The first prompts nowadays use GPT4-turbo so I’d warrant that’s what Cedric was using. After 20 or so back and forths it goes into the older models.
I do not understand this one-line rebuttal. Where did I say that ChatGPT-3.5 uses tools?
I think ChatGPT does not use tools, no matter the model. It’s only possible with API access, as far as I know. If used via Bing, it can also do searches.
https://help.openai.com/en/articles/9260256-chatgpt-capabilities-overview
I said:
I asked the infamous ChatGPT 3.5
1 Like
I asked it and this is what I got:
ChatGPT
Yes, ChatGPT integrates with various tools to enhance its capabilities. For example:
Math & Coding: It can run Python code to solve mathematical problems, perform calculations, and execute programming scripts.
Web Search: It can fetch up-to-date information from the web for current events, sports, and niche topics.
Image Generation: It can generate images based on text descriptions using AI models like DALL·E.
Text Editing & Collaboration: It provides a canvas tool for creating and refining documents or code collaboratively.
So, if you have a math problem, I can solve it directly or run code for complex computations! Want to try one? 😊
I also asked him for some specific things and it showed popup saying “Browsing the web” and afterwards gave links to the urls where it got the information.
2 Likes
That’s not “a combination of multiple systems that work together to provide or solve a prompt”, even though your original response was in agreement with modern LLM solutions using tools (Deepseek-R1 ‘Open Source’ Chinese AI Interview in English - #60 by cedric).
OK, so they’ve updated that. Interesting that they don’t advertise it on the description page.
1 Like
I would say this is hardly an indication of lack of logical thinking, as it could describe the answer of many humans learners to trick questions*. Actually, supposing some extra assumptions just to answer the question is a very human (or animal) thing to do when presented with an unsolvable problem**. At most, it’s an indication of a system that can do logical deductions but has some holes in its training/knowledge.
My opinion*** is that current AI systems, which are combinations of many sub-systems besides LLM, are just limited by their complexity/scale, which in turn is limited by their power consumption. The human brain is just orders of magnitud bigger, and consumes orders of magnitude less power. Once somebody solves that problem, real general AI will be here with us (let’s call it Multivac 
). This probably requires some non-incremental breakthrough, so it can happen tomorrow or never.
*Asking for a solution to a problem that can’t be answered without further assumptions is a trick question
**The worst thing you can do for your “survival” (in any sense) is freeze because the input is not completely defined
***Very personal view not anchored in any real science but only on gut feeling and general knowledge of the field
2 Likes
It could equally be because the person putting the question doesn’t realise that it is ill-defined.
Elsewhere, someone presented a hypothesis as to why the speed of the expansion of the universe is increasing. Needless to say, it was based on no known physics, and was contradictory to the physics we do know.
I responded that the expansion was due to an infinite number of leprechauns pushing at the boundaries and that as the boundaries expanded, the forces stopping the expansion weakened. Essentially, it was a take on “There’s a Dragon in my Garage” from Carl Sagan’s Demon Haunted World.
Yes, we can produce hypotheses about situations, and we can protect these using ad hoc auxiliaries. What we shouldn’t do is pretend that these are true unless we have adequate backing for them.
Is it? Or did you mean to say it was contradictory to the physics known at the time the question was posted?
“for the discovery of the accelerating expansion of the Universe through observations of distant supernovae”
(https://www.nobelprize.org/prizes/physics/2011/press-release/)
Edit: sorry, I misunderstood what you wrote. This someone presented their own hypothesis as to why the expansion is accelerating.
Correct. I was so focused on the part about whether the engine would do more than a simple addition (as previous versions did), that I did not consider the case when they are not arranged along the width of the table, on a straight line. I already admitted that here:
I’ll just stop here, before I make an even bigger fool of myself.
AFAIK that’s not how the current generation of ML operates. Logic per se is not encoded, it’s just that it was trained on inputs that follow some kind of logical reasoning, so it does that. If one trained it on illogical nonsense, they would get back the same kind of illogical nonsense, in impeccable style.
That said, watching a “reasoning” model like the o3-mini is eerie.
Yes, and soon the honeymoon period of users not paying the full price of that computational capacity and power will be over because it does not scale.
1 Like
On a lighter note, one of the comics I read regularly had something to say about AI too:
And I completely agree with this. But that’s not a problem of the AI, it’s a problem of the teacher that lets it run free and present itself as an infalible system
Oops (no context at all)
What did people expect??
The latest fun term is “distealing”, meaning “distilling” a model’s training to derive another model, without permission.
There’s a new model, S1, which was distilled from Qwen2.5; the latter was told to “think” before generating a response (so it holds an internal conversation with itself before generating the output exposed to the user).
2 Likes
There is now enormous amounts of public money going into ML. Even the EU joined the bandwagon with
which is at least open source. The prevailing attitude seems to be “just take my money please”, which was OK with private investors but may backfire with government spending.
1 Like
Relevant, perhaps. I just read the intro and summary on Bluesky.