Over many conversations and watching tumultuous market movements whipsaw sectors, I have come to the conclusion that a lot of people have a lot of opinions about Artificial Intelligence (AI) without actually understanding either these sectors or even the technology.
Firstly, what is AI?
Well, AI is a marketing term. The technology is not intelligent, but rather operates based on inferencing (they literally call the process of using AI as “inferencing”, as opposed to creating AI, which is “training”). The technology inferences from large data sets to arrive at correlations and using, effectively, mean reversion to fit the best correlations together depending on a user request or inputs.
I.e. Based on all the data Gemini has been trained on, when a user inputs “What is the best, super short, super easy recipe for chocolate cake.”, the strongest correlations (thus, the most likely “best” answer) is, on average, the below:
The 3-Ingredient Miracle Cake
- 3 large eggs
- 170g (1 cup) Semi-sweet chocolate chips (high quality makes a difference here)
- 115g (½ cup) Unsalted butter, cubed
Process:
- Prep: Preheat your oven to 180°C. Grease a small (6-inch or 15cm) round cake tin and line the bottom with parchment paper.
- Melt: Melt the chocolate and butter together in the microwave in 30-second bursts, stirring until smooth. Let it cool slightly.
- Whisk: In a separate bowl, whisk the eggs with a hand mixer until they double in volume and look pale and foamy (about 3–5 minutes). This is the “secret” to the lift since there’s no flour.
- Fold: Gently fold the melted chocolate into the eggs using a spatula until just combined. Don’t over-mix, or you’ll lose the air!
- Bake: Pour into the tin and bake for 22–25 minutes. The edges should be set, but the center should still have a slight jiggle.
Source: Gemini with Gemini’s source being all the chocolate cake recipes on the Internet.
A little statistical humour, but “AI” should rather be called a “Mean Machine”. The marketing teams at AI firms have convinced everyone that the statistical inferencing technology is “intelligent” and, to that, I give them credit.
Likewise, the cutely named “hallucinations” that AI can generate from time to time is just because the statistics sometimes point towards an answer (i.e. generative output) that is not actually true and/or does not actually exist. The correlation exists (or is implied from existing data) but it does not actually exist in the real-world (e.g. hallucinating legal precedents from courts cases that never happened).
Re-phrased, AI does not “hallucinate”, it just gets some things wrong and has no morals (because you need self-awareness and intelligence to know what these are) and, thus, does not just say “I don’t know”.
Once again, the marketing teams at AI firms have spun this fact into a lovely marketing phrase to make users more comfortable with it.
Last week, many stocks in insurance brokerages around the world sold off sharply like Brown & Brown and AJ Gallagher. Why? Because Insurify launched a ChatGPT app that can help users research, browse, compare and, ultimately, buy insurance quotes.
So who needs insurance brokers?
Yay! All the world’s problems are solved… But what if you have bespoke, complex risk needs and it is not just simple motor cover you need? What if you risks do not look like anyone else’s risks?
Likewise, last week, the stocks in financial advisory firms like Charles Schwarb, Raymonds James Financial, and Stifel Financial also tanked sharply. Why? Altruist announced the launch of a new AI-powered tax planning tool. This allows financial advisors to create personalized tax strategies for clients by automatically analysing tax forms, financial statements, and other documents within minutes.
Yay! All the world’s problems are solved… But what if it hallucinates, and you suddenly have huge tax penalties? Who do you sue? What if tax laws change, but the data it was trained on is out of date and its advise does not?
Digging deeper, and consultancy firms like Accenture, have seen their share prices sell-off steadily across the last year and a bit, because apparently you can spin up deep research pieces on any subject matter that will make consultancies redundant?
Yay! All the world’s problems are solved… But which CEO and CFO are going to pin their careers on a ChatGPT report suggesting massive corporate restructuring and directing huge capital decisions for a group when convincing their Board’s, shareholders and investors that this is the correct thing to do? (Even more subtly, to implement actual AI into a complex business that is not filled with AI and IT specialists, well, you need more consultants like Accenture and not less… Hence, we invested our funds in Accenture recently!)
Thus, what can AI do?
Well, it can (statistically) correlate vast amounts of data reasonably well and much, much faster than any human anywhere. Once correlated and, if trained and setup correctly, it can therefore automate some tasks and approximate processes wonderfully and at a faster speed and lower cost than humans. Used this way, and AI is spectacularly.
But, what can AI not do?
AI cannot give you assurance that it is correct and you can base your entire life/wealth/career/business on its outcome. AI does not do bespoke well (if the need differs from the data it was trained on, it risks hallucinations?). AI does not do creative well (if it doesn’t exist, AI cannot come up with it). And AI cannot form, maintain and manage trust amongst human beings (and, for economies to operate, people need to trust each other, for example, auditors issuing an audit opinion or advisors structuring your family trust that should out live you, or a rating agency placing their reputation on the line with a credit rating).
Or am I wrong? Write a comment below and let’s discuss this uncertain future…
ORIGINALLY APPEARED ON MONEYWEB.