First screen, captured 3 days ago. Open the video.

Amazon saves millions by replacing Sonnet 4.6 with Strands Decider

What HN thinks

The field completes another lap around the track and rediscovers the SVM, this time with a bigger invoice to compare against.

LLM classifiers are wasteful overkill
Many argue a classic classifier (SVM, Naive Bayes, BERT) would have been far cheaper and faster from the start, and that putting an LLM in every inference call is hype-driven.
LLMs are great for prototyping first
Others say LLMs let you ship a working classifier in minutes with no labelled data, then distil or replace it with a cheap model once usage justifies the effort, which is what Amazon did.
Specialist vs general models debate
A long subthread argues whether small specialised models beat large general ones, citing the Bitter Lesson, MoE, and the value of world knowledge.

Consensus: Most agree that at high volume a small dedicated model beats a frontier LLM on cost, and that using LLMs to generate labels or bootstrap a cheaper classifier is a sensible pattern. Many also note the post reads partly like an AWS advert.

Recommended in the comments

scikit-learn (library): Suggested as the obvious cheap way to classify tickets into a few dozen groups.
BERT (library): Commenters say a fine-tuned BERT-style encoder does this faster with less memory.
Gemma (library): One commenter replaced LLM pipelines with Gemma embeddings plus a classifier head.
Strands Agents ↗ (library): Linked as the clearer explanation of how the decision model was built.
Bonsai (library): One commenter used it with cheap models to orchestrate fast, controllable logic.

Summarised by Claude from the comment thread 3 days ago.

Points: 673
Comments: 233
Best rank: #1
Posted: 2026-11-10 by aloukissas
Category: AI & ML

strands-decider claude-sonnet decision-models aws-sagemaker ticket-classification llm-cost-optimization amazon

All captures:

Full page

Captured 3 days ago. Click for the whole screenshot.

Looks similar

Screenshot of naut.ca: The Hackintosh Reborn

The Hackintosh Reborn (naut.ca)

385 points by rubatuga 8 days ago | 223 comments

Blog post arguing AI coding tools like Claude Code are reviving the Hackintosh community, citing macOS ports to Snapdragon, ARM SBCs, and Nvidia GPU support.

hackintosh macos claude-code arm

Screenshot of linebender.org: JIT bitmap text rendering for microcontrollers

JIT bitmap text rendering for microcontrollers (linebender.org)

102 points by dmit 5 days ago | 22 comments

Raph Levien's Linebender blog post describing a JIT technique for fast proportional bitmap text rendering on the RP2350 microcontroller in his vic-2350 project.

rp2350 microcontrollers jit bitmap-fonts

JIT bitmap text rendering for microcontrollers (linebender.org)

5 points by raphlinus 6 days ago | 1 comments

Raph Levien's Linebender blog post describing a JIT technique for fast proportional bitmap text rendering on the RP2350 microcontroller for his vic-2350 project.

jit-compilation bitmap-fonts text-rendering rp2350

Screenshot of arstechnica.com: Amazon plans to charge for Alexa in June–unless internal conflict delays revamp

Amazon plans to charge for Alexa in June–unless internal conflict delays revamp (arstechnica.com)

34 points by goplayoutside 2 years ago | 39 comments

Screenshot of jonaserthal.com: Comparing Jensen to Steve Jobs Is Stupid

Comparing Jensen to Steve Jobs Is Stupid (jonaserthal.com)

3 points by apollopower 9 hours ago | 1 comments

Blog post arguing that comparing Nvidia's ouster of Jensen Huang to Apple's 1985 removal of Steve Jobs is wrong, recounting Jobs' redemption at NeXT.

steve-jobs jensen-huang nvidia apple

Screenshot of restate.dev: Durable Coroutines for TypeScript

Durable Coroutines for TypeScript (restate.dev)

8 points by calvinfo 9 days ago | 0 comments

Restate blog post introducing Durable Coroutines in its TypeScript SDK, using generators and a deterministic scheduler for low-overhead fault-tolerant agent workflows.

restate typescript durable-execution coroutines