Generative AI and the Near Future

Generative AI entered public life fast enough that sober description got drowned out by two loud voices: one promising imminent superintelligence, the other dismissing the whole thing as a party trick. Both are selling something, and both leave you worse at using the tools actually in front of you. This closing lesson is about calibration — holding real, measurable progress in one hand and genuine uncertainty in the other, so your expectations track evidence rather than headlines. The durable skill you carry out of this module is not a prediction about the next decade. It is a habit: asking what a system has been shown to do, where it fails, and who checks its work.

Demonstrated capability versus speculation

Draw a firm line between what a model has been observed to do reliably and what someone claims it will do. Generative models predict the next token given context, and that mechanism produces genuinely useful drafting, translation, summarizing, and code. It does not make them conscious, and it does not make them artificial general intelligence. Marketing tends to blur "can sometimes" into "can," and "might one day" into "will." Naming which claim you are looking at is half the work.

Progress is real but uneven

Capability has advanced quickly, but not smoothly. These systems are strong at fluent language and pattern completion, and weak or unpredictable at reliable reasoning, arithmetic, staying factual, and knowing what they do not know. Improvement in one area does not guarantee improvement in another. Drawing a straight line through recent gains is extrapolation, and this field's own history is a warning against trusting such lines.

The open questions have no settled answers

Reliability, employment, misuse, and regulation are all live and genuinely unresolved. Will hallucination be tamed? Which tasks get reshaped rather than replaced? How is deception or fraud constrained, and by whom? Anyone offering certainty here — utopian or apocalyptic — is offering confidence the evidence does not support.

A vendor says: "Our AI writes production-ready code." Break the claim apart. Demonstrated: it often drafts plausible code. Assumed: that the code is correct, secure, and maintainable. The test: run it, review it, probe the edge cases. Calibrated reading — a capable assistant that speeds up a competent reviewer, not a replacement for one.

The opposite error is just as costly. Someone declares, "It's only autocomplete, ignore it," and refuses to touch the tools. Meanwhile colleagues who verify their outputs reliably draft and summarize faster. Blanket dismissal is as miscalibrated as hype; it forfeits real, demonstrated value. The answer to overselling is not cynicism — it is calibration.

AI has been over- and under-hyped before. In 1973 the mathematician James Lighthill reported to the British Science Research Council that AI had failed its grand promises and worked mainly on toy problems; the report helped trigger deep funding cuts in the UK, echoed by reductions in the United States. That was the first "AI winter." Enthusiasm returned in the 1980s with commercial expert systems and specialized LISP machines — until that hardware market collapsed around 1987 and a second winter set in. Each cycle followed the same shape: bold promises, a boom, disappointment when reality proved narrower, then a bust that under-rated the genuine progress that had actually been made.

• "Fluency means understanding." Predicting tokens well is not comprehension or consciousness.
• "Recent speed guarantees a fixed trajectory." That is extrapolation; gains are uneven and the history includes reversals.
• "The hype is either all true or all false." The field is characteristically over- and under-hyped at the same time.
• "Once it is good enough I can stop checking." Reliability is unresolved, so verification stays in the workflow.

Sort a stack of real-sounding AI claims into demonstrated, plausible-but-untested, and speculative — then name the single check that would move each one to firmer ground.

Think Like a Maester: The AI field has been wrong in both directions before — promising too much, then dismissing too much. Bet on calibration, not on either crowd.

1. What is the difference between a demonstrated capability and a marketing claim?
2. Why is "progress is fast, so it will keep accelerating along the same line" an unreliable argument?
3. Name two genuinely open questions about generative AI and explain why they resist confident answers.
4. What pattern was common to the earlier AI winters, and what does it teach about today's hype?
5. What practical stance keeps human judgment in charge when you use generative AI?

Generative AI is a capable, fallible tool: real progress, uneven strengths, and open questions that honest people cannot yet close. The history of AI winters shows the field swinging between breathless optimism and blanket dismissal, and being wrong both ways. Calibrated expectations beat both. Use the tools, keep verifying, keep learning, and keep human judgment in charge — that stance ages better than any prediction.

• AI vs Machine Learning
• Artificial Intelligence