Trump humiliated by fact-check after claiming Canada used AI for Reagan ad
www.irishstar.com
The fact-check features a link to a transcript of Reagan's address from the Ronald Reagan Presidential Library, as well as full audio

President Donald Trump was once again fact-checked following an embarrassing tantrum over a Canadian advertisement.

The ad quoted former U.S. President Ronald Reagan from 38 years ago criticizing tariffs – a policy tool Trump frequently employs. The commercial includes audio clips from an April 25, 1987 radio address where Reagan stated: "Over the long run such trade barriers hurt every American worker and consumer.‘’

    • ms.lane@lemmy.worldEnglish
      11·
      7 hours ago

      From chatgpt-

      Prompt: What is a very rough ballpark figure of how long it would take a Symbolics 3600 to process this request on you?

      That’s a fascinating question — and quite a big gap in computing history!

      Let’s estimate very roughly:

      1. Symbolics 3600 performance

      Released around 1983.

      Based on a 10 MHz 68000-series CPU with extensive microcode for Lisp operations.

      Effective performance: roughly 1 MIPS (1 million instructions per second), though Lisp code was often slower due to dynamic typing and garbage collection.

      Floating-point and memory bandwidth were also very limited — on the order of kilobytes per microsecond.

      1. This request

      Let’s say your request (this chat exchange) takes me — a large language model with billions of parameters — maybe a few hundred billion floating-point operations (FLOPs) to process. For perspective, a modern GPU (like an NVIDIA A100) can do around 10¹⁵ FLOPs per second (1 petaFLOP) for AI workloads.

      1. Comparison

      If we very loosely assume:

      Symbolics 3600: ~1 MFLOP (10⁶ FLOPs/sec, being generous)

      Modern GPU: ~10¹⁵ FLOPs/sec

      That’s a factor of 10⁹ difference.

      So if a modern GPU takes about 0.1 s to process your request, then the Symbolics 3600 would take:

      0.1 s × 10⁹ = 10⁸ s ≈ 3 years

      And that’s if it could even hold the model in memory, which it couldn’t — it had only a few megabytes of RAM, versus the hundreds of gigabytes (or terabytes) needed.