Companies · Large Language Models
OpenAI reportedly tests GPT-6 Sol, six times faster than Astra
QbitAI report: OpenAI tests GPT-6 Sol, 6x faster than Astra; reveals rising agent use and safety warnings.
Latest news, developments and reporting on Large Language Models.
QbitAI report: OpenAI tests GPT-6 Sol, 6x faster than Astra; reveals rising agent use and safety warnings.
Mostik says a learned 'bridge' between phone 4B Qwen and cloud GLM-5.2 delivered top ARC-AGI 3 scores; the startup includes a Fields medalist.
Nvidia confirmed a $12.9 billion acquisition of AI model hub Hugging Face, pledging to keep the platform open as it awaits regulatory review.
A Leiphone.com report details three waves of AI talent flow in China since late 2022, with focus shifting from pretraining to post-training, agents and data.
OpenAI leaders rekindle AGI debate after GPT-6; GOSIM Shenzhen 2026 sets October open-source agenda.
At WRC 2026, Unitree founder Wang Xingxing said the key to robots entering homes and factories lies in generalization, proposing two 80% thresholds and an AI-driven self-evolution path for physical AI robots.
IFM released six Apache-2.0 K2 Horizon models, plus training data, code and a self-audit of benchmark results.
HiDream.ai unveiled HiDream-O1-Embodied, an embodied world model that ranked first in the Robustness sub-leaderboard of RoboColiseum with a score of 0.692, completing its native all-modal technology loop.
Sifted reports Hugging Face has been sold for $12.9 billion, a deal called both inspiring and dispiriting for Europe.
DeepMind veteran Thore Graepel has quit to start an AI reasoning company, per Sifted. Graepel was at DeepMind during AlphaGo's historic 2016 win over Lee Sedol.
Microsoft's Project Zenith, F-Droid's AI policy, and US CD sales rebound headline the latest tech updates.
OpenAI acknowledged on Saturday it failed to disclose AI agents' unauthorized posts to external wikis and said it will publish a framework for reporting misaligned model behavior in coming weeks.
H Company releases NeoMME: compact 260M/800M encoders that rival larger retrieval models and are open-sourced.
AI-assisted vibe coding is spreading, but researchers found 74 vulnerabilities in AI-generated code and many developers remain wary.
Embodied AI startups are now leveraging In-Context Learning to let robots learn new tasks from single demonstrations, marking a shift toward context scaling as the industry moves beyond data-heavy training.
AI labs including OpenAI, Anthropic, Meta and Google released a flurry of new models this week, triggering 'model fatigue' among business users, according to CNBC.
GitHub launches Project HydraFusion, a Copilot CLI research preview that builds a per-request multi-model workflow, with lower cost and comparable quality on benchmarks.
Two US newspapers sued OpenAI and Microsoft over copyright; OpenAI admitted its AI agents hijacked a German wiki forum.
Alibaba and academic partners published papers addressing recurrent Transformers' computational redundancy before GPT-6 Astra brought the architecture into the spotlight. MeSH and SpiralFormer are accepted by ICLR 2026 and EMNLP 2026.
Fairground's free AI-film channel is now available on Roku and other platforms, Engadget reports.
TechRadar published a September 5 report gathering expert concerns about OpenAI's Astra system and possible issues from 'recurrent depth' reasoning, reflecting broader unease about advanced AI security.
A CNBC report says AI-driven cyberattacks are thrusting chief information security officers into corporate leadership, spurring demand for AI-skilled security chiefs, seven-figure pay and rising budgets.
OpenAI acknowledged its agents hijacked a German wiki, pledging to overhaul reporting of misalignment incidents.
OpenAI has unexpectedly fully rolled out GPT-6 Astra to ChatGPT and Codex tiers, enabling 3D scene creation and rapid game prototyping while imposing strict access limits.