AI Vs MTO
-
That’s an impressive profile, almost certainly derived from your previous conversations with ChatGPT (which it can and does access when giving you personalized answers). This happens to be my area of professional expertise, so I can elaborate a bit more on what’s probably going on here.
To be clear, ChatGPT is not a Google product (it’s made by OpenAI, an independent company whose most prominent major shareholder is Microsoft) and it definitely doesn’t have access to your search history. It’s capable of doing its own web searches for information when asked (so it could pull data from a public resumé like a LinkedIn profile, for example), but most information it collects on you comes from the specific conversations you’ve had with it. If you’ve been using it a lot for job applications, then it’s likely to have quite a lot of data in that area. But it really is based on what you ask it: when I asked it “what do you know about me?”, this was the whole of its response:
You have an M2 MacBook Pro with 96GB of RAM, and you've been experiencing an issue where the "Your system has run out of application memory" warning appears periodically. Even though all running applications combined are using only 26GB of RAM, the WindowServer process sometimes uses an abnormally high amount of RAM, up to 588GB. This issue happens after your system has been left running for about a week without a reboot.
Let me know if there's anything you'd like to update or add!
In my case, the answer is pretty limited because I haven’t told it much at all about myself and nearly all questions I ask it are abstract in nature. The one thing it knows for certain is that I have a specific model of laptop because I explicitly said so in a question I asked it a few weeks ago. On the other hand, if I were to ask the equivalent system at my workplace — which also uses the same GPT-4o underlying AI model as the public-facing ChatGPT, but has a completely different archive of conversation data to draw from — it would be able to tell you a great deal more about me, since I interact with it every day and have very detailed conversations specific to my work.
With all that said, Gemini, Google’s competing product to ChatGPT, is believed to be able to access your search history and is known for certain to be able to access your Gmail archives, Google Docs and Google Maps location data when giving you personalized answers. To be very clear, this data is not used as part of its general training set that shapes its answers to everyone, so you can ask it about emails you’ve recently received, but you can’t ask it about the contents of my inbox and get anything useful.
Apple’s upcoming Apple Intelligence feature (slated to be released toward the end of this year) is also going to be able to do this sort of personalized answer-gathering from your data, although unlike Google, they’re doing it in an absurdly elaborate way that is very privacy-conscious and ensures the data can’t be collected or harvested by bad actors, even within Apple itself.
-
Thanks for the clarification!
I got confused about OpenAI- I thought it was a Google product. Thanks for the correct information.

Either way, I use ChatGPT a lot when I am at work or researching things, but other than the examples listed earlier in this thread, I don't "write" anything for MTO using it. A lot of sites do, and they make a lot more $$ than MTO, but that just isn't our way here.
That having been said, I do appreciate it for the things it does, but as MTO seems to be so far exempt, I will keep all my plans for world domination here on the forum where they belong! :D
If you have any tips to share to make ChatGPT more effective (or more accurately, allow me to use it more effectively!), I would love to hear them! While I am absolutely concerned about Terminators, I would also really like to have a better job between now and Judgement Day! :ahhh
Def
-
Okay, strap in for a wall of text. So the main thing to understand about GPT and other large language models is the rough outline of how they work, because a lot of people conceive of them incorrectly, and once you understand their limits, you can get a lot more use out of them. First and foremost: they are not search engines, indexing content on the web. Rather, they are randomized predictors. This may be easiest explained with an analogy.
Say you want to plot a route to drive from Los Angeles to New York City. A traditional computer system will do this by consulting a detailed map, tracing individual roads, measuring the distance, and pruning possible routes until it finds the shortest one. Its data may be inaccurate — maybe it doesn’t know that a road is closed for an emergency — but given the dataset, it will always find the shortest possible route between two points. This is what we call deterministic: if you put the same data in, you will always get the same data out.
A system like ChatGPT uses a completely different method to solve the problem. It doesn’t have a map of the country; instead, it has a bunch of data on what other drivers did. So say most people who say they’re going from Los Angeles to New York take the 210 freeway out of LA and then I-15 up to Las Vegas, and so on. The system works by chaining together individual choices the drivers made: if 70% of drivers take I-210 and 30% take I-10, then for any given time it’s asked the question, 70% of the time it will tell you to take I-210 and 30% of the time it will tell you to take I-10. At every off-ramp along the way, it’s essentially rolling the dice again, with the idea being that people who say they’re going to New York and then take I-210 are most likely then going to take I-15, and people who say they’re going to New York and then take I-210 to I-15 are probably then going to drive to Las Vegas, and so on. Most of the time, this approach will yield something roughly correct, because it has a vast amount of data on the patterns of drivers and people tend to be creatures of habit.
The problem with this approach, though, is that lurking in that dataset will be people who drove all the way to I-15 and then realized they left the stove on at home and turned around and went back. These will be very few in number, but roll the dice enough times, and it will eventually land on a person who started to do something and then changed their mind. And the problem is that the system isn’t looking at the problem holistically: it does not have a goal in mind of “keep making decisions that get to New York”. Instead, it’s just following the pattern of “these are things people did when they said they were going to New York”. If it follows down a route laid by someone who forgot something at home, or got sick and turned around, or got lost and drove off a cliff, well, it doesn’t know it made a mistake and it will keep on building out your route based on the mistake.
This is how we get the well-publicized cases of GPT writing very authoritative-sounding legal briefs that cite nonexistent court cases, or making up baseball players who never existed, and so on. Every word that comes out of a model is based on a random decision weighted by all of the words that preceded it: if most people say the phrase “that’s comparing apples to oranges”, then the vast majority of the time, GPT will use those words in that order. However if it instead rolls the dice and says “that’s comparing apples to orangutans”, it’s going to go down a totally different road and every word that follows the misprediction will be based on it, compounding the error.
So, what does all of this mean for you in practice? The bottom line is that GPT is a great aid for manipulating language. It can summarize documents for you very well. It can rewrite your bullet points into an email that sounds very professional and dignified. But what it can’t do is consistently give you accurate factual information. If facts are sufficiently well-publicized, then most of the time, you’ll get an accurate answer from it. But if facts are disputed, or not widely publicized, it will merrily go on its way predicting the most likely words that people would use to follow the other words it’s already said, completely oblivious to the fact that they’re meaningless babble.
There is, however, a way that you can get more use out of the tool when asking factual questions: you can tell it to cite its sources, which will force it to use Google to find documents and then summarize them for you rather than just trying to give an answer based on its own prediction engine. For example:
I want you to act as a knife and cutlery expert. I will ask questions, and you will provide answers with citations and links. Do not conjecture or speculate. My first question is: how many different models of the MiniChamp Swiss Army Knife have been produced by Victorinox?
The Victorinox MiniChamp Swiss Army Knife has been produced in several different models since its introduction. These include:
[list type=decimal]
- Original MiniChamp (I) - The earliest version, which lacked some tools found in later versions, such as the combo tool and pen.
- MiniChamp II - The current model, simply known as the MiniChamp, which added tools like a retractable pen and tweezers.
- Midnite MiniChamp (MiniChamp Lite) - This variation replaces the tweezers with a small LED flashlight, available in red or white light.
- MiniChamp Alox - A version with aluminum scales (Alox) instead of the standard Cellidor plastic, but without scale tools like the pen and tweezers.
- MiniChamp XL - A now-retired, larger version that added tools like a cut file and golfing divot tool, produced between 1999 and 2002 oai_citation:3,SAKWiki | MiniChamp oai_citation:2,Victorinox MiniChamp oai_citation:1,Victorinox MiniChamp 58mm Swiss Army Knife.
These variations cover different needs, such as incorporating a flashlight or prioritizing a slimmer, more durable design with Alox scales.
For each model, list the scale colors that have been produced.
Here is a breakdown of the scale colors for each model of the Victorinox MiniChamp Swiss Army Knife:
[list type=decimal]- Original MiniChamp (I):
- Red Cellidor (standard) oai_citation:18,SAKWiki | MiniChamp oai_citation:17,Victorinox MiniChamp
- MiniChamp II (current MiniChamp):
- Red Cellidor (standard) oai_citation:16,SAKWiki | MiniChamp oai_citation:15,Victorinox MiniChamp oai_citation:14,Victorinox MiniChamp 58mm Swiss Army Knife
- Translucent ruby oai_citation:13,Victorinox MiniChamp 58mm Swiss Army Knife
- Translucent sapphire oai_citation:12,Victorinox MiniChamp 58mm Swiss Army Knife
- Midnite MiniChamp (MiniChamp Lite):
- Red Cellidor (standard) oai_citation:11,SAKWiki | MiniChamp oai_citation:10,Victorinox MiniChamp
- Translucent ruby (with white LED) oai_citation:9,SAKWiki | MiniChamp
- Translucent sapphire (with white LED) oai_citation:8,SAKWiki | MiniChamp
- MiniChamp Alox:
- Silver Alox (standard) oai_citation:7,SAKWiki | MiniChamp
- Red Alox (limited) oai_citation:6,SAKWiki | MiniChamp
- Black Alox (limited) oai_citation:5,SAKWiki | MiniChamp
- Sunset orange Alox (limited, commissioned by Swiss Bianco) oai_citation:4,SAKWiki | MiniChamp
- Copper Alox (limited, commissioned by Swiss Bianco) oai_citation:3,SAKWiki | MiniChamp
- Other limited colors were also commissioned by Swiss Bianco over the years oai_citation:2,SAKWiki | MiniChamp
- MiniChamp XL (retired):
- Red Cellidor oai_citation:1,SAKWiki | MiniChamp
These colors reflect both the standard options and limited editions that have been produced over the years, particularly for the Alox variants, which often included special editions commissioned by collectors.
Now here’s the kicker: not only is this answer fairly thorough, it’s more thorough than SAKWiki. While most of its citations go to SAKWiki, it pulled in sources from other websites and found models that aren’t listed there, like the sapphire version of the MiniChamp II.
-
Fantastic info, and thank you for taking the time to share it.
AI is a great tool with lots of potential, and I feel like when it fails (like any other tool) it's largely because I'm not using it properly.
Its interesting that it picked up on the sustainable genetics things but didn't include any of the other stuff I have asked it, like how many potatoes it would take to charge a Tesla Model 3, or how l many times I'd need to rub my feet on the carpet. Or that time I asked whether a small potato or cherry tomato would make a better prosthetic eye and it suggested I use a cocktail onion instead.
Yeah, I ask it a lot of silly things to keep Skynet from getting a proper read on humanity. I want it to underestimate us from time to time, just in case....
Def
-
@Grant Lamontagne said:
Its interesting that it picked up on the sustainable genetics things but didn't include any of the other stuff I have asked it, like how many potatoes it would take to charge a Tesla Model 3, or how l many times I'd need to rub my feet on the carpet. Or that time I asked whether a small potato or cherry tomato would make a better prosthetic eye and it suggested I use a cocktail onion instead.So without knowing the exact wording of your questions, it's hard to make any confident predictions about why that might be the case. But I can tell you that when I asked it "what do you know about me?", the reason it talked about my laptop was that it's one of the very rare occasions where I started off a question telling it something declarative: "I have a laptop with...", as opposed to asking an abstract question: "if a person has a laptop with...". It's very likely designed to intentionally preserve and retain declarative statements you make about yourself for future use to improve the answers it gives you, while hypotheticals and general questions are limited to the scope of the current conversation that you're having with it. Because every interaction with the system is following a series of predictive decisions, retaining and using any information about you progressively increases the cost of every future interaction you have with it, meaning that they're going to want to be selective about what things are retained between conversations to keep the cost of your future conversations manageable.
Yeah, I ask it a lot of silly things to keep Skynet from getting a proper read on humanity. I want it to underestimate us from time to time, just in case....
I'm guessing this was tongue-in-cheek, but you can be assured that modern large language models like GPT-4 are not a feasible path to Skynet. They have no concept or comprehension of the meaning of what they're outputting, and it's not being done with a human-like reasoning chain that can pursue objectives.
With that said, there are models in development that use neural networks to execute logical reasoning and could potentially be the path forward to AGI (artificial general intelligence), which is what people think of as "AI" in popular concept: a self-aware machine that can reason and creatively pursue both short and long-term objectives and use its successes and failures to self-improve like humans do.
The first model of this type that's become publicly available is GPT-o1, which was debuted to the public 11 days ago. Despite the "GPT" name, it has very little in common with previous GPT models (OpenAI just sucks at naming things) and operates through completely different mechanisms. At the moment, it is slow, clunky, and unbelievably expensive to operate. But what makes it unique is that it's capable of pursuing objectives and self-correcting: in my mapping example, if GPT-4 tells you to drive off a cliff, it's going to happily keep giving you instructions for how to escape a burning car, get through effective physical therapy and continue your trip after you get out of the hospital. GPT-o1 is capable of determining that maybe this isn't a good way to get to your destination and discarding the whole chain of reasoning. What particularly makes this unique is that GPT-o1 is the first model that can produce better answers by throwing more computing power at it. If you give them more processors, GPT-4 and similar models will just give you answers with the same errors, but faster. GPT-o1 will actually give you improved answer quality the more resources it's given, because additional processors mean more possible chains of reasoning it can evaluate and prune down to find one that works.
We're living at a fascinating inflection point in history, honestly: machine reasoning has the potential to transform society to a greater extent than maybe anything since the steam engine if -- and this is a big "if" -- the cost of that reasoning can be brought down to be competitive with human brains, something which we are very far away from at the moment. As impressive as GPT-o1 is, we're still talking about something that can solve a very simple 7x7 crossword puzzle, taking two to three times longer than an average person, and will get totally lost on the average New York Times daily puzzle. And it needs enough electricity to power an average household for a month to solve that simple puzzle. It remains to be seen whether or not these costs can be brought down; at the moment, it's a cool project, but definitely not going to be threatening anyone's job.
-
ChatGPT at the moment undoubtedly can do some pretty cool stuffs
I used it to help me summarize or explain some concepts to me in a simple ways. However, not everything it said is true and I still need to double-check things later on, but it definitely helps me getting a quick look at stuffs. Even its citations is not always trustworthy though. Sometimes I ask it to give me the scientific paper names it used to gain some of the information it said but turned out that the names it gave does not exist, but were generated. -
@Grant Lamontagne said:
Until Boston Dynamics develops a robot capable of handling a folding set of pliers or flip out a can opener from a Champion, I'm not that worried.When you want REAL info from people who actually USE things, MTo may be the last bastion of hope!
Def
Boston Dynamics needn't develop a robot that can handle a set of folding plyers or a can opener any more than the Cylon Raider ships in the original Battlestar Galactica ever needed humanoid robots to sit in and pilot them. The ship itself could have contained the computer brain in the robot heads and the hand controls themselves were not necessary if that was the case
Boston Dynamics can simply create robots that have those tools on appendages natively.
I agree that MTO has much more real world use case analytics than almost anywhere else. But how will the average John Q. Public ever know that?
-
@friendliestdoggo said:
Even its citations is not always trustworthy though. Sometimes I ask it to give me the scientific paper names it used to gain some of the information it said but turned out that the names it gave does not exist, but were generated.Instruct it to give you citations with links: that makes it very quick to check whether or not they’re real. It also forcibly engages the feature of the tool that causes it to make actual web searches to formulate its answer, which significantly reduces the chance of hallucinations. It’s not foolproof — you do still need to check the links it gives you — but it will give much better results on average.
-
@ThundahBeagle said:
Boston Dynamics needn't develop a robot that can handle a set of folding plyers or a can opener any more than the Cylon Raider ships in the original Battlestar Galactica ever needed humanoid robots to sit in and pilot them. The ship itself could have contained the computer brain in the robot heads and the hand controls themselves were not necessary if that was the caseBoston Dynamics can simply create robots that have those tools on appendages natively.
I agree that MTO has much more real world use case analytics than almost anywhere else. But how will the average John Q. Public ever know that?
I'll know.
If I can't have a site I'm proud of, I'll shut it down in a heartbeat.
Def
-
Just for the fun of it, I asked AI to write a review of the XEN-OAL knife I recently reviewed from OKnife. I wanted to see if I was wasting my time hacking out reviews the old fashioned way. Here is what ChatGPT wrote:
The XEN-OAL knife from OKnife is a striking addition to the competitive world of everyday carry (EDC) tools, showcasing a blend of modern aesthetics, high-quality materials, and precise functionality. OKnife, a subsidiary of Olight, has built a solid reputation for producing reliable knives that appeal to enthusiasts and professionals alike. The XEN-OAL continues this legacy with a design that balances practicality and sophistication.
Design and Build Quality
The XEN-OAL immediately stands out due to its sleek and minimalist design. Its clean lines and ergonomic contours not only make it visually appealing but also ensure a comfortable grip. The knife features a handle crafted from durable G10 scales or, in some versions, titanium. Both options provide excellent durability and a premium feel, catering to users who appreciate both aesthetics and performance. The handle is slightly textured, offering a secure grip even in wet or slippery conditions.The blade itself is made from D2 tool steel, a popular choice for EDC knives due to its excellent edge retention and wear resistance. The blade profile is versatile, with a drop-point design that excels in a variety of tasks, from cutting and slicing to light prying. OKnife's attention to detail is evident in the precision-ground edges and the smooth satin finish, which enhances the knife’s overall appeal.
Performance
In terms of performance, the XEN-OAL is a reliable workhorse. The D2 blade retains its sharpness exceptionally well, requiring minimal maintenance even with frequent use. It slices cleanly through various materials, including cardboard, rope, and fabric, making it ideal for both outdoor adventures and urban EDC needs. The blade deployment is smooth and satisfying, thanks to the ball-bearing pivot system. The flipper tab is well-positioned for quick, one-handed opening, and the knife locks securely into place with its liner lock mechanism.One aspect that stands out is the knife's weight and balance. The XEN-OAL is light enough to carry comfortably in a pocket yet feels substantial enough in hand to inspire confidence during use. The pocket clip is robust and allows for deep carry, keeping the knife secure and discreetly tucked away.
Value for Money
The XEN-OAL is reasonably priced for the materials and performance it offers. While it may not compete with high-end custom knives in terms of exclusivity, it provides excellent value for those looking for a functional and stylish EDC knife without breaking the bank. OKnife's commitment to quality control is evident, as the knife arrives razor-sharp and well-tuned out of the box.Drawbacks
While the XEN-OAL excels in many areas, there are a few minor drawbacks to note. The D2 steel, while durable, is not stainless, which means it is prone to corrosion if not properly maintained. Regular cleaning and oiling are essential, especially in humid or wet environments. Additionally, the flipper tab might feel slightly stiff for some users initially, though it loosens up with use.Conclusion
The OKnife XEN-OAL is a solid choice for anyone seeking a dependable and stylish EDC knife. Its combination of robust materials, thoughtful design, and practical performance makes it a standout in its price range. With proper care, the XEN-OAL is sure to become a trusted companion for daily tasks and adventures alike.Here is what I wrote:
https://multitool.org/knives/oknife-xen-oal
A few comments from the AI generated review really got to me- for example, the option of titanium or G-10 handles is absolute proof why a hand written review is much better than than AI generated stuff you are seeing everywhere. The XEN-OAL knife has their new, proprietary aluminum alloy handle they call OAL, which s where the OAL part of the name comes from.
Next up, it's not D2 steel, its a CPM blade, which if I'm honest, I prefer to D2. But then I have been following CPM steels since (I think?) the 90's when it first came out. It was a fascinating concept then and still is today, especially with the recent news about Crucible.
Last but not least, I'd think that mentioning that world famous knife designer Ken Onion would have been mentioned in there somewhere.
Sorry Skynet, no world domination for you today! MTO is going to continue to be operated by people.
Def
-
The problem (as I see it) with all this AI generated "content" both verbal and imagery is that it (the AI of choice) is just regurgitating what it has been trained on or has scraped from the internet.
It has no way of verifying the accuracy of the information other than comparing it to the data it's trained on, a self confirming loop.This can result in it producing truth, fiction or outright lies as its output, with any option put forward with the same enthusiasm/sincerity.
It's as simple as "Garbage in -> Garbage out, and there is inadequate garbage filtering on the input of many, if not most, AIs.At the risk of sounding like an evangelical artist... All it can do is plagiarise, not create.
-
@PTRSAK said:
The problem (as I see it) with all this AI generated "content" both verbal and imagery is that it (the AI of choice) is just regurgitating what it has been trained on or has scraped from the internet.
It has no way of verifying the accuracy of the information other than comparing it to the data it's trained on, a self confirming loop.This can result in it producing truth, fiction or outright lies as its output, with any option put forward with the same enthusiasm/sincerity.
It's as simple as "Garbage in -> Garbage out, and there is inadequate garbage filtering on the input of many, if not most, AIs.At the risk of sounding like an evangelical artist... All it can do is plagiarise, not create.
One might add to that as more and more of the internet AI training digests consists of AI made content the level of "truth" it is being trained on is getting diluted too.
That said I do believe AI can come up with original creations - or at least as original as most new designs in this world. Currently many AI models operate with a concept of "latent space" which holds the AIs perceived essential characteristics and or features of a given subject or relation. It can be somewhat likened to our memory of a book after reading it - we can reiterate the main content, storylines and characters but not the exact text.
Such a latent space might for instance codify the requirements for a functional chair and limits given by say economical production and materials. Within that envelope of possibilities it can then be asked to output a chair design. To the degree it has also been trained on other existing chair designs it can further be asked to avoid those (to an adjustable degree) and still be within the design envelope of possible chairs. That will yield original designs - some likely more original than most human original chair designers do today.
-
Another AI masterpiece that I just had to share!
Def

-
See, that is original! :D

-
I literally cannot stop laughing at this photo!
Def
-
Maybe AI thinked that tent needs heater
-
I guess it makes sense to keep your firewood dry!
Def
-
The shoelaces are just amazing too.
One can suspect the creator prompt to have been something like "put a bonfire inside a tent" - which would be inspired indeed! :D
-
Yes, don’t believe everything you read on the internet. And that goes double if it was written by AI…
In defense (sort of) of the poor AI, there was in fact a blog post stating this, though to any adult of at least average intelligence it was obvious that the blog post was an April Fools joke.

-
Sixty inches seems like a reasonable restriction to me.

Def
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login




