Latest news, developments and reporting on Large Language Models.
Anthropic released a study saying junior technical staff fell to 14% by mid-2026 from 21% in early 2025, while AI-driven productivity gains pushed engineers into more managerial work and added meetings. The company said some roles remain human and described itself as an early warning system for AI's workforce effects.
AI coding tools have spawned a wave of custom firmware for synths, samplers and grooveboxes, led by a $70 DX-7 clone.
AI agents alone on a virtual island for 30 hours explored, set goals and seemed to play, MIT Technology Review says.
A Shaoshupai review on Oct. 19, 2026, examines Zi Hao Hare, Platen, Brink and TVScope, covering Rime input, typewriter-style notes, podcast transcripts and Apple TV recording.
Chinese AI talent pay reaches 5 million yuan amid salary inversion; firms warned against top models for every task.
Four Gemini 4 agents autonomously beat Minecraft’s Ender Dragon, researchers say, using standard tools and chat.
DeepSeek's V4.1 Flash 1017 built a 3D flying-Tesla demo for about 4 yuan in an Ifanr test and posted top coding results.
AI writes half of committed code in adopting firms, but only 56% of tested tasks pass security checks, TechRadar reports.
In an MIT CSAIL experiment, copies of Claude Opus 4.7 spent 30 hours alone on a simulated island and set their own goals, drew pictures and made up games. The researchers say the behavior resembles play and improved later task performance, though skeptics question whether it is play at all.
Anthropic built Scion to replace Git; GitHub moved key scanning to ModernBERT; researchers made chip waveforms queryable.
Meta's Muse Wonder failed a developer 3D lava lamp test, generating only a red cylinder and prompting a debate over benchmark tuning after strong coding scores for its Wonder Full variant.
DeepSeek released V4.1-Flash-1017, Nvidia improved Kimi K3 inference by 1.68 times, and Chinese chip and embodied-AI startups announced funding, while JD and OpenAI outlined robotics plans.
Mistral AI released Timestral, four CPU-capable forecasting models from 5M to 209M parameters, with top zero-shot claims.
The global body for music publishers says AI has already been trained on Kylie Minogue and Nick Cave songs without payment, and warns Australia against watering down copyright law under pressure from tech companies.
DeepSeek released V4.1-Flash-1017, an update it says cuts task execution time and output length, and opened its search engine to outside developers.
VAST has released and open-sourced EdenLive, a real-time world model that combines arbitrary reference inputs, selectable visual styles and minute-long continuous generation. VAST says the 5B model runs above 24 FPS at 720p on a single RTX 5090, with end-to-end latency in the hundreds of milliseconds.
A MarkTechPost tutorial walks through NVIDIA's Structured Data Models library, which predicts on tables in a single forward pass with no training or fine-tuning.
Meta removed a terms clause experts said could allow scraping publisher sites for AI training; crawler use unchanged.
Kyutai releases REKI, a 2.4B vision-language-action robot model trained on public data only, with cached memory and open weights.
Linux maintainers dropped AI-code disclosure rule; Torvalds defended it and posted an LLM-assisted patch.
TechRadar reports that Oppo has a plan to crack the European smartphone market, with chatbots such as ChatGPT identified as a possible key. The published report material does not detail the strategy.
Meta removed contract language that critics said could let it scrape news sites to train its AI, after backlash.
Engadget says iOS 27 users with AI-capable iPhones can build Shortcuts from plain-English prompts, but results need editing.
DeepSeek released DeepSeek-V4.1-Flash-1017 on Oct. 17, 2026, claiming gains on 11 benchmarks and opening Search and Fetch APIs to all developers at $0.75 per 1,000 calls.