#new online learning experience - based on LLM and computation

21 messages · Page 1 of 1 (latest)

toxic aspen
#

I'm working on an AI Tutor along the ideas of Stephen Wolfram...

Here is the thing:

Today

online learning - doesn't matter if it is a textbook that made available online or some video lectures where someone sat in front of the webcam - is content that is prepared ahead of time. If the learner doesn't understand something there is little they can do... hence the push for live, cohort based courses or selling access to some sort of online discussion forum. Neither of these scale well, especially not the live Q&As.

The Future

using generative AI there is now an alternative. Textbook content does not need to be prepared ahead of time - it can be generated on the fly - both text and images.

And with the conversational chat interface the AI is able to answer questions, do a socratic dialog and actually help with the learning process.

the challenge

Even though ChatGPT is really good... it seems to favor giving an answer rather then saying "I don't know". I suppose one can tweak these responses with prompt engineering... but the point is that some answers should come from clean, curated, up-to-date data set. And some answers should be computed.

So today we have two systems:

generative AI that is very good in understanding English questions and giving full sentence verbose answers

and the Wolfram technology stack that (despite their efforts) is weak in natural language understanding but can give quality answers

The idea is to create an "agent" that orchestrates and we bounce queries between the two, feeding in the factual data for GPT to answer in nice English sentences.

3298c3eb001bbed90f1d616da66708480096a0a1b6e81bd4f8a2d6e9b831d301.webp
lunar estuary
#

Good job and nice idea. If you can solve for the inaccuracies in the data synthesis, this could be a useful learning machine/platform.

are you planning to have a public or beta release soon?

toxic aspen
#

Right now it's available if someone is interested.
You do need to have Wolfram Desktop to run it and bring your own OpenAI API key

#

This version is built on GPT3 model.

#

I've started uplifting it to GPT3.5 as soon as it launched

#

learning is always the priority... if someone really wants to use my products and truly cannot afford it should just shoot me an email

night ivy
#

LOL found it funny i was literally in the process to blow its CPU to give me more ability to poorman tune it to answer my question. And i saw this post, gonna check ur stuff out right now. The idea seems sound

night ivy
#

I do have a question though, being that really deep down, it is more accurate to state that facts do not exist than to say there are very, VERY few "actual facts" 100% etc. When attempting to set a boundary such as this aren't you building the model on bias? For example ( I like the fact you chose history as the starting subject, major history fan) History, although being more of a search for truth and accuracy than what most people assume the study of History is, its as about a prime of an example as you can get, into how low the human can actually get. With an infinite list of vulnerabilities i.e. perversion, exaggeration, misleading, misunderstood, misinterpretation, lost in translation, actually lost, hyperbole, scheming, propaganda. I mean the list goes on and on, a prime example of the the nature of humans. All that being said, again I am simply wanting to pick your brain on the subject with zero intention of malicious intent, wouldn't the idea of letting a truly unbias model determine the most likely or closest to truth response based on its commutated results on its given dataset, perhaps be the most efficient route towards a closer truth? Why, do you suppose isn't this a more attempted method or even talked about, for that matter, for building and training effective AI?

toxic aspen
#

If I take the Greek - Persian wars as an example or the rise of the Ottoman Empire... I'm sure it's seen (and taught) differently in Europe than in Asia.
I've also found that the AI will not bring up subjects that a history book covers unless I ask about it specifically.
I think we shouldn't try to make more of this technology than it is...
I understand your question... but don't really have a good answer.
If an AI reads in text about how people suffered throughout history it won't really understand that suffering is bad because it's not sentient. It will at best just repeat like a parrot that it was bad, and we shouldn't do it etc.

#

If it one day does understand it... than we have other things to worry about than bias in historical story telling. No?

stiff vale
#

What about implementing an educational model aimed more towards science or mathematics as well these disciplines do not seem to require much emotional responsivity if any ? I would imagine once all the curriculum has been created once it’s and provisions and nuance has been completed . I would imagine one could then possibly almost definitively outsource other teaching methods that are useful to an intelligent being that could become relatively close to that of college model without all of the student debt.

toxic aspen
#

We have even bigger fish to fry ... thinking of the climate and ecological crisis, student debt will be the least of young people's problems

fervent sparrow
#

Is this built on langchain or purely on top of Wolfram?

toxic aspen
#

I've looked at langchain and took some ideas. It's purely Wolfram

toxic aspen
#

We have made some good progress.
GPT-3.5 has brought huge improvements and was relatively easy to switch to. And now GPT-4 takes it to the next level - it passed AP History test, finishing in top 10%.

Russia.png
toxic aspen
#

new online learning experience - based on LLM and computation

#

This demo learning session captures learning about end-of-19th-century Russia.
The model is upgraded to GPT-4 and it is just so much better:

  1. it does indeed pay more attention to how it is setup with the "system" role
  2. and man, it uses tools ( It's not even the ChatGPT plugins version ... though that is certainly super exciting)