Adam Skye Jones

From linkBryan Cantrill recalls a fast-growing OpenSolaris customer that wanted to buy Sun hardware in 2005. Sun had won the software strategy but failed to answer the customer, and then tried to sell the wrong product. Dell responded overnight, helped arrange pricing and leasing, and delivered servers within two weeks. Cantrill argues that Sun had grown bored with the ordinary work of selling and serving customers, a weakness that good engineering could not overcome.

From commentsReaders who bought hardware in that era echoed the frustration with Sun’s slow quotes and difficult sales process, while others said Dell has since acquired similar problems. Former Sun employees also pointed to internal rivalries, expensive hardware, and reluctance to adapt to commodity x86 and Linux. The discussion admired Sun’s technical legacy but treated operations and market shifts as intertwined reasons for its decline.

From linkFederico Viticci tests an M5 Ultra Mac Studio with 256 GB of unified memory against an M3 Ultra and an RTX 5090 PC. He reports much faster prompt processing and response generation than the prior Mac, making long-running local assistants feel responsive in his daily research workflow. The article also acknowledges the machine’s high upfront cost and the Nvidia system’s speed advantage in some benchmarks; privacy, quiet operation, and avoiding ongoing cloud bills motivate his preference.

From commentsHN readers focused on the benchmark details, especially how performance changes as context grows. Some praised the leap in prompt processing and the ability to keep large models in unified memory. Others questioned the RTX 5090 configuration and said optimized Nvidia inference can be substantially faster, while readers debated which model sizes and mixtures of experts make the comparison useful.

From linkPaul Bakker argues that drafting prose is a way to discover and test ideas, so outsourcing a document to AI can skip the very thinking the writer needs to do. He uses a strategy document as an example: a shorter, carefully reasoned piece may be more valuable than a polished but unfocused long one. He still uses AI for coding, research, data collection, and limited document help, drawing the line at letting it form the writer’s argument.

From commentsCommenters debated where assistance turns into substitution. Some agreed that research and proofreading can help while having a model structure an argument weakens the author’s own thinking. Others compared AI to earlier writing tools that made revision easier and argued that careful human direction can still produce useful prose. The thread repeatedly returned to whether the writer has worked through the ideas rather than merely accepted fluent output.

From linkThe mini-AGI repository describes an experimental byte-level language model trained from scratch on a single GPU with 8 GB of VRAM. It keeps most weights on disk, loads a working set onto the GPU, and grows or prunes expert capacity as it reads a continuous stream of text. Its author calls the current result a small toy-level model, says the first training pass is still underway, and has not yet published weights. The project is a reproducible attempt at personal continual learning rather than a claim of frontier capability.

From commentsHN readers welcomed the accessible hardware target and open code, but questioned whether the design really avoids catastrophic forgetting. One commenter noted that the shared trunk still changes and that pruning experts may discard learned knowledge. Others were curious about training from a stream and about keeping knowledge on disk. The author openly described using AI during development, which drew discussion of both experimentation and the gap between an interesting method and demonstrated capability.

From linkA community project is reviving Commodore’s Amiga Unix, or Amix, for 68040 and 68060 systems and modern accelerator cards. Its site shows a floppy-and-CD installer, a package client with a hosted repository, drivers for SCSI and Ethernet hardware, a cross toolchain, and an OpenLook desktop. Some pieces are working on real machines while graphics, packages, and installer polish remain in progress. The project documents its experiments in public.

From commentsHN reaction mixed delight in the technical resurrection with debate over how the project is presented. Some readers enjoyed seeing a historically unusual Unix system gain a modern installer and packages. Others found the site’s prose overly synthetic and wondered how much of the implementation had been reviewed by its authors. Several defended the value of an open, playful experiment while asking for clearer evidence of what works on physical Amiga hardware.

From linkA Consumer Rights Wiki page examines Disney+ language that permits some promotional material even on subscriptions sold as ad-free. It points to exceptions for live programming with embedded advertisements and for Disney’s own promotions, and discusses how those exceptions may appear to subscribers. The page itself carries an incomplete-sourcing warning, so its broadest interpretation should be treated cautiously. The concrete issue is how the service defines advertising and explains exceptions in its terms.

From commentsHN readers disagreed over the size of the change. Some argued the linked page stretches narrow exceptions for live content and in-app promotion into a claim that ordinary films will get pre-roll ads. Others said an ad-free subscription should also exclude promotions for the service’s own products and tiers. The thread broadened into complaints about similar definitions at other streaming services and called for clearer, plainer promises to customers.