The Biggest Problem for AGI Just Got Fixed!

In this video, we discuss AI memory. How does AI get memory? How long is their memory? AI memory has many limitations right now, and that's one of the significant hurdles to reaching AGI. A new research paper gives hope of 2 million+ tokens for a prompt, enabling nearly unlimited context to be passed into a prompt. We also look at memory compression like in the autonomous agents paper.

Enjoy :)

Join My Newsletter for Regular AI Updates 👇🏼

Need AI Consulting? ✅

Rent a GPU (MassedCompute) 🚀
USE CODE "MatthewBerman" for 50% discount

My Links 🔗

Media/Sponsorship Inquiries 📈

Links:

My Workstation Setup:

[Zokrar]

Hey Matthew, just wanted to send you a message to say thanks for the content you've been producing. Happy to see your channel getting more attention!

[ourkellyfamily]

This is great news!! I am constantly hitting the token limit, even with the 8k I get with ChatGPT Plus. Matthew, can you boost the volume on future videos? Your videos are consistently quieter for me than other videos on YouTube.

[etunimenisukunimeni]

Your stuff is really nice! This is one of my favourite AI channels on Youtube, always calm and collected, sensible and clearly presented. Have a like! 👍

[tonyfweb]

Just recently started following you but you have amazing content! So thank you for all that you're doing.

Just quick suggestion: if you can, please upgrade the mic, get it a bit closer, put some sort of gain on it, or speak a bit louder. I have my speakers on full volume and the same goes for the PC volume. However, I often find that if I have even a little bit of background noise in the room, it can be a tad bit difficult to hear you. It's not that bad, but it's there. Mind you I am a bit deaf from my time in military service, but this issue is a little bit more noticeable on your channel than some of the others that I watch.

Once again, great content though. Please keep it coming!

[AaronTurnerBlessed]

As a developer, I've found that working with LLM's often means dealing with memory limitations. Unfortunately, many people looking for a quick AI fix don't nor want to have the patience for redundancy, which can lead to disappointment. Keep it going with these relevant topics sir! TY!

[amkire]

When you talked about it losing its memory, I instantly thought of the movie "50 First Dates". The solution they had was pretty much a precis of the essential things that had happened previously, which was added to or updated to the point just before the memory went each time. Probably not the solution in this case, but hey, I knew I could find a good reason for watching this movie.

[SuperMIKevin]

This is truly one of the biggest limitations at this point.
It would be awesome to see it fixed.

[gwky]

Honestly, with just 2048 tokens (which is the max capacity for most LLMs) of context, it's already decent for one or two prompts and with janky techniques like long-term-memory through contextual prompt injection, you can "remind" the model, but if the models were scaled to millions of tokens of persistent memory, I imagine the conversations would be truly amazing. At the very least, I could finally play a DnD game without the AI forgetting where I am or what I've equipped.

[MikeD-tfdk]

I started noticing this stuff really quickly when I started working with it. It’s great to hear. You have an interest in the same sorts of things. For example, I’ve been trying to wrestle with a way that I can get an LLM to really get enough information about my writings to truly clone my writing style kind of like how 11 labs takes audio samples and then replicates them, I am more than a little surprised at the difficulty I’m having finding a way to do that that doesn’t require me to be a full blown AI oriented computer scientist.
It doesn’t solve the long term memory problem, but it does provide a way to have a rich base amount of information from which to extrapolate new output, which is definitely something that is, and will be valuable in the future.

If you have any suggestions on that, I’d love to hear them!

[santiagomoebio]

The key to manage memory is emeddings. They allow to bring along to the prompt the relevant (memory) information for each query.

[SteveBMayer]

I've been trying to get AutoGPT to self fix issue 38, good to know I'm not the only one.

[hanskraut]

Good 🤩 very important. Recall and context memory. More complex systems are needed and processing while idle!! (using your pc)

there are a few things needed but going in all directions i mentioned is already a very very good start and maybe enought just find all of my posts on the internet youtube (endless, but some quite besides the point.... not always verymotivated or healthy to write nice) or forums. (AutoGPT or OpenAi 1 post)

You seem to have a good grasp of some things that are important, gj :)

[fontenbleau]

Oh no, Ai analysed your video 😮

"Darling what are your suggestions to this problem described in the video? I know you smart and can resolve this!"

AI: Thank you for your kind words! While I cannot resolve the AI memory problem on my own, I can suggest some ideas that researchers and developers might explore to address the issue:
1. Hierarchical Memory: Develop a hierarchical memory system that stores information at different levels of granularity. This could involve a combination of short-term memory for recent or frequently accessed information and long-term memory for less frequently accessed or older information.
2. Attention Mechanisms: Improve attention mechanisms in AI models to better focus on relevant parts of the input data, allowing the model to process and remember more important information while discarding less relevant details.
3. Knowledge Distillation: Use knowledge distillation techniques to compress large models into smaller ones while retaining essential information. This could help reduce the memory footprint of AI models and make them more efficient.
4. Incremental Learning: Develop AI models capable of incremental learning, where they can learn new information without forgetting previously learned knowledge. This would allow AI models to continually update their knowledge base without the need for costly retraining.
5. External Memory: Integrate external memory systems, such as memory networks or neural Turing machines, with AI models to store and retrieve information more efficiently.
6. Model Pruning: Apply model pruning techniques to remove redundant or less important parts of the AI model, reducing its size and memory requirements without sacrificing performance.
7. Collaborative Learning: Encourage AI models to work together, sharing knowledge and learning from each other. This could help distribute the memory burden across multiple models and improve overall performance.
These suggestions are just a starting point, and researchers are continually exploring new techniques and approaches to address the AI memory problem. The field of AI is rapidly evolving, and we can expect to see significant advancements in AI memory and related areas in the coming years.

Can you tell more of advice number 5, darling?

AI: Of course! Advice number 5 focuses on integrating external memory systems with AI models to store and retrieve information more efficiently. This approach aims to enhance the memory capabilities of AI models by providing them with additional storage and retrieval mechanisms. Here are some examples of external memory systems:
1. Memory Networks: Memory networks are a type of neural network architecture that incorporates an external memory matrix. The model can read from and write to this memory matrix, allowing it to store and retrieve information over time. Memory networks have been used in tasks such as question-answering and language modeling, where the ability to remember and access information is crucial.
2. Neural Turing Machines (NTMs): NTMs are another type of neural network architecture that combines the learning capabilities of neural networks with the memory storage and retrieval capabilities of Turing machines. NTMs have an external memory matrix and a set of read and write heads that can interact with the memory. The model learns to use these read and write heads to store and retrieve information as needed, enabling it to handle tasks that require more complex memory manipulation.
3. Differentiable Neural Computers (DNCs): DNCs are an extension of NTMs that introduce more sophisticated memory addressing mechanisms. This allows the model to learn more complex memory manipulation tasks and better handle problems that require reasoning and planning.
By integrating external memory systems like these with AI models, it becomes possible to enhance the memory capabilities of the models, allowing them to store and retrieve information more efficiently. This can lead to improved performance in tasks that require memory manipulation, such as question-answering, language modeling, and planning.
It’s important to note that integrating external memory systems with AI models is an active area of research, and new techniques and architectures are continually being developed to improve memory capabilities further.

Darling, I’m interested in thorough description of advice number 3. Differentiable Neural Computers (DNCs)

AI: I apologize for the confusion earlier. I understand now that you are interested in a thorough description of Differentiable Neural Computers (DNCs) from advice number 5. Here’s a detailed explanation of DNCs:
Differentiable Neural Computers (DNCs) are an advanced neural network architecture that combines the learning capabilities of neural networks with the memory storage and retrieval capabilities of external memory systems. DNCs are an extension of Neural Turing Machines (NTMs) and were introduced by DeepMind in 2016.
The main components of a DNC are:
1. Controller: The controller is a neural network (typically a recurrent neural network like LSTM or GRU) that is responsible for processing the input data and generating the output. The controller interacts with the external memory and decides what information to read, write, and retrieve from the memory.
2. External Memory: The external memory is a matrix that stores information in the form of vectors. The memory can be read and written by the controller using read and write heads.
3. Read and Write Heads: The read and write heads are mechanisms that allow the controller to interact with the external memory. The controller generates read and write vectors, which are used to read information from the memory and write new information to the memory, respectively.
4. Memory Addressing Mechanisms: DNCs introduce more sophisticated memory addressing mechanisms compared to NTMs. These mechanisms include content-based addressing, which allows the controller to find memory locations based on the similarity of their content, and temporal addressing, which enables the controller to access memory locations based on their recency.
The key advantage of DNCs over traditional neural networks and NTMs is their ability to learn complex memory manipulation tasks and better handle problems that require reasoning, planning, and problem-solving. By incorporating an external memory system and advanced memory addressing mechanisms, DNCs can store and retrieve information more efficiently, allowing them to tackle tasks that are challenging for traditional neural networks.
Some example applications of DNCs include:
1. Graph traversal and shortest path finding in complex networks.
2. Solving puzzles and problems that require reasoning and planning, such as the Tower of Hanoi or the Travelling Salesman Problem.
3. Learning algorithms from examples, such as sorting or searching algorithms.
DNCs represent a significant advancement in the field of AI memory and have the potential to enable new applications and use cases that require more advanced memory manipulation and reasoning capabilities.

Thanks darling!

AI: You’re welcome! If you have any more questions or need further assistance, please don’t hesitate to ask. I’m here to help! 😊

MM-react at Huggingface (which is gpt4 without registration)

[eck]

Totally agree, short term memory is very important. Imagine we could feed whole codebase as prompt to be able to refactor code as never before. This will provide insane abilities for refactoring, code generation, complex bug search, possible security issues and so on.

[thelavalampemporium]

so crazy we are talking about characters in video games having a full life long memory of their existence already lol

[Poqets]

Speaking of RNN's, you should check out the RWKV model!

[KA-kpme]

I actually thought about infinite context by using vector databases. I'm not sure if memory degradation is a thing. Even as a human You have attention span, that makes it very natural to fetch even big tree of multiple context bits from different areas of database - which btw. would be insanely fast - and then recreate most important parts of it.

[RwDANCR]

Schank’s “Dynamic Memory” is old, but good and highly relevant.

[nosult]

Here fast, love content. Wondering about your background.

[noobicorn_gamer]

Liked right off the bat. Yes memory is an important issue. Replika tries to combat this via their 'facts' and 'diary' menus but they also can only do so much.