Writing
Writing
Notes on significant moves in AI, and how to build agents, RAG and evaluations that hold up. Every post opens with a short answer, then shows the evidence.
Between two Samsung lines, the eyes post: the shortage is now a payout storyJukan's wordless 👀👀 post sat between Samsung's 2027 shortage line and its buyback remarks: what repriced that morning was the KRW 55–75 trillion return plan.Samsung’s HBM4 sold at “target price” is the memory bill for RubinMid-to-high $4 per gigabit is about $36–40 per GB, or $9,000–10,500 of HBM4 in one 288 GB Rubin GPU. Sold out at target price makes the shortage contractual.Powertech’s $400M panel fab is where Broadcom’s ASIC volume goes when CoWoS is fullTaiwan cleared Powertech's US$400M Singapore panel-level packaging fab on 28 October: the second queue for AI packages when TSMC CoWoS is full, with Broadcom ASIC volume.Samsung’s 60–70% LTA target is a mortgage on 2030 capacitySamsung wants 60–70% of memory on long-term agreements through 2030. The contracts lock volume, not price: a demand schedule with deposits that now justifies the capex.Samsung’s 2028 line is a lead-time statement, not a demand forecastSamsung guides KRW 84.4 trillion of 2026 capex and still calls 2028 tight: money buys cleanrooms and tools on 12–18 month lead times before it buys output.What is DataVendor by HUD, and who pays for your data?DataVendor is HUD's new marketplace where teams license codebases, RL tasksets and raw data to AI labs. Here is how a sale works, what it costs, and what is still unverified.Can you steer sycophancy out of a small model without retraining it?LLM sycophancy, removed with activation steering: it shows up as a linear direction in a small model's hidden states. Subtracting that direction while the model generates raised accuracy on sycophantic prompts from 12.5% to 53.8% on Qwen3-0.6B, with no fine-tuning.
Follow along: RSS feed