Meta's New Model Lets AI Manage Context Like Editing Files, Boosting Accuracy by 11.4% · 1005-711

2026-10-05 00:00

Author: RadarAI Editorial Editor: RadarAI Editorial Last updated: 2026-10-05 Review status: Editorial review pending Brief 速报官方 AI动态开源

This week brought several key breakthroughs in the AI field: Meta proposed Context Language Models, allowing models to autonomously manage context like editing files, improving accuracy by 11.4% on BrowseComp-Plus while reducing FLOPs by 21.5% [1]; meanwhile, Trump signed an executive order requiring the term "AI" to be replaced with "Super Intelligence" (SI) and established the Super Intelligence Force, with Musk promptly confirming that SpaceXAI will be renamed SpaceXSI [3][19]. On the industry side, Strata's open-source solution enables a 125B-parameter model to run at approximately 100 token/s on an RTX 4090, further lowering the barrier for running large models on consumer-grade hardware [6].

Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.

## 🔍 Core Insights This week brought several key breakthroughs in the AI field: **Meta** proposed **Context Language Models**, allowing models to autonomously manage context like editing files, improving accuracy by **11.4%** on BrowseComp-Plus while reducing FLOPs by **21.5%** [1]; meanwhile, **Trump** signed an executive order requiring the term "**Super Intelligence**" (SI) to replace "AI" and established the **Super Intelligence Force**, with **Musk** promptly confirming that **SpaceXAI** will be renamed **SpaceXSI** [3][19]. On the industry side, **Strata**'s open-source solution enables a **125B**-parameter model to run at approximately **100 token/s** on an **RTX 4090**, further lowering the barrier for running large models on consumer-grade hardware [6]. ## 🚀 Key Updates - **Meta proposes Context Language Models** [1]: The model treats real-time context as editable files, used zero-shot, improving accuracy by 11.4% and reducing FLOPs by 21.5% - **Trump establishes Super Intelligence Force** [3]: Jay Clayton leads coordination across the federal government to ensure U.S. leadership in super intelligence - **Strata goes open source: running 125B models on consumer-grade GPUs** [6]: MIT licensed, approximately 100 token/s on RTX 4090, requires 12GB+ VRAM - **Kling 4.0 short film "The Beat" surpasses 5 million views** [9]: 30-second native continuous long take, Omni Reference supports up to 15 multimodal references - **Google proposes RRSI to prevent agents from memorizing test tasks** [11]: Uses shrinking edit budgets and strict critic review to suppress test memorization when harnesses are repeatedly rewritten - **NVIDIA releases Mid-Harness action sampling method** [14]: Pass@1 on TerminalBench-Lite rises from 50.0% to 68.0% - **NASA and IBM open-source lunar foundation model** [18]: Built on 17 years of orbiter data, one of the earliest open-source lunar science foundation models - **Google pauses product vulnerability submissions for open-source bug bounty program** [24]: Thousands of AI hallucination reports crowd out time for fixing real high-severity vulnerabilities ## 🔗 Sources [1] Meta et al. propose Context Language Models, where models self-manage context — https://aihot.news/items/tzi9tbjk38i07bf4np38x2o0z [3] Trump announces establishment of Super Intelligence Force, led by Jay Clayton — https://aihot.news/items/memfk03m6yv8cz4r17fi4abus [6] Strata goes open source: running Qwen3.8-Flash-Next (125B) at approximately 100 token/s on consumer-grade GPUs like RTX 4090 — https://aihot.news/items/ci1q47jrk7a61xbsi7in5938r [9] Behind the scenes of Kling 4.0 short film "The Beat" surpassing 5 million views — https://aihot.news/items/gf048arpz9hrjjq61wm1ce97l [11] Google research proposes RRSI to prevent self-improving agents from memorizing test tasks — https://aihot.news/items/ppoem158qon2xzi0swn3xp0xp [14] NVIDIA proposes Mid-Harness: scaling action sampling between model and harness to improve terminal agent success rates — https://aihot.news/items/raw0jq5w0on3s2p10zh6ey

This week brought several key breakthroughs in the AI field: Meta proposed Context Language Models, allowing models to autonomously manage context like editing files, improving accuracy by 11.4% on BrowseComp-Plus while reducing FLOPs by 21.5% [1]; meanwhile, Trump signed an executive order requiring the term "Super Intelligence" (SI) to replace "AI" and established the Super Intelligence Force, with Musk promptly confirming that SpaceXAI will be renamed SpaceXSI [3][19]. On the industry side, Strata's open-source solution enables a 125B-parameter model to run at approximately 100 token/s on an RTX 4090, further lowering the barrier for running large models on consumer-grade hardware [6].

🚀 Key Updates

Meta proposes Context Language Models [1]: The model treats real-time context as editable files, used zero-shot, improving accuracy by 11.4% and reducing FLOPs by 21.5%
Trump establishes Super Intelligence Force [3]: Jay Clayton leads coordination across the federal government to ensure U.S. leadership in super intelligence
Strata goes open source: running 125B models on consumer-grade GPUs [6]: MIT licensed, approximately 100 token/s on RTX 4090, requires 12GB+ VRAM
Kling 4.0 short film "The Beat" surpasses 5 million views [9]: 30-second native continuous long take, Omni Reference supports up to 15 multimodal references
Google proposes RRSI to prevent agents from memorizing test tasks [11]: Uses shrinking edit budgets and strict critic review to suppress test memorization when harnesses are repeatedly rewritten
NVIDIA releases Mid-Harness action sampling method [14]: Pass@1 on TerminalBench-Lite rises from 50.0% to 68.0%
NASA and IBM open-source lunar foundation model [18]: Built on 17 years of orbiter data, one of the earliest open-source lunar science foundation models
Google pauses product vulnerability submissions for open-source bug bounty program [24]: Thousands of AI hallucination reports crowd out time for fixing real high-severity vulnerabilities

🔗 Sources

[1] Meta et al. propose Context Language Models, where models self-manage context — https://aihot.news/items/tzi9tbjk38i07bf4np38x2o0z [3] Trump announces establishment of Super Intelligence Force, led by Jay Clayton — https://aihot.news/items/memfk03m6yv8cz4r17fi4abus [6] Strata goes open source: running Qwen3.8-Flash-Next (125B) at approximately 100 token/s on consumer-grade GPUs like RTX 4090 — https://aihot.news/items/ci1q47jrk7a61xbsi7in5938r [9] Behind the scenes of Kling 4.0 short film "The Beat" surpassing 5 million views — https://aihot.news/items/gf048arpz9hrjjq61wm1ce97l [11] Google research proposes RRSI to prevent self-improving agents from memorizing test tasks — https://aihot.news/items/ppoem158qon2xzi0swn3xp0xp [14] NVIDIA proposes Mid-Harness: scaling action sampling between model and harness to improve terminal agent success rates — https://aihot.news/items/raw0jq5w0on3s2p10zh6ey

← Back to Updates