All Updates
Page 558 of 644
February 23, 2026
SIE Breaks RL Env Scaling Bottleneck
Shanghai Jiao Tong Univ's SIE uses structured data like KGs for scalable, verifiable RL envs without expert labels. Models learn multi-hop reasoning, generalizing to math/logic puzzles. ICLR 2026 accepted; open-sources code.
1904 Telekino Pioneers Autonomous Waymo Cars
Ars Technica spotlights Quevedo's 1904 telekino as the first bold step in self-driving tech. This remote-control invention laid groundwork for modern autonomous vehicles like Waymo.
Undetectable AI Backdoor Hijacks Self-Driving Cars
Georgia Tech researchers discovered VillainNet, a dormant AI backdoor enabling hackers to hijack self-driving cars with 99% success. It remains invisible to all current security tools. This poses a severe threat to autonomous vehicle safety.
600+ FortiGate Firewalls Hit by AI Attacks
AWS reports Russian-speaking cybercriminals used off-the-shelf generative AI tools to compromise over 600 exposed FortiGate firewalls. The campaign hit 55 countries in just over a month. It demonstrates AI lowering barriers for cyber attacks.
ChatGPT Chats Turn Personal for Young Users
New OpenAI data shows users leverage ChatGPT for personal expression and venting as much as work tasks. Younger users especially treat it like a sounding board. This shift reveals evolving AI interaction patterns.
AI² Robotics Hits $1.4B Val in B-Round Mega-Funding
智平方 completes 5-round B-series financing exceeding 10B RMB, total 12 rounds in a year, valuation over 100B RMB. Dubbed 'China's Tesla' for end-to-end models and production focus. Funds expand GOVLA models and AlphaBot production.
AI Job Obsolescence Sparks Food Debate
Article questions societal feeding mechanisms if AI eliminates human jobs. It notes recurring fears since Industrial Revolution but urges serious debate. Most adults remain employed despite past tech shifts.
OpenAI Drops Flawed SWE-bench Verified
OpenAI no longer evaluates on SWE-bench Verified due to contamination and poor measurement of frontier coding progress. Their analysis identifies flawed tests and training leakage. They recommend switching to SWE-bench Pro.
EU Tightens Horizon Europe Ties with China
In 2026, Horizon Europe (€93B research framework) shifted to a guarded stance on global partnerships. Critical areas like AI, semiconductors, and quantum now face restrictions. It marks a strategic rebalancing away from open invitations.
Zhipu Shares Plunge 23% on Compute Crunch
Zhipu AI’s shares dropped nearly 23%, erasing HK$70B in market cap due to computing resource constraints. The firm seeks global compute partners amid user complaints on service quality. Investor frenzy persists despite issues.
Qwen 'Help Me' Hits 5B Uses in China
'Qianwen help me' has become a Spring Festival tradition for effortless one-sentence ordering. Over 4 million users aged 60+ have tried AI ordering, with nationwide users issuing 5 billion commands.
OpenAI and Jony Ive's Secret Project
OpenAI is collaborating with former Apple designer Jony Ive on a new undisclosed project. The article teases details on what they are building. It also previews The Rundown Roundtable discussing AI use cases.
AI Glasses Sales Surge 80% in Festival Boom
AI glasses saw up to 80% sales growth during China’s Spring Festival, driven by consumer demand and subsidies. Shenzhen’s Huaqiangbei market revenue rose 35% year-on-year, led by AI eyewear. Pent-up demand fueled the boom.
Cloudflare One Pioneers Full Post-Quantum Encryption
Cloudflare One now supports post-quantum encryption via hybrid ML-KEM from latest IETF drafts in its IPsec product. This upgrade extends protection across all major on-ramps and off-ramps. It positions Cloudflare One as the first SASE platform with modern post-quantum encryption everywhere.
OpenAI Launches Frontier Alliance Partners
OpenAI announces Frontier Alliance Partners to support enterprises. The initiative helps transition AI pilots to production deployments. It emphasizes secure and scalable agent systems.
Zhipu GLM-5 Fully Open-Sourced for Huawei
Zhipu has fully open-sourced GLM-5 technology, ensuring full compatibility with Huawei and other domestic chips. It incorporates a DeepSeek-like mechanism, drawing envy from US netizens.
Survey: 67% See Agent ROI, But Scaling Hits Inference Costs
DigitalOcean's survey of 1,100 devs/CTOs shows 67% productivity gains from AI agents, with 60% viewing them as top AI value. Yet only 10% scale to production due to 49% citing high inference costs. Agents target code gen (54%), ops (49%), support (45%).
32x Compression Outperforms by 25 Points
A new method achieves 32x compression on long texts while surpassing baselines by 25 points, cracking common compression failures. Featured at ICLR 2026, it stresses high-quality compression beyond mere deletion.
OpenAI Stargate Stalled After One Year
OpenAI's ambitious Stargate data center project has seen no progress one year after announcement. The team remains unassembled, and construction has not started at all. The article uses the proverb 'three monks have no water' to highlight coordination failures.
OpenClaw 2026.2.22: Mistral & Multilingual Memory
OpenClaw 2026.2.22 introduces Mistral provider with memory embeddings and voice, plus multilingual FTS memory search for Spanish, Portuguese, Japanese, Korean, Arabic. New features include auto-updater, Synology Chat channel, and config UI improvements. Fixes address CLI security and Discord voice dependencies.