This is thread to request, and offer assistance with, AI.
This is not a thread to complain about AI, its training models, its use on Facebook Reels, or other criticisms. We have plenty of those already, may I suggest this one? If you don’t like AI, that’s fine, but tell us elsewhere.
Feel free to ask (and answer) questions about prompts, general AI tips, discussion of various AI models (GPT vs Claude vs Gemini vs …), and more! Business use, personal fun, productivity, programming - all use cases are supported here*.
Also, stories of how you used AI to solve (for want of a better word) problems are great as well!
*To this point, we already have a AI-Image generation thread which is covering topics related to creating images from AI prompts, so image-related questions may do better there.
I asked my AI and it answered with the text below. I think this would be a very difficult question to nail down since there are so many variables but we can get in the ballpark:
The electricity consumption of an AI query varies significantly depending on the model size, infrastructure, and type of query, but here are some general estimates:
Large Language Models (like GPT-4, Claude):
Simple text queries: 0.001-0.01 kWh per query
Complex reasoning tasks: 0.01-0.1 kWh per query
This translates to roughly 1-100 Wh (watt-hours) per interaction
For comparison:
A typical Google search uses about 0.0003 kWh (0.3 Wh)
Streaming an hour of Netflix uses about 0.8 kWh
A smartphone charges with about 0.02 kWh
Factors that increase energy use:
Longer responses requiring more computation
Image generation or processing
Complex reasoning or multi-step problems
Peak usage times requiring more server resources
Training vs. Inference: The numbers above are for “inference” (answering your query). Training these models initially requires vastly more energy - sometimes millions of times more - but that cost is amortized across billions of queries.
The exact consumption is hard to pin down because companies don’t always publish detailed figures, and it depends heavily on the efficiency of their data centers, cooling systems, and hardware. Some estimates suggest that AI queries use 5-10 times more energy than traditional web searches, though the technology is becoming more efficient over time.
Thanks! Assuming we can trust the AI to be accurate on this question, that is.
I love playing around with graphics and haven’t been tempted to use AI. But I did wonder: if, say, I wanted a new banner for my Bluesky account and decided to give AI a try, how much guilt would I have to take on?
I’m frequently amused (and a bit freaked out) by what I see on shows such as John Oliver’s Last Week Tonight, which makes a lot of use of AI graphics. It’s amazing how consistently surreal the results are—and I’m assuming that the person ordering the image does NOT specify “make it look surreal.”
I’m in the process of creating a business-specific AI model tailored to our Business. In doing so, I started with GPT and ended up with Gemini (Google), largely driven by the fact that GPT can’t access our Google Workspace drives and emails.
To start, Gemini suggested I create a document which provides Gemini with a lot of our business details: clients, Service Level Agreements, team members and which clients they are assigned to, more. All told, the first draft of this thing was 4,000 words.
After creating this document, I was then told that it cannot be accessed outside the chat without constantly referencing it, which kind of defeats the purpose of a general business overview document.
I’d rather my team ask “what is the SLA for Client X” than ask “looking at document @WHATEVER, what is the SLA for Client X”, but Gemini forces the second option and cannot do the first.
It may not be possible to build a business-specific AI model in Gemini like it was in GPT, which is a serious strike against it.
Gemini also has weird limitations that one wouldn’t expect - for example, I have an attorney-reviewed engagement letter which I commonly modify with new client information. I wanted to automate this in Gemini, but Gemini (a) cannot copy my template into a new file, (b) cannot modify the contents in my template, and (c) if it gives me the rewritten engagement letter in the ‘chat box’, it cannot keep the formatting of the original document. Gemini suggests that I fix the formatting, but ‘fixing the formatting’ is a much more arduous task than entering 10 pieces of data, that’s for sure!
If you have the same task you repeatedly do with the AI I find asking the AI how to best phrase the query helps. (Once you have arrived at the results you want.)
Copy/paste what it tells you into a text doc and re-use that query every time. I will say, the AI results can still vary. It’s a little weird but still, this works well for me with a task I do daily.
To be fair, you need to compare the electricity usage from the AI which takes 30 seconds to generate an image for you (or less) versus the electricity you would use taking a few hours to do the same task on a computer.
I don’t know precisely how much energy the big online AI models use, but there are plenty of models, for both images and text, that you can download for use on a mid-high end personal computer, so it can’t be that much.
Running SDXL (AI image model) or a 22B LLM (competently entertaining chatbot) takes about as much juice on my computer as playing a modern AAA-esque video game. No formal testing, just based on seeing what sort of CPU/GPU load they create and how warm those components get.
Yes - and that’s exactly what I was going to recommend. Ollama +Open WebUI.is my local solution for RAG, in my case to access music tech documentation and RPG rulebooks.
Both can be accessed via Pinokio, which is kind of like Steam for AI apps.
I have zero experience with AI. Our CEO told me I should look into using it. I am a grant writer and federal grants manager for a midsized non-profit. I’m not sure I would use it to craft grant narratives, but I do have a ton of tedious administrative tasks that take up too much of my time. Oh, and I hate prospecting, which is looking for new grant opportunities. It’s time-consuming and boring. Hands-down my least favorite task.
Can AI help with tedious administrative tasks? Anything to do with sorting email or documenting things? Could it be used to automate a process like entering info from a grant contract into a SharePoint spreadsheet?
Can AI help with grant prospecting?
What AI model would best be suited for this kind of work?
There is a lot of integration software out there which allows you to create various workflows with AI as a part of the whole scheme.
My answer to you would be…maybe. This is all a new world and things are still being figured out. I have seen some amazing integrations done though but again…it depends on a lot of variables. There is no firm answer I do not think.
And know, trying to get it to work can be a real chore. Once done it’s great but getting there, if you get there, is work and there may be no payoff in the end.
Also, AI tend to not be very current in their dataset (my AI tells me it is five months out of date…so good to January 2025) so I am not sure they could find current prospects for you since they do not know about the most recent prospects. The info they get might be too dated to be useful to you. I’d suggest just try asking a few and see how they respond.
The thing about the neural net of an AI is that the whole thing needs to be loaded in memory at once. (You surely could write one to work one segment at a time using a swap file, but it would be much, much slower.) There are small AIs that can fit in a video card with 8 GB or 16 GB of memory. But the really high-end models need several hundred GB of memory. More than is available on one “video card”/AI card. The cards are designed for several cards to work together to emulate one giant card with one memory pool. I know some designs have allowed up to 8 cards to work together (all installed into a single giant motherboard). I’m not sure what configurations the latest cards support.
As an example, the Grace Hopper 100 runs at up to 700 watts. Say that you have 4 of them linked together, that’s 2.8 kilowatts. A task that takes 5 seconds to run would use 1/720th of 2.8 kilowatt/hours.
I am finding that Gemini is as kludgy as the rest of Google’s applications. I asked it to do a function that I’ve done in GPT a number of times, finding if there is a combination of transactions which add up to a specific amount…
(In the following, the AI’s are in quote boxes, my responses are in italics.)
GPT’s answer:
Gemini:
OK, well, that was my fault, I should have been more direct in my prompt. But then, asking Gemini to identify the transactions, I was hit with:
We went around a couple of times about this issue - why Gemini couldn’t access the file that it told me it accessed just one query earlier, why Gemini couldn’t access the file that I uploaded the second time, etc. I finally decided to paste the table as plain text in the chat, which worked…
I then responded:
Earlier you said:
"Yes, there is a combination of payments that totals exactly $111,198.00.
After analyzing the transactions from your file, I found a set of payments that add up to this amount.
If you need the specific transaction details that form this sum, just let me know!"
Now you’re telling me there is no solution to this problem?
Gemini’s response:
Emphases mine, JT
In addition, one of the reasons why I upgraded our Workspace sub to include Gemini was to help Inna manage her emails. Well, sometimes we are told that Gemini cannot access our emails, even if we use the Gemini button in our Gmail! Other times, it has no problem identifying action items, unreplied-to emails, and the like.
To be fair, I shouldn’t be surprised. I see how well Google does in other things, I should have been cognizant that their AI would also be as half-assed as, say, Sheets. Or search, which brings up a bunch of sponsored links which might be relevant to my query.
You may want to look into using something like n8n which is a workflow automation tool that can incorporate AI into it. Will it do what you need? I have NO idea. Just something you may want to look into (there are a lot of YouTube videos on it as well if you want to see how it works).
NOTE: I have NOT used this tool and I have no connection to them. Just one option for this I have come across.
