nutanc

Do any of the vector databases have support for bit embeddings. We have created bit embeddings[1] for sentences and they save a lot of space. Currently we are just using numpy and sometimes faiss to search through these bit embeddings. Would love for one of the vector dbs to support bit embeddings natively. Then we don't have to engineer that piece :)

[1] https://gpt3experiments.substack.com/p/building-a-new-embedd...

Learning by playing is the holy grail. This means that you will have fun and at the same time you develop a skill. What I have found is that play happens in different environments. For most kids now, these environments are virtual.

One of the best example of playing through learning for my kid was him learning the multiplication tables playing Roblox[1]

[1] https://www.roblox.com/games/10489578087/Multiplication-Game

You guard your API, I guard my data.

Agree with first part: LLMs will make it easier for teachers to make teaching more fun(it's easier than Googling and building)

Disagree with second part: Personalized learning is not here. Personalized learning does not mean I like rap and I want my lessons to be in rap. Even if that's the case, the LLMs are no where close to maintain context for a 45 minute teaching session.

We now have some tools to improve education but personalized learning is most certainly not here. Yet.

For the same reason OpenAI is not allowing text generated by its system to train competing systems. I want to maybe build a competing system with my data and don't want OpenAI to be a competitor in that space.

The most important question is, what new applications can be developed using GPT4 which couldn't have been developed using GPT3.5?

Around $250k for getting your own chatGPT. Sounds reasonable.

Of course this is for the enterprises. It's upto their managers to define how they can extract $250k worth of value from the investment which I think is easy given the AI frenzy.

Of course for startups they will have to use the direct OpenAI API on a pay per use model.

One question that is not clear, can I take a ChatGPT instance and provide an API similar to OpenAI API and charge for it?

Mermaid+chatGPT is a killer combination to create diagrams from plain text.

Describe your diagram in plain text and ask chatGPT to generate Mermaid syntax for said diagram. Then adjust the text generated by chatGPT to get the perfect diagram

The generative AI piece in these kinds of solutions is just the last step which is to take the list of nearest neighbour sentences or paragraphs and generate an answer from that set.

So I think these class of search engines should actually be called something other than generative AI search engines.

Maybe generative AI powered search summaries?

It's 2023 and the site does not have a neural search interface :)

For example pass the text from https://www.vcinsights.co/memos?recordId=recu0WgZCKeNxjVpf to chatGPT and get the following summary:

    1. The use of AI in sales email coaching: Lavender uses AI algorithms to score emails and suggest edits, as well as generate a personality profile for prospects.
    2. Strong team with deep understanding of end users: The co-founders of Lavender have a strong passion and focus in sales, and have designed the product with deep understanding of its end users, sales development representatives and account executives.
    3. Strong customer love and community: Customers have seen significant increases in cold email conversion, booking more meetings and reducing email creation time. Lavender has built a tight-knit community of sales leaders and SDRs, with customers even getting tattoos and recording music for the company.

I remember seeing a chatbot for fiction from the same website. Looks like it's the same tech but applied to a different book. As the comments have mentioned, it's a fun experiment, but lots of engineering has to be done still to stop out of context answers.

Agreed. That's why the approach we are taking is to answer the question 'and' show the links/paragraphs which answer the question.

Demo we have created for our website: https://speech-kws.ozonetel.com/ozosearch

It might be so. Even if the sample text was not part of training material the GPT-3 LLM is just correcting spelling and grammar based on previous words. So, there is no "intelligence", but gets the job done.

We have created a binary embedding to help us with the embedding sizes. I think in the future we will see a lot more research in reducing the model and embedding sizes.

https://medium.com/ozonetel-ai/compressing-bert-sentence-emb...

The search piece can be self hosted and can be compliant etc. But the generative piece is a little harder to do. The models are really big if you need good quality output and the best models are closed APIs.

In my experiment for our web site, https://speech-kws.ozonetel.com/ozosearch, the search piece is self hosted(know more links). But the generated answer is given by OpenAI API.

Another reason might be that ozonetel might not that popular in the embedding space. Other typos like "what is cloud cll center" seem to give the correct results as these are common words in the embedding space.

Yeah, looks like the embedding threshold has to be adjusted. I have kept it pretty high for now so that we get pretty valid matches only.

I used the huggingface embeddings(compressed using our algorithm, https://medium.com/ozonetel-ai/compressing-bert-sentence-emb...) and OpenAI API. Will do a blog post next week on how we built this step by step.

Its working for me. Is it just slow or the web page itself is not loading?

Trying something similar. Using a mix of embeddings and generative AI(davinci) to answer questions from scrapped data of website. Scrapped data for our website (Ozonetel.com) and created this site.

1. Scraping website. Used default node scraper. 5 mins.

2. Generated huggingface embeddings. 10 mins.

3. Use code AI to generate basic website. 5 mins.

4. Created prompt to limit to answers known. 1 min.

So in 30 mins we are able to create a site search powered by generative AI.

Disclaimer. This is still a work in progress.

http://speech-kws.ozonetel.com/ozosearch

Agree. This can be very useful for an online learning tool. Will see how I can leverage this. Thanks for this.

That would seem so right? I thought so exactly too. This works when your engineering team is small. But once you go beyond the first level of engineering and its mostly maintenance and updates, the people who apply for the job are not interested in learning new things. They have learnt one thing and want to apply that skill 9-5 and go home. For those roles hiring for react is easier.

We tried to use other frameworks but finally came back to React for only one reason. It's easier to hire for React. The talent pool for React is almost 10 times that of other frameworks.

And for now, this reason is good enough for us.

The following phrase seems to work in case someone wants to try out directly on SD DreamStudio:

"A black and white tattoo design of a bouquet of roses, in the old school style"

Basically structure is "A {black and white/color} tattoo design of {user prompt}, in the {style}"

So if I understand correctly, the web site seems to take input from the users, build a phrase from the selections(like maybe,"A black and white tattoo of a bouquet of roses, in the old school style") and then pass that phrase to stable diffusion and generate the images. If you have access to DreamStudio, I think we can try out phrases like this for free and see the output.

Thats a good point. Let me try turning those off.

Familiar with it. But as they themselves say,the fine tuned model(not released) achieves 5/7 rating for only 15% of the time. So 85% of the time the results are not satisfactory.

I "want" to teach. It's just that, teaching as a profession is not sustainable for me.

I am sure a lot of teachers are in the same boat as me.

The context is lost. We have tried all that. We tried that when there was free usage. Now it is too expensive. If you believe it can be done please do go ahead. There is a huge market out there for summarizing novels. No one has cracked it yet.

The question was about movie scripts and children stories. Not about paragraphs. Sure it works ok for paragraphs(even then it fails many times) but the moment you cross the 2000 token limit, it does not work. When GPT3 was first released we did a lot of experiments on summarisation. We discussed a lot in the OpenAI slack. But no one could come up with a right prompt for summarisation. Yes, it works for toy paragraphs and is fun to show off. But I wouldn't build a summarisation startup on top of GPT3. Yet.