Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlay for freeFury BloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

OpenAI posts 722 math papers as Anthropic touts Claude 5 quantum win

AISunday, October 18, 2026· 11 videos

Briefing

Audio player

OpenAI drops 722 math manuscripts

OpenAI said an AI system generated 722 manuscripts spanning 372 result families, including proposed solutions to longstanding open problems. The release, dated October 6, has intensified debate over whether frontier models are crossing from benchmark performance into original high-skill research. OpenAI said some proofs have already been formally verified, while others remain under review. The company framed the average compute per result as roughly three hours of reasoning in ChatGPT Pro, a claim likely to sharpen scrutiny of both methodology and validation.

Anthropic claims quantum circuit breakthrough

Anthropic reported on October 15 that Claude Mythos 5 discovered quantum circuit designs beating published human baselines on seven of eight benchmark primitives. Researchers said the model received plain-language task descriptions, generated candidate circuits in Python, and was scored by an external evaluator using qubit count multiplied by gate count. The standout result was a new modular inversion construction, underscoring how machine-guided search can exploit domains with cheap automated verification. The paper arrived alongside a separate advance from Google Quantum AI on lowering error rates for magic states, giving the field a notable one-two punch.

Anthropic readies Claude Fable 5.5

Evidence is mounting that Anthropic is preparing a release dubbed Claude Fable 5.5, with model names reportedly surfacing in developer tools and Microsoft Azure listings. Early-access labels including Orzada EAP, Vanilla EAP, Baron, and Oniiri suggest several checkpoints or variants are under active evaluation. Interest has centered on demonstrations of unusually strong spatial and world-building behavior, including interactive game-like outputs generated from simple prompts. If confirmed, the launch would raise pressure on rivals by pushing frontier competition beyond chat and into dynamic simulation and code-native media creation.

DeepSeek upgrades V4.1 Flash 1017

DeepSeek released V4.1 Flash 1017, positioning it as a faster, cheaper coding-focused update rather than a wholly new flagship. The company said the model improved from 31.2% to 42.3% on Terminal Bench 4.0, while also shortening outputs and speeding completion. DeepSeek simultaneously broadened access to its search and fetch APIs, pricing both at $0.75 per 1,000 calls and waiving supported server-side tool-call fees until October 31 at 16:00 UTC. The strategy points to a larger push toward agentic software development and end-to-end project execution.

Gemini becomes a desktop workspace

Gemini expanded its desktop app into a fuller agent workspace under Gemini 4 Argon, separating Chat from Task mode. Users can now connect local folders so the system can read, create, and edit files directly, with automatic backups attached to each task as a safeguard. Gemini also supports starting jobs from a phone that continue on a computer, extending local-file workflows beyond the desktop. The move positions Google more directly against agentic productivity suites that blend chat, coding, app control, and file operations.

Claude records skills from screens

Claude's desktop app now includes a screen-recorded skill creation tool that turns a demonstrated workflow into a reusable command. The feature lives in the app’s Customization area and can capture both on-screen actions and spoken instructions, then convert them into a saved skill. That shifts workflow setup away from handcrafted markdown prompts toward show-and-tell automation. The desktop-only design also hints at Anthropic’s broader push to make agent behavior teachable through observation rather than formal scripting.

Anthropic details safer agent architecture

Anthropic also outlined a managed-agent architecture built around a model in a loop with tools, separated from execution sandboxes and persistent session logs. The company argued that the surrounding harness is critical because model behavior changes across generations, citing differing context-window quirks between Sonnet 4.5 and Opus 4.5. It recommended isolated containers for each task to contain failures, buggy code, or unsafe dependency installs. The design reflects a growing industry shift from flashy demos toward reliability, auditability, and controlled long-running automation.

REA fuels reverse-engineering boom

REA, short for Reverse Engineer Anything, has emerged as a breakout open-source project, jumping from about 10,000 to 100,000 GitHub stars in six days. The tool plugs into coding agents such as Claude Code and Codex to inspect websites, apps, desktop software, and binaries, then produce structured maps of how they work. That output can be handed to orchestration layers to build custom lookalikes or MVPs with much less manual product analysis. Its rapid adoption highlights how AI coding is shifting from generating snippets to reconstructing and reproducing complete software workflows.

Videos covered

Previous briefings · AI