HN user

moatmoat

185 karma
Posts21
Comments3
View on HN
github.com 4mo ago

From Any ePub to Audiobook

moatmoat
2pts0
amathislab.github.io 5mo ago

Unlocking full-body human motor learning at scale

moatmoat
3pts0
inkslop.github.io 5mo ago

Testing Frontier Vision-Language Models on Mazes and Handwriting. They Failed

moatmoat
1pts0
github.com 9mo ago

Modern Lab Page Template

moatmoat
1pts0
www.anthropic.com 10mo ago

A postmortem of three recent issues

moatmoat
381pts116
actu.epfl.ch 1y ago

AI monitors wildlife behavior in the Swiss Alps

moatmoat
1pts0
dl.acm.org 1y ago

Live Coding in Classrooms: Why It's a Cognitive Nightmare for Instructors

moatmoat
2pts1
github.com 1y ago

A multi-view video behavior monitoring dataset of wild mammals in the Swiss Alps

moatmoat
1pts0
xiaotiansu.github.io 1y ago

Why Users Crave Control in the Age of AI

moatmoat
1pts0
github.com 1y ago

Open-source Notion alternative: Affine

moatmoat
2pts1
github.com 1y ago

[Google Research] Handwriting Conversion with Vision Language Model

moatmoat
1pts0
github.com 1y ago

Evaluating and Training Multi-Modal Large Language Models for Action Recognition

moatmoat
1pts0
github.com 1y ago

Modular Quantization Aware Training

moatmoat
1pts0
xiaotiansu.github.io 1y ago

Scalable Feedback Through AI and Visualization

moatmoat
1pts0
actu.epfl.ch 1y ago

Morphing robot turns challenging terrain to its advantage

moatmoat
40pts2
ethz.ch 1y ago

Storing CO2 Underground in Switzerland

moatmoat
1pts0
charlieleee.github.io 1y ago

InkSight: Offline-to-Online Handwriting Conversion by Learning to Read and Write

moatmoat
3pts0
www.theverge.com 1y ago

Elon Musk just offered to buy OpenAI for $97.4B

moatmoat
1pts1
actu.epfl.ch 1y ago

Bat wings boost hovering efficiency

moatmoat
3pts0
actu.epfl.ch 1y ago

Choosing the best medical treatment with the help of AI

moatmoat
3pts0
exbody2.github.io 1y ago

Advanced Expressive Humanoid Whole-Body Control

moatmoat
94pts25

TL;DR — Anthropic Postmortem of Three Recent Issues

In Aug–Sep 2025, Claude users saw degraded output quality due to infrastructure bugs, not intentional changes.

The Three Issues 1. *Context window routing error* - Short-context requests sometimes routed to long-context servers.

   - Started small, worsened after load-balancing changes.
2. *Output corruption* - TPU misconfigurations led to weird outputs (wrong language, syntax errors).
   - Runtime optimizations wrongly boosted improbable tokens.
3. *Approximate top-k miscompilation* - A compiler bug in TPU/XLA stack corrupted token probability selection.
   - Occasionally dropped the true top token.
Why It Was Hard to Detect - Bugs were subtle, intermittent, and platform-dependent.

- Benchmarks missed these degradations.

- Privacy/safety rules limited access to real user data for debugging.

Fixes and Next Steps - More sensitive, continuous evals on production.

- Better tools to debug user feedback safely.

- Stronger validation of routing, output correctness, and token-selection.

New iMac with M4 2 years ago

Apple giving us 16GB base RAM in the iMac is like finally getting a decent cup of coffee after years of instant