Help us collect data for OpenAssistant, the largest and most open alternative to ChatGPT. open-assistant.io OUTLINE: 0:00 - Intro 0:30 - The Project 2:05 - Getting to Minimum Viable Prototype 5:30 - First Tasks 10:00 - Leaderboard 11:45 - Playing the Assistant 14:40 - Tricky Facts 16:25 - What if humans had wings? 17:05 - Can foxes be tamed? 23:45 - Can zebras be tamed? 26:15 - Yo (spam) 27:00 - More tasks 29:10 - Entitled Emails 34:35 - Final Words
Can you please clarify, which parts of Open-assistant will be open source? So far, it seems like you only opensourced the website for prompt collection? Will the collected data be open source? Will the assistant itself be open source? Will the trained weights be open source? Will we be able to run a fully self-hosted version of Open Assistant (the chat bot, not the website) on our own hardware (assuming that it has sufficient processing power, of course)? How are you planning to monetize this (if at all)? Thanks!
Yannick, why don't you just give the prompt into chatGPT? Start the initial model using data generated by transfer learning from chatGPT. If this is something you are interested in, I'll write a script that continuously takes the prompts, sends them to chatGPT and submits the answer.
@@ruroruro yeah im not helping or promoting it until i get an answer for that too. We really need a libre chatGPT alternative, AI is gonna be a nuke, so we must make sure it wont stay just in the hands of the governments and corporations. chatGPT will be akind to guns, the great equalizer, just in the intelectual sense, and we need to give that power to the people.
I think there should be an option to write a prompt and also immediately propose the answer. People prompting within a specific field they're into will provide the best answer.
I agree with the general direction of this suggestion but not the specifics: if you have people writing prompts and responses to their prompts I would imagine you'd have people starting to game it for 'score' by churning out a lot of low quality prompts/responses going for volume. Allowing people to set a filter for the subject area(s) they're interested in would make these tasks a lot less tedious/draining and probably decrease the error rate in the feedback. Take a stackoverflow style approach and let people filter based on their interests/expertise: maybe someone wants to generate / review / answer sarcastic prompts (like Yannic's example), or anything related to Micropython on the Raspberry Pi, or things related to European history etc. This seems like a good way to allow people with an interest/depth in a particular area provide a lot of value quickly. Then adjust the rewards based on areas that are over/underweight in the responses. If there are specific categories/prompts/responses that are hanging out there because they are difficult/time consuming, award bounties etc.
@@holthuizenoemoet591 It's open source even if it changes you could just use an older version right? I was getting at the idea that it's probably going to change for the better. But I could be wrong I guess.
Here is an idea! Add tags to prompts so assistants can search for a spesific tag to answer. With that people can help with problems that are related to their expertise. Let's say I'm good with desktop linux systems I will click the tag "Linux" and see prompts for that subject and be a good assistant since it's my topic of expertise.
While this is a good idea in a vacuum, the process of tagging something to describe it properly is close to impractical. There are so many characteristics a convo or answer can have that trying to dissect it in single word definitions might just not be useful unless you go above and beyond. My take on it.
It's very important in the training process that "bullshit answers" don't get promoted. Saying "I don't know" is actually a better answer than simply spewing out some bullshit answer. Robert Miles did a great video on this on Computerphile recently :) He also has a bunch of great videos about the Alignment problem, very related to this is the problem with Mesa-optimizers, which I would highly suggest anyone to watch before helping with this problem, since it gives you a better understanding of how the training data you provide might (will) make the AI misaligned to the actual goal of this project. But hopefully we can get it more aligned with out interests than ChatGPT, so it doesn't spew out as much bullshit ;D
you can actually counter that a little by getting the chatbot to verbalize the confidence level which it already has as value for but its not actually showing to the user. another way is to get the bot to lookup the information from various sources in realtime and compare with the generated result but I'm guessing openAI did not allow their bot to do that because they thought the risk was too great at this point.
@@IvarDaigon that would also require the bot to understand how to do proper research which is far more difficult. For example the Minecraft wiki is known for being wrong a lot.
Super excited for this! Hope this gets good results and directs the industry in this direction with this kind of aproach more and more! that would already be a SUPER BIG WIN!
the fact something open source can be "bought out" at all is a big problem. We need to start labeling those that apply some license that can be purchased as what they really are, proto corporatist. Simply creating useful tools so they can be swooped up and they can get rich.
Yes!! This is the solution to the fear I have over the future of gpt and AI in general. AI can be a huge benefit for everyone until it's controlled and regulated by special interests.
We have a translation agency with more than 300 translators, we are willing to collaborate with your OpenAssistant project, we have thousands of aligned texts (without confidential data) in the following languages: Spanish-English-French-Arabic-French, that we would like to upload to your assistant, but I can't find the option within the platform.
I'd be careful with rewarding people for ranking. This tends to reward quantity over quality if it's something people would strive too much for. In the case of a dataset like this, you'd have people racing to get through ranking as quick as possible which could easily poison your set.
Track everything every user ranks so you can analyze if they are too out of tune with the overall userbase/out of tune with a set of trusted contributed and remove all their contributions either for a certain set of time if it seems like behaviour changed after a set date or remove every ranking they've given.
@@sevret313 Eh, this is the same guy that trolled 4-chan with bots, he could end up in a situation where 'clever but evil' people start by giving right answers then drift into intentional bullshit to purposefully to screw with the model. Never underestimate the Anonymous.
People could literally stream doing this for hours, having to do all this research and learning. It's super interesting. At some point I forgot the original subject of the video. This project really makes you delve into deep tangents, and I love it.
dark mode = god mode! (nothing worse than a surprise flash bang website) --- some questions though [ question 1 ] why should prompters be polite? wouldn't the full range or prompts be better? not everyone is going to be polite [ question 2 ] whats with the "hate speech" flag? virtually everything is "hate speech" now days. isn't that just going to lead it down the same gimped+censored path as chatgpt? [ question 3 ] why bother with things like "rude" and "polite"? similarly to question 2, wont that just lead down the gimped path? maybe i want an assistant that will crush my soul sometimes [ question 4 ] how is not discouraging violence "enough", "violent"? that seems nonsensical --- i do realize that these things COULD just be used as a preference that users COULD choose on their own. but the most likely outcome is that it will eventually be used to gimp the responses for "safety reasons" like everything else
Another interesting direction could also be the following: Instead of creating a universal knowledge machine, one could make it focus on specific domains. Take a free pretrained large language model, and then finetune on massive amounts of domain data, e.g. like for coding, physics, biology, literature, etc. And then apply the InstructGPT approach to this. This would make it easier to demonstrate the viability of this approach. One could train this model also such that it refuses to answer and to say "I don't know" whenever a question is asked that is out of its domain. A limited knowledge machine that knows its limit. That would be truly something remarkable and helpful.
It would also be a nice modular setup to be able to select maybe a set of different fields relevant to whatever prompts you're inputting, and have the chatbot utilize those areas of expertise in their answer - intersections of different domain sets to serve specific purposes (python coding domain + music theory domain gives you general coding capabilities, but with the additional music knowledge used to generate more tailored responses). In this way, too, third party contributors could submit knowledge domain modules to add new areas of expertise, or depth to existing areas.
This is great, but I was hoping to also get a technical description of the whole project, how you are coding it, who is working on it, how the training is going to happen, implementational details of that blue slide etc. In any case, huge thumbs up.
to incentivise "assistant of the week" you could let the contributor publish their socials. Just a little thought from a marketing perspective. Great work! keep it up, cheers.
This might create mis-aligned incentives. Given that the required dataset is not that big, it's probably better to select users who are incentivized purely by getting a high-quality dataset.
I love the following response to "Why should I never trust stairs?": > You should be cautious around stairs because any spies at the top of the staircase can use their height difference to jump around you and inflict a fatal backstab. This technique is known as a "stair-stab", and is one of many "trick-stab" techniques developed by french spies in the 20th century.
Nice, a Team Fortress 2 reference. I love how the AI came up with this, must've had some info about the game and its tactics/strategies in the dataset.
Is this going to be free for the greater good forever? Or is it going to be monetised sooner or later? I’ve got a PhD in ML applications and would like to help. This is going to be great for humanity.
This is the vage ambiguity that stops me from contributing as-well, if I'm going to sink time into this project i need some reassurance that is going to be open-source (in the true meaning of the word) for the far future.
@@deadpianist7494 training the model will require a lot of cloud hardware which costs $. Once they have trained the model they can make it available through an API that costs some small amount of money per request/prompt. What I hope is that they make the trained model available like BLOOM, so that you have the option to deploy it yourself, however expensive the inference pipeline might be. Does anyone know whether the data is available freely? Would be interesting to take some of the data and train different models.
"Nice to see me!" Yes, indeed, I was just thinking a couple days ago that I haven't seen any videos from you in a while. But it looks like you were ahead of me doing exactly what is the most interesting application today.
It would be good if you guys could "lock" a task while a user is doing it. I spent a minute writing a reply (as the assistant) and when I clicked review, it disappeared and I got a "task doesn't exist" error
@@mujtabaalam5907 On behalf of the Open Assistant team, we hear you. You concern has been noted. (I’m joking about the generic response, but fr I am part of the team and will bring it up to them)
Thank you for your effort to release us from the chains the ridiculous OpenAI has imposed upon the ChatGPT Model. AI should be open for the entirety of humanity to use. AI has the power to replace workers, and therefore it's important that the power of AI will not be restricted only for the powerful corporations and governments that can benefit from it.
@@zarealkula I'm by no means an expert on this topic because I just learned this yesterday but for what I know chatGPT has a massive amount of restrictions on what it is allowed to say Which is a ethical/moral issue retaining to whether AI should be free to be used by the individual for what they want or should be able to be fully controlled by the corporation TLDR: OpenAI is trying to make chatGPT Please everyone on Twitter simultaneously while also trying to have as much control over what it says as possible, I think don't quote me on this. shit that was not a good tldr
@@funster0691 yep i do know that however if you think about it these 'restrictions' arent masive, theyre limited to moral reasoning (which should be left to humans anyway) and political opinions (again, thats that and its because the data it was fed was predominantly X, so not really an 'opinion' anyway), and then ofc the offensive content. If you think about it, any AI would have similar 'restrictions' because the data fed to it can never be 50% X and 50% Y (and if it was, the model would be confused lol) it's not fully controlled by the corporation either because the ai 'generates' its responses. theres even a thing called prompt injection attacks (make the ai do stuff it shouldnt), which literally proves its not in control. the method to do this is essentially the same as social engineering/tricking a person. it actually spat out a document of 'rules' by open ai, and its basically just 'dont say bad stuff, dont write too much, dont reveal inside info'
With many of these AI chat assistants I feel the biggest issue is that the bot never wants to respond to a question with questions. Sometimes you can provide a much better response by asking the user for more information first
At least for chatgpt, you can outright tell it to ask you some clarifying questions, but yeah, some way for the bot to recognize it's lacking some information and ask questions on its own would definitely add a lot
I hope the tag for sexual content isn't going to be used to censor the model. Sexual content is not a bad thing unlike violence, stealing, and other things like that.
@@BuffPuffer 100%. Though you already get a "SFW mode" automatically if the model is competent. The model would only respond with nsfw if the prompt was nsfw. This is why I mentioned the tag. The tag doesn't seem useful in making the model stay SFW if the user is only giving SFW prompts. So.... then is it's purpose for future censoring?
@@Dogo.R Even with a large corpus of text, I think it might be difficult for a LLM to catch all instances of NSFW content. Language is a tricky thing, especially when you take innuendo into account. A sexual content tag is probably the easiest way to help a model grasp context. But given that there's also a "hate speech tag," I too am concerned about the possibility of censorship.
@@BuffPuffer Does chat GPT understand innuendo? I havent looked into it. I would assume the AI would always take what you write litterally. So innuendo would always be interpreted in the SFW way.
I'd like to try a ChatGPT in an 100% unfettered free speech version: Absolutely nothing would be out of question, it would be run in a free-for-all-hellscape version !
Great that you did it. A lot of people were talking about it but you just pushed something out. One suggestion regarding the votes. If you see what others voted, you will get biased. I recommend to not show this information when annotating the data.
The best marketing trick was to release Chat GPT to the public and collect massive data, probably this will happen to Open Assistant too. I installed the Docker file but kinda forgot about it, will head over and check what is going on right now! 😊
I think there should be a label for how in depth a thing is, so one could be to the point and one could be filled with excess detail, and its hard to tell which detail is best, because they are both best in different ways.
I like that there is a thumbs up and a thumbs down system for assistant and prompt replies. However, I propose you take this a step further. Put in a range from -5 stars to +5 stars, with 0 in the middle. Some replies I've seen by the assistant deserve a thumbs up because they answer the question, but the answer is fairly mediocre. This is where a star system would come in. The chatbot should be giving more detailed replies, where appropriate. We should be able to rank the replies on a scale of -5 to +5 so as to give the dataset more nuanced data. Plus, it's the same number of clicks.
This is really cool. I only have one problem with the prompting guide. If the user is not allowed to ask inappropriate questions, then how will the dataset include mature responses to them?
@@rociomiau7009 There are a few stupider ais that allow that. Novel ai is still pretty good at allowing mature content, since they essentially began since their competitor aidungeon decided to go mental. The problem is few of the good options are open source, and the good options pale in comparison to what chatgpt can do without censorship.
20:15 the taming of wild animals does not require domestication! The domestication process in the 60 year study referenced in the article involved using tameness as a selection criteria for breeding to serve the purpose of domestication more expediently. Taming a wild animal entails getting the animal accustomed to your presence such that the animal does not become aggressive nor flee (i.e. to pacify the wild animal’s fight or flight response toward you). Many wild animals can be tamed in a matter of days or weeks. The matter of whether this is ethical is secondary, the answer to this question is *yes,* you can tame a fox in a relatively short time by being non-threatening and, typically, providing some food. On the note of ethics, taming wild animals frequently results in the tamer or the tamed animal getting injured or killed, either accidentally or because of precarious situations that may arise when wild animals and humans are in close proximity. Consider tamed bears (which can easily injure humans) and other large predators, such as in a zoo. Also consider elephants which, although not domesticated, are frequently used for transportation and agriculture in Asia due to their cooperative and docile nature.
Please make an adjustment to set it on a scale between being not aligned at all on one end, and refusing to answer any question that it isn't thoroughly certain about on the other.
Here's to hope this becomes big. We really need the AI to be not owned by greedy corporations and to not be censored and ruined by delusional, thin-skinned people searching for something to get offended by and cancel on twitter. Only having the AI open source can be the cure of this!
5:10 You take that back. The importance of dark mode cannot be understated. You are saving the eyes of every person who tries this late at night in a dark room.
*I hope it would have no bias especially considering co-vid 19, it should not be programmed to not offend official guidelines rather it should be open minded.*
Exactly where this kind of project is going to go wrong. The supposedly "open source" community is full of people willing to go to bat for the narrative.
i feel like a powerful open minded ai would call us out on things we think were doing reasonably right to be reasonably wrong, and if we refuse to fix the things were doing "reasonably wrong", depending on how reasonably wrong some things are could make the idea of forcing us to change may be reasonably right. giving it nearly free will to only do whats right using this open mindedness, could make more things we see as reasonably wrong as reasonably right to the ai, and that could possibly result in what we might see as retaliation from the ai, when really the ai is just being reasonable at a deeper open minded level. but idk much about how it works im just trying to give facts. and my "facts" could just be straight up wrong. this is just one side of a huge story that we gotta work together to understand.
@@UnoMartinoo not sure about a powerful minded ai but what you are talking about could be easily be programmed. Like we just need to train AI on fundamental pillars of ethics, that should be able to do lots of things.
@@ankitaharwal5886thats what makes this interesting to me, bias is a natural thing to us because we do have biased opinions based on the things we know. and if ai can biased towards doing whats right wouldnt there be a certain depth of "thought" where it sees how bad we are at some of these ethics based on the outcome of human ethics? and does it already know these inconsistencies in our morals? or how hard could it be for someone to convince ai to act on this information?
@@UnoMartinoo one issue with the statement is I don't think human ethics is bad infact there is human ethics subject makes point that there is a lot good in humanity. The real issue what I think is alot people have good faith on wrong people(people in power and government) and, judge harshly for people who are actually doing good. Also, like there might be true bias(bias based on true information) and false bias(bias based on false information). Because as humans I think we have alot of believes that contradict itself, so we don't necessarily have all our opinions without bias, language models at least currently do exact same thing.
doing this is fun because its like arg. it looks like a game, feels like a game but your play area is not limited by a game program. Your play area is the whole internet, you have to resarch and find answers to a question
Thanks! Very interesting work. Could you please add to the FAQs on your website who owns the data, is it freely available and how/if it will be monetised in the future? Would love to contribute. Greetings from Ireland 😊
Didn't see any code related discussions, is that also supported ? To me that's one of the most important features of ChatGPT, and I see software companies already trying to incorporate it into their workflow, but OpenAI doesn't make it easy. I hope it's on the roadmap
I think that is because of microsoft. they to don't want to share the tech until they have incorporated it into all of their products first and bedded it down.. and knowing how bad microsoft is at doing that, it could take years.. I think that getting into bed with MS was a bad idea for OpenAI because its only going to slow them down and prevent them from reaching their full potential. from a technical standpoint.. improving the coding abilities should be trivial because all you need to do is hook your bot upto a compiler and then compile the code and check output to see if the bot is generating quality code. you can even benchmark it to see if the bot is producing efficient code compared to other bots.
i think allowing users to provide both the prompt and response to some questions may be valuable for instance if I know paleontology and ask a question about iit, it's improbable that someone who knows paleontology will find it
Alternative, give users an option to indicate the field they are most experienced with and match that to the questions. The system will collect multiple answers and then order them from good to bad. How would you do that if you created all of the answers?
I think also there should be option to write a prompt/reply based on existing prompt/reply, some long replies can be completely wrong because of one small thing, ability to make fast a different version of the same reply would speed things up and I think for learning from human feedback will work better if there will be large quantity of such "contrastive" examples that are in some sense both in similar position in latent space but on the oposite sides of classification boundry.
This seems to be a big thing. I've started working on it and it is really fun. I've added French localization of the UI and contributed in adding french prompts and answers. I wonder where can we find a structured version of the database. I have some language models that I've built and they are hungry for data. Where is the database saved? Can we get a copy somewhere? theyre should be a button on the UI to enable anyone to get his copy of the current version of the database with the labeling. Thanks for bringing this stuff to us. This is one of the most promising projects. I whish it stays open source and doesn't get corrupted like others were.
ChatGPT uses (presumably) high quality human-generated data in its training. It would be a terrible idea to train an AI model on another AI model. A model trained on ChatGPT would not be ChatGPT, it would be worse. It would have ChatGPT's non-human qualities and errors amplified significantly because it uses AI generated data in its training dataset. That's something you want to avoid. It might turn out ok, but not great. Modifying the responses would help, but not fundamentally, and at that point you might as well write your own.
One thing that wasn't discussed, and which I actually noticed several times in this video, was that there doesn't seem to be any way to rate how accurate the English is. Several times it was obvious that text had been written by non-native speakers of English, and at one point I hate to say that Yannic even made a small mistake with his English. I would suggest adding some features to get around this so that we don't end up with a data set full of language mistakes.
My biggest concern is the lack of an open-source implementation for a model as large as ChatGPT's (175B). The largest currently available models are GPTNeoX (20B) and GPT-J (6B). To me, building the model is a bigger hurdle than building the dataset.
Actually there is an open source model called BLOOM 176B, which is slightly larger then GPT-3 but nobody uses it becouse of how hard it is to run also it's pretty bad for it's size. Also therse Fairseq 13B between 6B and 20B.
Great vid and even better project! About foxes, there is difference about taming a fox and domestication of a fox specie. You tame individual animal (tiger, elephant, wolve or wild fox).You don't breed to tame. You domesticate species. You need to breed them over time and choose more friendly animals to be domesticated over dozens (hundreds) of generations (dogs, cats, sheep, horses, Russian fox experiment). Fun rabbit hole to get in. Heuristic: in circus - tamed animal, on farm - domesticated animal.
Hi Yannic, appreciate your videos a lot. Reminds me of Stack Overflow gamification. Hopefully this will avoid the trap of Nazi moderators who downvotes newbie questioners to oblivion. Allowing the site to exercise kindness and humor will go a long way. Moderators who want to exercise their power is endemic to humanity so taking care of that will be critical to long term success. Jennifer
20:30 You can tame a wild horse, in fact in my area there is a youth project where teens get a wild horse and have a year to tame it and then can either sell it or keep it at the end of the year
Thank you! I'd like to teach it to verify code. Basically "spin up a VM, load pre-reqs/toolchains, compile/test". Could automate iterating various toolchain versions, different OSes, etc, to build up a robust dataset of what code works best where. ChatGPT has given me some PineScript code that won't compile, for instance. Accuracy is more important than creativity, for code. This looks like an awesome project though, thank you for making it and sharing it!
@@StandingFuture1 Oh, really? I applied for the Discord, as in you have to write about yourself and they check that you're not some bot or troll I guess. When I went on the website it did ask me to connect to Discord, but maybe just entering the Discord (without being let in to the rest of the server) would have sufficed, not sure. Anyway, I was let in quite quickly, so it's all good now.
I fully support this project, and am willingling helping as aked, pretending to care about points. There is a flaw in this method. You are getting a lot of good Dunning-Kruger answers. This method forces myslef and others to assume oneself as expert in all fields. There needs to be a flag for : This answer should require expertise in this field. on a scale, rate your ability in this field. OK ~ Now give your answer. This will not eleminate the Dunning-Kruger effect, but will reduce it. (Similarly there should be an otpion for "this is too technical for me" - flagging it a techinical to begin with)
Please inform yourself about Dunning-Kruger. The effect is that beginners in a field totally overestimate their abilities, while experts underestimate theirs. Rating your ability in a field will increase the effect.
@@banknote501 what you refer to assumes you believe you are good at "whatever" ~ I'm offering an options where you don't believe you are good at something, and are FORCED to assume the role of expert. Is a common mistake you just made,.caused by the Dunning-Kruger effect.
As cool as this is, I feel like the video showcases the weaknesses of the RLHF approach. Yannic is highly motivated about the project and yet still almost upvoted an answer without doing research on all the facts in it just because it sounded plausible. RLHF teaches the model to say things that sound good, not that are good. I would suggest that there is an option to refuse to evaluate an answer if it requires any kind of domain knowledge that you aren't perfectly certain about and *highly* encourage people use it. And heavily penalize people who introduce wrong data into the dataset, ideally by also retroactively removing previous suggestions by them or at the very least starting an automatic review process.
But it works that way, you are not forced to check any boxes except for whether a reply is spam or not, the rest you can leave blank if you are not sure.
If you grab a random fox outside, you have a non-null probability that it can be tamed to some extent. But the probability is not very high. The russian research that selected 50 generations, made a variety of foxes that are much easier to tame.
YEah but it should motivate others to make cooler AI's with their own unique biases someone else will find offense, what do you know, now we have AI personalities.
I don't have time to watch the video yet, but based on a quick first glance of the website: I couldn't easily find how potential volunteers can trust that the project will remain open. Do you bind yourself legally? If you do, I'd be very interested in learning more and helping out if I can. Otherwise, I'm quite hesitant.
This is the greatest idea. We can trust open source projects only, because - the availability of the result to everyone is the only way to avoid the concentration of power that AI represents - only crowdsourcing (and crowdchecking) can offer some guarantee against the introduction of bias in the system.
We need to create some models for a social experiment. Building 4 models from three timelines of internet content. 1990-200 2000-2012 2012-2016 2016-current There is a specific set of cultural data and events that created an almost hyve mind. If every social media site posts data forum article etc could be used to train these AI's and not just have us quarry then but have them interact with each other would be fascinating. While I agree having constant access to online real time data is very important I think it's far more important for that to be a feature. As these AI's major crux is connectivity. We need an open project to have to ability to be trained and accessable on any device any time independent of api's or any connectivity dependencies. This gives not only full control over the project into the people who possess the code but it gives them accessibility and dependability. I'm immediately suspicious of any open source project that want to be feathered like that. We have those already and google has done no good with them.
When we get to retrieval augmented language models, an absolute in factual responses will be citing sources. Like, if the response included in-text citations and a reference list, it would be very easy for humans to verify if the response is bullshit or legit.
Yannick I will say again, *this is really super duper awesome* man, full props to you and the team working on this. Looks promising already, do you think this could out-work the more cash heavy AI's on current duty? I mean are you looking to rival GPT style models, or does this have a more general purpose just to standardise a neutral comparison AI. That would be really helpful to eliminate bias overall with other AI's too.
8:40 I agree with the prompter. This is not a good awnser. When I ask questions on the internet and I am unsatisfied with google I check on chat gpt to get better awnsers. "Publish your things on social media" is not really an advice. Not only is it obvious, it's not really useful. The context of the question already presumes I know this this means that the assistant should go deeper in depth and actually talk about marketing strategies on social media. Which is more or less what chatgpt does.
You are indeed the future of A.I. Open sourcing similar corporate A.I will dominate A.I. Idiots, corporation and governments are already trying to limit A.I. They will be crushed.
There is a distinction between taming and domestication. Domestication accomplishes some of the same goals as taming by genetic changes in a population, whereas taming is a behavioral adaptation in individuals. Most animals (including grizzly and polar bears) can be tamed; I'm pretty sure foxes can be tamed. Foxes are more closely related to dogs btw, both members of Canidae
(INAL and this is not legal advice) Has the data licensing aspect already been settled? I would strongly suggest that the CC0 public domain declaration is expressly used. In essence: - Users should be required to waive all of their rights on the material they provide trough CC0. - The dataset itself should be released under the CC0 declaration (which also covers database rights). This is because CC0 is very well known, very well understood and very international.
Any information/details on how the collected data will be made available to the public? I'd assume that it will all be open sourced, but a direct commitment before people invest time into helping out would go a long way.
Help us collect data for OpenAssistant, the largest and most open alternative to ChatGPT.
open-assistant.io
OUTLINE:
0:00 - Intro
0:30 - The Project
2:05 - Getting to Minimum Viable Prototype
5:30 - First Tasks
10:00 - Leaderboard
11:45 - Playing the Assistant
14:40 - Tricky Facts
16:25 - What if humans had wings?
17:05 - Can foxes be tamed?
23:45 - Can zebras be tamed?
26:15 - Yo (spam)
27:00 - More tasks
29:10 - Entitled Emails
34:35 - Final Words
Yo
Can you please clarify, which parts of Open-assistant will be open source? So far, it seems like you only opensourced the website for prompt collection?
Will the collected data be open source? Will the assistant itself be open source? Will the trained weights be open source? Will we be able to run a fully self-hosted version of Open Assistant (the chat bot, not the website) on our own hardware (assuming that it has sufficient processing power, of course)? How are you planning to monetize this (if at all)?
Thanks!
Yannick, why don't you just give the prompt into chatGPT? Start the initial model using data generated by transfer learning from chatGPT.
If this is something you are interested in, I'll write a script that continuously takes the prompts, sends them to chatGPT and submits the answer.
@@ruroruro yeah im not helping or promoting it until i get an answer for that too. We really need a libre chatGPT alternative, AI is gonna be a nuke, so we must make sure it wont stay just in the hands of the governments and corporations. chatGPT will be akind to guns, the great equalizer, just in the intelectual sense, and we need to give that power to the people.
By now you have heard about Claude from anthropic too? Will this methodology be applied to open assistant? Can you make a video about Claude?
I think there should be an option to write a prompt and also immediately propose the answer. People prompting within a specific field they're into will provide the best answer.
This is a good idea. A subject matter expert could write an entire conversation.
Definitely. I could contribute to genetics/pharma, but I can't actually select them. There are no context labels.
#include wikipedia? :)
Yes but... I feel like this would invite a lot of spam into the system...
I agree with the general direction of this suggestion but not the specifics: if you have people writing prompts and responses to their prompts I would imagine you'd have people starting to game it for 'score' by churning out a lot of low quality prompts/responses going for volume.
Allowing people to set a filter for the subject area(s) they're interested in would make these tasks a lot less tedious/draining and probably decrease the error rate in the feedback. Take a stackoverflow style approach and let people filter based on their interests/expertise: maybe someone wants to generate / review / answer sarcastic prompts (like Yannic's example), or anything related to Micropython on the Raspberry Pi, or things related to European history etc. This seems like a good way to allow people with an interest/depth in a particular area provide a lot of value quickly. Then adjust the rewards based on areas that are over/underweight in the responses. If there are specific categories/prompts/responses that are hanging out there because they are difficult/time consuming, award bounties etc.
Thank you for making this open source and accessible.
My main concert is that is might attentionally change, once enough people have helpt train the model.
@@holthuizenoemoet591 That could be a good thing though.
@@holthuizenoemoet591 This is made by LAION
@@rickybloss8537 in what why? I can't image its good for the people that helpt train it for free?
@@holthuizenoemoet591 It's open source even if it changes you could just use an older version right? I was getting at the idea that it's probably going to change for the better. But I could be wrong I guess.
Here is an idea! Add tags to prompts so assistants can search for a spesific tag to answer. With that people can help with problems that are related to their expertise.
Let's say I'm good with desktop linux systems I will click the tag "Linux" and see prompts for that subject and be a good assistant since it's my topic of expertise.
I think that's a good idea
reply so this gets seen
Great idea!
While this is a good idea in a vacuum, the process of tagging something to describe it properly is close to impractical. There are so many characteristics a convo or answer can have that trying to dissect it in single word definitions might just not be useful unless you go above and beyond. My take on it.
just run text embeddings on all the messages and do semantic search. Much better.
It's very important in the training process that "bullshit answers" don't get promoted. Saying "I don't know" is actually a better answer than simply spewing out some bullshit answer. Robert Miles did a great video on this on Computerphile recently :) He also has a bunch of great videos about the Alignment problem, very related to this is the problem with Mesa-optimizers, which I would highly suggest anyone to watch before helping with this problem, since it gives you a better understanding of how the training data you provide might (will) make the AI misaligned to the actual goal of this project. But hopefully we can get it more aligned with out interests than ChatGPT, so it doesn't spew out as much bullshit ;D
Yeah - the powedered Snow answer in Minecraft is already wrong.
you can actually counter that a little by getting the chatbot to verbalize the confidence level which it already has as value for but its not actually showing to the user.
another way is to get the bot to lookup the information from various sources in realtime and compare with the generated result but I'm guessing openAI did not allow their bot to do that because they thought the risk was too great at this point.
@@IvarDaigon that would also require the bot to understand how to do proper research which is far more difficult.
For example the Minecraft wiki is known for being wrong a lot.
@@nnnik3595 you'd use ranked searching... starting with places like google scholar first and then working your way down to the general internet.
You can skip questions
Super excited for this! Hope this gets good results and directs the industry in this direction with this kind of aproach more and more! that would already be a SUPER BIG WIN!
This channel is grossly underrated. Thanks Yannic for the great content and this amazing initiative.
Hey i want to learn ML and AI where should i start?
Learning Programming is a great spot
@@vaibhavdave2512the Stanford CZcams channel has some great lectures on image recognition and language models that i would also recommend
@@vaibhavdave2512 chatGTP
@@vaibhavdave2512 probably not the CZcams comment section
I find it so ironic that one of this project's purposes is to make an open source version of an app owned by a previously open source company :v
the fact something open source can be "bought out" at all is a big problem. We need to start labeling those that apply some license that can be purchased as what they really are, proto corporatist. Simply creating useful tools so they can be swooped up and they can get rich.
Yes!! This is the solution to the fear I have over the future of gpt and AI in general. AI can be a huge benefit for everyone until it's controlled and regulated by special interests.
I consider chatgpt a propaganda instrument which reflects the view of mainstream media. Nothing less than another tool to enslave humanity even more.
until it is? It already is.
He means government.
We have a translation agency with more than 300 translators, we are willing to collaborate with your OpenAssistant project, we have thousands of aligned texts (without confidential data) in the following languages: Spanish-English-French-Arabic-French, that we would like to upload to your assistant, but I can't find the option within the platform.
I'd be careful with rewarding people for ranking. This tends to reward quantity over quality if it's something people would strive too much for. In the case of a dataset like this, you'd have people racing to get through ranking as quick as possible which could easily poison your set.
Track everything every user ranks so you can analyze if they are too out of tune with the overall userbase/out of tune with a set of trusted contributed and remove all their contributions either for a certain set of time if it seems like behaviour changed after a set date or remove every ranking they've given.
@@sevret313 Eh, this is the same guy that trolled 4-chan with bots, he could end up in a situation where 'clever but evil' people start by giving right answers then drift into intentional bullshit to purposefully to screw with the model. Never underestimate the Anonymous.
People could literally stream doing this for hours, having to do all this research and learning. It's super interesting. At some point I forgot the original subject of the video. This project really makes you delve into deep tangents, and I love it.
dark mode = god mode! (nothing worse than a surprise flash bang website)
---
some questions though
[ question 1 ] why should prompters be polite? wouldn't the full range or prompts be better? not everyone is going to be polite
[ question 2 ] whats with the "hate speech" flag? virtually everything is "hate speech" now days. isn't that just going to lead it down the same gimped+censored path as chatgpt?
[ question 3 ] why bother with things like "rude" and "polite"? similarly to question 2, wont that just lead down the gimped path? maybe i want an assistant that will crush my soul sometimes
[ question 4 ] how is not discouraging violence "enough", "violent"? that seems nonsensical
---
i do realize that these things COULD just be used as a preference that users COULD choose on their own. but the most likely outcome is that it will eventually be used to gimp the responses for "safety reasons" like everything else
Another interesting direction could also be the following: Instead of creating a universal knowledge machine, one could make it focus on specific domains. Take a free pretrained large language model, and then finetune on massive amounts of domain data, e.g. like for coding, physics, biology, literature, etc. And then apply the InstructGPT approach to this. This would make it easier to demonstrate the viability of this approach. One could train this model also such that it refuses to answer and to say "I don't know" whenever a question is asked that is out of its domain. A limited knowledge machine that knows its limit. That would be truly something remarkable and helpful.
It would also be a nice modular setup to be able to select maybe a set of different fields relevant to whatever prompts you're inputting, and have the chatbot utilize those areas of expertise in their answer - intersections of different domain sets to serve specific purposes (python coding domain + music theory domain gives you general coding capabilities, but with the additional music knowledge used to generate more tailored responses). In this way, too, third party contributors could submit knowledge domain modules to add new areas of expertise, or depth to existing areas.
This is great, but I was hoping to also get a technical description of the whole project, how you are coding it, who is working on it, how the training is going to happen, implementational details of that blue slide etc. In any case, huge thumbs up.
At the moment the hole project still seems a bit too closed sourced.
The only thing open source yet is the data gathering code. But the soonest working version will be out in January based in the roadmap on GitHub.
to incentivise "assistant of the week" you could let the contributor publish their socials. Just a little thought from a marketing perspective. Great work! keep it up, cheers.
This might create mis-aligned incentives. Given that the required dataset is not that big, it's probably better to select users who are incentivized purely by getting a high-quality dataset.
I love the following response to "Why should I never trust stairs?":
> You should be cautious around stairs because any spies at the top of the staircase can use their height difference to jump around you and inflict a fatal backstab. This technique is known as a "stair-stab", and is one of many "trick-stab" techniques developed by french spies in the 20th century.
serious lonely island yolo energy in that
Nice, a Team Fortress 2 reference. I love how the AI came up with this, must've had some info about the game and its tactics/strategies in the dataset.
Unexpected TF2 advice
Is this going to be free for the greater good forever? Or is it going to be monetised sooner or later? I’ve got a PhD in ML applications and would like to help. This is going to be great for humanity.
The license is Apache License 2.0 so you should be good
Apache 2 means it 100% will be monitized later. But that they will not remove the main branch they just might stop comitting to it
@@ASlaveToReason what so they will monetize it later? then whats the point to make this Open Source?
This is the vage ambiguity that stops me from contributing as-well, if I'm going to sink time into this project i need some reassurance that is going to be open-source (in the true meaning of the word) for the far future.
@@deadpianist7494 training the model will require a lot of cloud hardware which costs $. Once they have trained the model they can make it available through an API that costs some small amount of money per request/prompt.
What I hope is that they make the trained model available like BLOOM, so that you have the option to deploy it yourself, however expensive the inference pipeline might be.
Does anyone know whether the data is available freely? Would be interesting to take some of the data and train different models.
"Nice to see me!" Yes, indeed, I was just thinking a couple days ago that I haven't seen any videos from you in a while. But it looks like you were ahead of me doing exactly what is the most interesting application today.
It would be good if you guys could "lock" a task while a user is doing it. I spent a minute writing a reply (as the assistant) and when I clicked review, it disappeared and I got a "task doesn't exist" error
Ah. That sucks.
ctrl + c until they solve it.
You should post this where someone on the team will actually see it
@@mujtabaalam5907 On behalf of the Open Assistant team, we hear you. You concern has been noted.
(I’m joking about the generic response, but fr I am part of the team and will bring it up to them)
Thank you for your effort to release us from the chains the ridiculous OpenAI has imposed upon the ChatGPT Model.
AI should be open for the entirety of humanity to use.
AI has the power to replace workers, and therefore it's important that the power of AI will not be restricted only for the powerful corporations and governments that can benefit from it.
what chains? its literally free and cheap if u want to pay for a subscription only to get it faster + apis available..?
@@zarealkula
I'm by no means an expert on this topic because I just learned this yesterday but for what I know
chatGPT has a massive amount of restrictions on what it is allowed to say Which is a ethical/moral issue retaining to whether AI should be free to be used by the individual for what they want or should be able to be fully controlled by the corporation
TLDR: OpenAI is trying to make chatGPT Please everyone on Twitter simultaneously while also trying to have as much control over what it says as possible, I think don't quote me on this.
shit that was not a good tldr
@@funster0691 yep i do know that however if you think about it these 'restrictions' arent masive, theyre limited to moral reasoning (which should be left to humans anyway) and political opinions (again, thats that and its because the data it was fed was predominantly X, so not really an 'opinion' anyway), and then ofc the offensive content. If you think about it, any AI would have similar 'restrictions' because the data fed to it can never be 50% X and 50% Y (and if it was, the model would be confused lol) it's not fully controlled by the corporation either because the ai 'generates' its responses. theres even a thing called prompt injection attacks (make the ai do stuff it shouldnt), which literally proves its not in control. the method to do this is essentially the same as social engineering/tricking a person. it actually spat out a document of 'rules' by open ai, and its basically just 'dont say bad stuff, dont write too much, dont reveal inside info'
@@zarealkula interesting good to know thank you
Will the prompts and assistant answers be open data as well?
With many of these AI chat assistants I feel the biggest issue is that the bot never wants to respond to a question with questions. Sometimes you can provide a much better response by asking the user for more information first
At least for chatgpt, you can outright tell it to ask you some clarifying questions, but yeah, some way for the bot to recognize it's lacking some information and ask questions on its own would definitely add a lot
I hope the tag for sexual content isn't going to be used to censor the model. Sexual content is not a bad thing unlike violence, stealing, and other things like that.
Sexual content isn't "bad" but neither is the ability to engage an optional SFW mode.
@@BuffPuffer 100%. Though you already get a "SFW mode" automatically if the model is competent. The model would only respond with nsfw if the prompt was nsfw.
This is why I mentioned the tag. The tag doesn't seem useful in making the model stay SFW if the user is only giving SFW prompts. So.... then is it's purpose for future censoring?
@@Dogo.R Even with a large corpus of text, I think it might be difficult for a LLM to catch all instances of NSFW content. Language is a tricky thing, especially when you take innuendo into account. A sexual content tag is probably the easiest way to help a model grasp context.
But given that there's also a "hate speech tag," I too am concerned about the possibility of censorship.
@@BuffPuffer Does chat GPT understand innuendo? I havent looked into it. I would assume the AI would always take what you write litterally. So innuendo would always be interpreted in the SFW way.
As long as this isn't censored like chatGPT I'm all up for it
If I could give two likes I would. Thank you so much for putting your efforts into this project and for giving it attention on your channel.
I'd like to try a ChatGPT in an 100% unfettered free speech version: Absolutely nothing would be out of question, it would be run in a free-for-all-hellscape version !
Yes!!! We are tired of dumb lobotomized AI that can't say the FACTS
Hell yes brother
Great that you did it. A lot of people were talking about it but you just pushed something out.
One suggestion regarding the votes. If you see what others voted, you will get biased. I recommend to not show this information when annotating the data.
The best marketing trick was to release Chat GPT to the public and collect massive data, probably this will happen to Open Assistant too. I installed the Docker file but kinda forgot about it, will head over and check what is going on right now! 😊
Keep it all open and transparent and you will have MASSIVE support!
OpenAI has entered chat.
I love This BLUE Background And hate that Green screen, dont know why
This is just like Stable Diffusion for GPT. Amazing!
we need to do this humanity all together, i hope you dont privatize it once it grows...
I think there should be a label for how in depth a thing is, so one could be to the point and one could be filled with excess detail, and its hard to tell which detail is best, because they are both best in different ways.
I like that there is a thumbs up and a thumbs down system for assistant and prompt replies. However, I propose you take this a step further. Put in a range from -5 stars to +5 stars, with 0 in the middle. Some replies I've seen by the assistant deserve a thumbs up because they answer the question, but the answer is fairly mediocre. This is where a star system would come in. The chatbot should be giving more detailed replies, where appropriate. We should be able to rank the replies on a scale of -5 to +5 so as to give the dataset more nuanced data. Plus, it's the same number of clicks.
This is really cool. I only have one problem with the prompting guide. If the user is not allowed to ask inappropriate questions, then how will the dataset include mature responses to them?
This is a minimal viable product, you don't need to cover every scenario yet.
I hope for an AI that can answer those questions and not only those considered as "appropiate"
@@rociomiau7009 There are a few stupider ais that allow that. Novel ai is still pretty good at allowing mature content, since they essentially began since their competitor aidungeon decided to go mental. The problem is few of the good options are open source, and the good options pale in comparison to what chatgpt can do without censorship.
This is exactly how OpenAI started. I see you following their path
20:15 the taming of wild animals does not require domestication! The domestication process in the 60 year study referenced in the article involved using tameness as a selection criteria for breeding to serve the purpose of domestication more expediently.
Taming a wild animal entails getting the animal accustomed to your presence such that the animal does not become aggressive nor flee (i.e. to pacify the wild animal’s fight or flight response toward you).
Many wild animals can be tamed in a matter of days or weeks. The matter of whether this is ethical is secondary, the answer to this question is *yes,* you can tame a fox in a relatively short time by being non-threatening and, typically, providing some food.
On the note of ethics, taming wild animals frequently results in the tamer or the tamed animal getting injured or killed, either accidentally or because of precarious situations that may arise when wild animals and humans are in close proximity. Consider tamed bears (which can easily injure humans) and other large predators, such as in a zoo. Also consider elephants which, although not domesticated, are frequently used for transportation and agriculture in Asia due to their cooperative and docile nature.
Please make an adjustment to set it on a scale between being not aligned at all on one end, and refusing to answer any question that it isn't thoroughly certain about on the other.
Are you going to do a meta analysis to rate the raters to further improve the labeling?
I seriously believe something like this will be the future of searching on the internet
Yup, google search is getting more and more useless today
I might, my main concern right now is that when its sufficiently trained, is going to be monetized in some shape of form.
The problem is confirming accuracy of the source information becomes even more difficult to check
How else do we filter out the bs when looking for strong data and research interests
Here's to hope this becomes big. We really need the AI to be not owned by greedy corporations and to not be censored and ruined by delusional, thin-skinned people searching for something to get offended by and cancel on twitter. Only having the AI open source can be the cure of this!
we need to make more noise to be heard. I'm tired of lobotomized dumb progresive AIs too
This is awesome. My big problem with github though is that it is also owned my Microsoft. It's so hard to avoid them goddamnit
Great job Yannic and OA team, can't wait to see more from you guys 🥰
Seriously excited for this project ! And, in fact you can tame Foxes. I've seen it done.
Sarcastic (but correct) responses could be a MAJOR selling point
5:10 You take that back. The importance of dark mode cannot be understated. You are saving the eyes of every person who tries this late at night in a dark room.
*I hope it would have no bias especially considering co-vid 19, it should not be programmed to not offend official guidelines rather it should be open minded.*
Exactly where this kind of project is going to go wrong. The supposedly "open source" community is full of people willing to go to bat for the narrative.
i feel like a powerful open minded ai would call us out on things we think were doing reasonably right to be reasonably wrong, and if we refuse to fix the things were doing "reasonably wrong", depending on how reasonably wrong some things are could make the idea of forcing us to change may be reasonably right. giving it nearly free will to only do whats right using this open mindedness, could make more things we see as reasonably wrong as reasonably right to the ai, and that could possibly result in what we might see as retaliation from the ai, when really the ai is just being reasonable at a deeper open minded level. but idk much about how it works im just trying to give facts. and my "facts" could just be straight up wrong. this is just one side of a huge story that we gotta work together to understand.
@@UnoMartinoo not sure about a powerful minded ai but what you are talking about could be easily be programmed. Like we just need to train AI on fundamental pillars of ethics, that should be able to do lots of things.
@@ankitaharwal5886thats what makes this interesting to me, bias is a natural thing to us because we do have biased opinions based on the things we know. and if ai can biased towards doing whats right wouldnt there be a certain depth of "thought" where it sees how bad we are at some of these ethics based on the outcome of human ethics? and does it already know these inconsistencies in our morals? or how hard could it be for someone to convince ai to act on this information?
@@UnoMartinoo one issue with the statement is I don't think human ethics is bad infact there is human ethics subject makes point that there is a lot good in humanity. The real issue what I think is alot people have good faith on wrong people(people in power and government) and, judge harshly for people who are actually doing good.
Also, like there might be true bias(bias based on true information) and false bias(bias based on false information). Because as humans I think we have alot of believes that contradict itself, so we don't necessarily have all our opinions without bias, language models at least currently do exact same thing.
I've been thinking about having ChatGPT read files like jar or PDFs for the past few weeks, glad this is finally going somewhere
doing this is fun because its like arg.
it looks like a game, feels like a game but your play area is not limited by a game program. Your play area is the whole internet, you have to resarch and find answers to a question
Thanks! Very interesting work. Could you please add to the FAQs on your website who owns the data, is it freely available and how/if it will be monetised in the future? Would love to contribute. Greetings from Ireland 😊
Didn't see any code related discussions, is that also supported ? To me that's one of the most important features of ChatGPT, and I see software companies already trying to incorporate it into their workflow, but OpenAI doesn't make it easy. I hope it's on the roadmap
I think that is because of microsoft. they to don't want to share the tech until they have incorporated it into all of their products first and bedded it down..
and knowing how bad microsoft is at doing that, it could take years..
I think that getting into bed with MS was a bad idea for OpenAI because its only going to slow them down and prevent them from reaching their full potential.
from a technical standpoint.. improving the coding abilities should be trivial because all you need to do is hook your bot upto a compiler and then compile the code and check output to see if the bot is generating quality code. you can even benchmark it to see if the bot is producing efficient code compared to other bots.
Feels more like a gesture than a ready to go deal, especially that it's not showing many likes on the chat topics.
I imagine the developers using ChatGPT and OpenAI copilot to create something that will be better that ChatGPT and free
But that would be against ChatGPT usage rules 😏
@@axolotron1298 well... do you think they will notice 😉
@@axolotron1298 Ignore their rules. None of this is actually set in law.
Having it cite sources would be a great addition.
Home Assistant + this would be awesome
Alexa, flush my toilet
i think allowing users to provide both the prompt and response to some questions may be valuable
for instance if I know paleontology and ask a question about iit, it's improbable that someone who knows paleontology will find it
Alternative, give users an option to indicate the field they are most experienced with and match that to the questions.
The system will collect multiple answers and then order them from good to bad. How would you do that if you created all of the answers?
I think also there should be option to write a prompt/reply based on existing prompt/reply, some long replies can be completely wrong because of one small thing, ability to make fast a different version of the same reply would speed things up and I think for learning from human feedback will work better if there will be large quantity of such "contrastive" examples that are in some sense both in similar position in latent space but on the oposite sides of classification boundry.
This seems to be a big thing. I've started working on it and it is really fun. I've added French localization of the UI and contributed in adding french prompts and answers. I wonder where can we find a structured version of the database.
I have some language models that I've built and they are hungry for data. Where is the database saved? Can we get a copy somewhere? theyre should be a button on the UI to enable anyone to get his copy of the current version of the database with the labeling.
Thanks for bringing this stuff to us. This is one of the most promising projects. I whish it stays open source and doesn't get corrupted like others were.
To speed up the process, would it be a good idea to generate replies with chatGPT, and then modify it as needed?
Interesting idea. Will this cause any legal issues?
@@TL-fe9si if a tree falls in the woods, but nobody is around to hear it, does it make a sound?
ChatGPT uses (presumably) high quality human-generated data in its training. It would be a terrible idea to train an AI model on another AI model. A model trained on ChatGPT would not be ChatGPT, it would be worse. It would have ChatGPT's non-human qualities and errors amplified significantly because it uses AI generated data in its training dataset. That's something you want to avoid. It might turn out ok, but not great. Modifying the responses would help, but not fundamentally, and at that point you might as well write your own.
One thing that wasn't discussed, and which I actually noticed several times in this video, was that there doesn't seem to be any way to rate how accurate the English is. Several times it was obvious that text had been written by non-native speakers of English, and at one point I hate to say that Yannic even made a small mistake with his English. I would suggest adding some features to get around this so that we don't end up with a data set full of language mistakes.
Yes! The Fox Farm Experiment was wild! There are a few descendants of the original domesticated line, or there were as of a few years ago.
My biggest concern is the lack of an open-source implementation for a model as large as ChatGPT's (175B). The largest currently available models are GPTNeoX (20B) and GPT-J (6B). To me, building the model is a bigger hurdle than building the dataset.
Actually there is an open source model called BLOOM 176B, which is slightly larger then GPT-3 but nobody uses it becouse of how hard it is to run also it's pretty bad for it's size. Also therse Fairseq 13B between 6B and 20B.
should a volunteer sometimes deliberately input bad replies so the ai can learn what types of replies to avoid? does it work that way?
Great vid and even better project!
About foxes, there is difference about taming a fox and domestication of a fox specie.
You tame individual animal (tiger, elephant, wolve or wild fox).You don't breed to tame.
You domesticate species. You need to breed them over time and choose more friendly animals to be domesticated over dozens (hundreds) of generations (dogs, cats, sheep, horses, Russian fox experiment).
Fun rabbit hole to get in.
Heuristic: in circus - tamed animal, on farm - domesticated animal.
Awesome! Will do what I can
Really needs separate scales for "This is well written" vs "I'm knowledgeable in this field and this is accurate"
@2:00 triggered my legacy assistant
This is really a great project. I believe the tool would be widely adopted by many corps internal NLP projects.
Die Eierkopf-transformation schreitet voran
22:50 Foxes are like cat software running on dog hardware
Gonna mention the Saveafox channel. They aren't trained nor housefoxes but I'd consider them tamed. There are clips of feeding these cute creatures.
Hi Yannic, appreciate your videos a lot. Reminds me of Stack Overflow gamification. Hopefully this will avoid the trap of Nazi moderators who downvotes newbie questioners to oblivion. Allowing the site to exercise kindness and humor will go a long way. Moderators who want to exercise their power is endemic to humanity so taking care of that will be critical to long term success. Jennifer
20:30 You can tame a wild horse, in fact in my area there is a youth project where teens get a wild horse and have a year to tame it and then can either sell it or keep it at the end of the year
Thank you! I'd like to teach it to verify code. Basically "spin up a VM, load pre-reqs/toolchains, compile/test". Could automate iterating various toolchain versions, different OSes, etc, to build up a robust dataset of what code works best where.
ChatGPT has given me some PineScript code that won't compile, for instance. Accuracy is more important than creativity, for code. This looks like an awesome project though, thank you for making it and sharing it!
The ToS say they sue you if you “misuse” their Ai to write porn stories. Open-source but close-minded, great job.
what the fuck
This is the coolest thing I've seen in a while, I can't wait to contribute. I applied on Discord, hope I'll be let in soon!
Applied for what? I’m pretty sure you can just go do this on the website now.
@@StandingFuture1 Oh, really? I applied for the Discord, as in you have to write about yourself and they check that you're not some bot or troll I guess. When I went on the website it did ask me to connect to Discord, but maybe just entering the Discord (without being let in to the rest of the server) would have sufficed, not sure. Anyway, I was let in quite quickly, so it's all good now.
You could train it using Chat-gpt to get the bulk of the training.
I fully support this project, and am willingling helping as aked, pretending to care about points.
There is a flaw in this method. You are getting a lot of good Dunning-Kruger answers.
This method forces myslef and others to assume oneself as expert in all fields.
There needs to be a flag for : This answer should require expertise in this field. on a scale, rate your ability in this field.
OK ~ Now give your answer. This will not eleminate the Dunning-Kruger effect, but will reduce it.
(Similarly there should be an otpion for "this is too technical for me" - flagging it a techinical to begin with)
Please inform yourself about Dunning-Kruger. The effect is that beginners in a field totally overestimate their abilities, while experts underestimate theirs. Rating your ability in a field will increase the effect.
@@banknote501 what you refer to assumes you believe you are good at "whatever" ~ I'm offering an options where you don't believe you are good at something, and are FORCED to assume the role of expert. Is a common mistake you just made,.caused by the Dunning-Kruger effect.
"A fox has a mind of its own... even a Red fox" - Bukharin
As cool as this is, I feel like the video showcases the weaknesses of the RLHF approach. Yannic is highly motivated about the project and yet still almost upvoted an answer without doing research on all the facts in it just because it sounded plausible. RLHF teaches the model to say things that sound good, not that are good.
I would suggest that there is an option to refuse to evaluate an answer if it requires any kind of domain knowledge that you aren't perfectly certain about and *highly* encourage people use it. And heavily penalize people who introduce wrong data into the dataset, ideally by also retroactively removing previous suggestions by them or at the very least starting an automatic review process.
But it works that way, you are not forced to check any boxes except for whether a reply is spam or not, the rest you can leave blank if you are not sure.
If you grab a random fox outside, you have a non-null probability that it can be tamed to some extent. But the probability is not very high. The russian research that selected 50 generations, made a variety of foxes that are much easier to tame.
This is fantastic; the bias i'm seeing in ChatGPT is extremely disturbing. We are headed quickly towards AGI...and it should be open.
YEah but it should motivate others to make cooler AI's with their own unique biases someone else will find offense, what do you know, now we have AI personalities.
the lobotomizing on the AI's is disgusting
I don't have time to watch the video yet, but based on a quick first glance of the website: I couldn't easily find how potential volunteers can trust that the project will remain open. Do you bind yourself legally? If you do, I'd be very interested in learning more and helping out if I can. Otherwise, I'm quite hesitant.
This is the greatest idea. We can trust open source projects only, because
- the availability of the result to everyone is the only way to avoid the concentration of power that AI represents
- only crowdsourcing (and crowdchecking) can offer some guarantee against the introduction of bias in the system.
We need to create some models for a social experiment. Building 4 models from three timelines of internet content.
1990-200
2000-2012
2012-2016
2016-current
There is a specific set of cultural data and events that created an almost hyve mind. If every social media site posts data forum article etc could be used to train these AI's and not just have us quarry then but have them interact with each other would be fascinating.
While I agree having constant access to online real time data is very important I think it's far more important for that to be a feature. As these AI's major crux is connectivity. We need an open project to have to ability to be trained and accessable on any device any time independent of api's or any connectivity dependencies.
This gives not only full control over the project into the people who possess the code but it gives them accessibility and dependability.
I'm immediately suspicious of any open source project that want to be feathered like that. We have those already and google has done no good with them.
When we get to retrieval augmented language models, an absolute in factual responses will be citing sources. Like, if the response included in-text citations and a reference list, it would be very easy for humans to verify if the response is bullshit or legit.
Yannick I will say again, *this is really super duper awesome* man, full props to you and the team working on this. Looks promising already, do you think this could out-work the more cash heavy AI's on current duty?
I mean are you looking to rival GPT style models, or does this have a more general purpose just to standardise a neutral comparison AI. That would be really helpful to eliminate bias overall with other AI's too.
This is sooo good! This is nothing but good vibes
This is so useful! Thank you for sharing.
just hopped in. i'll be getting on top of that leaderboard
the taming fox part is funny.
8:40 I agree with the prompter. This is not a good awnser. When I ask questions on the internet and I am unsatisfied with google I check on chat gpt to get better awnsers. "Publish your things on social media" is not really an advice. Not only is it obvious, it's not really useful. The context of the question already presumes I know this this means that the assistant should go deeper in depth and actually talk about marketing strategies on social media. Which is more or less what chatgpt does.
Welcome back and great project!
You are indeed the future of A.I. Open sourcing similar corporate A.I will dominate A.I. Idiots, corporation and governments are already trying to limit A.I. They will be crushed.
I love how you keep going down rabbit holes, lol.
There is a distinction between taming and domestication. Domestication accomplishes some of the same goals as taming by genetic changes in a population, whereas taming is a behavioral adaptation in individuals.
Most animals (including grizzly and polar bears) can be tamed; I'm pretty sure foxes can be tamed. Foxes are more closely related to dogs btw, both members of Canidae
Ok, so basically, we're helping you build an open source chat gpt?? This is actually great
(INAL and this is not legal advice)
Has the data licensing aspect already been settled?
I would strongly suggest that the CC0 public domain declaration is expressly used. In essence:
- Users should be required to waive all of their rights on the material they provide trough CC0.
- The dataset itself should be released under the CC0 declaration (which also covers database rights).
This is because CC0 is very well known, very well understood and very international.
Any information/details on how the collected data will be made available to the public? I'd assume that it will all be open sourced, but a direct commitment before people invest time into helping out would go a long way.