State of DAISI
Daisinet
The network has been up with 100% uptime for the past three months. Usage is fairly low, with our own tools comprising the bulk of the 40M+ tokens processed. Right now, it is running a subset of open weight models that can run on low VRAM boxes, but there are a few hosts running larger models in the 30B category so more capability is available.
We've moved away from focusing on the network as a product, as was obvious when we went away from the Open Source model. The primary reason for this was that our labs are having major breakthroughs every week on our DaisiDb, Telos Smart Engine, Telos OS, and more products that have not been announced. We think that Telos will become the de facto model for the network fairly soon.
Telos Is Different
Telos is a thinking machine. That may sound derivative in today's world of AI, but it isn't. Attention models today don't do much thinking really. They are trained on massive datasets and require massive hardware to run them.
Telos is different in that it thinks first, then researches, then answers honestly. It doesn't hallucinate or give you answers that are not backed by what it found. Yes, it can give you a wrong answer if the data that it researched was incorrect, but we have curated a list of trustworthy sources for its research, so that is very rare.
Logic, thinking, and math underpins everything that it does. There is no training round (pre or post) and it doesn't need a GPU to run at all. It uses the new DaisiDb on the backend for long term memory and has in-memory caches to keep things quick. After it learns something, it always remembers it in context to what it was asked to do. Intent is always a driving force in how it organizes it data.
It's something different. It's going to be a different experience, as such, but in many ways, it's proving to already be better. For example, in math situations (and coding) it turns out that it can build its own tools and instrumentation to infer an accurate result. This is done in extensible tools that it can stack. So, if one problem requires something that is a part of another problem, reuse isn't an issue. Always, it find test cases that validate the results that it received are the expected results. Proofs and unit tests galore so it doesn't misuse them.
Telos Coder
It's still very nascent, but I have been able to produce a limit coding application that runs on a local box's CPU without any frontier LLM. It learns the user's intents over time, generalizes similar intents across numerous users, and evaluates itself against public repos and commit histories to fully understand the problem sets. In a way, it could be considered training, but not like LLMs are trained. Still very early, but it can build basic CRUD websites, mobile apps, and CLIs without much problem from a natural language request of the user.
Next steps are to focus on architectural framings so that it can encompass harder problem sets. There's also a "fine tuning" for little things that users will want to change like "make this column sortable", which seems fairly easy, but it hits every layer of the application if done correctly, so therefor has to understand each layer to remain useful and complete.
Crowd Editor
What started out as a tool that I wanted to be able to build and share my fiction writing with the world in a new, interactive way morphed (read scope creep) into a story reader. The idea that I want to unfold is that the daisinet will help extract characters, define traits and personalities, and then use text to voice models to read the prose in a narrative voice, but the characters themselves come alive. Easier said than done, but the structure is there. Added to that, I want the general public to be able to recommend edits, so they can fork my work, edit it, then submit for review. It's very close to the git PR system actually. Writers can review editor's work and accept it or not, rate them or not, and all scoring and metrics are tracked for everyone to see in real time.
URL Rewriter
Simple clone of other tools out there that take your URL and shorten it. I had a few short domains that I wasn't using, so why not. Generates QR codes and tracks all usage metrics that you'd expect. Easy, free.
Daisi CRM
We use this in our actual businesses, so we eat our own dog food here. It's great and easy to use. Includes basic lead tracking and customer relationship management (SMS, Email, etc), but also has a full digital signature implementation that I've found useful in many ways - and avoids the CRAZY costs of the online tools that are available in this space today. Robust set of automations and conditional routing as well.
Where this is going
What I want to do is have a basket of tools that any small business would need. You don't have to leave our ecosystem, everything works on a single sign on, and your whole team works with a reasonable month price. Adding up everything that I've spent for software in my previous companies, I spent well over $5,000 per month on various things. I want this all-inclusive, full-team solution to be under $250, including unlimited inference on daisinet. Doable, but many more apps needed to show the value.
Conclusion
That's all I've got for now. Just plugging away, writing about 10 apps at a time using Claude, Codex, and Grok. What a great time to be alive!