The Problem with AI Is About Power, Not Technology

The reason it takes so much energy is that memory is separate from compute in most architecture s, so you end up getting an expensive bottleneck that is almost always the limit to performance on a given task. Most of the silicon area and power go into trying to minimize the hit of that bottleneck to raw performance.<p>Alternatives exist, they&#x27;re radically different to program, but I strongly believe offer very cheap Petaflops for the masses.

&gt; The cost for training ChatGPT-4 came in at around $78 million; for Gemini Ultra, Google’s answer to ChatGPT, the price tag was $191 million.<p>Yeah... the cost to create ChatGPT-4 isn&#x27;t the cost to train it. You would need like $1-4b to create a GPT-4 style model.<p>This is like people who say the iPhone is mAdE iN cHiNa, when you buy it for its software, which is all made here. Its cost isn&#x27;t the price you pay up front, it&#x27;s your LTV with Apple.

Yeah, all the pearl-clutching over Skynet-style doom scenarios seems to be:<p>1. Product aggrandizement. (&quot;It&#x27;s so awesome, it might even take over!&quot;)<p>2. Creating a fictional worry&#x2F;scapegoat so that individuals and consumers aren&#x27;t on-guard for more-immediate problems caused by human&#x2F;corporate decisions.