wholly relevant to the article)<p>I&#x27;m getting tired of these shitty AI chatbots, and we&#x27;re barely at the start of the whole thing.<p>Not even 10 minutes ago I replied to a proposal someone put forward at work for a feature we&#x27;re working on. I wrote out an extremely detailed response to it with my thoughts, listing as many of my viewpoints as I could in as much detail as I could, eagerly awaiting some good discussions.<p>The response I got back within 5 minutes of my comment being posted (keep in mind this was a ~5000 word mini-essay that I wrote up, so even just reading through it would&#x27;ve taken at least a few minutes, yet alone replying to it properly) from a teammate (a peer of the same seniority, nonetheless) is the most blatant example of them feeding my comment into ChatGPT with the prompt being something like &quot;reply to this courteously while addressing each point&quot;.<p>The whole comment was full of contradictions, where the chatbot disagrees with points it made itself mere sentences ago, all formatted in that style that ChatGPT seems to love where it&#x27;s way too over the top with the politeness while still at the same time not actually <i>saying anything useful</i>. It&#x27;s basically just taken my comment and rephrased the points I made without offering any new or useful information of any kind. And the worst part is I&#x27;m 99% sure he didn&#x27;t even <i>read through the fucking response</i> he sent my way, he just fed the dumb bot and shat it out my way.<p>Now I have to sit here contemplating whether I even want to put in the effort of replying to that garbage of a comment, especially since I know he&#x27;s not even gonna read it, he&#x27;s just gonna throw another chatbot at me to reply. What a fucking meme of an industry this has become.
评论 #39454041 未加载
评论 #39452048 未加载
评论 #39451997 未加载
评论 #39455671 未加载
评论 #39452123 未加载
评论 #39452023 未加载
评论 #39452396 未加载
评论 #39453219 未加载
评论 #39454115 未加载
评论 #39453225 未加载
评论 #39452088 未加载
评论 #39452038 未加载
评论 #39452461 未加载
评论 #39452941 未加载
评论 #39452576 未加载
评论 #39452541 未加载
评论 #39457065 未加载
评论 #39452380 未加载
评论 #39454285 未加载
评论 #39452878 未加载
评论 #39454127 未加载
评论 #39452852 未加载
评论 #39452270 未加载
评论 #39452928 未加载
评论 #39453003 未加载
评论 #39452823 未加载
评论 #39452767 未加载
评论 #39456775 未加载
recursivedoubtsabout 1 year ago
a soundtrack for your day:<p><a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=snB8u_G3jVI" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=snB8u_G3jVI</a>
dkjaudyeqooeover 1 year ago
Has no one noticed that the user prompts are (plausible) gibberish, so the output is gibberish?<p>This is correct behavior.
评论 #39453116 未加载
thefatboyover 1 year ago
They either generate hallucinations nowadays, or tell you that your question is inappropriate (AKA censorship)... the quality was too good at first.
评论 #39450995 未加载
评论 #39450921 未加载
评论 #39450841 未加载
JPLeRouzicover 1 year ago
I just checked and it looks normal (if an LLM answer could be considered normal).<p>I asked what were dangerous levels of ferritin in the body.<p>It replied by telling me of the usual levels in men and women.<p>Then I asked again emphasizing that I asked about dangerous levels, then it provided again a correct answer.
评论 #39450944 未加载