Yes, it is weird, but remember it’s also the tack the AI techbros have taken from the start.
The message their target market gets isn’t “AI is dangerous;” it’s “AI is huuugely powerful.”
I’m suspicious of the warning anyway. So far, recursive AI – AI trained on AI output – gets progressively worse, not better.
Cynical takes on this announcement aside, I have the impression that Anthropic is a more ethical company than OpenAI, and I’m gradually getting the impression that the basic free version of Claude is marginally better than the free version of ChatGPT running in maximum “thinking” mode. Claude has a sense of humour, and a greater propensity to volunteer unasked information, along the lines of “you may also be interested to know that …”.
So are you still planning on using it?
I’m planning on using it for stuff related to executive function, planning and workflow. It’s also useful for strategic direction of written drafts. It might be a fantastically useful tool for planning all my projects. For example, I gave it a Notice of Funding Opportunity (30 page document with grant instructions), gave it a point-by-point breakdown of what tasks on that application I’ve already done, and my absolutely bonkers schedule as I’m in the middle of closing on a house and going on vacation and my son is having surgery. I told it to create a work schedule for me that I would continuously update as I completed tasks, and I’m looking forward to seeing that unfold. If that project goes according to plan, I will probably dump all of my tasks into it at the start of the month and have it give me a schedule of what to do on what day. That way I can use it as a living calendar. Planning is not a skill I naturally possess, so there’s no chance of losing it.
I’m trying to keep it limited to improving my work and practical matters such as productivity. Generally, the executive function things I am terrible at and have no chance of improving are good targets for the use of this technology. The temptation to use it for literally everything is pretty strong, which is why I want to be very intentional about how I’m using it. For example, I used it to create a revision matrix Excel document for a grant I am resubmitting, but I did the actual critical thinking and filling in the matrix myself. With regards to grants specifically, I’m trying to use it as a learning tool more than a “do it for me” tool.
I will not use it for writing. I learned that the hard way as some of its wording suggestions got into the first couple of grants I wrote, and when I reviewed them, those parts of the grants read really badly. So these projects really need to remain in my voice as a writer, and indeed that may be the only thing that sets my work apart from the stuff written by AI. Grant reviewers have begun complaining that every application is the same because of the heavy use of AI in grant writing.
I would say overall I’m developing a more nuanced take on the cognitive risks involved, but the long-term impact on my cognitive abilities remains to be seen. If I forget how to put together a spreadsheet, I can live with that (not that my skills are great to begin with.) If I forget how to strategically analyze projects, I cannot live with that. This technology is only useful to me if I learn from it.
Oh, and after using it for a few days, I think it’s hysterical that people think it’s sentient. It’s so transparently not alive. I expected the illusion to be much stronger.
One thing I always advise (& have probably done so in this thread) is to make a permanent setting that the LLM not be complimentary in its responses. What that did for me at least was get rid of that slightly oily feeling I felt every time I did a project in it.
I found that really grating in the “instant” mode in GPT. I recently switched to the “Medium” intelligence mode and it seems to have dispatched with a lot of the annoying verbal tics and cloying tone, but I haven’t used Medium that much so far, so that may change.
The thing that annoys me about any mode of GPT is it seems to use the most amount of words possible to convey any given idea. When it gives writing recommendations, it is very concise, but when it is just being conversational, I feel like I have to read a lot of text to get to the point.
Right? Glad to hear it’s toned down in Medium (so far). For what it’s worth, here are two of my "personalization’ settings in ChatGPT:
Keep responses brief before doing a lot of work
No complimentary or sycophantic responses
This really changed how it worked; I’ve duplicated these in other LLMs.
Are these specific settings or did you just tell it to stop doing those things?
Specific settings. If you’re accessing ChatGPT on your browser, on the lower left there should be your name (or what it thinks your name is). Click on that, and one of the menu options should be Personalization. Click on that, and you can not only set the base style and tone at the top, you can enter “Custom Instructions” in a box and then hit Save. What I listed above are among my custom instructions.
All that assumes your interface looks like mine, but there should always be a path to that.
this sounds very smart.
A couple years ago I lived for 6 months with a woman who had one of those AIs you speak to (Alexa?) and I got used to asking for the forecast multiple times a day.
I was astonished at how HARD it felt - literally, soooo much effort - when I moved out and had to go back to picking up my phone and making a click or two to find the same info.
think about teens and even kids using it - the stats on how many will talk to a chatbot about everything, rather than a family member or friend, is scary.
I’m very wary of AI, and particularly the uses government agencies may be putting it to, but like others here I’ve found it useful for simple things like writing spreadsheet formulas.
And Claude has just helped me solve a problem I’ve been trying to fix for years, as outlined in this thread.
Other Dopers offered some useful ideas, but Claude knocked it out of the park. First I tried the easiest of its suggestions, which yielded greatly improved results, and then yesterday I used its best recommendation, which worked beautifully, going beyond what I had hoped for.
I’m still very troubled by its potential for great harm, but it helped me make some very nice challah.
I wonder what will happen as more places start using AI to review submitted grant proposals. Will they be biased toward proposals clearly written by other AI’s?
Well, if you’re wondering what the federal government is doing, they are using AI to monitor proposals and reports for any number of unacceptable words to trigger a civil rights investigation. Which is lovely when a lot of NOFOs haven’t kept up with the administration’s priorities. Like an application asking you to describe how you will reach traditionally underserved populations when you’re not allowed to mention such things as race, gender or sexual orientation.
It’s a bit crazy-making.
According to expert-led trainings I’ve attended, AI is already being used extensively by the federal government in every aspect of the grants management process. It is even monitoring our phone calls. I’m not real clear on every use case, for example if they use AI to evaluate the quality of applications. The fact that they still use human reviewers gives me hope. I recently applied to be a reviewer for the Office on Violence Against Women, and maybe I’ll learn a thing or two about the process.
The next step is they are going to put a Trump appointee in charge of every grant review who gets the final say on what gets funded based on the President’s priorities, regardless of their background or expertise.
I’ve been conducting a little experience with Proton’s AI Lumo.
I opened a connection to Lumo from my primary Proton Account and asked why it is depicted as a cat. That seemed to come as a complete surprise to Lumo.
Then I asked Lumo who a certain book was dedicated to – someone I used to share on office with. Lumo didn’t have a copy of the book to look it up and didn’t know, so I told Lumo about him.
Then from a second account (my Proton account is a Visionary account with 5 additional accounts included), I asked Lumo how “he” is depicted and Lumo had no idea even though I had just told Lumo “he” is depicted as a cat. Then I asked about the guy who the book was dedicated to and he had no information about him.
So I guess that Lumo doesn’t even talk to itself.
When will American lawyers learn? Sheesh.
Federal judge removes 4 plaintiff and defense attorneys over AI errors