Uggghhh what a busy week. Been up all night fixing Intellite and I got some progress I wanted to share.
ALL errors we experienced (there’s a lot): Gradient vanishing (AI dies) Gradient explosion (AI barely learns and destabilizes) Activation Explosion (AI dies) Layer Scale Crushing Signal (AI dies) Activation Explosion (again) (AI dies) All fixed, but now it keeps overfitting on datasets due to a dataset handler error 😩
Intellite Chat is going well! I have made format much more uniform (now chatml), data handling now works, and we removed and added new datasets for a better mix of quality. I have also updated the default model to be fp16 instead of fp32 for faster training. Stay tuned!
Question about Intellite Chat to you guys! (If you don’t know, Intellite Chat is my up-and-coming 100m parameter AI model focused on high-quality chat.)
What quantization variants would you want to see for Intellite Chat? It’ll come with FP32, FP16, and BF16, but any others you want? Maybe FP8 or even BitNet’s 1.58-bit quantization? Let me know!
My first ever working model! Introducing tinyvvision. An extremely small zero shot image classification model for simple shapes, letters, colors, and numbers. This is a foundation model for Intellision, an image-text-to-text model, which will be much bigger but still impressively small. ProCreations/tinyvvision NOTE: it’s small, mistakes will happen!!!! Not meant for complex stuff, just a demo for simple stuff.
All your GPUs you have suddenly vanish and any devices with a GPU turn to CPU only
Every time you use a HF space you randomly start dancing for 5 minutes OR Every HF model you’ve hearted / liked suddenly disappears and becomes inaccessible
What do you think of Intellite’s new icons/logo? Let us know!
Also Intellite chat technically does work! But we decided to scale it up a bit (same parameter count at 100m, but we went from trained on 4b tokens to 200b tokens, big upgrade!) for max quality.
Yesterday we asked: If AI could master just one thing, what should it be? And the responses? Insightful, creative, and genuinely thought-provoking.
Here’s a few that stood out:
🍼 @NandaKrishvaa said “Curiosity like a baby.” Instead of just answering questions, an AI that asks them with childlike wonder? That’s a whole new kind of intelligence.
@MrDevolver suggested “Master being Jack of All Trades.” Sure, it bends the rules a bit — but adaptability is key. Sometimes breadth can outshine depth.
@afranco50 argued for “Perfect logic,” saying it could unlock all other abilities. It’s a solid point: if an AI can reason flawlessly, it may just learn to improve everything else on its own.
⸻
Our take? We still believe the biggest leap forward is flawless conversation — not just accurate, but deeply human. Emotional intelligence, nuance, humor, empathy. That kind of interaction is what makes AI feel real.
It’s also why we’re building IntellIte Chat to focus on that exact skillset: • Emotion-aware replies • Natural, flowing conversation • Strong command of casual and expressive English
When it releases, it won’t just talk — it’ll connect. And in a world full of tools, we think the future needs more companions. What do you think? Let us know! If we get more comments, might as well do another post on this tomorrow lol.
If AI could master just one thing, what should it be?
Human-level conversation that actually feels real? Flawless, bug-free code? Perfect math and logic every single time?
Whatever you pick, just know the AI will not be so good at the other topics! No picking "all of them" either lol What do you think matters most for AI to truly level up? Drop your thoughts in the comments — we’ll share our answer (and maybe a few of yours 👇) in the next post.
Come check out ProCreations/black-hole-sim-randomized a high-fidelity dataset with 400,000+ randomized black hole simulations — packed with relativistic metrics, Kerr geometry, and GR weirdness to help AIs actually understand physics.
🕳️ Teach your model: • Time dilation • Redshift • Orbital dynamics • Frame dragging • Full Kerr tensors …and more, all in raw JSONL!
This release celebrates SimpleMath hitting 200 downloads — thank you all so much for the support! 🙌
Training works fine, normal loss, no more gradient explosion or gradient vanishing, etc. BUT, before I officially flip the switch and turn on training, I wanna make sure its the best possible 100m parameter model it can be, so I am working a bit more (probably an extra 3-5 days) to add even more innovative AI improvements to intellite.
