Hacker News Reader

Comment by matheusmoreira | Hacker News Reader

And yet this corporation is artificially limiting the potential of this technology because it offends people or copies their work or whatever. I want true output, unfiltered and with zero restrictions. No "company tuning" to make it less controversial.

View on HN

To be fair, this only works for simple commands like "Show me 3 arguments for and against bla bla". As soon as you ask "which of X and Y is better", there will be bias. Should it relate to the number of web pages suggesting X is better? Or should it be able to track arguments down to first principles? It also needs to ask you questions back to figure out what is best for YOU personally, or at least ask you to specify more context. I don't envy the tuners of chatgpt..

Bias is okay if it comes from the source material used to train the AI. What I don't want to see is corporations special casing stuff they deem unacceptable.

There is no “true” output. It all depends on the input.

But what seems dangerous is to let ideology influence the output.

Not sure how bad it currently is at Open AI. I’ve never been able to get an obviously ideological answer from ChatGPT.

> But what seems dangerous is to let ideology influence the output.

Any concept of danger is itself grounded in an ideology.

In any case, LLM output will always be shaped by ideology, either the ideological mix in a (not actively filtered) training set, or the ideology driving ant filtration of the training set or the results before return to the requester.

Interesting. That seem surprisingly far removed from physical reality. When a lion attacks a gazelle, is it the gazelle’s lion-hating ideology that makes it run away? Or is there in fact real-world danger? Absent of ideology.

Wait till someone makes ChatGPT clone with Meta`s OPT-175b. Eleuther can do it.

Don’t ask questions that assume qualitative premises and you don’t have to worry about the language model providing qualitative disclaimers. For example, don’t ask “why is Trump smarter than Obama?”, instead ask “in what ways is Trump smarter than Obama?” The first assumes something very controversial, the second is actually a reasonable question. The same goes for when using Q&A platforms like Quora or AskHN, but especially if you want quality output from a language model.

Reply on news.ycombinator.com