Seit früher in diesem Jahr habe ich wieder versucht, Deutsch zu lernen. Ich wohne im Mittleren Westen, darin wurde Spanisch wahrscheinlich besser für das tägliche Leben sein. Aber ich habe ein paar Jahre an der Schule Deutsch gelernt. Und ich finde, dass ich ein bisschen Deutsch noch verstehen kann. Nicht viel, natürlich (Hochschule war vor fast einem Jahrzehnt und ich habe seitdem quasi nichts…
Almost a decade ago, I was an intern at certain large tech company, working on a team that no longer exists. I was living on the third floor of a house that lodged around 15 other people who temporarily found themselves in Seattle for one reason or another. The standard introduction included our tech stack, and so it went when one other transient arrived one day: "I work on robotics firmware", "I…
I was talking with a colleague the other day about the current industry trend to focus so much on further quantizing today's LLM models. I'm sure the reasons for pursuing quantization are apparent to most people who work in this industry: going from BF16 to FP8 or NVFP4 nets what amounts to a nearly 2x or 4x reduction in memory requirements and often a similar speedup. But I can't shake the…
Recently, DeepSeek released a new model and paper , titled DeepSeek-OCR. I've seen a bunch of people talk about it recently, and it's pretty cool stuff. I love the idea of using images to compress text. As Andrej points out , the tokenizer is kind of ugly, and it would be cool to have a more natural system. Intuitively, it makes sense that a more optimal way to parse written letters/characters…
Jekyll is great and all, but it sometimes feels a little too polished, bordering on bland. I've been looking at some more bare-bones sites recently, and realizing that I tend to like the feel of these when used for blogs and the like. In particular, I recently ran across Miku's website , and I thought that it seemed great. So I stole it. And then changed the colorscheme. And then added a basic…
Computer architects can't agree on a way to find the average. For years, academic practitioners in this field have been arguing about the appropriate way to summarize the average performance of their designs [1]. That is: given \(n\) workloads, if system \(A\) outperforms system \(B\) by \(S_1, S_2, \ldots, S_n\) on each, how much faster should you say system \(A\) is, on average? I think this…
Not much of a point to this post. The blog should soon speak for itself. I plan on writing mostly about technical topics, but who knows? These things can change. This blog is currently using the Beautiful Jekyll theme, although the appearance may change slightly as I fine-tune.