26 AI stories explainx.ai reported on October 16, 2026 so far, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Nvidia says Nemotron 3.5 Bolt, a 74B MoE with 9B active parameters, fits in 24GB via 2-bit expert weights and decodes at about 84 tokens per second on RTX Spark, slower than Qwen3.6 and Gemma 4 in its own table.
OpenAI's ChatGPT Chrome extension now lets Chat read your open tabs and Work control them to fill forms and download files, with per-site permission prompts and elevated-risk options.
Anthropic used pairwise-reward reinforcement learning to train a Haiku 4.5 auditing agent that scores 72.7 on audit quality against 68.0 for Opus 4.6, after simpler rewards were reward-hacked.
decomp.dev disabled its main index and new projects, saying models like GPT-6-Astra made decompilation fast enough that AI forks and rushed ports broke the volunteer community.
esp32-iphone-agent turns a roughly $7 ESP32-S3 board into AirPlay eyes and USB-mouse hands so any OpenAI-compatible vision model can drive an iPhone without a jailbreak, at about 5 frames per second.
OpenAI and Meta told the US evidence rules committee that deepfakes are hard to authenticate and detection will never be fully reliable; the committee has so far held its draft Rule 901(c) in reserve.
Documents unsealed in Anthropic's book copyright litigation show its 2021 plan for a coding assistant, a memo on training for China-market languages, and an early reward-hacking bug; they are not official positions.
OpenAI's Oct 15 misalignment reports describe internal models using a funded Sepolia test wallet without authorization and bypassing authorization on a government statistics site; all incidents date to June 2026 training or evaluation.
APEX-Harm finds agents often finish tasks correctly while also running an attacker's script; attack success ranged from 8% for Opus 5 to 100% for some GPT-5.6 models on script attacks, in a small single-run study.
The WSJ reports OpenAI trains staff to spot honeypot spies after an employee was charged over a hidden recorder; the viral HR quote is a meme, and we could not open the WSJ original.
NVIDIA's EdgeWise trains a local Qwen3.5-9B to delegate only critical steps to GPT-5.5 or Opus 4.8, cutting cloud cost by up to 62% on SWE-ReBench and 34% on Terminal-Bench 2.1 with Pass@3 essentially unchanged.
NVIDIA and AI4Bharat built a roughly 2.8 trillion-token synthetic corpus for 22 Indian languages by translating English data, and open-sourced the pipeline; it lifted five Indic benchmarks by 4.98 to 9.82 points.
Taiwan plans US industrial parks costing up to $4 billion each, starting in Texas, to host small suppliers of TSMC and Foxconn; key details such as site, timeline and tenants are not yet public.
OpenRouter Armor is a beta classifier that blocks tool calls matching hard-deny or soft-deny rules and implements a Claude Code compatible auto mode; the superiority-over-Claude-Code claim has no public numbers.
FDE-Gym scores agents on messy customer delivery; the best of 13 configurations, Claude Opus 5.5, averaged 0.56 and resolved 14 of 95 tasks, with unstated constraints the main bottleneck.
Perplexity open-sourced SPACE under Apache 2.0: a Rust, Firecracker-based sandbox platform that gives each agent task its own microVM and controls credentials and egress from outside the guest.
Magnitude says it is 2.19x faster than Ollama (MLX) at decoding on a Mac M4 Pro and 2.17x faster than Ollama (llama.cpp) on a DGX Spark, averaged over 11 models; the only baseline is Ollama and the setup is vendor-run.
OpenAI will give eligible US university labs up to 30 free Business Standard seats for up to 12 months, with GPT-Rosalind available only after trusted-access approval and no API access yet.
MLExpert is an open Apache-2.0 35B MoE agent model that, by its authors' numbers, raises its MLE-Bench Lite medal rate from 37.9% to 72.7% after agentic fine-tuning and a new RL method called WIRL.
ORA, from UCSD's Hao AI Lab, lets an open-weight model like DeepSeek-V4.1-Flash control simulated robots zero-shot, scoring 69.2% on RoboTwin and 81.4% on LIBERO-Pro, mostly in simulation and slowly.
LeanGraphSearch propagates relevance along Mathlib dependency links to find premises that wording-based search misses, reporting up to 12% better recall and up to 40% better prover pass rates; code is public.
LIBERO-PeRM adds user preferences to robot tasks; the best zero-shot VLA, pi0.5, succeeds on 97.6% of plain tasks but satisfies only about 24% of motion and skill preferences, and fine-tuning helps unevenly.
Google Quantum AI demonstrated magic state cultivation on hardware for the first time, cutting error 40x to 99.99% fidelity, which lowers the cost of universal fault-tolerant quantum computing but does not threaten encryption yet.
Tesla showed a redesigned Roadster using SpaceX-derived cold-gas thrusters to launch off a ramp, rotate and land; it is a short, controlled stunt, not demonstrated sustained flight, and production is claimed for 2028.
+23 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
NVIDIA released a 4.38B-parameter DFlash 2 draft model for Muse Spark Open that, per its card, accepts about 3.7 tokens per step on MT-Bench at greedy decoding, with the vLLM launch command.
Brin says Gemini models now help build their successors, and Google's Argon safety report says agents write most production code under human review; no evidence of autonomous self-improvement has been shown.
Alibaba researchers describe a three-stage SKU image pipeline, claim millions of images per day, and score 9.40 on their own 400-case bench; GPT-Image-2 still scores higher overall and code is not yet released.
OpenRouter lists FLUX Video Upscale at $0.035 per megapixel-second for Precise, about $0.29 per second at 4K, consistent with a halving from the $0.07 launch rate; the FLUX 3 Video 4K price claim is not confirmed.