It's getting smarter, but it still makes a lot of mistakes, while financial markets are expected to fall..
By: Andy Kessler, businessman, investor and author / The Wall Street Journal
Translation: Telegrafi.com
Like indoor plumbing or avocado toast, artificial intelligence [AI] is set to have a huge impact. But don't expect a linear growth. Here's the good, the bad, and the ugly.
• Good. AI is very smart. ChatGPT scores in the 93rd percentile on the SAT reading and writing tests [Scholastic Assessment Test - tests used to measure the readiness of high school students to attend college in the United States] - [scores] 710 out of 800. And, 700 in math. Can also achieve a score of 3 or higher on most AP exams [Advanced Placement - standardized tests given to high school students in the United States in May]. If he could volunteer to build huts in Costa Rica and play the clarinet acceptably, ChatGPTcould be accepted to the University of Michigan.
Generative AI can also pass all levels of the Certified Financial Analyst exam [Chartered Financial Analyst - globally recognized professional qualification], the Unified Bar Examination in the United States [Bar Exam Uniform] with 97 percent accuracy for 2023, as well as the American medical licensing exam [United States Medical Licensing Examination] for the year 2023. Sure, if I could remember the relationships between every word written in the last 500 years, I'd probably pass these tests too. But would you believe it? ChatGPT-any of these professions? Hmmm. Our tests have their limitations.
Fraud with ChatGPT is so widespread that professors have reverted to the old four-page notebooks for exams. What's next, parchment paper? That's not progress. But, notice, analytical thinking will increasingly be done through commands [Prompt] for ChatGPTand verification of the results. Can you evaluate this? Not yet.
So much of today's education is based on learning to pass tests. Quizzes, tests, lectures - most of education is upside down. Rote learning is out of fashion. Lectures too. The future is self-paced, participatory learning on projects where AI can help and assess at the same time. Build a human-powered (virtual) rocket to Mars - learn calculus, physics, biology, economics, psychology, and materials science to make it happen.
Also impressive is the speed with which companies are adopting AI for customer service. Virtual agents using text, and increasingly voice, can answer questions. In my experience, anywhere from 30 percent to 70 percent of answers are generated automatically before difficult questions are passed on to real people. Most tell me that costs are reduced by about 30 percent. After chatbots, this may be the biggest adoption of AI. Be warned: it still makes mistakes.
• The Bad. Yes, “hallucinations” are a big problem. Even the term “hallucination” is misleading. Just tell it like it is: AI makes mistakes. Glitch. Flaws. It can be stupid. Because generative AI is a statistical model, it often doesn’t hold up from one query to the next. Even in customer service, companies are careful. Mistakes hurt. Brands can be tarnished. Liability is a big issue for externally-driven AI.
For coding, Mark Zuckerberg from Meta says their AI is equivalent to “mid-level engineers.” Programmers like “interesting” tools [Vibe Coding] as Cursor and repeat. Company AI anthropic provides many of these tools, but few trust the results, which are often full of errors [Bugs]. Programmers, whether mid-level or not, have become testers who fix AI code. More productive? Yes, but differently.
Current large language models are trained through brute force. With many GPU [Graphics Processing Unit] chips Nvidia-s related to terabytes of memory. DRAM memory prices, which typically fall by 30 percent per year, are expected to rise by 15-20 percent in 2025. AI is sucking up every joule of free energy. In Silicon Valley [Silicon Valley] there's talk of kilowatts per token as a unit of measurement for AI production. That won't last. Keep an eye on memory prices as a warning system for demand. I'd also watch Japanese bond yields as they trade with [Yen Carry] could be the source of hot air in the stock markets. Even for cryptocurrencies.
More efficient AI methods are coming. Check out the new $12 billion company from the former CTO OpenAI, Mira Murati - Thinking Machines Labs, which trains existing AI models to become experts at narrow tasks. Teaching AI how to learn is a lot like rethinking human education. The efficiency could reduce demand for data centers.
• The Bad. The stock market is crazy about AI. But is investment in AI data centers ahead of real, revenue-generating demand? It could be. We’ve seen the superstructure in every cycle: memory chips, telecoms, fiber optics. On the horizon are low-power chips and more efficient AI algorithms. The brute-force game could change.
The planned capacity of over a trillion dollars by 2028 is far ahead of any near-term revenue opportunity. And, markets are volatile. Jeff Bezos likes to recall that in the early 2000s, Amazon shares Amazon- fell from $113 to $6 even though the company met every forecast. Markets have priced AI shares based on expectations for three to five years out. This can be reduced to a more normal timeframe of six to 18 months. As in the case of Oracle infrastructures -it last week. It happens in every cycle: What have you done for me lately? New supply and innovation mean that AI prices could fall faster than expected. So could stocks.
Short-term ugliness is good news in the long run. In Silicon Valley, prices always fall. When they don't, I worry. Lower prices mean that the elasticity of demand is coming into play: new applications are emerging to take advantage of the cheaper functionality. Cheaper laser printers, digital cameras, smartphones, and now AI. Just follow the zigzag line. /Telegraph/
