Superintelligence Isn't Close
I do not believe that Superintelligence - AI with 1000x the intelligence of the smartest humans across domains - is around the corner.
We hear from the CEOs/promotors of OpenAI, Anthropic and others that SI is 2-3 years out, or will be here “soon”. Really? “Yes” AI industry promotors say. They point to the fact that SSI founder Ilya Sutskever rejected META’s $32 billion offer as evidence. “Ilya must have seen something”, they say. What they don’t say is that SSI’s CEO, Daniel Gross, took the money and is now part of META’s AI group. I guess Daniel didn’t see what Ilya saw.
If SI is around the corner, where is GPT 5?
Why was the last wave of frontier models only incrementally better than the previous crop?
Why do the frontier models still hallucinate?
You can’t reliably use LLM’s for complex calcs. I trust my calculator far more than an LLM to run numbers.
SI feels like it is 50 years out, which is actually the time horizon that LLMs return when queried about SI.
Everyone is getting played. NVDA will continue to benefit as the AI arms race continues, but this game is about LLM training and inference, not actual SI.
Will LLMs continue to improve? Yes, but not exponentially with each release. That’s not SI. I don’t believe that a single language model ever gets to SI. Perhaps a complex network of LLMs, ML-powered data sets and the like, each with different specialties, but not a single language model.
Be prepared for the definition of SI to change. I think we have one more big AI generation in us, but I don’t see MSFT for example investing $160-200 billion in AI in FY 2027. Not for the minimal Revenue and EBITDA lift that AI has delivered thus far, unless your name is NVDA.
Ad
FinModel.app: 12 quarters of income statement history in less than 3 seconds.