The truth lies somewhere in the middle. And it's more complicated than either side wants to admit.
This Year's Model Explosion
OpenAI launched GPT-5 in August as its new default model—unified reasoning and multimodality in one package. Google responded with Gemini 3 Deep Think in December. Anthropic took the lead in coding with Claude Sonnet 4.5, topping the SWE-bench. Grok 4.1 surprised everyone with a 2M token context window, while Mistral Large 3 proved that European AI can still compete.
But the biggest story was DeepSeek. In January, the Chinese firm launched R1—an open-source reasoning model that matched American frontier models. Training cost: $5.6M. Nvidia's stock dropped 18% in a single day. DeepSeek proved you don't need billions to build competitive AI. This has massive consequences for the entire industry.
The Race Toward AGI: Who Asked For This?
We are in the midst of an arms race between tech giants. The goal is AGI (Artificial General Intelligence). But who actually asked for it?
Tristan Harris from the Center for Humane Technology puts it bluntly: We are building systems smarter than ourselves without knowing how to control them. And we are doing it at a pace where no one has time to consider the consequences. His point is simple: there is no definition of wisdom that doesn't involve restraint. That applies to technology too.
Yoshua Bengio, Turing Award winner and one of the "godfathers of AI," is so concerned that he founded LawZero in June with $30M to develop provably safe AGI. His argument: You don't want superintelligence controlled by one person, one company, or even one government. You need checks and balances. In October, he joined Geoffrey Hinton in signing a declaration calling for a ban on superintelligence development until we know how to control it. These aren't activists; these are the scientists who built the technology.
The Job Debate: What do the Sources Actually Say?
I have reviewed what the WEF, ILO, Federal Reserve, and EU are actually saying. The picture is more nuanced than the headlines suggest.
World Economic Forum (WEF): Their Future of Jobs Report showed a positive net result (more jobs created than lost by 2030). However, 4/10 employers plan to reduce their workforce due to AI, and nearly 4/10 skills will be obsolete by 2030. These new jobs aren't 1-to-1 replacements; they require different skills and occur in different locations. The WEF also warns of an "AI precariat"—millions losing not just income, but identity and purpose.
ILO and the UN: Their research shows 1 in 4 jobs is exposed to GenAI, but transformation, not replacement, is the most likely outcome. Women are hit harder because they are overrepresented in administrative roles.
The Federal Reserve: Jerome Powell noted in November that job creation is "pretty close to zero." He linked this directly to AI—companies are producing more with fewer people. Amazon cut 14,000 middle managers; Verizon cut 13,000; IBM has frozen hiring. Economists call this "The Great Freeze"—a market where companies grow without hiring.
From Prompt Engineering to Context Engineering
A technical shift is occurring that matters for practitioners. In 2023–2024, it was all about writing the perfect prompt. In 2025, the industry realized that was the wrong focus.
Tobi Lütke (Shopify) and Andrej Karpathy agree: It's not about the prompt—it's about providing the context that makes the task solvable. In serious LLM applications, the prompt is a tiny part. The rest is conversation history, retrieved documents (RAG), tool outputs, and agent state.
The takeaway: Most agent failures are context errors, not model errors. Invest in RAG systems, MCP (Model Context Protocol) adoption, and memory management. Think in systems, not single prompts.
Human-Centric AI: Still the Best Path
Research points in one direction: the biggest gains lie in augmentation—AI that enables humans to do new things, not just the same things cheaper. McKinsey finds that organizations with high AI ROI prioritize augmentation over pure automation.
What Should You Do?
Build Hybrid Intelligence: Focus on resilient skills like interpersonal communication, creativity, and cross-disciplinary thinking. Nurses and plumbers are among the safest jobs for a reason.
Learn Context Engineering: Shift your focus from "prompts" to "systems" (RAG, MCP, memory).
Be Realistic About the Transition: A large portion of your current skills may be obsolete in five years. This isn't a reason to panic—it's a reason to plan.
The Short Version
2025 has been a year of exponential acceleration. The models are better than ever, but the transition will be uneven and painful. Net positive in the medium term? Likely. Transformation over replacement? For most, yes—but not for everyone. You don't need AGI to create value; you need context engineering and a human-centric approach.
Stefano Vincenti is GenAI Product Lead, External Lecturer at the IT University of Copenhagen, and co-founder of BotTellMe.
Have questions about AI strategy? Contact me – I'm happy to help navigate these complex questions.