frontpage.

The question is less about the classic MoE/Quantization debate, but more about the trained behaviors for a model.

It feels like the most ideal model for an LLM provider is the one where user have to follow up with another prompt to clarify or improve the result and make it behave like this maybe 50% of the time to avoid excessive churn.

Someone Has to Fly the Plane

TimeCapsuleLLM: LLM trained only on data from 1800-1875

Show HN: Reversing YouTube's "Most Replayed" Graph

Playwriter: Browser extension MCP uses 90% fewer tokens than Playwright

Show HN: AI Motion Control – Transfer any motion to any character with Kling AI

VeilNet: A Different Approach to Overlay Networks for Dynamic Infrastructure

I put Economic rules within silicon

Apple Warning–iPhones Must Now Restart

New York begins cargo drone trial between Manhattan and Brooklyn

Show HN: ETLR – An alternative to Zapier/Make/n8n for developers (YAML-based)

Dead shrimp. Living machines. Necrobotics is here

Founding AE/GTM Role

Are you hiring designers & engineers?

LLM remembers every past conversation (no embeddings, no RAG) [video]

Things that Janet and Clojure do better than each other

Show HN: Convert bank statement PDFs to Tables

Streamer Spend to Top $100B for First Time in 2026 – Report

Interactive SHA-256 Visualizer

Keyword-Match Domains Still Matter for SEO

Python UCP Client (Universal Commerce Protocol)

Show HN: Create LLM-optimized random identifiers

Influencers and OnlyFans models are dominating O-1 visa requests

Building a No-Tracking Newsletter

Betterment Unauthorized Access

Show HN: From Weekend Hack to the Serial Console Adapter We Enjoy Using

The Middle Binomial Coefficient

Combining In-Shuffles and Out-Shuffles

Resilience vs. Fault Tolerance

Toward single-cell control: noise-robust perfect biomolecular adaptation

Windows 2000 still earning its keep running a rail ticket machine in Portugal

Ask HN: Are LLM providers making LLMs worse on purpose?