Shashank Jain

Codex wrote ABAP knowledge into random Qwen weights. No training loop or gradient descent-direct weight edits. Manual checks: 0/25 → 18/25 (7 recall cases). Inspired by: rikiyat.github.io/blog/weigh… My code: github.com/allthingssecurity…

for Tarak Mehta Ka Ulta Chashma(TMKOC) lovers, Claude created this chess game.

Built a Rube Goldberg machine using Opus 5.5. Every domino, ball and brick is a real physics sim in Blender: 444 objects, real masses, gravity. A ball rolls down a tower, tips 181 dominoes, which grow into a 3.7 m slab, which sets off a pendulum, a loop, a jump through fire, a funnel and a wrecking ball.

Making @typesafeai Jev play the game of Doom. How it works: the game engine feeds Jev exact state (where each monster is, which way to turn, how fast it's closing in). Jev never sees pixels. Every move is Jev's call. Code only turns its answers into button presses.

I connected a fruit fly brain simulation to a 9×9 Sudoku game. The video shows three puzzles with progressively fewer clues: 69, 51 and 36. Beside the board, the gold flashes show recorded spikes from the simulation. The starting point was FlyWire, a Princeton-led collaboration that reconstructed the wiring of an adult female fruit fly brain. Researchers combined electron microscopy, automated reconstruction and extensive human proofreading to map 139,255 neurons and roughly 50 million synapses. A connectome is that wiring diagram: which neurons connect to which others. FlyWire project My processed dataset contains 138,639 neurons and about 15.1 million weighted connections between neuron pairs. Those connections can each represent multiple synapses. I kept the wiring fixed and added a simplified model of neurons accumulating input and firing. Each board position was converted into input signals. The network ran for 40 milliseconds per decision, and saved the resulting spikes. I then trained two small output models using recordings from 120 generated Sudoku puzzles. The first learned to decode which numbers were present from spike counts. The second learned to score possible placements using those responses, grouped by row, column and 3×3 box. It chose the empty cell where its leading prediction had the clearest margin. I built that Sudoku structure into the output layer. Training changed the output models, not the connectome’s connections. During play, the system had no access to the answer, no backtracking and no Sudoku solver. A solver was used beforehand to check that the generated puzzles had unique solutions. All three puzzles were completed correctly. I checked all 87 placements against the saved spike events and trained models. The fly’s walking is animated using NeuroMechFly anatomy. The number choices and firing display come from recorded model activity. Fewer clues is how we defined these stages, rather than a formal difficulty rating. This is a small, engineered experiment. What is interesting here , is being able to start with an openly available brain map, run an experiment and follow the chain from input to recorded activity to a visible decision. Full Credit to the FlyWire Consortium and its many contributors, including teams at Princeton, HHMI Janelia, the Allen Institute and Google Research, and to the NeuroMechFly/FlyGym team for the body model

Using the Princeton-led FlyWire Consortium’s brain map or a fruitfly , I simulated a fly taking a football shot and tested how neural pulses change its path. No training. Navigation and walking controls are coded. A small experiment in connecting mapped brain wiring to movement.

Open AI Astra has some decent robot hand movement skills. Used GPT6 Astra + MuJoCo to make a robot arm write a Python program on paper. With a pen. And a five-fingered hand. The program prints prime numbers.

Asked OpenAI’s GPT-6 Astra to recreate my photo in Paint using computer controls, starting from a blank canvas. Here’s the 58-second timelapse with music. The resemblance still needs work, but watching it come together stroke by stroke is pretty fun. #OpenAI #GPT6 #DigitalArt

Built a 60-second badminton animation using GPT-6 in Codex, Blender, MediaPipe for pose tracking, and FFmpeg for the final edit. Started with two match clips and added 3D players, slow-motion replays, and racket and footwork sounds. The hardest part was getting the body movement and shuttle contact to look believable.

Built a 3-minute Chinese New Year walkthrough with GPT-6 in Codex + Blender 🏮 Animated lanterns, flower markets, festive food, red envelopes, lion dance and fireworks, all from a first-person view. Rendered in 1080p at 24fps, with an editable 3D scene.

Got access to Open AI Astra and tried to create Namma Auto simulation with it. I can easily say this is the best I got so far, including physics etc. Open AI rocked with this one.

An Indian village one shotted by Fable 5.1 Medium thinking mode kishanpur-gaon.jain-sm.worke…

Microduck dance

I have just released a new Udemy course: From Autocomplete to Assistant: RL for Beginners. The course explains how a pretrained language model becomes an AI assistant through instruction tuning, reward models, RLHF and reinforcement learning. I cover PPO, GRPO, RLOO, DPO, RLVR and reasoning models, but without starting from equations. The focus is on understanding why each technique was needed, what problem it solves, and how the pieces fit together. It is meant for anyone who wants a clear conceptual understanding of modern LLM post-training before going deeper into papers and implementations. Course link: udemy.com/course/from-autoco…

Left Claude Fable and Kimi to produce an experience of an Indian village. Both one shotted some real good simulations of the same. Not sure which one to choose out of the two. The Chinese labs are closing in fast it seems.

Recursive Self Improvement

Now qwen 3.8 max did this chat.qwen.ai/s/deploy/t_1427…