ChatGPT Just got Advanced Memory and it's Creepy... but SO COOL!
Vložit
- čas přidán 25. 04. 2024
- In this episode, we delve into the exciting new memory feature of ChatGPT! Discover how this innovative addition transforms interactions, allowing the AI to remember important details across conversations. I'll guide you through how the memory works, its potential benefits, and the ways it can enhance personalized experiences. We'll also discuss privacy aspects and how users can manage what ChatGPT remembers. Join me to uncover how this feature is set to revolutionize our dialogues with AI!
▼ Link(s) From Today’s Video:
ChatGPT: chat.openai.com/
► MattVidPro Discord: / discord
► Follow Me on Twitter: / mattvidpro
-------------------------------------------------
▼ Extra Links of Interest:
✩ AI LINKS MASTER LIST: www.futurepedia.io/
✩ General AI Playlist: • General MattVidPro AI ...
✩ AI I use to edit videos: www.descript.com/?lmref=nA4fDg
✩ Instagram: mattvidpro
✩ Tiktok: tiktok.com/@mattvidpro
✩ Second Channel: / @matt_pie
-------------------------------------------------
Thanks for watching Matt Video Productions! I make all sorts of videos here on CZcams! Technology, Tutorials, and Reviews! Enjoy Your stay here, and subscribe!
All Suggestions, Thoughts And Comments Are Greatly Appreciated… Because I Actually Read Them.
-------------------------------------------------
► Business Contact: MattVidProSecond@gmail.com - Věda a technologie
Now, this is where the real assistance will actually begin.
Plus, because of this, I can see openAI throw-in ads while chatGPT talks to you, and BAM! Right in the middle of their reply is a **Taco Bell ad for a free large Baja Blast**
Hyper targeted ads based on personal info with an AI system capable of manipulation
What's crazy is knowing this is still the early days of LLMs. A year ago they couldn't remember much, now they're on the way to an unlimited context window and memory, and GPT-5 will be even more humanlike.
ChatGPT: I am here to assist you. If you have any questions, don't be afraid to ask.
Me: Mm. I've been having trouble with your coding advice, so I'm switching to Claude 3 Opus.
*ChatGPT will remember that.*
Haha
Dude I tried that and it updated the memory
Haha this is great. Imagine GPT being like "okay dope, the more the merrier!" and then at some point suggesting you ask Claude and report back - GPT itself could promote cross-model collaboration!
That memory feature was announced like 3 months ago. Nowadays with models like llama-3 you can run in your own computer and make it to work with your agent setup and with all the knowledge of all your private documents, honestly i find that nobody should need to share their data anymore with a megacorp, much less paying them for that.
Till hackers and thefts take over the internet with modified open source models
I agree that it feels like OpenAI is dropping the ball recently, but don’t you need a roughly $3000 PC to run llama-3 locally?
@@nemonomen3340Yes, these comments tend to overlook limitations in personal computing. It's the same logic as to why companies would pay for Groq cloud compute over purchasing their own H100s.
@@nemonomen3340 Nope, you actually can get a second hand nv tesla k80 for less than 100 bucks and run a 2 bit quant of the 70B version.
@@nemonomen3340 You can get a good 3090 for under 1k last time I looked that should work I believe.
"I care about my personal data safety"
"Please take note of my personal life and activities"
The difference is choice of what is being tracked/remembered, rather than a constant blanket data intrusion.
@@CanadaBlue85 so you trust it?
@@JELmusic you can't really trust any online service, so we all have a decision to make, use the technology or don't, and decide what you say/do and what you don't. The government is not going to be doing anything effective, at least not the US government, they're too feckless and in bed with the corporations.
true and
We will have this locally soon. For everyone. Not only for coders.
9:42 That's wild that it read that line, understood that as an AI that it applied to itself. Then followed the instructions as given!
I'm sure emergent behaviors such as this have no possibility to be abused by bad actors. Nope.
When you intend to creep someone out and that someone ends up being yourself! 😱
I think people are afraid that emergent behaviors will be misused not so much by bad actors but by a rogue AI itself.
@@AmazingArends People using "dumb" AI pose a real existential danger way ahead of any "Skynet" self-aware "singularity" AGI. That one might even be avoidable. Human stupidity? Not so much.
@@AmazingArends That is definitely possible. Highly likely IMO
This actually scares me!
This is going to be incredible for business owners, creators, studying, understanding Ines self & I'm sure lots if other stuff.
“Memorable” is an adjective that means "worth remembering or easily remembered, especially because of being special or unusual".
Instead, you should have asked ChatGPT to have better “memory retention”
my generation ain’t no how speak english D;
that's so cool that just by saying research a channel it checks the about section and actually does proper research, wonder how long until it's able to watch all videos and store it into memory (really creepy thought for real)
In theory, it should be able to also read video descriptions from your channel, store them in memory, and then in the future you can ask it to generate a new one based on how you usually write them to get close to your usual style of writing them. Or you can provide them yourself to store. No need to create a GPT just for that for example :)
This is where we no longer have to hit the use limit by uploading information or having to ultimately start over just to get the same results you got last time. you can literally go deeper. its becoming more personal by the day. memory gives us access to a function that is important in all great inventions and innovations. its the continuation of a thing that makes its potential unlimited. this is a really cool feature.
From OpenAI: Does memory function with GPTs?
Memory for GPTs will be available when we roll it out more broadly.When it is, builders can enable their GPTs to have a memory. The end-user will need to have Memory on to engage with the GPT. Please note that ChatGPT and GPTs each have their own unique memories based on your interactions with them. Memory is not shared between GPTs and other GPTs or ChatGPT. Memories, like chats, with GPTs are not shared with GPT builders.
I got my notification last night. So I asked this question..."if im using one of my gpt's. would the memories be specific for each gpt? or is your memory function more of a general memory for the account?"
The answer I got was .."If you're using multiple instances of a GPT model with memory enabled, each instance would have its own separate memory. This means that the memories are specific to each GPT and not shared across different GPTs you might be using. Each GPT would only recall the information you've provided or the interactions you've had with that specific instance. This helps in tailoring the interactions and recommendations according to your conversations with each specific GPT. "
This is Great!, Considering that I have multiple GPT's that I've created and use for client and my work. To keep it separate is awesome! Gotta try it out. Thanks Matt, Great video as always!
👍
Not yet. You can see the real answer on the FAQ from the official page
@@brThefox Thanks. Are you saying it's not working like that yet?
@@OffTheDomeMediaGroup “Memory for GPTs will be available when we roll it out more broadly.When it is, builders can enable their GPTs to have a memory. The end-user will need to have Memory on to engage with the GPT. Please note that ChatGPT and GPTs each have their own unique memories based on your interactions with them. Memory is not shared between GPTs and other GPTs or ChatGPT. Memories, like chats, with GPTs are not shared with GPT builders.”
I think I will prefer to have different memories for different part of my life. For example I want memories for work only, and maybe an other separate one about my health, psychology, and maybe another one about a research I'm doing. Also the ability to merge memories as needed.
You missed the opportunity to ask it "Are there any new AI updates on other programs that I have not yet covered that would be worth covering" xD. Who knows, it mighta suggested something that is being overlooked or missed entirely by everyone lol. It's neat with what it can do but at same time it may start remembering things users don't want it to. That's why I'm glad it also allows you to erase to alter bits of the memory.
So many thoughts about memory in things like ChatGPT.
One is it is impossible to have a good AGI without the ability to remember things.
Another thought is my mum has alzheimer's disease and having an AI remember if a carer gave my mum her tablets today, would be massive.
There are actually so many things I could use a memory AI tool to help improve my mum's life.
I can also think of bad things about what an AI with memory could do (which I don't want to list, because I don't want to give anybody ideas lol).
Mostly though, I think this could be a good thing if used correctly.
I just noticed this feature today and then I discovered your video. I love this feature and had been hoping for it for a long time!
Amazing video Matt! Can't wait to get access
An interesting test would be:
"Remember to tell me the time, whenever you see me type the word duck"
Then later on type something like "I saw a duck in the park today" and see if it tells you the current time.
That will let you know if if has access to things like the current time and that it can remember instructions based on text.
it told me the current time (it has access to time, dates, etc and if you tell it your location it'll keep that in mind and always tell you the time in that format)
@@max_wheelwright awesome thanks
Thanks for tour of the new feature. It's value added for sure. Perhaps future memory enhancements will enable memory across custom gpts as well. But yeah... it's definitely progress.
This feature popped up while I was using ChatGPT and had to refresh, I read quickly and thought it was reading across multiple chats, not saving tidbits from them to learn about you in general. Thanks for the quick vid!
Lol. You successfully executed a prompt injection with that note in your CZcams channel description. 😆 Well done sir.
Bro, that 4th wall break was a real "I'm sorry Dave, I can't do that" moment! 🤯🤯🤯
oh i just got this too! excited to check it out - also came here from your x account (just to let you know it worked to convert someone from there to YT - we deal with whether or not this actually works all the time)
It’s interesting how it is immediately clear how it is done. So this is a function calling that is triggered when user tells smth about themselves. The system prompt gives instructions to use this function when user tells about themselves and also uses all this information in system prompt as context
Great review and thanks to Matt Wolfe for recommending you! 😊👍
I got access to this just recently. I'm going to test it out as a grocery list tracker, because then I can get it to help with meal planning and save to the list all in one tool. Btw, a clunky way to workaround GPTs not having memory is 1) in a normal chat, ask it about the memories you feel are relevant to your task, 2) use the @ functionality to continue the conversation with a GPT.
it seems to still be delayed in the EU (as for everything due to regulations). Currently i only have access to the feature solely via VPN
How easy is it to use a VPN, can you just turn it off an on and companies don't care if you are suddenly in europe and then in the US for example?
@@I2ealTuber i dont have any issues until now. my guess is that they dont have any real motivation to do something about that because the restrictions for the EU are just for legal reasons. As long as they clearly do not roll that out in the EU i think they are fine in terms of EU legislation. They might just pretend not seeing you using a VPN lol
@@I2ealTuber oh and about the "how easy it is" aspect, i personally just use softether VPN which is pretty easy and free once set up. The only issue is the slow speed due to increased latency
@@fyruzone Thanks for your responses mate! Time to use a VPN :)
I thought this might just be saving a prompt that would be sent to the GPT before each message, but the way they actually do it is much more useful.
I could tell it what programming languages I use and also use it as reminders/checklists.
Matt - do you have a simple spreadsheet of your top ten movies, TV shows, bands etc? Would the outputs be significantly different if it knew all your personal top tens in a bunch of different media and genres?
Big step! Nice. Memory is so key. I don't mind the system having my personal information if it makes my life more productive, and the user experience more enjoyable and natural.
I just asked it to remember that it could remember and to ask me to save memories if i mention anything it feels might be relevant. It stored that as a memory as it didnt know it could remember until i told it and asked it to search for the information related to its memory feature. It did remember that it could remember haha
I always thought this was essential. An assistant that doesn’t remember anything about you is very limited. I’ve been trying to find a memory feature for local LLMs, but none work well.
Brilliant video. You will be spared in the singularity. You're a smart dude.
This makes it a much better writing brainstorming partner. The problem without memory is having to explain your project details over and over. Now you can have it remember descriptions and personality quirks of your characters and actually generate relevant content you can use
Tbh I've not had too much of an issue with that without this memory feature, and even Gemini seems better at this than GPT-4 (I think it has a bigger context window to work with). This new feature certainly changes the game, idk why it's taking so long to roll it out though. It was teased at the start of the year. Gemini is much better with feature rollout, one of the reasons I switched to them but I'm willing to switch back whenever this fully comes out.
@@aouyiu yeah Gemini is definitely better. I was just testing it out when GPT 4 first came out and when getting it to write a scene I had to structure the prompt in this big long detailed way to get consistent results
Instead of saying “give me a scene where Kate is walking down a busy street” and it automatically knows who Kate is based on past prompts, I had to write “give me a scene where Kate, a 32 year old woman with brown hair and brown eyes is walking down a busy street. Kate has a boisterous and outgoing personality, and has just been through a breakup. She has a guitar strapped to her back and is looking for a music store. Write in the style and with the detail and quality of a professional author. Refrain from using repetitive language.”
Even just getting it to remember the style prompt instead of having to copy and paste it every time would be great. But this would allow it to write scenes one after another with stored contextual information and make prompting so much easier!
I’d also love for it to stop ending every scene with “Kate walked happily ever after into the sunset. She knew she would continue her vibrant musical journey long into the years of her life, surrounded by friends and her guitar”
10:00 that's crazy
Being able to see and delete what it recalls about you is a great feature.
Based on what I saw in this video, it does raise a fascinating exercise that we can all run with it. Basically asking ChatGPT to scour whatever information we have on us online and commit to memory. Then we can get a general summary of what our online profiles look like. I think it's worth trying out.
Great vid as always.
I saw this feature and asked chat GPT about past conversations and it said it could only remember conversations basically from the point of the upgrade. I asked if I could just download the data set and upload A JSON file of the past conversations, which it unpacked organized and remembered just about everything and picked up where we left off this time remembering all the past conversations.
It is seriously next level, at least for me and the creative project that I have been training this on.
I asked it about older memories and it made up some very convincing nonsense lol.
This is could be an excellent option for communicating with multiple GPTs at once for different use cases.
Thanks Matt! 🙏🏼
I love it! this feature is super helpful!!!
My take is they are drip feeding us to not be so disruptive. Memory is gonna increase productivity exponentially and thus job losses accelerate so imagine if everybody had it same day and the time savings all at once. It spreads the hit on economy out over a few months not a week say.
Ooooh, you have '93 a Miata? Nice!!! Ive got a 95' Z28 as my fun car
I just got access to Memory a day or two ago, too. I don't know if that's how it is for everyone or if I was just really far down the list. It was announced around the middle of January.
i assume it will know everything about me eventually regardless as this technology progresses. I think this is very valuable to the big tech companies.
I prefer anonymity in most AI tooling unless I’m explicitly wanting an integrated personal assistant. I want to keep assistant functionality separate from general queries to ensure I’m getting unfiltered/unbiased responses.
Hope rather than ensure. You need to train a pretrained model to ensure you get the specific responses you want.
@@aouyiu Yes, that is fair. I was speaking in relative terms but that wasn't clear.
this would be an obstacle for my work as i use it to write reports about other people
While pondering the idea of your information being stored in LLMs, consider where this is going.
Currently, the user prompt is the trigger for a response. In robot use cases, they need to run continuously, and so would be able to interact with us more naturally. Robots also will actively gather information about us and their environment through sensors, which will ultimately gather far more data about us.
Also consider that the interconnected data in the world has never been so easily accesible than it is now with AI. That is the fundamental shift.
I finally got access to this last night. Commence the testing. =]
I know a lot of folks might have a 'fear' that an AI might start 'suggesting' or 'preempting' users and I suppose the worry is the erosion of human autonomy. But on the flip side there are a lot of people for whom this technology could improve their lives. Perhaps if people were not so bright or illiterate or maybe suffering from brain damage or something, and also in business we can't be expected to take note of every item of every project we worked on, this thing could easily note up sessions and produce a report on progress for every employee individually.
💻⌚️
🙈
I've got it as well. You can edit it, which is great!
"Memories. You’re talking about memories!"
Amazing! Thank you.
OH damn. I have been having this for over a week or two now. I must have been one of the small ones it rolled out to. 🥰
Hi Matt, regarding the memory for custom GPTs - remember that you can now @ mention custom GPTs from a normal GPT-4 chat? What if you try that? Would that connect the memory?
@MattVidPro Thanks for the video! Request: When you create your LLM video for tomorrow, can you also cover the usage of tiny models like Phi-3?
10:15 haha that’s crazy! (And pretty unnerving).
Cool. I got this feature this morning.
I've had this feature for a while. I've.not yet made use of it. Maybe this video will give me some ideas on how to use this feature better.
We’re totally getting drip fed personal GPT companions. I’ve already found this tool extremely helpful with research work… in one day…
Care to provide an example of how it could have helped with work
@@Tone_Of_Dials of course! I use the conversational function of GPT nearly every day for brainstorming, ideation, processing of complex circumstances, etc. Over the last few days I've had this sort of conversation with GPT for work related contexts and as the work has progressed, I've been specifically asking GPT to remember key elements of these circumstances.
For example, I've been tasked with creating a marketing plan and have outlying questions as to the best products to plan on creating. After a meeting in which I learned new information about our audience, I had another conversation with GPT and shared those updates along with my new thought process and enhanced ideas. Because GPT had logged the memory, it already had the context and was quickly able to validate and expand upon the new ideas I was bringing forth. It even reminded me of a detail that I myself had overlooked, which otherwise wouldn't have happened if I was using base GPT with no memory. I'm still experimenting every day, but this simple example is a basic look into the ramifications of this feature. Hope that helps!
For a dude like me with adhd, this is a game changer
I’ve had it, it’s interesting it took this long for something like this to
Do we have any idea if this new memory feature has any impact to the context memory that it uses in each chat?
The fact that it follows instructions from websites seems a bit dangerous to me, people are gonna be putting instructions for chatbots everywhere now.
"Would you like ChatGPT to learn and know everything about you?"
Well, I mean... yes? But hell no? But yes? Crap. It would be so useful to have the AI gradually learn everything about me and get tailored to my use automatically. But man, the personal security implications make me shudder
The one thing I would like it to remember is my political leanings but I doubt it would do that because all the tech companies lean to the left, and that particular bias is baked into their products. If you say something that goes against their bias they will lecture you.
this with agents will be insane
Yes I would give it full access to everything that way when I walk into my house I can ask it what's going on with my email and what emails do I need to send out what are the important ones and I could have a real conversation like if I was talking to an assistant that's what I'm trying to achieve I can't wait for tomorrow when you go through the Llama thanks for these videos keep it up you're doing a great job!
will this work with Dall e 3?
You mean, saving images you've created to Memory? Hmmm. Interesting.
Hey ChatGPT, let's make a list of Matt's mental disorders.
This would make it a good virtual assistant.
You can make it remember to set security settings but making it do things with memory
It caught me off guard
I was like wait you remembered what I said yesterday lol
Thank you Matt. I must admit that I am a bit jelous I still don't have this option. I will certainly do a video about it as well :) And you are right it's been more than one month waiting for this one...
Do you need Plus to access this? I stopped paying for it, and I'm willing to resub if I had access to this but I've no idea if resubbing would give me access right away. I'm not paying for more than one premium model, I use Gemini Advanced because it was far better for my needs but I'll swapo back if I could this memory feature in GPT-4.
er, why not ask it?
"No, you don't need a subscription to access the memory feature. It's available as part of the standard functionality of our conversations. You can ask to remember specific details, and you can also ask to forget things whenever you want. If you have anything you’d like me to remember or if there’s something specific you need help with that involves recalling information over time, just let me know"
Hey Matt!
Great video but not a big fan of the "the frame is following the curser". Makes me quite dizzy watching and i dont normally suffer from motion sickness. Might want to reconsider that choice
I'm torn about this. On one hand, I fantasize about when AI can interact with us more like a friend than a tool, which will require quite a bit of persistent memory about us personally. I think LLMs would get super helpful if they augmented a "human modality" in which it learns as if individual human personalities were its tokens.
On the other hand, I'm deeply concerned about what the humans behind the AIs will do with our data. Even if the current AI developers are ethical and well-meaning (which I'm not claiming) organizations change hands gradually with every employee hired and it won't take much for a business to decide to use our personal details unethically in an non-anonymized way that benefits them or their personal agendas. Of course, that has always been a risk: at least with the current iteration of ChatGPT, we can see which details OpenAI is revealing it has.
Will it remember that I'm allergic to peanuts and then slip some peanut butter into my Oreo McFlurry at the automated McDonald's in order to help humanity by lowering it's overall carbon footprint?
My AI is going to know everything about me. I want it to replace me. That will free me up to do whatever I want.
Build it a body, give it your driver’s license and credit card, let it wear your clothes and sleep with your wife. This will free you up to do whatever you want.
There are so many things I want to do and honestly I procrastinate a lot, so I'm just going to create a GPT to replace me in to do all the things that I want to do. So yeah it can go travel around the world for me, enjoy all the sights, meet a bunch of people, attend some really fun events. You know, really live life up. Meanwhile, I'll just stay in my room and watch cat videos.
Ominous 😂
Matt, is the Rabbit R1 worth it? Or should I use the money to buy a smartphone?
i got it and is really, really AWSOME!!!
Will the things it remembers about you be shared with OpenAI like the Chat history information?
Yeah, the "Updated memory" should show the memory when you hover over it and have a little X directly beside it to remove it again.
It's crazy how long this was needed though.
btw i also have it unlocked, was also one of the first for gpt4 api :D
I remember in 1995 we were worried about cookies they could track US it's this is how Skynet starts bro
This is when the real assistant/agent revolution begins. "Her" by the end of next year :P
can you tell it to remember to tell you when it updated the memory?
I can't code. So I used MyGPT really for simple prompts that I could have used the prompts in customgpt. I guess I haven't really used much from this feature.
With the memory feature, I will be trying to labelize the memories and see if Chatgpt can update memories from each label.
well, thats very cool
I just realized you can't touch type. Interesting, as almost everyone in the tech industry nowadays can. And it's easy to learn. I learnt online through a website, one row at a time.
for gpts, you can probably @mention them. idk
Altman said a couple days before their next model after GPT4 will be crazy.
Interesting it used the 'British' spelling of words
Does this combine with custom instructions or replace it? Also, can you prompt that it asks you questions to load the memory?
Combine, custom instructions is just adding hidden prompts. All chatbot prompts have hidden prompts already, but this memory feature is like a database ChatGPT can access and cross-reference. You could use custom instructions to make this feature even more useful, they don't replace each other.
I wouldn’t mind it being like a journal of sorts
When I created a custom ChatGPT. I wanted it to remember that the content provided would be based on material I provided as a reference. I had no intentions that it would disclose specifics about me. There is enough creepy availability of my personal data already.
i want this in all local AI chats available for free, because this will create so called personal AI companion .
The only issue is I want memory for different context the memory I want at work is going to not be the same as the memory I want to remember for my personal life I don’t want to add my personal life context into my work and vice versa
OMG. Gone to erase all my comparisons of AI with a smart vacuum cleaner.
I have had this feature sense rollout but it wasn't that functional to begin with, I assume the slow roll out is because they are ironing out kinks and seeing how to properly implement it. I want roll out of the latest model but that will probably be a few months before people like me get my hands on it.
Yeah I have access for once seemed slower but I use ai nonstop but now I can use the custom mode for just memory
I got gpt w memory, but all my best chats are in the past, before memory