In December 2024 I spent two weeks in San Francisco talking to builders across labs, clouds, and startups. The same patterns I saw there have crystallized in 2025.
The 2025 inflection: agents + reasoning + open models
The conversation has moved beyond chat. The hottest race now is for reliable agents — systems that can plan, take multi-step actions, and operate software on our behalf. Even the biggest platforms are saying it out loud: Amazon’s AGI group is prioritizing agents over raw LLM size because doing things matters more than talking about them.
Reasoning-first foundation models have arrived too. OpenAI’s o3 family is optimized for long, careful thinking and is now broadly available via ChatGPT and API. On the open side, Meta’s Llama 4 and 3.1 (up to 405B) pushed the ceiling for openly available models, and Qwen has been iterating fast with Qwen2.5/Qwen3 and large MoE variants. There’s also a wave of open agentic research like Moonshot’s Kimi K2.
Under the hood, hardware capacity is exploding — NVIDIA Blackwell systems roll out through 2025, enabling cheaper inference and larger context, though demand remains intense. And on the edge, Apple Intelligence is mainstreaming on-device AI with a privacy-by-design architecture developers can tap into across iPhone, iPad, and Mac.
Why Europe can win this wave
Two structural advantages stand out:
- Trust by design. The EU AI Act is now in force, with prohibitions/AI-literacy obligations already active (Feb 2025), GPAI duties applying from Aug 2, 2025, and full applicability by Aug 2, 2026 (with some transitions to 2027). Building to these standards is a global trust signal.
- Industry depth. Europe owns complex verticals — telecom, energy, health, manufacturing — where reliable agents plus tight governance beat raw benchmarks every time.
My 7 rules for AI engineers in 2025 (the playbook I brought home)
Design for agent reliability, not demos.
Add eval gates that block deploys when plans/actions regress (schema checks, tool-use validation, safety rails). Benchmarks are nice; passing CI with real tasks is better.Measure unit economics from day one.
Track cost per request (input/output tokens × model pricing), p50/p95/p99 latency, cache hit rate, and error budget. This is how you scale without surprises.Compliance is a feature.
Treat EU-AI-Act alignment like performance work: data lineage, audit logs, PII handling, human-in-the-loop where risk is high. Teams that can show compliant by construction will close bigger deals faster.Leverage open models strategically.
Llama and Qwen families give you sovereign options: fine-tune locally, serve on your infra, and mix with closed models when you need peak reasoning (e.g., o3 for edge cases).Be hardware-aware.
Build for the 2025 stack: longer contexts, MoE routing, paged KV, quantization. If you can show a 30–50% latency or cost drop on Blackwell-class nodes (or smart batching on current GPUs), you’re speaking the language of production.On-device is a first-class path.
Privacy-sensitive features belong on the device when possible; use secure fallback to cloud for heavier tasks. Apple’s model tiers and Private Cloud Compute make this an easy story to tell customers.Ship thin slices into real workflows.
Pick a vertical (telecom, energy, health), automate one painful multi-step task end-to-end, and instrument the results. Repeat. Careers are built on shipped systems — not whitepapers.
A note from San Francisco
What struck me in San Francisco wasn’t just the pace — it was the clarity. Teams that win obsess over reliability, cost, and trust. In 2025, Europe can add something special to that equation: deployment in regulated, high-impact industries. That’s where agentic AI stops being a demo and starts changing how the world runs.
If this resonates, let’s connect — I’m building systems-first AI, open to collaborations.
— Riku Lauttia
#ArtificialIntelligence #MachineLearning
Top comments (0)