finally back from my trip. i didn’t post the last update because the entire scope of what i’m doing just completely escalated.

to recap: my first ship attempt for thalamus got rejected. the feedback was that the UI looked too much like a generic AI wrapper boilerplate, the README felt AI-generated, and it needed unique features. instead of just doing a superficial skin change or tweaking the CSS to look pretty, i am using this as a reason to completely redo the underlying engineering and turn this into a true L4.5 or L5 agent that destroys things like claude code.

i attached a photo of my handwritten notes to this post. it maps out the exact problem statements i’m tackling and the major flaws of currently existing solutions out there. my handwriting is messy, but the technical breakdown is all there.

here is the actual game plan for the AI routing and training infrastructure:

right now, i am updating the app to route claude models via puter.js. to be clear, this isn’t for the user interface—puter.js is strictly my temporary backend AI router. it acts as a fallback bridge while i build out my own custom models. i’m going to benchmark my self-trained models against every single task in my 9-agent pipeline. until my model is 100% foolproof on a specific task, puter.js will keep routing that specific agent’s requests to claude. but the second my custom model qualifies and beats the benchmarks for code mode, puter.js stops getting requests for it and the pipeline switches over entirely to my model.

the actual training project is running parallel to this:

• i’m fine-tuning a 32B model optimized explicitly for pure coding and logical tasks.
• training is happening via QLoRA at FP8 with a 4k/8k base context, but i’m extending the context window all the way up to 800K using RoPE and YaRN with 4-bit compression, flashattention-3, and chunked prefills so it can digest entire codebases at once.
• i’m doing the actual training on kaggle kernels using the 30 free weekly hours on their TPU v5e-8 clusters.

based on my calculations, this model will completely fail at things like history or literature, but it doesn’t need to be good at them. my gemini researcher agent team already handles live web search and pulls fresh information for those blind spots anyway. on pure code, this thing should easily beat sonnet 4.6 and compete with opus 4.6.

the token economics are insane if this works: about $2 per million input tokens and $6 per million output tokens based on max output possible per hour versus GPU compute costs. that gives me sonnet/opus level intelligence at the price of haiku 4.5.

the real bottleneck right now isn’t the data or the training—it’s hosting. deploying a 32B model with an extended context window means i need to secure an 80gb VRAM per parellel prompt configuration. figuring out how to source or fund that infrastructure without incinerating cash is the main issue i’m working through while compiling the training scripts.
Today i will just do the rerouting to puter.js, then start working on fine tuning the 32b model. Wish me luck.