The Coming AI Crash: Faking Intelligence Versus Real Thinking

 The Coming AI Crash: Faking Intelligence Versus Real Thinking

Yesterday's mini-crash of AI stocks is not surprising. AI related stocks have been priced to perfection reflecting the view that AI is a revolution on par with the industrial revolution or the transition from horse to combustion engine for transportation. But the crash is just the first symptom of a greater problem, namely that the technology does not live up to the hype. The large language models are impenetrable black boxes due to their nature; they are nonlinear statistical models with huge number of parameters. It is the huge number of parameters that make them power hungry beasts of requiring GPUs to estimate them. For details on the crash itself, read https://thetechcapital.com/tech-stocks-tumble-as-deepseek-triggers-1-trillion-market-crash/. 

The term 'training an AI model' means exactly the same thing as estimation of a large nonlinear statistical model like the one below. I am a time series statistical analyst by training so these models are familiar to me. The large language models are gigantic versions of the relatively simple neural net example below. Because they are Black Boxes no one knows how or why they work or their weaknesses or strengths. Nonlinear models with lots of parameters can approximate any mathematical function with arbitrarily high precision. This makes them great for summarizing data. So estimate a model on all known human language references to lemon trees can provide a useful summary of what the human race knows about lemon trees. 

But as people have discovered, the models fail to solve simple riddles that they have not encountered because they are just data summaries. ChatGPT has failed to solve this riddle: Sally has two sisters. Her mom is Florence. How many daughters does Florence have? 😃 So my concern is that these AI models into which vast sums are being poured are not really thinking machines, but just statistical models good as sniffing out relationships or finding patterns. Another example of a riddle stumping a massively expensive AI program like ChatGPT is below on the left hand side. 

If I am correct, then an excessive amount of money is being poured into this line of research and that means a crash is inevitable including related infrastructure like data centres and even some subsea cable investments. Large language models have limited reasoning power and in fact so little that they appear to rather ape intelligence as opposed to demonstrate real intelligence. Intelligence involves being able to infer new relationships. |


Simple Neural Net AI Model

Example of a Riddle That ChatGPT Could Not Solve


Comments

Post a Comment

Popular posts from this blog

Breaking Story: Facebook Building Subsea Cable That Will Encompass The World

Image
Several sources have whispered in my ear that META is planning a new 16 fibre pair cable that will encompass the world going from the US East Coast to the US West Coast via the Atlantic, Indian Ocean, and the Pacific. The most ambitious subsea project ever undertaken. I do not know the exact routing. I know that the cable will launch from the American East Coast and will go down the West African Coast to South Africa and then head straight to Mumbai. It is not clear if Europe will be online or not. From Mumbai it will head straight to Australia and then up to the US West Coast. I speculate that there may be branching units to Singapore, Malaysia, and Indonesia. But the basic routing is US/South Africa/India/Australia/US. What is not clear is if there will be branching units to add more countries to the cable.  This semi-secret cable reflects META's desire for network resiliency given the four month Red Sea down time that AAE-1 and other subsea cables suffered during the first half ...
Read more

Facebook's Semi-Secret W Cable

Image
 A few days ago I wrote that Facebook was planning a new subsea cable directly linking the US East Coast to South Africa, then up to India, onward to Australia, and finally to the US West Coast. Clearly resiliency is a big theme. It avoids Lisbon, Marseille, Egypt, the Red Sea, and Singapore. All choke points due to their telecom hub status. Note that the map below shows lots of branching units, but I don't have insider confirmation although they make perfect sense. The cable is 16 fibre pairs.  This cable may be heavily influenced by AI considerations. AI data centres require lots of space and power. The US, South Africa, India, and Australia are good candidates for AI data centres in terms of space. All of them except for South Africa have modern power grids. But South Africa is still the best place on the continent for large scale AI facilities. An AI theme makes particular sense given the cable's high latency. An AI data centre does not directly serve custo...
Read more

How To Calculate An IRU Price For a 100G Wavelength

The IRU price should reflect your forecast of future lease revenes. If the current MRC is high and prices have been stable and are expected to remain so, then the opportunity cost of selling an IRU is high. You are sacrificing a lot of lease revenue if you sell that capacity. In this case the upfront IRU price should equal the three year MRC multipled by 4 to 6 years. So if the three year Faster cable MRC is $17K, then 48x to 72x times that number is fair value. So 48*$17K=$816K. Plus a standard 4% annual fee to cover the wavelength's share of maintenance and operations.. So strong markets and tight capacity mean that leasing is more attractive than IRUs to the supplier. Vice versa, the buyers will be desperate to lower their long term costs via an IRU. In my example, $816K divided by 15 years equates to $4500 per month excluding O&M. Bad for the seller. Good for the buyer. That's why so many of you are looking for IRU capacity to resell to your clients, but finding few se...
Read more