Every wave of AI progress invites two unhelpful extremes: “nothing meaningful has changed” and “AGI arrives next quarter.” Both are wrong. Progress is real. AGI—in the sense of broadly human-level general intelligence across the range of tasks humans can do—has not been reached. The useful conversation sits between those poles: what is working, what is fragile, and how we should build and govern systems now.
This is an expanded set of notes on that middle path.
Where we actually are
Modern machine learning—especially large models for language, vision, and multimodal tasks—has pushed capability on generation, translation, coding assistance, summarisation, and tool-using agents. The user experience shifted: natural language became a viable interface to software.
That is not the same as understanding, common sense in the human sense, or reliable autonomy in open-ended environments. Systems still hallucinate, miss context, inherit data biases, and fail under distribution shift. They can be extremely useful with boundaries. They are dangerous without them.
Holding both ideas at once is the adult stance: ambitious about applications, sober about limits.
Jobs, displacement, and democratization
AI will automate tasks inside jobs before it “removes jobs” in a clean line. That still matters. Task automation changes staffing mixes, junior apprenticeship paths, and leverage per person. Some roles shrink; new ones appear around building, evaluating, integrating, and governing these systems.
Democratization is the other side: smaller teams can ship capabilities that once needed large specialised groups. That is empowering and destabilising. Markets will reward people who can combine domain judgment with AI leverage—not people who only paste prompts, and not people who refuse to learn the tools.
The humane response is not denial. It is reskilling, redesigned workflows, and clarity about which human skills become more valuable: problem framing, verification, taste, accountability, and cross-functional leadership.
Safety needs defence-in-depth
Safety is not a single filter. It is layered:
- Data access controls and minimisation
- Model and tool permission boundaries
- Input/output checks and monitoring
- Human approval for high-impact actions
- Incident response when systems behave badly
- Organisational policies that match the technical controls
Defence-in-depth assumes failure. Prompts will be attacked. Models will be wrong. Integrations will be misconfigured. Design for containment.
Bias, copyright, and oversight
Bias. Training data reflects the world and the institution’s history. Outputs can amplify harm even when nobody intended it. For decisions that affect people, measure slice performance, keep humans in the loop, and be willing not to automate.
Copyright and IP. Generative systems raise unsettled questions about training data, output ownership, and similarity to protected works. Practically, organisations need clear policies on what data may be sent to external models, how outputs are used commercially, and how provenance is tracked—especially for code and creative assets.
Human oversight. Autonomy without accountability is negligence. Oversight should be risk-based: more human judgment where stakes are high, lighter touch where errors are cheap and reversible.
Regulation: inevitable and uneven
Regulation will expand—product liability concepts, sector rules, privacy regimes, transparency duties. It will be uneven across jurisdictions. Builders should assume that “move fast with customer data in a black box” will get harder, not easier.
The constructive approach is to engineer for auditability early: logs, versioning, data lineage, decision traces where appropriate. Waiting for the final statute before adding observability is how you get expensive rewrites.
How I think organisations should respond
- Separate hype use cases from durable ones. Prefer workflows where AI assists with clear evaluation criteria.
- Invest in data and governance as much as in models.
- Teach critical use — staff should know how these systems fail.
- Prefer modular architectures so models can be swapped and controlled.
- Put ethics and safety in delivery gates, not only in values posters.
- Watch labour impacts inside your own teams; redesign learning paths as tasks automate.
Closing
The current state of AI is powerful, incomplete, and unevenly understood. We have not reached AGI. We have reached a point where software can generate and act with unprecedented fluency—and where that fluency can mask error.
The leaders who navigate this well will not be the loudest predictors of utopia or doom. They will be the ones who ship useful systems with defence-in-depth, human oversight, and honesty about what the technology can and cannot do.