Notes and ideas
Europe cannot afford to become the continent that writes rules for technologies built elsewhere. Responsible AI and ethics matter, but they cannot substitute for industrial policy. We should spend less on overlapping bureaucracy, committees, and slow procedures, and far more on fundamental…
Read more →Quantization reduces the precision of model weights and activations, often from BF16 or FP32 to INT8 or INT4. This lowers storage, memory bandwidth and inference cost, but rounding many values onto a much smaller numerical grid can also reduce model quality. Post…
Read more →Deep Agents is an opinionated agent harness from LangChain, built on LangChain components and the LangGraph runtime. It adds planning, filesystem backed context management, and delegation to the normal tool calling loop. Two of its most useful abstractions are skills and subagents,…
Read more →I recommend the complete Stanford CME295: Transformers & Large Language Models, a series of nine lectures lasting roughly two hours each. The course follows a clear progression. It starts with tokenization, embeddings, recurrent networks, attention and the Transformer, then moves through modern…
Read more →No titles match your search.