Thread by @Moonalice

🧵
Today I was on @SquawkStreet on @CNBC to talk about the bubble in AI stocks. @GoldmanSachs put out a report warning capex for AI was way too high, given lack of high value use cases. Goldman is right on, but does not address other crippling issues facing AI. Thread.
1/15

https://twitter.com/squawkstreet/status/1810359584063705144

.@GoldmanSachs focuses on capex for compute and data centers. Numbers are >>$100B so far, rising by >$10B a month. Key spenders — MSFT, nvidia, Google, Amazon, Meta — have tons of cash, and dominate S&P500. If they are wrong … disaster for market.
2/15
goldmansachs.com/intelligence/p…

Goldman is not the first Wall St firm to raise concern. Barclay’s was. The VC firm @sequoia also issued a report on capex in AI, saying that industry needs $600 million in revenue to justify the existing investment in compute and cloud.
3/15

sequoiacap.com/article/ais-60…

When firms like @sequoia and @GoldmanSachs tell me there is a bubble, I pay attention. When their argument is based narrowly — on discretionary capex — alarm bells go off. Why? Because there are lots of other reasons to think generative AI is going to disappoint.
4/15

AI issue #1: power. Generative AI is breaking the power grid nationally and accelerating climate change. This would be a problem if AI created big value, but it doesn’t. MSFT and GOOG promised to be carbon neutral in 2030. MS carbon is up 30% in 2 years; Google 45% in 5.
5/15

AI issue #2: water. Water is required for semiconductor fabs, training data sets, and cloud services. One report suggests 1/2 liter of water is ruined every time you query a chatbot. Many fabs and data centers are in desert areas.
6/15

AI issue #4: copyright. GenAI companies have stolen huge amounts of copyrighted content to train LLMs. Their goal is take away jobs from the people whose content they stole. There are several problems with this.
7/15

AI issue #5: privacy. When we started using cloud services for email, word processing, spreadsheets, storage, social, etc. we did not give permission for our content to be used to train LLMs. And yet that is what is happening.
8/15

AI issue #6: security (esp. NatSec). Big Tech LLMs create vulnerabilities for anyone who uses them. It is almost certain that agents of foreign powers have infiltrated AI vendors. Huge asymmetry of risk relative to
China.
9/15

AI issue #7: disinformation. LLMs were designed to pass Turing test by fooling humans. Optimized for plausibility, not factuality. Volume of nonsense is a huge threat to search and to quality of online information.

Disinfo may be most valuable use case of LLMs.
10/15

AI issue #8, part 1: toxic use cases. Four best use cases so far are disinfo, deepfake porn, spam, plagiarism. Positive use cases are still mostly speculative. One that looked good — programming — has turned out to be a mixed bag, at best.
11/15

spectrum.ieee.org/chatgpt-for-co…

AI Issue #8, part 2. The early efforts to apply generative AI to web development have been disappointing.
12/15

baldurbjarnason.com/2024/new-web-d…

AI Issue #8, part 3. Thanks to the GenAI mania, people are trying to apply LLMs to everything, including a wide range of use cases where it is guaranteed to fail. People are going to be harmed, some in ways from which they will not recover. Safety should be a priority.
13/15

Conclusion: LLMs are not intelligent. They use statistics to find the most suitable next word, paragraph, image. They only know their training set. What they do best is to BS you.

More evidence supports the view that LLMs are a scam than the Next Big Thing. Beware.
14/15

🧵Coda: BigTech dominates the S&P500. When the AI bubble bursts, Big Tech will not go broke, but all the forecasts will collapse, taking stock prices with them.

No way to know when the bubble will burst, but it will happen. Be prepared.

Thank you, @SquawkStreet!
15/end