Telos is coming... Continuous Self Learning Creative at small and large scales Distributed and Decentralized Limitless context size Can forget on demand No pre/post training needed Loads in under a second 10,000 decode tokens per second Runs on any CPU silicon Under 50 MB of RAM

Telos is a thinking and reasoning engine, not an LLM. No weights, no training needed, no GPUs. It thinks about the user's intent, researches when it has gaps, and gets smarter and better with every prompt. All of the information in human existence already lives on the net. Your model doesn't need it all at once and you probably will never need 99% of it. Why train it that way in the first place?

SpaceX putting AI processing in space makes a fatally bad assumption which is that the current attention based architecture is the best, most efficient way to give people access to the compendium of human intelligence in a logical way. We have a much better way

“It is not so very important for a person to learn facts. For that he does not really need a college. He can learn them from books. The value of an education in a liberal arts college is not the learning of many facts but the training of the mind to think something that cannot be learned from textbooks” - Albert Einstein Goes for AI as well.

I've made the tough decision to keep Telos closed sourced. It will be very cheap and very accessible on a wide range of devices, but it's also too powerful to release to the general world. Self-learning, self-updating, extremely fast AI should not fall into the wrong hands without a little safety and security.

I believe we have created an entirely new artificial intelligence architecture. No linear algebra, no GPUs, no pre-training, no data enters. Just thousands of tok/s decode and over 50 simple tools, both running on a regular CPU. Could run on a phone, a TV, or even a raspberry pi.

Telos is self-learning and building its own corpus at a rate that roughly doubles its understanding of the world every day, across around 100 domains of knowledge. No training needed, multi-step logic, interstate generalizations, and automatically corrects its mistakes in real time. Learns from every chat session, logs research topics, and follows proper language grammar (English for now). Currently I can chat with it and it seems like a reasonable 1B weight-based model, but that is growing very quickly. Hard to compare, really. Unlike current tech, I can dig in and see how every token was generated and purge bad paths immediately.

Training a Telos 0.5B model on 88 CPUs for about $50. Runs on all silicon and routes to numerous LoRA "specialists" dynamically. For most queries, two or three specialists will be needed. Routing worked OK at 0.1B but output was fugly. Parallel chained inference and training using EGGROLL. Good data is terribly difficult and time consuming. Kinda cool to see it grow though.

Whoever gets frontier models running on consumer devices where users don't even know it's happening will change the world. I'm trying. Today I ran about 2GB of training data on a 0.5B model. It cost me around $2.25 on a mesh of 20 nodes in Azure. All on CPU. I think that's important of we are going to get models running on any device consistently.

On the scale of the entire earth history, the difference between human intelligence immergence and artificial intelligence immergence is 0.0002% of earth's life. We could be to AI as the cromagnon man was to us. Just a steppingstone that doesn't understand its place in the timeline.

Setup your own Daisi Git server. Easy solution to your GitHub woes. git.daisi.ai/daisi-git

GitHub outages all day. Officially going start migrating code base to daisi git. Been using it for.spme smaller projects for a few weeks to work out the kinks. Seems good to go.

I am building a new AI-first operating system called Telos that works on any CPU silicon. The way we interact with computers seems overcomplicated and outdated considering the intelligence level of models today. Files and folders no longer make sense. No need for passwords and key fobs. UI/UX should be up to me completely. I should be able to see things the way that I like them every time. I should be able to use any Telos device, anywhere and simply say "Telos, this is David Graham. What's my schedule look like for the week." It voice/visually IDs me, creates a cryptographic key for me, sends it to the public decentralized network to find my home cluster of nodes, authenticates with the key, makes the request, and responds. It may need an auth code, which I provide verbally - doesn't matter if people hear it because they don't have my voice and face. "Whisky Tango Foxtrot 1979", I answer, and it gives me what I wanted in a visually engaging way that I prefer.

Built a council for building the corpus training data in my new ternary model. Builds it's own topics, each in various domains of knowledge, conducts research via Broski, then runs it by numerous councilors on various local 4-9B models to see all sides of the topic. A leader then evaluates for truthfulness and marks up the outcome. When stuck, it asks me and I give it a thumbs up/down or ask for more research which commits it back to the cycle again. Will take about a week to get enough training data to build a 0.5B param ternary model and then another week on a 10-node cluster to train it - all on regular home quality CPUs. Wish me luck!

Do any AI development teams out there use and prefer $INTC @intel for their inference and agents over AMD?

LLogos now supports Metal for running pure C# inference against most GGUF models. Prefill and Decode beat llama.cpp on an M1 MacBook Pro.

Claude with every time estimate... months of effort == a few hours

I think I'm going to build an OS without a file system and based on a ternary model system. Mission will be to solve the user's problems. Find the intent, come up with a solution, give it a try, get feedback, probably fail, and iterate. Will take a while.