I was researching on the topic on the Engineering viability in terms of processing power needed to build something of what can be objectively be called as a True AI Agent, that can prove that its intelligent either through speech or navigation or just mix of 2 or more skills (Modalities) to prove one AI agent can merge learnings across its modalities and draw provable conclusions from its knowledge.
The best Computation you can extract from a Azure / Amazon HPC Instance per second: 1 PFLOP
Total Computation power of Brain per Second : 10^15 ( 1 Peta FLOPS.)
Seems like if Computation is not a problem ?
Is it a tech problem ? Meaning we havent cracked the algorithm yet which can extract data from any kind of problem by itself without the need for giving it humongous amounts of data to it?
From a semiconductor perspective we can make enough compute for the purpose, but the missing magic is the right code.
When that right magic comes along compute ability/technology will already be far more than what would be required (i.e. your AI will run far faster than a human).
Current models alread learn with far more information than a stadium full of people can hold, understand and process. So, that magic algo will also be capable of learning with far less data.
Think about how much information as a digital equivalent of text a human needs to become “intelligent”…
What’s to say we haven’t already cracked it, but the algorithms are slow learners? It takes years for a human to display significant intelligence. Very expensive.