TinyNews

AI news for engineers: quick to scan, deep when you want to learn more.

25,792 articles · fetched 7:01 PM PDT25,792 articles · last fetched 10/19/2026, 7:01:39 PM PDT

7 PM PDT·3

arXivNian Ran, Sebastian Brulin, Markus Olhofer, Mauricio A. Álvarez, Richard Allmendinger13:05score 15.7

LLM Replaces Hand-Tuned Rules for Planning Optimization Evaluations

Researchers replaced a hand-tuned rule-based planner in optimization with a large language model, which decides what to evaluate next given cost and budget limits. It matches or beats the human-designed method and adapts its choices across different conditions, suggesting LLMs can be flexible planners.

arXivSavan Doshi11:24score 15.6

Medical fine-tuning shifts AI risk rather than removing it

New evaluation scores AI medical answers for dangerous language like dosing advice and overconfidence, not just factual errors. The finding: medically fine-tuned models still carry risk, they just trade one kind for another, so accuracy alone can't measure patient safety.

arXivMeng Shu, Wei He, Hang Zhou, Runke Liu, Kai Han, Yunhe Wang07:34score 15.5

MetaSkill makes AI agent skills run in a fixed order

MetaSkill is a way to wire AI agent skills together so they run in a set order and pass results along, while each skill still decides its own work. It matters because multi-step agent tasks become auditable instead of improvised.

6 PM PDT·1

Hacker Newsbcongdon07:52score 23.8

AI Agents Take Over Production System Monitoring

Software teams are assigning AI agents to watch individual system components around the clock, reading logs and charts like a dedicated engineer. These agents catch subtle regressions that traditional alerts miss, making monitoring cheaper and more precise.

47HN3HN cmts

4 PM PDT·1

OpenAI05:00score 35.0

OpenAI's new models get much better at old languages like COBOL

OpenAI's latest models are now far better at writing and converting aging languages like COBOL, Fortran and Visual Basic, which still run most banking and ATM systems. That helps companies modernize critical software nobody left knows how to maintain, cheaper.

1 PM PDT·2

Hugging Face17:00score 24.1

Hugging Face models now run natively on AWS Trainium chips

A new plugin lets PyTorch and Hugging Face libraries treat AWS Trainium AI chips as a normal device, so standard model code runs unchanged. This removes the old compiler-style workarounds, making cheaper Trainium hardware far easier to use.

Cloudflare AIRyan Skidmore12:12score 23.6

Cloudflare Cuts AI Code Review Costs 68% Through Prompt Caching

Cloudflare cut its AI code review cost from about $2.46 to $0.77 per review by fixing how prompts are cached and reusing shared context across sessions. This shows that tuning inference infrastructure, not just swapping models, can slash AI bills.

2HN

12 PM PDT·quiet

7 AM PDT·1

Hacker Newserikkannike06:00score 27.4

Small language model retrained to strip out AI writing clichés

A developer fine-tuned a small open language model so it stops writing like a chatbot, cutting telltale tics like em dashes and stock phrases to near human levels. It's a proof of concept that the fix works without special prompts or runtime tricks.

145HN108HN cmts

6 AM PDT·1

Hacker Newsisuruwijesiri05:07score 26.9

Free Book Teaches AI Agent Building From Scratch

A 16-chapter online book explains how AI agents work by building a small one in plain Python, treating the model as the brain and the surrounding code as the body. It demystifies agent frameworks by showing each feature fixes a concrete failure.

128HN64HN cmts

4 AM PDT·1

Hacker Newsankitg1222:55score 27.2

Using tmux to run and manage multiple AI coding agents

A guide to running AI coding assistants in tmux, a terminal tool that keeps long-running agents alive after you disconnect and lets scripts type into or read them. It matters because it enables managing several agents at once cheaply.

149HN114HN cmts

12 AM PDT·quiet