Pasi Tuomainen’s Post

Last week I watched an underwriter spend 40 minutes opening 6 different tools to review a single submission. 😮💨 PDF reader for the slip. Outlook for the broker email. A separate viewer for the loss run video walkthrough. Excel for the schedule. Teams for the recorded call. And a chatbot that only does text. 📄📄📄 Then NVIDIA dropped Nemotron-3-Nano-Omni this week. 🔥 One model. Text, audio, video, documents — all in a single turn. • 30B-A3B MoE, hybrid Mamba-Transformer • 9x throughput vs other open omni models • 1,200s native audio, long video, 100+ page docs • Agentic CUA built in • BF16 / FP8 / NVFP4 The lesson: the bottleneck in enterprise AI was never the model's intelligence. It was the seams between modalities. Every handoff between PDF → audio → video → spreadsheet was a place context died. 👀 What to actually do with this: • Audit one workflow this week. Count the modality switches. That's your real backlog. • Stop building 4 separate pipelines for 4 file types — one omni model collapses them. • If you're on closed APIs, start benchmarking open omni now. The cost curve just bent. • The agentic CUA piece matters more than the spec sheet — that's where workflow automation actually lives. Which workflow in your business still dies at the seams between modalities? 👇

You can add a tiny "side car" LLM that reads near the bottom of a model, and runs backwards to inject near the top. This allows even small models to focus on accurate coding or other tasks - as the secondary model can prevent "jitters" in the output that cause the model to go astray. https://lnkd.in/gTTmqXkf

 For small and medium enterprises, the path forward is rarely about building multi-million dollar models from scratch. The true competitive advantage lies in taking specialized, highly efficient open-weight models, grounding them in your unique internal company data, and choosing a compute structure that insulates your balance sheet from compounding variable token fees. GALAXAI Solutions for more info pls contact at info@galaxai.ae

Most people use AI models. Last week, I fine-tuned one locally on my own laptop. I built and fine-tuned a coding-focused LLM using: Qwen3-0.6B QLoRA Hugging Face Transformers PEFT bitsandbytes RTX 3050 Laptop GPU The interesting part wasn’t just the training. It was understanding how much real AI engineering happens BEFORE trainer.train(). Some things I learned during the process: → Raw datasets are messy → Chain-of-thought traces can actually hurt small-model fine-tuning → Data cleaning matters more than most people think → Quantization makes local AI genuinely practical → LoRA/QLoRA completely changed what consumer GPUs can do I went through: CUDA setup VRAM optimization dataset preprocessing chat template formatting LoRA adapter training inference debugging Hugging Face deployment The model was trained locally on Windows using an RTX 3050 and then published to Hugging Face. One of the coolest moments was seeing the model’s behavior actually shift after fine-tuning: more structured responses, engineering-style outputs, and copilot-like formatting. This project gave me a much deeper appreciation for: efficient fine-tuning edge AI open-source LLM ecosystems practical ML engineering workflows We’re entering a phase where running and customizing AI models locally is becoming accessible to individual developers, not just large companies. And honestly, that future feels incredibly exciting. Check It Out Yourself: https://lnkd.in/dZwfPDwn #AI #MachineLearning #LLM #HuggingFace #QLoRA #OpenSource #GenerativeAI #Python #DeepLearning #LocalLLM #Transformers #AIEngineering

Spent about $3.50 on a single RTX 4090 to figure out why recent papers on hybrid AR + diffusion language models keep contradicting each other. Some report that adding an autoregressive planner improves diffusion-model reasoning. Others report it degrades. Both are right, in different projections. Starting observation: on LLaDA-8B, prepending the literal string "Plan: " to a GSM8K question costs 8pp of accuracy. No plan content. Just the word. That single number forced a decomposition. Hybrid AR/DDLM reasoning fails along at least three orthogonal axes: Interface-format brittleness — how much accuracy drops from any plan-shaped scaffold, content-free or otherwise. Planner-content trust — how much the model uses upstream plan content once the prefix shape is absorbed. Sampling-diversity preservation — whether fine-tuning collapses or expands the stochastic branches that consensus mechanisms rely on. A small (r=8) prefix-robustness LoRA flattens axis 1 from 8pp damage to within 1pp. Axis 2 turns out to be capacity-dependent in opposite directions across planner sizes — previously unmeasured. Axis 3 unexpectedly expanded under format-augmented training rather than collapsing, the inverse of the standard encoder-collapse story. The consensus-distillation track was the most instructive part. A late-block LoRA designed to distill majority-vote into a single forward pass plateaued at 70.5% across a 3.25x capacity bump. Looked like architectural impossibility. It wasn't — two design errors were masking each other. Fixing both recovered accuracy to 79%, within sampling error of target. Generalizable lesson: parameter-efficient distillation of sampling-based inference mechanisms requires the surgery to match the temporal structure of the original mechanism. A plateau across capacity is not, by itself, evidence the distillation is impossible. Workshop-scope, not main-conference. Single seed, N=200, GSM8K only. Limitations flagged honestly in the appendix. Total compute under $4. If you work on hybrid AR/diffusion or parameter-efficient distillation, especially if you've seen similar prefix-shape damage on other DDLMs, I'd be interested to compare notes. #MachineLearning #LLM #DiffusionModels

Here's an interactive pipeline visualization — click any stage (or the input nodes on the sides) to see a detailed explanation in the panel below. A few things to notice in the diagram: The fragment shader (amber, center) is the focal stage — it runs in parallel for every fragment the rasterizer produces. Uniforms (coral, left) and Varyings (pink, right) are its two input streams — global GPU data and per-fragment interpolated data respectively. The red dashed discard path shows how a fragment shader can kill a fragment entirely, bypassing blending — but at the cost of disabling early-Z optimization. The depth test sits after the shader in the standard pipeline, though GPU drivers often promote it to before the shader (early-Z) when the shader doesn't write depth or call discard.

The TTRPG "Manual Entry Tax" is officially dead. To the publishers at Chaosium, Modiphius, and Monte Cook Games: Your legacy back-stock isn't a "portfolio"—it’s a goldmine sitting behind a wall of manual labor. If you’re still paying humans to transcribe stat blocks from 20-year-old PDFs, you're hemorrhaging ROI. GM-Copilot v17.8.059-Universal has arrived. Our Module Ripper and Universal Translator don't just "read" PDFs; they dismantle them. We use a Swarm Architecture to extract lore, mechanics, and encounters into deterministic, system-agnostic data. Roll20 and Alchemy RPG: Your marketplaces should be flooded with content. The bottleneck isn't the creators—it’s the conversion. We are the bridge. The era of "printing money" from your archives has begun. Who wants the demo? #TTRPG #BusinessDevelopment #Automation #GMCopilot #VTT #GameDev

Over the last year I’ve been building something that started inside real-world infrastructure work through SlabWorx and evolved into a much bigger concept around machine-readable environments and infrastructure intelligence. I’ve been spending a lot more time documenting and refining the patent and system architecture in NotebookLM and internal R&D workflows, and I wanted to share a small high-level explainer. This is not just another ‘AI app.’ The focus is turning real-world conditions, visuals, measurements, documentation, and field data into structured operational intelligence that can actually support decision-making in the physical world. Right now we’re actively exploring strategic conversations with investors, VCs, PE groups, operators, and people interested in infrastructure, AI, robotics, spatial computing, and physical-world intelligence systems. This is still early, but the vision is very large. Small explainer and overview below. More information: patent.barichholdings.com

Post processing motion capture. It's important in understanding how to maintain the best quality of the raw data, without destroying its fidelity. In this video, i focused on the post processing tools in CapturyLive. The Adaptive and Butterworth Filters.

Reverse engineering is evolving beyond manual reconstruction. We are developing an AI-driven tool focused on transforming scan data and point clouds into intelligent 3D reconstruction workflows. The ability to extract usable structure from complex spatial data opens new possibilities for reconstruction, analysis, and digital asset generation at scale. Point clouds are no longer just raw data — they’re becoming interpretable geometry. Coming soon to the FutureSnap platform #ReverseEngineering #PointCloud #3DScanning #3DReconstruction #SpatialComputing