Copyright, Terry Bisson, 1991 Originally published in OMNI, 1991, and featured in HARPER’S and around the internet since. It has even made its way into several books on consciousness and brain science. I’m surprised, pleased, and proud. But please do not reprint, perform, alter or adapt in any way without first checking with the author. Thanks.
Pedestrians wearing headphones are exposed to an increased risk of accidents. In an effort to reduce collisions with cyclists, Škoda Auto, in collaboration with scientists, introduces an innovative bicycle bell whose sound can penetrate even active noise cancellation systems. In doing so, it helps prevent injuries to both pedestrians and cyclists.
In early March, I noticed approximately $180 in unexpected charges to my Anthropic account. I’m a Claude Max subscriber, and between March 3-5, I received 16 separate “Extra Usage” invoices ranging from $10-$13 each, all in quick succession of one another. However, I wasn’t using Claude. I was away from my laptop entirely and was out sailing with my parents back home in San Diego.
00:00:00 Intro00:00:59 The Revenge of Amiga by Reality00:07:22 Noctivagant by Copernicium00:11:04 SQUAPHEREEN by ERNIE00:13:22 Make πrogress not propaganda b...
We present MegaTrain, a memory-centric system that efficiently trains 100B+ parameter large language models at full precision on a single GPU. Unlike traditional GPU-centric systems, MegaTrain stores parameters and optimizer states in host memory (CPU memory) and treats GPUs as transient compute engines. For each layer, we stream parameters in and compute gradients out, minimizing persistent device state. To battle the CPU-GPU bandwidth bottleneck, we adopt two key optimizations. 1) We introduce a pipelined double-buffered execution engine that overlaps parameter prefetching, computation, and gradient offloading across multiple CUDA streams, enabling continuous GPU execution. 2) We replace persistent autograd graphs with stateless layer templates, binding weights dynamically as they stream in, eliminating persistent graph metadata while providing flexibility in scheduling. On a single H200 GPU with 1.5TB host memory, MegaTrain reliably trains models up to 120B parameters. It also achieves 1.84$\times$ the training throughput of DeepSpeed ZeRO-3 with CPU offloading when training 14B models. MegaTrain also enables 7B model training with 512k token context on a single GH200.
Constructing a skyscraper is a massive undertaking. You need architectural blueprints, council permits, and safety audits before the first piece of steel is even ordered. It requires hundreds of people coordinating over months or years. You can’t just throw up some drywall and hope the building holds weight. Then there is the backyard shed. No blueprints, no permits, no audits. You just grab some timber, a saw, and start hammering. It might be a little drafty, and the roof might leak if it rains too hard, but you built it yourself in a single weekend.
I spent years building one of the most popular LED strip visualizers on GitHub. It's the kind of project that's easy to get started but diabolically hard to master.
C# 15 introduces union types — declare a closed set of case types with implicit conversions and exhaustive pattern matching. Try unions in preview today and see the broader exhaustiveness roadmap.
Karpathy recently posted about using LLMs to build personal knowledge bases — collecting raw sources into a directory, having an LLM "compile" them into a wiki of interlinked markdown …