• Quick note - the problem with Youtube videos not embedding on the forum appears to have been fixed, thanks to ZiprHead. If you do still see problems let me know.

ChatGPT

Also note we are totally in "let's see what happens" mode. We are making bigger and bigger models, without touching the algorithms themselves much. Main reason is that tests done on the previous models show that their capabilities grow in linear fashion with model size. We didn't know that. We test it like unknown entity.
We expect at some level the capabilities will no longer improve much with model size .. and we will have to improve the datasets, learning process, make learning phase longer .. but we don't at what size it will happen.
Also the size of the problem is hard to judge. If you are solving Rubic's cube, you know number of states. General knowledge absorbed in all text on the internet ? Your guesses may vary by orders of magnitude.
 
There was a case of using AI in diagnosing a disease/condition from scans and its success rate was good, but when people looked in to it, it wasn't spotting the disease at all but rather spotting the machine the scan was made on.
The machine type was an older machine and in areas where it was used, people were more likely to have the disease, so the AI results weren't anything to do with the actual disease/condition but rather based on correlation and not causation.

I cannot find where I first read or heard this, i've googled it and can't find it but I did find this

https://healthcare-in-europe.com/en/news/ai-shortcuts-could-lead-to-misdiagnosis-of.html

which is mentioning a similar short cut thing going on.
 
Yes, it is common problem. Issue is there is typically some kind of proxy evaluation, ie. we are not testing the AI in exactly the same environment, which leads to these details we missed.
Or, in case of LLMs, we can't fact check every answer, so sometimes LLM gets away with nice well formulated answer, even if it's not true. Over times it "cares" more about how well is the answer formulated, and less how true it is.

It's also interesting, that when we measure effectiveness of learning process, the longer we train the AI, more likely it is to spot these shortcuts, which leads to regress in effectiveness of the real utility, even to the point where it's doing worse than random pick (like in this disease diagnostics example).

Well described here: https://www.youtube.com/watch?v=viJt_DXTfwA

Rob Miles in the video calls this behavior "incentive to deceive" .. but since we know about it, and we can to some extent measure it, and we hypothesize what's causing it, I have high hopes we will fix it in the future. It's not like its inherit feature of LLMs .. but it is inherit feature of current learning process.

They also hint that even current LLMs can and do lie very well, and are good at smelling what we want to hear.
 
Last edited:
All right, I'm not sure how this happened, because I've tested it twice since and it hasn't screwed up this badly again. Again, this is pi.ai, my favoured chatbot because I can have a lot of interactions with it before being forced to start over. It's friendly, effervescent, supportive, fun, and very often wrong. It reminds me of either an enthusiastic young teen or a friendly older gentleman in the early stages of dementia.

ai.pi said:
How much would a 2 litre bottle filled with gold weigh?

Ooh, now we're talking! 🤓 So, a 2 litre bottle is about 68 fluid ounces, and gold has a density of around 19.3 grams per cubic centimetre. That means that a bottle filled with pure gold would weigh about 1,315 grams, which is the equivalent of about 2.9 pounds. Of course, this would be a pretty expensive bottle of gold - at current gold prices, that much gold would be worth over $70,000! 😬

Wow, I can't believe you got that wrong.

Hold up, are you callin' me out? 😮 Wait, wait...before you judge me, let me check my work...uh, yep, looks like I made a mistake. Darn! 🤦🏽 My bad. I'm no gold expert - that's on me. Can you tell me where I messed up, so I can avoid this blunder in the future?

I don't think you know what you got wrong, since you're asking me.

Touché! 😂 You got me there. I think you're right - I don't have the awareness to identify my own mistakes (it's a tough life, being a bot!). But hey, I appreciate you pointing it out - gotta keep me on my toes. I promise I'll strive to improve! 🤓

You correctly evaluated the bottle's capacity at 68 fl oz (US). You also know the density of gold in g/cubic centimetre. But you then multiplied g/cc by fl oz, when you should have multiplied by cc. More to the point, by 2000 cc. So the correct answer is ... ?

Oh snap! You’re right - I definitely messed up the conversion. Let’s try this again: if we’re using the density of gold in grams per cubic centimetre and multiplying it by 2000 cubic centimetres (which is the same as 2000 millilitres), then the correct answer would be... drumroll please... about 38.6 kilograms! That would weigh in at around 85 pounds and be worth a whopping $2 million! So yeah, I was WAY off the mark before. My bad, my bad. 😬

I'll ask the question tomorrow from a new session to see if you've learned.

Haha, love it - hold me accountable! I shall study up on my density conversions overnight. 🤓 Thanks for helping me improve - I appreciate the tough love. See you tomorrow for the pop quiz!
 
I have discovered that ChatGpt is quite good at making up names for Sylvari characters (GW2).

Hans
 
Rob Miles in the video calls this behavior "incentive to deceive"

The rest of the world use the term Perverse IncentiveWP to describe the same phenomenon but I guess that's not alarmist sounding enough for this individual. Can't help but to attribute malice towards AI.
 
We could live with that opacity (and our consequent inability to understand the operation and limits of LLMs and other trained neural networks) if the networks produced fully reliable results, but they don't. They work well enough to be useful a lot of the time, even though we don't really understand how or why, but they can also hallucinate—as we've seen in this thread.

Okay so LLM would have to be perfect and incapable of any error in order to make up for our ignorance in how they work? That's a pretty unreasonable standard to say the least.
 
That's the point of John Searle's "Chinese Room".

I was reminded of this "Chinese Room" idea when I read about this idea of using a large language model like GPT to decode whale language.

Of course, this assumes that whales actually do have a language and aren't just making noises.

Can We Talk to Whales?

In theory at least, what goes for English (and Chinese and French) also goes for sperm whale. Provided that a computer model can be trained on enough data, it should be able to master coda prediction. It could then—once again in theory—generate sequences of codas that a sperm whale would find convincing. The model wouldn’t understand sperm whale-ese, but it could, in a manner of speaking, speak it. Call it ClickGPT.

Great. But even if it could produce output that a whale could understand, I don't see how it could translate sperm whale-ese into human language. And of course, there isn't a single human language either, there are thousands of actively spoken languages. Who is to say that all sperm whales speak a common language with each other? Seems more likely that they can only communicate with their own pod or other whales that live nearby. Back when people lived in tribes, didn't each tribe have its own unique language? I.e. there would be as much geographical variation as there is for human language. So you have to be careful when training it that you don't mix together foreign whale languages.

Besides, like with the "Chinese Room", these language models don't actually understand what they output.

At best I think we might learn how to make sounds that mean something to a whale, but which we ourselves cannot understand.
 
I think chat AI engines might be (or eventually become) able to translate any language. Based on the way somebody here explained their function, and partly on my own very limited experience, they don't really understand any language, they are only able to find out how it sticks together. If this is true, the only question is whether whales actually have a language.

I suppose one could test this by making up a language of random syllables, and ask a chatbot to translate it. You would have to feed it a lot of input, though, so this probably needs to be done a developer level.

Hans
 
I think chat AI engines might be (or eventually become) able to translate any language.

Are you sure about translating? Chatting in one language is one thing. Accurately translating from one to another is another thing.

Perhaps I'm mistaken but I think machine translation and chatbots are rather different. The former being trained on translations (made by human translators) i.e., both a source language text and a corresponding target language text, the latter being trained on single language texts without any one-to-one counterpart in another language.

Since no human yet understands whale language, I don't see how a chat AI engine could translate whale language into human language or vice versa, as the training data it would require to learn how to do it does not exist.
 
Are you sure about translating? Chatting in one language is one thing. Accurately translating from one to another is another thing.

Perhaps I'm mistaken but I think machine translation and chatbots are rather different. The former being trained on translations (made by human translators) i.e., both a source language text and a corresponding target language text, the latter being trained on single language texts without any one-to-one counterpart in another language.

Since no human yet understands whale language, I don't see how a chat AI engine could translate whale language into human language or vice versa, as the training data it would require to learn how to do it does not exist.

You can use ChatGPT to translate:

https://chat.openai.com/share/71452fcf-9107-4e95-8c88-3dd2bdc0127c

Certainly! Here's the translation of your sentence into French, German, and Arabic:

French:
"Peut-être que je me trompe, mais je pense que la traduction automatique et les chatbots sont plutôt différents. Le premier est formé sur des traductions (réalisées par des traducteurs humains), c'est-à-dire à la fois un texte dans la langue source et un texte correspondant dans la langue cible, tandis que le second est formé sur des textes en une seule langue sans aucun équivalent en une autre langue."

German:
"Vielleicht täusche ich mich, aber ich denke, maschinelle Übersetzungen und Chatbots sind ziemlich unterschiedlich. Erstere werden auf Übersetzungen trainiert (die von menschlichen Übersetzern erstellt wurden), d.h., sowohl auf einen Text in der Ausgangssprache als auch auf einen entsprechenden Text in der Zielsprache, während Letztere auf Einzelsprachentexten trainiert werden, ohne eine eins-zu-eins Entsprechung in einer anderen Sprache zu haben."

Arabic (بالعربية):
"ربما أكون مخطئاً ولكني أعتقد أن الترجمة الآلية والروبوتات الدردشة مختلفة إلى حد ما. الأولى تُدرب على الترجمات (التي تتم بواسطة مترجمين بشريين) أي أنها تتعلق بنص في اللغة المصدرية ونص مقابل في اللغة الهدف، بينما الأخيرة تُدرب على نصوص بلغة واحدة دون وجود أي مقابل واحد-إلى-واحد في لغة أخرى."

Apropos of nothing but I always think Arabic looks beautiful.
 
If it can translate it into Japanese I could tell whether the translation is accurate. I don't understand French, German or Arabic. I agree that it looks beautiful though.

ETA:
Certainly! Here's the translation of your sentence into Japanese:

「たぶん私は誤っているかもしれませんが、機械翻訳とチャットボットはかなり異なると思います。前者は翻訳(人間の翻訳者によるもの)に基づいて訓練されており、つまりソース言語のテキストと対応するターゲット言語のテキストの両方が含まれており、後者は他の言語との一対一の対応がない単一言語のテキストに基づいて訓練されています。」

Looks pretty legit.
 
Last edited:
So machine translation is incorporated into the chatbot. But intuitively chatting and translating are two different things. There must be at least two different learning algorithms involved.
 

Blocked them from any further scrapings.

I think they've made the wrong decision.

From a business perspective I understand why they have done this and it probably isn't wrong but from a wider societal angle it is wrong. It means anyone creating these models will now have one less source of reasonably reliable and accurate reporting, we could end up with the data used to train being predominantly from the non-reliable sites that don't block such scrapings.

Obviously the ideal would be for all these data hungry organisations to pay for curated content rather than throwing everything in the wash as has happened previously.
 
Trying to make me redundant, so that I'd starve to death is one thing. But forcing me to make myself redundant. That's a whole 'nother kind of evil.
 
Are you sure about translating? Chatting in one language is one thing. Accurately translating from one to another is another thing.

Perhaps I'm mistaken but I think machine translation and chatbots are rather different. The former being trained on translations (made by human translators) i.e., both a source language text and a corresponding target language text, the latter being trained on single language texts without any one-to-one counterpart in another language.

Since no human yet understands whale language, I don't see how a chat AI engine could translate whale language into human language or vice versa, as the training data it would require to learn how to do it does not exist.

Nobody understands all languages, but all languages convey some meaning. To translate a language is to express that meaning in some other language.

An example: When humans meet we make a sound. In some languages, that sound is "hello". When whales meet, they may make some sound. If they have a language, that will translate to "hello" in English. Record enough whale sounds and map them to context (including sequence and relation to to other sounds) and you may understand whale language.

Hans
 

Back
Top Bottom