i've noticed that there's a lot of fruitful #ai development around purely treating #LLMs as black boxes and focusing on prompt engineering + the ReAct pattern. Simply forcing the LLM to draw out it's thoughts over more text increases it's accuracy, and if you also interleave that with input from the user or calls to external services, e.g. Google, you can achieve very interesting results
@mnl i originally assumed that the best answers would be in fine-tuning it. But the more I see, it doesn't actually buy you that much and it's really hard. Further, something clicked for me, when we teach kids more complex concepts like math, we also break the problems down, effectively creating a language problem out of a math problem, which is what's going on with Chain of Thoughts. My intuition is that there's a whole lot more we can squeeze out of this without improving the models
Most of the middle part of the country would be an intolerable place to be a woman at this point. I left ND after high school and never really went back for much and I used to feel guilty about it but now I just feel bad for the women that are still there without much say in, really anything.
Having once worked there, and knowing what a shell of a "tech" company that place was 15 years ago, I am aghast that anyone is paying any attention to what IBM says about #ai and/or their hiring practices today.
So there's been a Bluesky vs. Masto convo in my mentions for the last couple of days, and it's made something very clear.
White men have this strange loyalty to Mastodon but not a better fedi itself. And they project this loyalty onto anyone that even considers Bluesky, no matter what cultural background they come from.
I have my own reasons for not considering Bluesky, but there's this nasty implication that Black and Brown people should be grateful for Masto despite the fact people loyal to that project have repeatedly run us off with no remorse.
This massive cultural gulf is at the core of why the fedi can't get out of its way and be the next evolution on public social media.
There's enough room here for everyone to have their communities, but the willful ignorance that often leads this discussion and projects in the fedi has us running in circles.
@Are0h I don’t have any opinions in this area, but I’ve certainly noticed the fedi being very resistant to change, especially (specifically) around race
@ElDiabloBlanco Welcome! I subscribe to hashtags, like the introduction one you used. Definitely subscribe to hashtags, but also follow lots of people. A lot of people think Mastodon is a lot more engaging than Twitter. It's kinda fun
It’s been 3 years or so since I’ve ridden Amtrak on an overnight route. To kick things back off in a royal way, I’m taking the Empire Builder to Chicago, transferring to the Capitol Limited with a destination of Pittsburgh.
I enhanced the ChatGPT story teller for the 7yo. Now, I get ChatGPT to generate a caption that I push into Stable Diffusion. I use imgcat to display the image that Stable Diffusion generates directly in the terminal
my prompt engineering is a bit off. I used to get full 5-6 paragraph stories, but after demanding that it also gives me a caption, it only gives me single paragraph stories.
ohhh, interesting. My "wrapper prompt" instructs ChatGPT to only include a caption if the prompt is a story. Well, I asked it for a story that was essentially a math problem (3 friends plus 3 divide by 2, etc.) and it seemed to forget that it was telling a story because it didn't include a caption.