Tech Launch AI Digest

The US is framing AI governance at the G20 Innovation Ministerial as an opportunity for hands-off rules and infrastructure growth, while others stress...

This gift guide spotlights desk items that pair eye-catching designs with practical features, moving beyond pure novelties.

• 8Bitdo Numpad: Retro...

Can a credit-card-sized USB-C controller deliver high-quality mobile gaming without Bluetooth drawbacks?

• Plug-and-play USB-C: Direct connection...

OpenAI's new methods that reduce chain-of-thought transparency threaten a fragile AI safety tool, as highlighted by Gary Marcus citing the 2025 paper...

Graph memory underperforms flat vector retrieval on LongMemEval, scoring token F1 0.42 versus 0.47, with the gap widest on questions needing specific prior turns. Entity extraction discards surface forms that simpler methods preserve.

Unified multimodal models can turn coexistence into genuine synergy when properly architected.

• Representation level: Generation enriches visual...

StudentSim turns sparse real-student data into individualized LLM simulators that mirror responses and adapt under guidance, outperforming GPT-5.4 on...

ASUS's AI City shifts urban systems from reactive monitoring to predictive, cross-domain action.

• Data Limits Exposed: Smart City 1.0 dashboards...

Coding agents are evolving from fixed harnesses to self-improving systems that sustain multi-day projects.

• HoH layers iterative...

The next AI breakthrough may come not from bigger models but from organizing thousands of specialized agents, memory systems and orchestration layers...

Acer's relaunch of Packard Bell at IFA 2026 with its 'Tech it easy' range of straightforward, budget gadgets like the DotBook laptop and DotLine...

Amazon's new Update Me When feature in Alexa for Shopping sends personalized alerts for tempting events like new product launches or tours, moving from reactive queries to anticipating purchases before users ask.

E-Commerce Bench simulates a full 365-day year running multiple online stores to test whether agents sustain business performance over time, evaluating 18 frontier models.

Long-running AI agents now deliver exponential gains in self-organized work, widening the gap with public understanding and requiring a fresh approach to capability, supervision, and risk.

Can objective rewards reduce subjectivity in AI-driven scholarly reviewing?

InternReviewer and InternAdvocate use agentic RL with a unified metric...

Washington Middle School permits AI tools solely when teacher-directed, using vetted district options that safeguard entered data. Students engage in...

• Users run Qwen3.5 and GLM models on M2 Mac Mini via Ollama for private screenshot OCR and description, using ~10GB RAM and auto-unloading after...

Salesforce made Claude its default AI layer, sending shares up ~20% as vendors lock in frontier models.

• OpenAI targets an internal AGI system by...

Before its 2015 debut, Apple aimed to create a smartwatch people genuinely wanted to wear by blending advanced tech with high-fashion appeal,...

美国在G20峰会推动Carolina Principles轻触式AI监管。

会议吸引日本、德国、法国、印度和韩国等国部长出席，马斯克、奥特曼、黄仁勋等领袖参与讨论。

这是否促成全球AI治理协调，还是加剧分歧？