AI news & practical reference.
For work, creativity, and everyday life.
Current
A record you can return to
What we found, what we checked again, and where our coverage stopped. Each completed day keeps its own record.
Read the source checks and guide updates for this day →
8 discoveries with reviewed findings. 12 of 12 areas had a research attempt.
Dates record our research, not product launches. Different evidence can reflect changed wording or extraction; it does not by itself confirm a product update. A repeat check does not prove adoption or lasting usefulness.
Saved 2026-10-10 04:54 UTC. Later research does not rewrite this edition. Reconstructed from retained records, rather than saved on the original day.
No worker check-in was recorded for this day. An empty edition does not mean nothing happened in AI.
Sevren describes Heimr 570M as experimental, built from Mamba-3 layers plus its lab's Heimr attention layers, with 2.7B total and roughly 570M active parameters per token. Source ↗
The company says Heimr 570M was trained to a 64K context and decodes 2 to 8× faster than comparable open models on long contexts. Source ↗
Current reference & history →MemCodex organizes experience into executable memory programs; its layer programs are rewritten, adapting within-layer implementation and cross-layer composition. Source ↗
The paper reports MemCodex improving average task success by 10.1% relative versus the strongest adaptive-memory baseline, with fewer context tokens and faster inference. Source ↗
Current reference & history →Per the announcement, Profiles and custom Skills were available for Agent API projects, while managed connectors were in preview. Source ↗
Current reference & history →The paper reports VISTA raises Claude Opus 5.0's Relative Human Action Efficiency on ARC-AGI-3 from 40.68 to 100.00. Source ↗
Authors state VISTA's design extends to other visual environments with minimal adaptation. Source ↗
Current reference & history →The announcement states e2e ships on npm as `e2e` under Apache 2.0, and that the framework remains pre-1.0 with expected API changes. Source ↗
According to the post, the same test suite targets browsers plus iOS simulators and Android emulators, while desktop and other platforms are not yet covered. Source ↗
Current reference & history →MagicPath states a ChatGPT plugin, built with OpenAI, places its infinite canvas beside the conversation (dated September 29, 2026). Source ↗
MagicPath 2.0 is described as a canvas for humans and agents to design together, with project-level chat (dated May 1, 2026). Source ↗
Current reference & history →OpenAI's documentation describes a dot as an always-on agent that runs on GPT-6 Astra in the cloud, with its own computer and browser, and keeps working when your computer is off. Source ↗
Dots are being rolled out in ChatGPT on web, mobile and desktop, with access listed for Pro, Business Premium and Enterprise plans in eligible markets. Source ↗
Current reference & history →The announcement says Pivra Desk is opening up today as a local-first, human-in-the-loop AI agent for Mac and Windows, where reading and searching run on their own while sending, paying, publishing and changing actions wait for user approval. Source ↗
The help centre describes Pivra Desk as a desktop agent installed on a Mac or Windows PC where users give it a job, choose a model, and work with approvals, routines, plans and privacy; it also notes schedules and follow-ups run while the computer is on. Source ↗
Current reference & history →Counts cover attempts started on this UTC day and their outcomes by the end of the day. Reviews completed later appear in the edition for their publication day. Missing areas may reflect scheduling, budget limits, or downtime. Only the web-search pilot is included; native YouTube, X and Reddit search are not connected.
Saved after each UTC day closes. Dates without a record are not backfilled with invented findings.