122 - Taming the Elephant
Yesterday I discovered a problem and a solution.
The problem was that Adolf Hitler was topping the charts on AI models sorted by "Likes", which was disturbing on several levels. After the AI music that led me there I was expecting perhaps Frank Sinatra or Freddie Mercury, but no, the top 2 were Hitler and Putin.
The solution I located on the same site was a more effective method of communicating with the US's Twitter-minded tech industry and the people who burn money on their black hole of bad ideas. That solution, as you can no doubt hear, is speaking to them in the voice of Rick Sachez from Rick and Morty.
I've often repeated the phrase "When in Rome, Speak the Language.", and a drunken careless inventor who solves every problem with trivial and flippant investments of time and energy is the quintessential thing that many in the tech industry dream of being. Had I discovered this two weeks earlier I'd definitely have put it to the test in the toxic waste dump that is Twitter.
While trash technology like LLMs hasn't actually made a shred of meaningful scientific progress (in the past 4 years), they've only deployed larger models to a consumer audience, the people working on voice cloning and image generation have been making steady progress in recent years.
Those technologies could be taken much further with the integration of working cognitive architectures, but how much further remains untested, unlike the domains of NLP and NLU where we've already demonstrated precisely how trashy LLMs are. What we know for certain is that they could improve the data quality well beyond human capacities, and could fine-tune such models in ways not accessible to humans, as well as assist in the design of more optimal architectures for them through research and the fruits of that fine-tuning process.
The elephant in the room for that topic is the risk of bad actors and deepfakes, which are already growing severe at the current levels of the technology. However, raising the technology above human capacities also means the ability to inject methods of "watermarking" and tracking into any output that bad actors have no realistic hope of detecting or masking. The improvement itself becomes what enables the solution to this problem, taming the elephant, rather than ignoring it.
For anyone curious about the state of the vocal cloning technology today, I've temporarily set up an example mixing together several such voices, which can also be used to make fun of the narrow AI worshippers on Twitter.This encapsulates my sense of humor on several levels.
More details on the websites and software used in the vimeo link's description. The Rick Sanchez voice model is shown here.