Kruel.ai V5.0 - Api companion with full understanding running 16k thanks to advanced memory system

that is pretty cool. I have nothing in IP, I build private system outside of ai for industrial plants, electrical mechanical and sensor stuff, pretty much plant optimization stuff, as well software systems for hardware. Kruel.ai was built because of memory loss in my family which I wanted to have an ai system that works with me daily that is my backup memory system. the idea behind it was that it can recall everything I need no matter how old I get. The other side effect the persona system was designed so that all users and ai’s are stored the same in the system meaning that if a user was loaded as a persona the Ai could in essence become the user, with all its characteristic and the likes along with even a cloned synthetic $$$ voice. That way the system could continue the legacy of the work that it had been doing with the user and itself. haha that is just a side effect of how the data memory system is setup. 1.7 years of talking to an ai is pretty neat to see what happens when you flip it.

3 Likes

Here’s a great series of lectures on Cognitive Neuroscience. I would encourage anyone working in this space to watch those lectures:

And here’s a must read book on the subject:

Brain and Behavior: A Cognitive Neuroscience Perspective https://a.co/d/5yJS8wu

5 Likes

That’s called a Memex Machine and has long been the dream of many people. So I say keep going:

3 Likes

I sometimes wish we were all in the same company with equal stakes and equal contributions so we could talk freely about all this crap :laughing:

I’m just saying while biomimetics is good, it would be a gigantic loss if we didn’t also meld in the advantages that digital systems bring to the table.

1 Like

I completely agree that a replication of the human brain at the cellular level is the wrong approach. A synthetic brain doesn’t need most of the systems an organic brain has anyway. Why does it need survival instincts or fight or flight?

I’m going down the path of a hybrid architecture. Leverage brain architecture where you can but employ standard software architectures everywhere else.

3 Likes

The core features of an organic brain that you want to mimic in a synthetic brain are memory, emotions (only a subset,) attention, and consciousness. Pretty much the entire limbic system you can ignore unless you’re trying to build a robot.

2 Likes

watching that youtube now, hah didn’t know it had a name but yeah same idea like an archivist. After seeing Fletch on Apple tv I was telling my wife that there is what I am building.

2 Likes

I have about a 90% accurate simulation of memory and about a 70% accurate simulation of emotion. I’m able to get GPT-4 to develop its own likes and dislikes. I’m starting to think about attention and consciousness but they’re trickier to model.

2 Likes

I’ll disagree with you on the consciousness part, depending on how you define consciousness.

I think that’s where we will run into serious trouble.

1 Like

For my purposes I define consciousness as goal creation & seeking and planning. These models are horrible at planning but I have an effective workaround for that.

1 Like

Yeah, I think personal mission finding is dangerous and unnecessary. But I guess that’s more of a philosophical issue.

But finding a mission is also just a mission, so I don’t think that’s technically hard in any specific sense.

1 Like

I’ll go ahead and share my workaround for the planning issue as it’s not part of my core IP.

It’s true that these models are horrible at planning but they’re really good at mixing so what you have to do is re-shape the planning problem as a mixing problem.

What you can do is show the model the goal and a list of human authored plans that are semantically similar to the goal but not necessarily a perfect match. The model will use these human authored plans as inspiration and then generate a new plan that’s better aligned at achieving the desired goal. This technique works amazingly well

3 Likes


still a baby I can’t even render it, it can only display a fraction of it. sadly it wont display.
I think we are testing around 150k messages per month currently and still ramping up to more as we get close to testing a full work week in the field to see how it keeps up.

we meaning me and the ai lol. We are a team after all :slight_smile:

2 Likes

What’s the central node? :thinking:

Very curious. Our engram topology is more of a hypersphere onion so you can only sorta really visualize shallow slices… I’m envious :laughing: !

maybe one day you will become one :rofl:

1 Like

that Center node is 1 persona lol. I have about 6 personas in this system, and 53 test users with myself as 3 test users from different systems that talk to it. it has multiple inputs, from machines that talk to it on electrical level, to twitch.tv with 300 viewers and other systems including the voice input. so it has a lot of things happening at once all recording everything all at once. no realtime vision though too costly and I think they cap us per minute with like 100 blinks a minute for lack of a better way to stay it lol

2 Likes

aaah. nice. for us, each persona is a completely separate group of container volumes lol.

question: do you compartmentalize information? i.e., is there information that a persona can share with one user, but not another?

1 Like

The persona and users chat directly to each other, however the ai when in a room setting like twitch tracks its own memory of the conversations between people so that it can respond and remember group things, well users maintain privacy. on the one on one discussions. so its compartmentalized.

the idea was one day I want to get this into a body so I am trying to think ahead of the things I will have to tackle from an automation perspective. having it have interpreters for multiple types of inputs using ai understanding = smart machines that can do everything you want with data and understanding.

1 Like

Very cool :slight_smile:

I have a slightly different approach to tooling but can’t talk about it yet :confused:

Regarding compartmentalization, I’m thinking that the agent should be capable of learning from multiple sources and integrate all that information - without blabbing to people it shouldn’t. We have Read/Write permissions on memory streams, but I think that’s still a little bit… …rudimentary.

3 Likes

this is the message app outside of voice that I built. the application allows user to sign in as well to generate wallpaper for the app. background. currently does not support code input yet, or uploads. both those are on my agenda.
Example of what I mean that the memory system is very dynamic there are memories going back a year along with short memory 7 days which is default for understanding current time frame. all in 16k that’s the impressive part more than anything if you ask me haha. I am still trying to put more into that 16k. this is why I hope to see a 32k model. that would give me so much more room to play. 128k gpt4 though is crazy. I opened my memory system up for that for an hour of play, well expensive it blew my mind and warranted a lot of calls in the middle of the night with excitement. Id run 128k all day if I was Rich, but then I would be shelling out over 1k / month out of my own pocket for that much data intelligence.

2 Likes

Have you considered running a dynamic model selector?

not every task needs 128k gpt-4. You could have the power of gpt-4 in your back pocket, but use 3.5 for small talk.

1 Like