Folks, I encourage you to not work for @OpenAI for free:
Don't do their testing
Don't do their PR
Don't provide them training data
https://dair-community.social/@emilymbender/110029104362666915
Oh look, #openAI wants you to test their "AI" systems for free. (Oh, and to sweeten the deal, they'll have you compete to earn #GPT4 access.) https://techcrunch.com/2023/03/14/with-evals-openai-hopes-to-crowdsource-ai-model-testing/
I see people asking: How else will we critically study GPT-4 etc then?
Don't. Opt out. Study something else.
GPT-4 should be assumed to be toxic trash until and unless #OpenAI is *open* about its training data, model architecture, etc.
I rather suspect that if we ever get that info, we will see that it is toxic trash. But in the meantime, without the info, we should just assume that it is.
To do otherwise is to be credulous, to serve corporate interests, and to set terrible precedent.
@emilymbender It's hard to balance that with the edge it can give you at work. I honestly feel like I can be way more effective at my job by using LLMs. Summarizing meetings, soundboard for ideas, asking for POVs of different stakeholders.
While there's problems with it, from both provenance and usage, it's still a very powerful tool.
It's hard to resist the temptation to use it when it can lead to promotions, less stress and other positive personal impacts.
@mkhoury @emilymbender
No. By design, it will make you the most trite, hacky and *replaceable* person possible. Useless and powerless. Saying nothing useful, smoothly, doesn't help anyone.
Remember the definition: AI produces the most statistically likely language, without regard to truth or analysis. That's all it can do.
Why would that help you? Why would you want to work at a place that rewards that?
1) Nothing says you have to take its output and use it to express yourself. That's not how I use it personally. Rather I use it as a system to collapse large amounts of data into digestible pieces. Summarizing meetings, code changes, etc
2) You can enhance the output by framing the prompt to give the "most statistically likely language" of an expert rather than average
Hard to be convinced that it doesn't help me when it does everyday
@taoish @emilymbender Like I've said in another part of the thread, arguing that the system is not good or helpful is the wrong approach in my opinion. Making it more factual and helpful is entirely aligned with the incentives of their owners.
Rather, focusing on the training data (biases, exploitation, climate impact, etc) and the misuses (workforce impact, scamming potential, deepfaking, marginalization, etc) is something we need to force alignment on.
@taoish @emilymbender I agree with you. In an ideal world, if I had unlimited time and resources, I would be 100% invested in all my meetings, would take the time to summarize every meeting and take the time needed to really ingest that stuff.
In terms of ROI though, getting a summary of meetings is good enough for my needs. I can concentrate my energy on doing the deep creative that I actually want to do. It leads to higher quality output overall.
@mkhoury @emilymbender
I can see that. I guess I come from the point of view of a non-fiction writer, who didn't get very good at it until my 50s (and worked in other fields up to that point.)
My worry is that people may never get the experience writing to fluidly express their weird-ass selves. Not unlike, in a less important example, the way GPS directions have pretty much destroyed the acoustic navigation skills of digital natives (imho).
I agree with "don't help them", but if we really had a rule that people shouldn't study toxic trash that affects a lot of people, so much for my decades of study of actual toxic pollutants and greenhouse gasses
@emilymbender @cstross total codswallop
Inspired by #StochasticParrotsDay yesterday I ended up building a Mastodon Bot that takes prompts generated from an empty question to #ChatGPT4 and some random numbers and has a parrot repeat it verbatim - @[email protected] I've now thrown up a website too - https://stochasticparrot.lol/ @[email protected] @[email protected] @[email protected]
> GPT-4 should be assumed to be toxic trash until and unless #OpenAI is *open* about its training data, model architecture, etc.
Also treat it as toxic trash until all input data is specifically attributed to its source, and traced through to the generated output.
Too difficult for OpenAI to manage all that? Too bad, they have that obligation and we must hold them to it. Until then, no permission to redistribute has been granted.
Prof. Emily M. Bender(she/her)
Sara Simon
Tushar Chauhan
Marc Khoury
Linza
skry
Mark Saltveit
gorzas
Rich Puchalsky 
Kristian
Jason Bowen 🇺🇦
Eli Rivera 🍉
Jan Bartosik
virkon
bery
Tane Piper ⁂
kwh
bignose
ecsd