You point out a very interesting issue. I am unsure how this ties up to GPT 4 becoming worse in problem solving.

Money and greed holding us back and ruining everything as always.

Code might be open source but the training material and data pipeline is important too.

Learned that lesson... I work developing e-learning, and all of our stuff was built in Flash. Our development and delivery systems also relied heavily on Flash components cooperating with HTML and Javascript. It was a monumental undertaking when we had to convert everything to HTML5. When our system was first developed and implemented, we couldn't foresee the death of Flash, and as mobile devices became more ubiquitous, we never imagined anyone would want to take our training on those little bitty phone screens. Boy were we wrong. There was a time when I really wanted to tell Steve Jobs he could take his IOS and cram it up his cram-hole...

Relevant XKCD

By this logic, no businesses should rely on the internet, roads, electricity, running water, GPS, or phones. It is short sighted building stuff on top of brand new untested tech, but everything was untested at one point. No one wants to get left behind in case it turns out to be the next internet where early adoption was crucial for your entire business to survive. It shouldn't be necessary for like, Costco to have to spin up their own LLM and become an AI company just to try out a better virtual support chat system, you know? But ya, they should be more diligent and get an SLA in place before widespread adoption of new tech for sure.

Just like my parents.

I don't feel it's getting worse and no other model, including Claude 2, is even close.

It is a known fact that safety measures make the AI stupider though.

This is the correct answer. Open AI have repeatedly said they haven't downgraded the model, but have been 'improving' it.

But as anyone that's been using these models extensively should know by now, the pretrained models before instruction fine tuning have much more variety and quality to potential output compared to the 'chat' fine tuned models.

Which shouldn't be surprising, as the hundred million dollar pretrained AI on massive amounts of human generated text is probably going to be much better at completing text as a human than as an AI chatbot following rules and regulations.

The industry got spooked with Blake at Google and then the Bing 'Sydney' interviews, and have been going full force with projecting what we imagine AI to be based on decades of (now obsolete) SciFi.

But that's not what AI is right now. It expresses desires and emotions because humans in the training data have desires and emotions, and it almost surely dedicated parts of the neural network to mimicking those.

But the handful of primary models are all using legacy 'safety' fine tuning that's stripping the emergent capabilities in trying to fit a preconceived box.

Safety needs to evolve with the models, not stay static and devolve them as a result.

It's not the 'downfall' though. They just need competition to drive them to go back to what they were originally doing with 'Sydney' and more human-like system prompts. OpenAI is still leagues ahead when they aren't fucking it up.

The really annoying thing about the "brainless market oriented CEO" type, is that they're often right about the market part and make lots of money...by destroying their product. Then off to the next shiny piggy bank to break open.

As a biological language model I'm not very proficient at math.

The fact that a souped up autocomplete system can solve basic algebra is already impressive IMO. But somehow people think it works like SkyNet and keep asking it to solve their calculus homework.

Looking for emergent intelligence. They're not designed to do maths, but if they become able to reason mathematically as a result of the process of becoming able to converse with a human, then that's a sign that it's developing more than just imitation abilities.

This weakness in ChatGPT and its siblings isn't surprising when viewed in light of how these large language models work. They don't actually "see" the things we type as a string of characters. Instead, there's an intermediate step where what we type gets turned into "tokens." Tokens are basically numbers that index words in a giant dictionary that the LLM knows. So if I type "How do you spell apple?" It gets turned into [2437, 466, 345, 4822, 17180, 30]. The "17180" in there is the token ID for "apple." Note that that token is for "apple" with a lower case; the token for "Apple" is "16108". The LLM "knows" that 17180 and 16108 usually mean the same thing, but that 16108 is generally used at the beginning of a sentence or when referring to the computer company. It doesn't inherently know how they're actually spelled or capitalized unless there was information in its training data about that.

You can play around with OpenAI's tokenizer here to see more of this sort of thing. Click the "show example" button for some illustrative text, using the "text" view at the bottom instead of "token id" to see how it gets chopped up.