AI Coding Live News — Real-Time 𝕏/Twitter Updates & Breakthroughs

👀9

🔥88

Cua + Scale AI Launch Cua Driver Bench: Background Desktop Agents vs Undisturbed Users; Vendor macOS Full-Pass ~58% for cua-driver

On 2026-11-06, Cua (@trycua) and Scale AI AgentEnv launched Cua Driver Bench (Bench v2): same Claude Sonnet 5.5, tasks, and limits with only the computer-use driver swapped, scoring task done, GUI-only, and eight undisturbed-user criteria. Scale macOS background full-pass: cua-driver ~58% (23/40), agent-desktop ~20%, browser-use ~3%; Cua-run Codex Computer Use ~33%. Tasks and trajectories are public to rerun.

Key Takeaways

Source: Cua official blog + @trycua thread on 2026-11-06; Scale ran macOS/Linux; Cua ran Windows and some controls
Method: 10 Bench v2 tasks; background keeps a simulated user typing; full pass = done + GUI-only + 8 undisturbed criteria
macOS background full-pass (Scale, vendor): cua-driver 58%, agent-desktop 20%, browser-use 3%; Codex CU (Cua-run) 33%
Foreground (Scale): cua-driver ~78%, agent-desktop 70%, browser-use 64%; gaps are mostly disturbance
Rerun: GitHub blog/libs/cua-driver-bench; HF Space/dataset earakely-scale/CuaDriverBench-AgentEnv

Read full story

Linked Models & Agents:

⚡ Claude Sonnet 5.5 🤖 cua-driver 🤖 codex 🤖 agent-desktop 🌐 Cua 🌐 Scale AI 🌐 OpenAI 🌐 Vercel

#Cua Driver Bench#Cua#Scale AI#AgentEnv#computer use#Benchmark#Cua driver

Scale AI Research@huggingface·Collected 17d ago

👀138

🔥95

Scale AI Open-Sources Sauron: Evidence-First Agentic Verification Slashing Evaluation Costs by 85% with 0.97 Recall

Scale AI Research introduced Sauron (arXiv:2610.17724), an open-source evidence-first agentic verification framework. Overcoming the severe failure modes of standard LLM-as-a-judge approaches—which hallucinate justifications and guess outcomes over long agent trajectories—Sauron enforces bounded, adaptive interrogation of agent traces with cryptographic-grade execution receipts. Evaluated on the PA100 benchmark across eight backbones, it delivers 0.94 precision and 0.97 recall at roughly 15% the cost of state-comparison baseline Archipelago, completing 10-episode grading 4.2x faster.

Key Takeaways

Scale AI open-sourced Sauron (arXiv:2610.17724), an evidence-first verification architecture eliminating LLM-as-a-judge hallucinations
Employs bounded adaptive queries and verifiable evidence receipts to extract targeted execution proofs without scanning full transcripts
Sweeps top F1 scores across eight backbones on PA100, reaching 0.94 Precision and 0.97 Recall at 1/7th the cost of snapshot comparisons
Completes 10-episode benchmark runs in 17 minutes vs 71 minutes (4.2x speedup), with full open code on GitHub

Read full story

Linked Models & Agents:

#Sauron#Scale AI#Agent Verification#LLM-as-a-Judge#Process Evaluation#OpenSource

LMSYS Org@arena·Collected 18d ago

👀72

🔥95

View on 𝕏 ↗

Meta Unveils Gigawatt-Scale Prometheus Supercluster as Muse Wonder Secures #6 on WebDev Code Arena

Meta officially disclosed its gigawatt-scale AI supercluster Prometheus in New Albany, Ohio, housing hundreds of thousands of frontier GPUs with 85%+ pretraining goodput and 90%+ in reinforcement learning. Powered by this infrastructure, Meta's latest coding flagship Muse Wonder Full surged to #6 globally on LMSYS Code Arena: WebDev with 1726 Elo points (surpassing Claude Sonnet 5.5 and Gemini 4 Argon), lifting Meta to the #3 AI research lab worldwide.

Key Takeaways

Meta unveiled its gigawatt-scale Prometheus supercluster in Ohio with hundreds of thousands of GPUs, delivering 85%+ pretraining goodput
Exceeded 90%+ goodput on RL workloads, recording ~47,000 GPU-hours MTBF during Muse Wonder's training cycle
Muse Wonder Full scored 1726 Elo on LMSYS Code Arena: WebDev (#6 globally), outperforming Claude Sonnet 5.5 (+16) and Gemini 4 Argon (+48)
Elevated Meta to the #3 AI lab worldwide; confirmed Muse Glimmer 30B open weights capable of 4-bit local execution on a single consumer GPU

Read full story

Linked Models & Agents:

#Meta#Prometheus#Muse Wonder#LMSYS#WebDev#Supercluster

CISA@CISAgov·Collected 48d ago

👀88

🔥93

NSA, CISA, and FBI Issue Joint Security Advisory on Industrial-Scale AI Model Distillation

In response to campaigns aimed at extracting proprietary capabilities from frontier models, the NSA, CISA, and FBI published a joint cybersecurity advisory. It mandates real-time semantic fingerprint watermarking, synthetic prompt topology filtering, and anti-distillation rate limiting on enterprise API gateways.

Key Takeaways

First multi-agency advisory addressing industrial-scale capability theft via API distillation
Recommends dynamic imperceptible watermarking and query entropy anomaly detection
Elevates model weight protection and synthetic data exfiltration to national cybersecurity priority

Read full story

Linked Models & Agents:

🌐 NSA 🌐 CISA 🌐 FBI

#CISA#NSA#AI Security#Knowledge Distillation#Cybersecurity#IP Protection