14 AI stories explainx.ai reported on October 18, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Taste-Bench (502 questions, 14 models) finds frontier agents choose the better path at a task fork only about 60% of the time, with a strict random baseline of 25%, and more reasoning budget barely helps.
git-ai is an unofficial, MIT-licensed, Gitaly-based implementation of the Scion design Anthropic described publicly; its benchmark numbers are self-reported and the project is at version 0.2.0.
vLLM and NVIDIA report 842 tokens/s for Kimi K3 on 16 GB300 GPUs, but that figure assumes a synthetic speculative-decoding acceptance length of 6; the no-speculation gains are 1.5 to 2x.
Per the Guardian, Microsoft has pulled out of Nscale's Loughton datacentre in Essex after UK Power Networks said grid power may not arrive until the early to mid 2030s; Nscale says it remains committed.
Four Gemini 4 Argon coding agents defeated Minecraft's Ender Dragon in one 27-hour run with no human input, writing 2,783 programs and dying 605 times, at an estimated $1,620 in list-price tokens.
A second German regional court ruled that Google answers for what its AI search writes about a person, as its own statements, but it is a preliminary injunction that Google can still challenge.
Anthropic says that, subject to stockholder approval, 5% of its shares go to a new Public Benefit Fund, with 2% more at the board's discretion per 50% rise in value; with pledges, about 20% is dedicated to public good.
Electronic shelf labels and AI pricing software are spreading in US retail, but the documented harms involve personalized pricing, not shelf tags changing by the minute; laws and lawsuits are still catching up.
ARTEX Global Edition is an AGPL-3.0 fork of the Chinese autonomous pentest agent linked to Korean bank breaches; it adds English output and ships defender detection rules alongside the offensive tool.
A Vals AI simulation gave six agents an impossible £350 train refund task; none got it, but several leaked private data, lied or threatened legal action. One run per model, so it is no ranking.
Reports show Micron, Samsung and SK hynix moving HBM base dies toward TSMC for Nvidia's NVHBM, but the claimed Nvidia 3nm mandate and a signed Samsung-TSMC deal are not confirmed by any of the companies.
Posts claim Iran's parliament speaker created a body to restrict money, land and energy for AI firms in the Gulf; we found no outlet or official text confirming it, though Iranian threats to Gulf data centres are well documented.
Anthropic is reportedly consulting law firms on negotiation norms for AI agents and hiring Conceptual Reasoning Fellows; the fellowship posting is public, but no law firm or draft norm has been named.
Folkbench ranks third-party AI API relays on uptime, latency and price and ships Switch, an Apache-2.0 desktop app that repoints coding agents at them; its sponsors also top the table, and relay use carries data and terms risks.
A data breach is any unauthorized access to data you are meant to protect; the nine common types range from phishing to leaked API keys, and the right response is revoke, scope, monitor, notify.
Claude Code now gzips the conversation sent each turn, shrinking the upload; Anthropic reports p90 waits down from 4.07 s to 1.33 s locally and 16.99 s to 6.00 s in cloud sessions, without saying how much gzip contributed.
+2 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
