Thank you for your explanation. Would you mind telling me which subscription tier you need for this kind of work?
As noted in our documentation, local AI is severly hardware-limited. A 16B parameter model requires at least 24GB RAM (and that’s pushing the limits) but that’s still a small model. For trivial tasks, e.g., tagging, simple summarization, etc., a smaller model can be employed. gpt-oss-20B is a small-midsized good model but you’re looking at a 32GB+ Mac to run it comfortably. If someone is more serious about local AI, I would recommend no less than a 48GB RAM machine, bearing in mind, you’re still in the mid-size range and it still can’t compete with commercial models for speed and “accuracy”.
Hi AndeasE:
I started out on the Free tier to try it, then did a basic subscription. Now I’m on the Business Tier, which is more costly, but it gives me things I need – more memory to use (I can upload chapters, articles, previous chats, other documents, etc., into a file that the assistant can read, and there’s a Library for even more material that I can ask it to refer to (like a whole book).
Cost: You can always change which tier you are on, or cancel your subscription without penalty – no yearly subscription – which means you can turn it on if and when you need it.
Most important it has several Modes it can work in governing the speed and depth of the research: Instant for quick questions, Medium for most things, High for deeper research, and now they’ve introduced Pro for going way down the rabbit hole, but that’s capped so use is limited. A Medium answer doesn’t take much time at all, High could take 5-10+ minutes, but the answer is obviously more comprehensive and there are citations for every source it looked at. The citations matter: I can check accuracy for myself, and they often lead to information or ideas I hadn’t thought to ask about.
I have the “Tone” set to Professional, and in custom instructions I wrote that I wanted it to act like an experienced editor of historical fiction, and it’s supposed to advise, not offer suggestions about how to rephrase something, or write. I told it exactly what kind of book I am writing and what my writing style is – it is instructed to check for anachronisms (and the date of your story matters), and prosody (it will suggest when I should read it out loud to check the rhythm and find anyplace where a reader might stumble), and told it “praise is nice; critique is useful”, which cuts down on flattery. (You have to be brave).
And I challenge Chat’s thinking (which means I’m thinking), so I might say to look at something from a different point of view, or remind it of some aspect of the story they didn’t take into consideration, or rarely, that the answer wasn’t what I was looking for, and try again. A lot of the latter are due to my not phrasing the query in a way that got me the answer I’m looking for (learning to write a query is important).
I engage it in conversations about historical events or characters, which have been fascinating.
And I have it set so ChatGPT does not learn from my material. I looked into how private my work is (even consulted my intellectual copyright attorney), and it’s basically like Las Vegas: What happens in Vegas, stay in Vegas). Nobody, not even Open ai, has access to my material.
Basically, you decide how you want to use it, but that means you have to put in the time to set all the preferences to make it work for you. And you have to accept the limitations and learn to work around them. It’s a complex tool; it takes a learning curve to use it effectively. OpenAi has lots of help and tutorials, and of course more help is all over the internet.
2 Likes
Thank you!
Whoa now.
I think we should take a step back. Many of us are overestimating/misunderstanding what “AI” can really do. It’s become so easy to fall into that trap.
First of all “AI” is not remotely actual artificial intelligence.
Can an algorithm check syntax? Certainly. (One just corrected my spelling here.) But can it think? Can it analyze? Absolutely not.
It’s easy to be seduced by its apparent facility. But we should always be cognizant of what its actual capabilities and limitations are.
“Summarize this for me” maybe. “Explain this to me” absolutely not.
1 Like
Frontier AI models can absolutely explain things at a doctorate research level in most fields.
With sentence-level citations to document source/accuracy.
1 Like
Accurate but superficial, in my experience. Marginally useful starting point for further research.
1 Like
Have you tried Opus 5 or Fable 5 at Extra or Max effort?
Neither is available without a Pro subscription. The free options have not convinced me to make the investment.
That may explain your perception of AI capabilities.
No question AI can make errors. I would never trust its output on either simple or complex matters without verification.
That said - Frontier AI models, with proper guardrails, are unquestionably useful tools for thinking and problem-solving at the highest levels of just about any academic area, any profession, or life in general.
2 Likes