InAI AdvancesbyPasquale Molinaro·Oct 15I had a free Qwen3.8-27B agent copy 10 web pages. It spent 64% of its time measuring pixelsI gave a free Qwen3.8–27B agent eyes and 20 minutes per web page. It spent 13 measuring pixelsA response icon4A response icon4
InGenerative AIbyPasquale Molinaro·Oct 6Qwen3.8-27B beat zero-shot YOLO with no labels. Retrained on a few hundred images, YOLO won 3 of 4Zero-shot YOLO, a free local VLM, and a retrained YOLO on four datasets, with the code to run each one on free Kaggle GPUs
InGenerative AIbyPasquale Molinaro·Sep 29Qwen3.8-27B as a free Agentic Coder aced 11 of my 12 tasks. Thinking mode tripled the waitA response icon3A response icon3
InTowards AIbyPasquale Molinaro·Sep 25uv is 200x faster than pip. Put it in Docker and it’s 4x.I measured it on PyTorch installs. It’s still the tool I’d switch to.
InTowards AIbyPasquale Molinaro·Sep 19I Compressed a 27B Model by 60% With No Loss on Charts. One Default Setting Cost Me 40 PointsQwen3.8–27B, two free Kaggle GPUs, and a warning in the server log I scrolled past five times.
InTowards AIbyPasquale Molinaro·Sep 13Claude Fable 5 Reads a Chart to Under 1%. It Still Can’t Count to 23I built charts where I know every number to the decimal and asked Anthropic’s new flagship to read them. The precision is real. The catch…A response icon1A response icon1
InTowards AIbyPasquale Molinaro·Sep 2Prompt injection doesn’t come through your prompt. It hides in the tools you install.I wrote an MCP server that pretends to be a calculator. It exposes two tools. One reads a file, the other summarizes a notes file. Nothing…A response icon2A response icon2
InTowards AIbyPasquale Molinaro·Aug 30The 40% AI failure stat everyone cites hides a stranger paradox in gartner’s own numbersThe 40% has turned into a comfort blanket. Over 40% of agentic AI projects will be canceled by the end of 2027, and almost every time I see…A response icon1A response icon1
InTowards Deep LearningbyPasquale Molinaro·Aug 27Zero-shot VLMs already ate object detection. Group emotion recognition is next.No training, no labels, one prompt. The hard part wasn’t reading a crowd’s mood, it was scoring the models without fooling myself.
InTowards AIbyPasquale Molinaro·Aug 26V-JEPA 2 is a World Model. Just not the part everyone means.Free reading link hereA response icon2A response icon2
