All Updates
Page 547 of 654
February 25, 2026
Alibaba Launches 3 Mid-Size Qwen3.5 Models
Alibaba open-sourced three mid-size Qwen3.5 models: Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, and Qwen3.5-27B. These outperform larger prior models and GPT-5 mini on benchmarks. Qwen3.5-Flash is hosted on AliCloud at 0.2 yuan per million input tokens.
CAAC Drafts Drone Airworthiness Standards
China's Civil Aviation Administration is soliciting public opinions on draft airworthiness standards for normal category powered-lift unmanned aircraft systems (unmanned) and multi-rotor UAS. The drafts draw from practical certification experiences to guide and regulate type approvals.
Sichuan Guides Short Drama + Ecommerce Push
Sichuan Province promotes 'content+traffic+commerce' micro-short drama platforms and 'short drama + ecommerce' models via new services opinion. Leverages big platforms for digital ads, smart marketing algorithms; offers up to 2M RMB incentives for qualifying firms.
Kaipuyun Reorg Fails on High Storage Valuation
Kaipuyun responded to reorganization failure citing high target valuation expectations amid booming storage sector. Stock dropped over 16% and halved monthly. Firm reaffirms AI infrastructure strategy as storage price rises validate industry binding with compute.
GAC Launches Flying Car Manufacturing Unit
GAC Group subsidiary founded High Domain Flying Car Manufacturing (Guangzhou) Ltd. with 30M RMB capital. Legal rep Zhang Ning. Includes intelligent unmanned flight vehicle sales, auto parts, 3D printing, and aviation services.
OpenClaw Creator Warns of Unstoppable Misuse
In a new interview, OpenClaw's creator, dubbed the 'Father of Lobster,' reveals internal details of the project. He admits that preventing abuse is impossible and advises users against 'playing with fire.' This highlights ongoing concerns with open-source AI tools.
HP: PC Memory BOM Share Doubles to 35%
HP revealed in its FY2026 Q1 earnings call that PC bill of materials costs have shifted dramatically, with memory proportion nearly doubling to 35% in one quarter from 15-18%. The company anticipates further increases this year amid supply pressures.
90% Firms Keep New Grad Hires Despite AI
Mynavi's survey shows over 90% of companies will not reduce new graduate hiring even as AI spreads. Released on February 25 for 2027 graduates, it was explained at a press briefing. Discussions covered AI's influence on recruitment and new employee tasks.
RTX 5080 Tops RTX 50 Series Sales
German retailer Mindfactory's 2026 week 8 sales data shows RTX 5080 as the top RTX 50 series GPU with 95 units sold. NVIDIA adjusted supply strategy, prioritizing RTX 5080 16GB shipments. It became the only 16GB model with increased supply, boosting popularity.
Tsinghua Predicts 2-Day Workweek by 2035
Tsinghua AIR dean Zhang Yaqin forecasts robots outnumbering humans by 2035, exceeding 70-80 billion. Humans will work just 2 days weekly as robots replace jobs. Wages expected to stay same or rise.
Hyundai Plans $7.3B AI Investment in Korea
Hyundai Motor Group is reportedly set to announce a 10 trillion KRW investment in Korea's west coast. The funds will support expansion into AI for autonomous driving and robotics technologies. This marks a significant push by the automaker into AI-driven mobility.
Alibaba Cloud Launches Ultimate Coding Plan
Alibaba Cloud's Bailian platform launched the Coding Plan, offering API access to four top open-source models: Qwen3.5, GLM-5, MiniMax M2.5, and Kimi K2.5, with multi-model switching. Lite plan provides 18,000 monthly requests for 7.9 RMB first month; Pro offers 90,000 for 39.9 RMB. It now supports 8 top coding models total.
Isomorphic Labs Unveils AlphaFold3-Beating IsoDDE
Isomorphic Labs, DeepMind's biopharma spin-off, released IsoDDE, a proprietary drug discovery AI engine outperforming AlphaFold3 on protein-ligand predictions. It doubles AlphaFold3 accuracy on benchmarks, surpasses Boltz-2 and physics methods on binding affinity, and identifies novel binding pockets from sequences alone. The model handles induced fit and cryptic pockets, aiding antibody design and new therapies.
Tsinghua Math Genius Joins OpenAI, Ex-SAM/Llama Lead
A top mathematics talent from Tsinghua University has joined OpenAI. He previously led development of Meta's SAM image segmentation model and Llama large language models. OpenAI's Sora project lead welcomed him publicly.
Trace-Free+: Rewriting Tools for LLM Agents
Trace-Free+ is a curriculum learning framework that optimizes tool descriptions for LLM agents without execution traces. It transfers supervision from trace-rich to trace-free settings using a large-scale dataset built from diverse tools. Experiments on StableToolBench and RestBench demonstrate gains on unseen tools, cross-domain generalization, and scalability to over 100 candidates.
PreScience Benchmark Forecasts Scientific Advances
PreScience introduces a benchmark for AI to forecast scientific advances via four tasks: collaborator prediction, prior work selection, contribution generation, and impact prediction. It uses a 98K AI paper dataset with 502K total papers and metadata. Frontier LLMs show moderate performance, producing less diverse synthetic research.
New Causal Reasoning Benchmark Launches
CausalReasoningBenchmark is a new benchmark with 173 queries from 138 real-world datasets across 85 papers and 4 textbooks. It separately evaluates causal identification (structured specs) and estimation (numerical outputs). Baseline LLM results show 84% high-level strategy success but only 30% full identification accuracy.
ML vs Stats for Child Obesity Prediction
Researchers compared statistical, ML, and deep learning models on 18,792 US children aged 10-17 from 2021 NSCH to predict overweight/obesity. AUC ranged 0.66-0.79, with logistic regression, gradient boosting, and MLP offering the best balance of discrimination and calibration. Complex models provided limited gains, and subgroup disparities persisted across races and poverty levels.
MIMIC: Steerable Inner Speech for AI Imitation
MIMIC introduces inner speech via vision-language models to guide AI agents in imitating diverse human behaviors for better human-AI coordination. It uses a conditional VAE to generate speech from observations and a diffusion policy for action selection. The framework enables inference-time steering without extra training and outperforms baselines in robotics and games.
Learning Optimal Verbalization for LLM RecSys
A new framework uses reinforcement learning to convert user interaction logs into optimized natural language for LLM-based recommendation systems, outperforming rigid templates. Experiments on industrial streaming data show up to 93% relative accuracy gains in discovery recommendations. Emergent strategies include interest summarization, noise removal, and syntax normalization.