I’ve written this blogpost, this isn’t just “AI is bad”, but it’s mostly “AI is almost all the time bad”.
The focus is on the usage as a tool, and maybe bringing some more context to the conversation with the pointing out that the problem maybe is on how the tool were created, and how it’s beying pushed. Not sure.
Let me know what you folks think about it.
You must log in or # to comment.
The creation is also the problem. Pillaged training data - problem.
The older AI society destroying models – the ad and recommendation algorithms – are also bad, as is their their training data.
The problem is capitalism and what it’s using AI for. LLMs and their pattern recognition could be very useful if we used them for what they are good at, but capitalists want to replace people with it to avoid paying wages.
There are use cases that have previously been impossible in coding, but are now relatively simple to resolve due to LLM. Like text categorization and summarization, which were previously near impossible to code. Nothing “ai” about that, it just uses the statistical nature of LLM’s.
text categorization and summarization, which were previously near impossible to code
Not saying that there aren’t any coding challenges that were impossible before AI, but these are bad examples. 10+ years ago Reddit was already infested with “summary” bots that summarized articles with high accuracy, often better than the AI generated summaries I see nowadays… especially when you consider the computational cost. AI summarization requires a $1000+ GPU. “Dumb” summarization algorithms can run on a phone from 2015.
Gonna need some references on that. To my knowledge there is no other algorithms/tech to do that kind of summarization than these kind of token prediction. Yes, reddit might have been doing that kind of things, but under the hood that is 100% same base tech than in GPT, just sooner version of it, with similar kind of GPU consumption.
There’s “LexRank” from 2019 that uses a graph-based approach, similar to the PageRank algorithm that originally made Google so successful. An older version called TextRank has been around since 2004: https://github.com/crabcamp/lexrank
SMMRY has been around since 2009 and I believe was behind most of the tl;dr style bots that were common on Reddit in the 2010s. The original implementation was rules-based instead of a transformer architecture, but it appears the company has pivoted into AI in recent years. Here’s an article about it from before they made the switch: https://medium.com/@mplaut929/smmry-the-algorithm-behind-reddits-tldr-bot-c268722a4c27
Neither of these use(d) the GPT architecture or needed a GPU to run.
Sure LexRank works, but it cannot reword the text, it can only reorder/remove sentences or words. It has use cases, and it cannot hallusinate, because it must just reuse the parts of the input. Unfortunately a good summarization requires often rewording.
But, I stand corrected, I did not know reddit used it.
Unfortunately a good summarization requires often rewording.
Agree to disagree, I guess.
Summarization is lossy no matter what, but I’d much rather the lost data be deterministic, and the preserved data be guaranteed to represent the original text. AI summarization is like a bad game of telephone, and it’s hard to be sure when it’s given you a genuine summary or injected its own bias, missed key details, etc. And that’s assuming it doesn’t just completely hallucinate.
Can you link such an example summary?
