Generative AI can also pass all levels of the Chartered Financial Analyst exam, the 2023 Uniform Bar Exam and (with 97% accuracy) the 2023 U.S. Medical Licensing Examination. Sure, if I could memorize the relationship between every word written in the past 500 years, I’d probably ace these tests too. But would you trust ChatGPT in any of those professions? Hmmm. Our tests have their limits.

Cheating with ChatGPT is so rampant that professors have reverted to those silly blue books for exams. What’s next, parchment paper? This isn’t progress. But note, analytical thinking will increasingly be done through the prompting of ChatGPT and checking results. Can you grade that? Not yet.

So much of education is teaching to the test. Quizzes, exams, lectures and most of education is upside down. Memorization is out. Lectures are out. The future is both self-paced learning and participatory education on projects that AI can assist and evaluate all at once. Build a (virtual) manned rocket to Mars—learn calculus, physics, biology, economics, psychology and material science to do it.

Also impressive is the fevered pitch of enterprises adopting AI for customer service. Virtual agents using text, and increasingly voice, can answer queries. Anecdotally I’ve heard numbers ranging from 30% to 70% of automatically generated answers before hard questions are kicked to real, breathing humans. Most people tell me costs are lowered by about 30%. After chatbots, this may be the biggest adoption of AI. Warning: It still makes mistakes.

• Bad. Yes, hallucinations are a big problem. Even the name hallucination is a deceit. Just say it: AI makes mistakes. Errors. Gaffes. Flaws. It can be stupid. Because generative AI is a statistical model, there’s often little consistency from query to query. Even with customer service, enterprises are cautious. Mistakes hurt. Brands can be tarnished. Liability is a huge issue for outward-facing AI.

For coding, Meta’s Mark Zuckerberg says its AI is as good as “midlevel engineers.” Software folks like “vibe coding” tools such as Cursor and Replit. Anthropic’s AI enables many of these tools, but few trust their output, which is often riddled with bugs. Programmers, midlevel or otherwise, have turned into testers fixing AI’s code. More productive? Yes, but different.

Today’s large language models are trained via brute force. Tons of Nvidia GPU chips connected to terabytes of memory. DRAM memory prices, which normally drop 30% a year, are up 15% to 20% in 2025. AI is sucking up every spare joule of energy. Folks in Silicon Valley talk about measuring kilowatts per token to judge AI output. This won’t last. Watch memory prices as an early warning system for demand. I’d also watch Japanese bond yields as yen carry trades may be the source of the stock market’s hot air. For crypto too.

More-efficient AI methods are coming. Look at former OpenAI chief technology officer Mira Murati’s new $12 billion valuation company, Thinking Machines Labs, which trains existing AI models to become experts at narrow tasks. Teaching AI how to learn is much like rethinking human education. Efficiency may cut data-center demand.

• Ugly. The stock market is cuckoo for Cocoa Puffs over AI. But is AI data-center investment ahead of revenue-producing demand? Likely. We see overbuilding every cycle: memory chips, telecom, fiber. Low-power chips and power-efficient AI algorithms are visible on the horizon. The brute-force game may change.

The $1 trillion plus of planned capacity by 2028 is way ahead of any near-term revenue possibilities. And markets are fickle. Jeff Bezos likes to remind everyone that in the early 2000s Amazon stock went from $113 to $6 even though the company hit every forecast. This is the zigzag vs. a straight line up. Markets have been pricing AI stocks based on expectations three to five years out. That may shrink to a more normal six to 18 months. Like Oracle last week. It happens every cycle: What have you done for me lately? New supply and innovation mean AI pricing can drop faster than planned. Same for stocks.

Near-term ugly is long-term good news. In Silicon Valley, prices always go down. When they don’t, I get nervous. Lower prices mean elasticity kicks in: New applications open up to take advantage of the cheaper functionality. Cheaper laser printers, digital cameras, smartphones and now AI. Just watch for the zigzag.

Write to kessler@wsj.com.