www.mckinsey.com/mgi/our-rese...

We are shifting from individual contributors to human-agent teams. Our accreditation, standards, and teaching practices need to reflect that reality immediately.

The report is clear that traditional degrees are too slow. We need policy that supports flexible, "earn-as-you-learn" apprenticeships and verified credentials. Schools and employers need shared programs so students can learn and prove they can manage an AI workflow.

"AI Fluency" needs to be a foundational standard from primary school onward. Fluency being the ability to audit an agent, question a result, and spot bias. We need policy that mandates this literacy across every subject as an overarching skill.

Education needs to look at the "Skill Change Index." It shows that digital and information-processing skills are the most exposed to automation. Judgement, evaluation, and discernment are likely increasingly important skills to have going forward.

Technological progress often appears steady until it reaches human equivalence, at which point the rate of change can seem abrupt. This offers a compelling perspective on the current trajectory of artificial intelligence.

Highly recommend the full read for those interested in behavioral trends: openrouter.ai/state-of-ai

Despite some "glass slipper" effects I feel that LLM providers are becoming commodities. The real value is occurring at the product layer. Locking into a single provider is a strategic error; organizations must prioritize flexibility and demand transparency on model updates.

Asia's share of token consumption doubled to 31% in a single year, with Singapore ranking as the second-highest source of usage. Useful, considering my context.

Agentic inference is taking over. The median request is now part of a structured loop, with tool invocation hitting ~15%. Programming tasks specifically jumped from 11% to ~50% of traffic. Notably, Anthropic dominates coding spend despite OpenAI’s broader awareness.

Average prompt length grew 4x (to ~6,000 tokens) while completion length only tripled. We are feeding massive context for analysis rather than just asking for open-ended generation. This largely mirrors my use as well.

Open source is sustainable, consistently holding 20-30% of total usage. It is also diversifying: the market is no longer dominated solely by DeepSeek. Qwen, Minimax, and MoonshotAI all saw significant adoption shifts in mid-2025.

...and Google Workspace just dropped this. If this can work even moderately well, however you feel about it the economic reality is simple. The workspace drudgery that kept a whole level of entry roles is going to get demolished.

workspace.google.com/studio/

It is a good reminder to build your own, reproducible benchmarks on your own tasks, instead of relying on vendor benchmarks or marketing hype.

None of this is to say year-over-year model progress has not been genuinely impressive. Just be cautious, and run your own tests.

Whether it is Gemini supposedly solving full math problems with work with NanoBanana Pro or generating accurate isometric and perspective views directly from floor plans (literally the work I am doing right now—it makes mistakes a lot and is not consistent at all).

In the piece I walk through the setup, limitations, and what this means for homework, in-class work, and “authentic” tasks in secondary math.

At that level, standard take-home math is no longer AI-safe in any meaningful way. As a classroom teacher, that matters more to me than another benchmark. If a student wants to outsource their worksheet, friction is low and errors will not expose them quickly.

Short version: o1 pro only hit the 4/4 reliability bar on 40/60 questions (67%) and got 177/240 attempts correct (74%). Under essentially the same framework, GPT-5 hit 48/50 questions at 4/4 and 392/400 attempts (98%). We moved from 1-in-4 errors to about 1-in-50.