Aditya Inamdar·8h agoBuild an n8n Agent That Can Explain Why It Refused a RefundWire a chat model to two tools, inspect a wrong decision, and fix it by changing what the model is allowed to pass between them.
Aditya Inamdar·9h agoHow to Spend $4 of AMD’s $100 GPU Credit on a Benchmark Worth KeepingRun your first ROCm experiment on an Instinct MI300X and find out what batching does to vector-search throughput.
Aditya Inamdar·10h agoStable Diffusion Starts With 32 KiB of Noise. Where Does the GPU Memory Go?Follow a single prompt through Stable Diffusion v1, from 77 text positions to 4,096 image positions. Along the way, a few lines of Python…
Aditya Inamdar·1d agoYour Stable Diffusion Autoencoder Can Pass Its Test With the Wrong LatentsFollow one image, one prompt, and one noise prediction through Stable Diffusion v1. Then reproduce a scaling bug that an image round trip…
Aditya Inamdar·1d agoA Diffusion Model Sees the Number 1. Should It Push Left or Right?Build a denoising diffusion probabilistic model from its training target to its sampling loop, then debug it against a distribution whose…
Aditya Inamdar·1d agoHow a Feature With Zero Weight Can Break an 8-Bit Linear LayerInside LabML’s GPT-NeoX implementation of LLM.int8(), with a CPU lab that lets you watch a projection lose its answer and recover it.
Aditya Inamdar·2d agoWe Tracked the Shape of a Neural Network as It Learned. The Shape Didn’t Tell Us Enough.Persistent homology can reveal how neural-network weights reorganize during training. But can those geometric changes distinguish learning…
Aditya Inamdar·2d agoQwen Got 31 Fewer Math Problems Right When Attention Heads Dropped Out in BurstsTwo runs can lose the same planned amount of attention and take different paths through a calculation. A paired experiment on 250 math…
Aditya Inamdar·3d agoThe User Is Testing Your Code. Should the LLM Keep Their Conversation in Memory?An LLM server has to decide which conversation it can afford to forget. Intent-aware caching turns the pause between messages into an…
Aditya Inamdar·3d agoThe AI Changed Its Answer 25.8% of the Time. Nobody Changed the Question.Researchers tested a 421-million-parameter decision model by rearranging the same answer choices. The model often changed its mind…
