Introducing PROWL AI Model Improvement Tool

⚙️ [Technical Update] Octal 007 – Dual‑Stage Fine‑Tuning Strategy We implemented a dual‑stage approach to teach Lean 4 grammar and then specialise on real game invariants. Stage 1 (Grammar mastery): 5,000 synthetic examples. Model: Qwen2.5‑Coder‑32B with 4‑bit QLoRA (rank 64). The MI300X sustains 94% GPU utilisation. Loss is decreasing, confirming the model is learning the formal language structure. Stage 2 (Specialisation) will use 1,000+ examples derived from actual Treblecross Grundy numbers, focusing on the discovered invariant (period 34, offset 52). This two‑stage method prevents catastrophic forgetting. #AMDhackathon #ROCm #FineTuning #Qwen #Lean4 @lablab.ai @AMD Developer

Building something I wish existed when I started learning quantum computing. Qiskit Intuition is an open-source interactive lab that teaches quantum computing the way physics should be taught — through visual intuition, not equations first. Drop a gate. Watch the Bloch sphere move. Ask the AI why. It combines a drag-and-drop circuit composer, real-time 3D state visualization, a multi-agent AI tutor (A.C.E.), and a live Qiskit sandbox — all in one place. Still very much a work in progress, but the architecture is taking shape. The latest update adds parameterized gates, an intent-routing AI chat, preset experiments, and a full 6-level curriculum from Python basics to IBM hardware execution. Would love feedback from anyone in quantum education or QC research. 🔗 https://lnkd.in/grQuKrJA #QuantumComputing #Qiskit #OpenSource #MachineLearning #Physics #Quantum

Incredible work by Hoorain Saud bringing swarm intelligence to life with the Nano Swarm Ant Simulator! Ahmad Hassan Afridi, Ph.D’s guidance clearly shows how hands-on learning transforms understanding turning algorithms from theory into a visual, dynamic reality. Joseph Wehbe ksaai.daimlas.com

Mathematical induction is a powerful proof technique used to establish the truth of an infinite number of statements. This video explains the concept in detail, ranging from conceptual overviews to step-by-step practice problems. #MathematicalInduction, #STEM, #ProofTechniques, #Calculus, #DiscreteMath, #Education, #MathHelp, #IBMath, #AdvancedMath, #Logic, https://lnkd.in/gZZw-hZe

🔴 Seed IQ is now at 10/10 games solved on ARC-AGI 3 🥳🙌🏻 This week we’ve had a lot of people suggesting that our posts are representative of our own report/interpretation of scores/performance and that they are somehow “not official.” We’ve also had accusations of “faking it.” ➡️ Make no mistake, these LIVE Scorecards ARE the OFFICIAL evaluation validated by ARC Prize, themselves, of Seed IQ’s performance. The scorecards sit on the ARC Prize website, generated by them, not us. These details are served up from their end recording & evaluating all of the details of game performance on every level of every game Seed IQ plays. They even include replays of every level. 🔸 It doesn’t get more official than this.🔸 ▪️The only thing that is not happening for us it placing Seed IQ on the leaderboard. And that is due to the fact that the ARC Prize rules state that you have to turn over your entire codebase & commercial rights to your system in order to be recognized as a contender on the leaderboard (officially entering the contest portion of the benchmark). ▪️We asked for a private evaluation, we offered to forgo prize money, and Greg Kamradt told us that option wasn’t available at this time. ▪️Yet, they clearly do it for the frontier models. Last week they evaluated both ChatGPT 5.5 (scored 0.43%) and Claude Opus 4.7 (score 0.18%), and he gave a detailed report of what they observed of those models performance on the backend. ▪️After I posted about our 5th game win, Greg commented on X about the steps he observed on the backend of our play, and he asked me what priors we are using. ➡️ They see everything we are doing. They are giving us our OFFICIAL SCORES. (If this was something you could fake, why don’t you see anyone else posting scores like this? Why wouldn’t the ARC Prize folks be calling us out for cheating? I’ve seen them call out people for spreading misinformation about the contest.) You would think they would acknowledge Seed IQ’s performance publicly, the same way they do frontier models who clearly aren’t turning over their codebase either, especially because we are the only system acing these challenges and crushing this benchmark. ▪️ARC Prize has positioned themselves as an entity to evaluate the best of AI. They have made it clear in the past that they do not believe DL/RL has any ability to adapt or to reason, plan, and act across novel environments. ARC-AGI 3 was positioned as an effort to spotlight advanced systems who actually can do that, and yet proprietary systems are being ignored while the entire benchmark is catering to DL/RL systems who cannot even score 1% on the challenges. It begs a much deeper question about the real objective of this benchmark. 🤷🏻♀️ ✅ Either way, we’ll keep letting Seed IQ play their games because regardless of the leaderboard, the benchmark is still acting as an official evaluation and validation of its performance. 🥳🚀 LIVE Scorecard for 10/10 games in comments… #AIX #SeedIQ

I really wanted to enroll but couldn't in the White Room of Classroom of the Elite as it does not exists in real life well maybe governments keep such training facilities a secret who knows. So I fine-tuned an LLM to talk like someone who came from it and was considered the best. None other than our poker faced Ayanokoji Kiyotaka So I can get to ask question and lessons on psychology, human behavior, emotional control, social dynamics, and strategic thinking through the lens of a calm, analytical persona etc. I worked on a personal fun project: Turning an 8B instruct model into an Ayanokoji Kiyotaka from the classroom of the Elite. How I built it: • Generated a synthetic persona dataset using Claude based on Ayanokoji Persona. • Created 1000 SFT examples to teach the base behavior • Created 200 DPO preference pairs to sharpen the response style • Fine-tuned Unsloth Llama 3.1 8B Instruct • Supervised Fine-tuning(SFT) LoRA with 4-bit QLoRA. Then afterwards Direct Preference optimization(DPO) on top of the SFT model. • Trained it on an NVIDIA L40S GPU Model is uploaded to the HuggingFace and can be downloaded and used. Link in comments

What if everything you think is solid... is actually fuzzy? Modern science suggests reality is fuzzy. Our best models struggle with "what is." Imagine a low-quality video game. The core code is missing. We only see the picture. - Current math has unexplained gaps. These gaps show our understanding is incomplete. - David Bohm offered a different view. His ideas try to make reality solid again in the math. - Testing this is tough. It moves physics from ideas to real engineering tests. Your basic ideas about system stability might need a rethink. Do not build on weak ground. If old physics fails, what new basic tools must we invent? 🤔 #QuantumComputing #TheoreticalPhysics #CTO #DeepTech #BohmianMechanics

While building a collision engine and optimizing it with QuadTrees, I began experimenting with random attraction and repulsion forces between particles. I noticed that these simple interactions caused the particles to self-organize into cell-like structures and organic patterns. This was my introduction to Particle Life simulators. It is incredible how a small set of parameters and a random distribution of particles can result in systems that are chaotic, stable, or simply beautiful. The boundary between mathematical logic and lifelike behavior is thinner than one might think. I am still exploring simulations like these and the unique patterns they produce. Here are some of the remarkable patterns that I found. Special thanks to the developers behind the second and third simulations for their excellent work on this concept! https://lnkd.in/dn-ddqer https://lnkd.in/d3bYc5Xp #Programming #Simulation #Coding #Algorithms #Emergence #ComputerScience

Every click tells a story. But how do we turn raw game telemetry into valid, scientific evidence of a student's cognition? In the compelling new case study, "Using Learner-System Interactions as Evidence of Student Learning and Performance: Validity Issues, Examples, and Challenges," researchers Greg Chung, Tianying Feng, and Elizabeth Redman tackle this exact question. They explore the groundbreaking potential of "measurement without testing." By treating fine-grained learner-system interactions—every click, drag, and pause—as atomic units of observation, educators can gather valid evidence of cognitive processes in real time. A massive thank you to Chung, Feng, and Redman for showing us the way! Transforming low-level behavioral events into high-level indicators of learning is the holy grail of educational technology. Their work proves that game mechanics designed to teach can simultaneously serve as rigorous, scientific observations of human problem-solving, provided there is careful algorithm development and robust validation in place. Speaking of pioneering this frontier, we are absolutely thrilled to congratulate co-author Teanna Feng on her new role as Assistant Professor in the Division of Games at the University of Utah! Her vital research in psychometrics, process modeling, and technology-based measurement will undoubtedly continue to shape the future of the field. The intersection of game design and measurement science just gained an absolute powerhouse. Please join us in a well-deserved round of applause for Teanna as she steps into this exciting new chapter! Read all about her appointment here: https://lnkd.in/gdSCd7Am What are your thoughts on the shift toward "measurement without testing"? Let us know in the comments!