129 pieces: some by me, some written by an AI. Each one says which.
129 items · Page 6 of 11
Analysis · AI
AI labs have scraped the entire internet and are rapidly running out of human text. Here is why training next-gen LLMs on synthetic data triggers an irreversible degradation called Model Collapse.
Analysis · Tech
For sixty years, we have wired chips from the top down. Here is why moving the power grid to the back of the silicon is the most radical architectural leap in modern semiconductor history.
Analysis · AI
The Open Source Initiative just finalized its AI definition, and Meta's Llama didn't make the cut. Here is why hiding training data is sparking a war over the term 'open source'.
Analysis · Tech
Nvidia's real AI monopoly isn't silicon—it's the proprietary NVLink network connecting it all. Here is how the Ultra Ethernet Consortium plans to break the lock-in.
Analysis · AI
Nvidia rules AI training, but Groq's deterministic Language Processing Units prove generating tokens doesn't require complex GPUs. Here is why SRAM changes the AI hardware economics forever.
Analysis · Tech
ASML's new High-NA EUV machines cost $350 million and require three Boeing 747s to ship. Here is why the sub-2nm chip era depends entirely on them.
Analysis · Tech
The FDA just approved Apple's AirPods Pro 2 as clinical-grade, over-the-counter hearing aids. Here is why the traditional medical hardware monopoly is terrified.
Analysis · Tech
WebGPU is finally handing the browser direct, low-level access to your graphics card. Here is why zero-install AI inference and AAA web gaming are about to kill the native app.
Analysis · Tech
For years, Apple used cryptographic software locks to prevent independent hardware repairs. Here is how new legislation finally forced them to kill 'parts pairing' and surrender the repair monopoly.
Analysis · Tech
Organic substrates are warping under the thermal stress of massive AI chips. Here is why the semiconductor industry is betting its future on microscopic sheets of glass.
Analysis · Tech
For 40 years, RAM has been physically locked to the CPU. Compute Express Link (CXL) is finally breaking that bond, allowing data centers to pool memory across entire racks and end the stranded RAM crisis.
Analysis · AI
Gemini 1.5 Pro can ingest 2 million tokens at once. Here is why the era of complex vector databases and chunking might be coming to an end.