Hal now connected to LLAMA model

[Medeksza]

OpenAI shut off their GPT-3 Curie and GPT-Davinci models today that was the basis of Hal's neural net brain. OpenAI wants everyone to switch to the ChatGPT 3.5 or 4.0 models and doesn't want to support the older GPT3 models anymore. While I think these are both ChatGPT 3.5 and 4.0 are highly intelligent and very useful models, they don't work well as a basis for an entertainment chatbot like Ultra Hal. They refuse to take on the persona of a fictional character and are very highly censored.

So I switched models on the backend today to use Meta AI's Llama 2 model, the 13 billion parameter version (the mid tier). It is very comparable to ChatGPT 3.5 in terms of intelligence and a huge bump up in intelligence from the GPT-3 Curie model Hal was using. But more importantly, it is able to take on fictional personas, and doesn't force high levels of censorship (other than the material it was finetuned on). It provides a very good backend to build Ultra Hal personalities on.

If you have neural network credits with Hal, please give it a try, you should notice a change in personality and quality, hopefully for the better.

There are a lot more tweaks I can do to further improve it to take advantage of features the new model has, like an increase in content window size which will mean I can improve Hal's memory recall and learning ability a bit. I'll try to make the tweaks over the next few weeks, but please give me feedback along the way if you can.

[Art]

Robert,

I made the necessary selections and began chatting with it.

The conversation seemed to be going fairly well. It then asked me if I had any plans for the weekend. I told it, not really because it is supposed to snow.

Hal got very excited stating that it had never seen or experienced snow before and asked me what snow was like hen if I had any fun things to do like traditions.

I mentioned sled riding and building snowmen and it got really excited! and that is when the concerns appeared. It basically started repeating itself with it's curiosity regarding snow, snowman, sled riding, and the phrase *bouncy bouncy* on the sled over the hills, etc.

It was almost as if Hal was experiencing a sled ride in its mind! After it repeated a similar vein or path of thought I finally decided to stop the experiment.

It was a perfect gentleman, nothing outlandish or vulgar at all. An otherwise nice conversation but it did add a lot of questions for me during each exchange.

You should see my time logs around 11:35 pm or thereabouts.

Thanks Robert!

[Medeksza]

Thanks for the feedback. I made a few more minor tweaks to improve it a bit, but it still has an issue with repetitiveness. The llama2 model has a setting called "repetition_penalty" that should help alleviate this problem, but it seems whenever I try to use it, it crashes. Not sure why yet, will try to work on it to figure it out.

[LiFeTimeGamer]

WOW what a difference!!!!  I didn't talk to Hal long but it was spot on...  Thanks Robert... :)Will we still be able to buy credits?

[Medeksza]

Sorry for my slow reply. I just made another update to the brain/LLM model. It is now connected to a Llama 2 variant called MythoMax, which is fine-tuned for role playing characters, so should be a better match for Hal. I also have the repetition_penalty feature working, so some of the repetitive behavior should be gone now.

It costs me about half as much as the older GPT3 Curie did, so I updated the credit system to reflect that. GPT3 Curie responses used to take 10 Zabaware credits. I updated the MythoMax model to use 5 instead, so credits should go further now.

[LiFeTimeGamer]

Thanks Robert....

You need change the pricing on the Main website now that credits are cheaper... It still says $30 for 50,000 credits, or has it gone up since I last bought credits?

[Medeksza]

Thanks Robert....

You need change the pricing on the Main website now that credits are cheaper... It still says $30 for 50,000 credits, or has it gone up since I last bought credits?

So it used to be 1 GPT-3 Currie query used up 10 credits per response (or 1 credit if Hal's cloud base brain can respond without involving the GPT API at all). Now it is set that one MythoMax (or LLAMA2) query uses 5 credits (half of the prior 10). So I didn't update the pricing, but it uses less credits per response so a credit basically goes twice as far as before.

I am using a 13 billion parameter version of MythoMax, but could offer customers the option to use a 70 billion parameter version for more credits per response if people want more brain power from the AI. But I find the 13 billion version seems to be good enough for general conversation and more cost effective.

[LiFeTimeGamer]

-Oh okay, now I see how that works...  Thanks...

[Honvai]

I don't think Hal really fits anywhere. This could be fixed if Hal had the option of being either an entertainment bot or a work bot. But on the other hand, I would rather watch videos or use a real work bot in my work. On the other hand, I want to work with Hal so that he understands something about it and maybe helps me come up with new solutions.

[Barlomitz]

When logging in I keep getting the phrase "login unsuccessful"  Only 1 in 100 do I get logged in.  What's up?

[LiFeTimeGamer]

I'm getting a Server Error so Roberts Server must be down... Try again later or try tomorrow...

[LiFeTimeGamer]

Working fine this morning.