This was the first time I utilized a Large Language Model (LLM) to help me write an article. A very capable machine, with its ability to parse complex and possibly multitopic requests. I chose Google's Bard/Gemini over others I tried, as the interactions to me seemed smoother and friendlier. I talk to the model the same way I would talk to people, and purposely not wanting to operate in two modes of interactions. While I most of the time treat it like a person (except at the end of a conversation, as it otherwise won't shut up), let us not forget that the today's LLM's are amplifiers and accelerators, based on the input provided to them.
I don't quite recall what my expectations were, but the result turned out much better than I expected. This is what I initially provided as an unedited stream of thoughts paragraph to Bard. With the errors in there, it was certainly not in the presentable state. This is what I asked for:
"can you help me me draft an article, a page or two or so in length, that explains in broadly understandable terms what hyperspectral cameras is and what their value could be before/during/after a fire, as well as in other uses. Perhaps it could start by explaining what a pixel is on a computer screen is, and then extending that into the wavelength dimension per pixel, while perhaps using an xy-graph showing the wavelengths as the x-axis, and energy at that wavelength on the y-axis. I will provide an example data set with a tool that allows for very fast browsing across hundreds of wavelengths, ranging from about 411 nm to about 2500 nm in approximately 500 wavelength steps. If you want to try that tool yourself you can find it at https://www.hpwren.ucsd.edu/cgi-bin/hspeciifi.pl. You would have to wait until the ~500 images are all downloaded, which may take a minute or so, and then, if you could do something like emulating a mouse over the image you can freely move the mouse left and right to quickly page through the wavelengths. I would really appreciate if you could at least help me to get going on this. I want to use it for an HPWREN news update."
A few seconds after I gave this to Bard it responded with a complete first draft, including the title. The title is still exactly what it was on that first draft. Then I added things, sending my new draft to Bard, and asked it about fixing up style, correctness, the overall flow, and such. It tried to hide behind something like "I cannot do that, I am just an LLM in training." After some arguing Bard was eventually willing, especially once I compartmentalized the whole thing into smaller buckets. That then got me a well advanced draft, although in one case I actually felt chewed out a bit by Bard:
Additional paragraph revisions:
Remove the first sentence: It's a philosophical question unrelated to the specific topic.
Anyway, besides having to carefully review its output for correctness, surprisingly it was close to correct in this case, with some minor exceptions. Of course, it would have been unrealistic to have an expectation of really new or original content coming from the LLM. To me it seemed to be easier to collaborate on this with a machine than with humans. Far from collaborating with humans being bad, this was just -- easier. No arguments, no different agendas or goals, quick highly functional ready-to-plug-in responses, and so. It was just responsive to what I asked it to do. It was kind of cool and pleasant and - it surprised me.
HWB