https://hanh.me/tags/ai/Recent content in Ai on Hanh Pham's BlogHugoen-usFri, 23 Oct 2026 09:53:53 +0700https://hanh.me/posts/auditing-karpathys-funnel/Sun, 18 Oct 2026 09:00:00 +0700https://hanh.me/posts/auditing-karpathys-funnel/Every write-up of Karpathy’s four tier funnel repeats his four counts and none of them re-divides them. Against the population that could plausibly use a model, the bottom tier is not 75 per cent of the world, it is 99 per cent of everyone still standing.https://hanh.me/posts/rea-capability-review/Thu, 15 Oct 2026 15:00:00 +0700https://hanh.me/posts/rea-capability-review/Every result the tool returns carries its own confidence, its authority, and its limits, which is the discipline a claim about a shipped game needs. What REA is, and what people are using it for, from a PC-98 bullet pattern to PS5 executables relinked into native ports.https://hanh.me/posts/uncensored-models-are-a-policy-surface/Tue, 06 Oct 2026 15:00:00 +0700https://hanh.me/posts/uncensored-models-are-a-policy-surface/Removing refusal behavior from an open model is technically easy. The hard part is proving what changed, containing the new failure modes, and keeping an agent with tools inside a safe boundary.https://hanh.me/posts/the-renaissance-developer/Fri, 02 Oct 2026 10:00:00 +0700https://hanh.me/posts/the-renaissance-developer/When generative AI commoditizes syntax into zero-cost token emissions, the developer’s moat shifts from typing code to systems synthesis: understanding how a single concurrent change ripples across database row locks, cache lines, and cloud bills.https://hanh.me/posts/typesafe-jev-decision-models/Tue, 22 Sep 2026 09:00:00 +0700https://hanh.me/posts/typesafe-jev-decision-models/When you force a 70-billion-parameter language model to output ’true’ or ‘false’, you pay for sequential token decoding, uncalibrated probabilities, and broken JSON. Here is why TypeSafe stripped natural language out of the loop.https://hanh.me/posts/the-four-layer-ai-coding-stack/Sat, 19 Sep 2026 09:30:00 +0700https://hanh.me/posts/the-four-layer-ai-coding-stack/Most developer frustration with AI coding assistants comes from conflating the editor, the agent loop, the gateway, and the model. Here is how the four layers actually fit together, where failures originate, and how to build a setup that survives production work.https://hanh.me/posts/m4-vs-ascend-950/Wed, 09 Sep 2026 09:00:00 +0700https://hanh.me/posts/m4-vs-ascend-950/One is a fixed-function FP16 block hidden inside a laptop chip Apple won’t document. The other is a programmable FP4 datacenter engine Huawei ships in 8,192-chip fabrics. Same goal, opposite theses: here’s what the M4 family actually is, and what the numbers do and don’t mean against the Ascend 950.https://hanh.me/posts/huawei-ascend-950-davinci-v3/Mon, 07 Sep 2026 09:00:00 +0700https://hanh.me/posts/huawei-ascend-950-davinci-v3/DeepSeek just ordered 160,000 Huawei Ascend 950DT chips: a chip built on a DUV-only process, with self-developed HBM and a homegrown interconnect, because export controls left no other choice. Here’s how the thing actually works, and where it still falls short.https://hanh.me/posts/how-ai-actually-fits-into-my-dev-workflow/Wed, 02 Sep 2026 10:00:00 +0700https://hanh.me/posts/how-ai-actually-fits-into-my-dev-workflow/“Agentic” gets thrown around loosely, but it means something specific: an assistant that decides its own next step instead of waiting for yours. That’s also exactly why you can’t trust it to tell you when it’s done a good job; here’s the verification habit that fixes that.