🗞️ Latest AI News 2025

• China: Mandatory labeling of AI-generated content (transparency and combating disinformation).
• InvestAI (EU): program of ~€200 billion including 4 "AI gigafactories" GPU.
• EU AI Act: prohibitions for "unacceptable risk" in effect; GPAI/transparency obligations and sanctions active in 2025.
• Datacenters: explosion in energy demand; optimization of priority inference.
• IFA 2025 Berlin: AI everywhere: PC, gaming, AR, smart home.
• Funding: Record H1 2025 for AI startups.

Agentic AI: autonomous agents

Planning, tooling, execution, self-verification. Immediate use cases: product briefs, QA catalog, customer service follow-ups, programmatic purchases, competitor monitoring.

1. Define a "safe zone": tasks, guardrails, stop criteria.
2. Trace prompts, decisions, sources, costs, and errors.
3. Increase latitude after KPI validation.

Edge AI: Small, local models

Objectives: low latency, confidentiality, reduced costs. 2025 strategy: local-first with targeted cloud scaling.

• Distillation and quantization (int8/4-bit) for mobile/PC.
• Encrypted knowledge cache and scheduled synchronization.
• A/B Tests: Quality vs. Cost/1000 Requests.

Chips, HBM and data centers

HBM memory dominates heavy training and inference. Selective shortage, record data center CAPEX, required load optimization.

• Profile the pipelines and remove redundant inferences.
• Specializing models by task rather than having "generalists" everywhere.
• Track an energy KPI: kWh / 1000 inférences .

Regulation: AI Act and InvestAI

2025 = compliance shift. Map risks, document training data, usage limits, evaluations, governance, and logging.

1. Classify your use cases by AI Act risk level.
2. Set up an AI technical file and a risk management process.
3. Follow the implementation schedule and sector-specific guidelines.

Humanoid robotics

Cheaper sensors, multimodal vision, learned policies. Pilots in logistics, inspection, retail and healthcare.

• Choose 1–2 repetitive physical tasks with a short ROI.
• POC "humanoid + supervision", KPIs: safety, MTBF, cost/hour.

Moving from SEO to AEO

• Clear H1, 40–60 word introduction, table of contents, and anchor links.
• JSON-LD schemas: Article , FAQ . Accurate data.
• Descriptive alt text, lazy loading images, fixed sizes, WebP if possible.
• Deep internal links and outbound links to reliable sources.
• Q&A blocks, tables and lists: easy extraction by agents.

Questions for your readers

1. Which task should I delegate to an AI agent this week?
2. What small, local model can replace a recurring cloud call?
3. Are your use cases mapped to the AI Act framework?
4. What energy KPI do you track for AI?

At a glance