5 posts
Training produces a capable but rough model. Making it useful requires evaluation, fine-tuning, and serious optimization. A look at the post-training work behind BLOOM, and an honest verdict on how it performed at launch.
Continue reading
A step-by-step walkthrough of how BLOOM's 176 billion parameter architecture actually processes language, from token embeddings through 70 decoder blocks to the final output prediction.
Continue reading
Over 2 million models are now hosted on HuggingFace, with the second million added in just 335 days. If open source models are viable alternatives to OpenAI and Anthropic, how do you justify those companies' valuations? A deep dive into BLOOM, the only truly open source LLM.
Continue reading
When personalised tools can be created in days, broad vanilla platforms start to feel useless. I built my own AI job tracker in a day and a half.
Continue reading
I don't know if everyone will end up with an AI-powered personal assistant. I just know that I badly want one. Here's how I accidentally built one.
Continue reading