Completely agree. I do believe calling current run of ML/DNNs "AI" is stretching it (and I expect to be hit with the "AI is a moving goalpost" argument but seriously? is it a moving goalpost for those who know how the current generation of solutions work?).
> There's also no reason to think that 'intelligence' can be exponentially scaled or that any system can recognize how to improve itself -- or have the capacity to do so in situ.
There's no reason to assume otherwise. Exponential scaling (i.e. "intelligence explosion") may be a contestable assumption, but really, we're in big trouble if a general AI reaches human levels. We know this level is possible because we exist. Even if AI did just that, we'd have to contest with a new sapient being, and give it ethical consideration. But I see nothing preventing that AI from being as smart as 10 or 100 humans combined.
The rest really follows from nature of technological artifacts - if you can make one, you can make one thousand (applies to both hardware and software) and make it work as one; you can iterate on it, and if you can, then a human-level AI will be able to do it too.
But, thus far, our best AIs are simply pattern recognition systems strung together like Christmas lights.
There's also no reason to think that 'intelligence' can be exponentially scaled or that any system can recognize how to improve itself -- or have the capacity to do so in situ.