Is AI getting dumber?
Notes From the Desk: No. 16 - 2023.11.30
Notes From the Desk are periodic posts that summarize recent topics of interest or other brief notable commentary that might otherwise be a tweet or note.
Lobotomized AI, is AI getting dumber?
It seems there is a recent trend in LLMs performing much worse than usual. Users of several different LLMs from different companies are reporting similar behavior.
Most suspect this is a side effect of continued “safety” training. The LLMs are now giving unusually brief answers while often telling the users to do the rest of the work themselves :-)
These are going to be major issues for business adoption of this technology. Indeterminate behavior as a baseline combined with new indeterminate behavior on every upgrade is in direct opposition to the needs of most businesses that must have a high level of reliability for their processes.
As stated by Brian Roemmele
“This prompt drift issue with OpenAI products has rapidly devalued its products.
Now imagine you worked on a spreadsheet for months and came to it one day and everything is broken, because of “safety” and alignment?
Welcome to the quagmire of cloud AI—you have no control.”
More questions about LLM Intelligence
If you spend any time researching how well LLMs are performing for people you will quickly see extreme contrasting opinions. They are seemingly both awesome and terrible. It seems rather difficult to discern what they are actually good at performing reliably.
A new paper raises more questions about their ability.
In spite of being successful at tasks that are difficult for humans, the most capable LLMs do poorly on GAIA. Even equipped with tools, GPT4 does not exceed a 30% success rate for the easiest of our tasks, and 0% for the hardest. In the meantime, the average success rate for human respondents is 92%.
LLMs are great at creating fiction. So scams and marketing are not so much hindered by hallucinations or inconsistent results. However, businesses that are hoping to use LLMs as a means to fully automate their processes may not be getting any closer to that being a reality.
Fake AI generations intersect with the real world
A tech conference called DevTernity seems to have used AI to generate fake participants which resulted in Microsoft and Amazon dropping out of the software conference as the implication seems that DevTernity used AI to make the participants appear more diverse for the acceptance of the tech crowd.
There are also livestreaming influencers that are totally AI generated. Welcome to the fake AI world of e-commerce.
“a swarm of Chinese startups and major tech companies have been offering the service of creating deepfake avatars for e-commerce livestreaming. With just a few minutes of sample video and $1,000 in costs, brands can clone a human streamer to work 24/7”
If you haven’t yet had a chance to read “AI accelerates post-truth civilization”, it is a much more in-depth coverage of deepfakes and their consequences to society. Most importantly, I cover what will likely be the suggested solutions that may pose additional threats to your liberties.
No compass through the dark exists without hope of reaching the other side and the belief that it matters …