Did a little more playing with Bard. It really isn’t beyond English. You can ask it to create a story, or a dialogue, but seemingly can only spit it out in English. I tried German and Spanish with no luck.
I believe the Bard you have in beta is still a very limited version of its capabilities. They will open it step by step. They didn’t want to release it but they don’t have any choice with chatGPT out there. If they don’t do something now, when they will open it, chatGPT will be already above everyone else.
Do you mean run the chatGPT entirely locally or do you mean some code that hits the APIs?
This is exactly what I have been doing for the last week.
Every day I have my list of words from my SRS (in my case anki, but you can use lingQ’s SRS).
I have been putting the ones I got wrong initially into a note on the site and when I’m finished, I have been getting chatGPT to write me short stories for those words in the list.
I feel like it is producing stories targeted exactly at my level rather than too low or too high. And it is reinforcing exactly the vocabulary I want it to reinforce (on a daily basis!) rather than me having to wait for zipf’s law to provide me with examples of the words I don’t know but need to know.
I’m really hoping this is the secret sauce to get over my current grind. I feel like it is working. But I don’t have enough stats to say for sure. In a month or so if I am able to keep up the volume (I’m able to read 4-5 X the volume), I will have stats proving or disproving that doing this will work to increase vocabulary learning and retention.
Running locally is the holy grail of course, the closest you can get currently is probably by using LLaMA (GitHub - facebookresearch/llama: Inference code for LLaMA models) by Facebook Meta. You need to apply to obtain the weights, or you could also use the internet to download them but that would be naughty.
The whole thing is a bit on the heavy side, afaik the largest 65B parameter model requires over 240 GB GPU memory. For a CPU implementation look at (GitHub - ggerganov/llama.cpp: Port of Facebook's LLaMA model in C/C++) through the use of quantization the requirements are reduced significantly, so for example the 7B model nominally requires 13GB when quantized to 4bit requirements go down to 3.9 GB.
To recreate a chat experience look into Alpaca (GitHub - tatsu-lab/stanford_alpaca: Code and documentation to train Stanford's Alpaca models, and generate the data.)
Due to inadequate hardware (only 16GB RAM) I was thus far only able to test 7B and 13B it’s about as slow as ChatGPT at peak times and the responses didn’t impress. No language capabilities beyond English as far as I see. This might need further investigation / training. I’m sure the community will come up with an ecosystem around it, for example I saw GitHub - ymcui/Chinese-LLaMA-Alpaca: 中文LLaMA&Alpaca大语言模型+本地CPU/GPU训练部署 (Chinese LLaMA & Alpaca LLMs) that adds Chinese capabilities on top of LLaMA.
I feel we’re currently still at the start of this revolution, so everything is quite technical and rough around the edges. Also the pace of development is rapid and it’s hard to keep up (at least for me).
So when you say “inadequate hardware (only 16GB RAM)” do you mean GPU RAM or normal RAM?
I tried it on a 2019 Macbook with 16GB RAM and 6 core Intel i7 processor, running on CPU only, using llama.cpp, no GPU. Super slow unfortunately.
@bamboozled:
I tried it on a Macbook with 16GB RAM, running on CPU only, using llama.cpp, no GPU.
Interesting. Do you think RAM is the bottleneck? So if you ran it on a cheap server with say 128GB of RAM it would be faster?
Today I was approved by Google to try it (BARD) out for giving prompts. Though I had to change my VPN to the USA and was not currently available in Germany. It is so “undercooked” that I can not even compare it with ChatGPT in the same breath. Its explanations were simple for explaining grammar concepts in German. However, it could not even come up with a simple story in German. As a German would say “Wie schade!”
I’m no expert, but here there are multiple things going on: In my case the CPU is the bottleneck, provided the models fit into RAM (7B or 13B), larger models would require swap which is probably terrible. I believe a more modern Apple Silicon Mac would be much faster (using llama.cpp). I know nothing about GPUs and don’t even own one currently.
I just tried it again and asked it “what is the meaning of life?”, after about 30 seconds it gave me this piece of wisdom: “Life is short, so make it sweet”. Hope this helps.
LLaMa 30B has been optimized to work on ~6 GB RAM, using mmap.
Probably… 
Fake news was debunked further down in the discussion, some tools reported nonsense. Also: https://twitter.com/jeremyphoward/status/1642726595436883969?s=46&t=LFe0jyu2TCnvXZNEvnb71g
I’m still on an “ancient” version of the alpaca fork, need to look into how to convert the models at one point (expecting it to be painful). But here is a demonstration on how to build a website in 10 steps: testing alpaca 7B - YouTube
I think this a very accurate prediction. The willingness of people to invest time into learning a skill will decrease the better AI gets in doing that skill.
In one sense this is impressive. It takes a speaking approach and discounts the comprehensible input approach though, which seems to have significant research backing it. Can it comment on “Is the comprehensible input approach a good way to learn a language?”
“what is the meaning of life?”
Wasn’t the correct answer:42?
We’re doomed if even the AI avoids numbers (like my students)
I tried the new vicuna (https://vicuna.lmsys.org/):
bamboo: What is the meaning of life?
vicuna: The meaning of life is a philosophical question and has been debated by scholars, thinkers, and individuals for centuries. It is a subjective topic that can have different meanings to different people. Some people believe that the meaning of life is to seek happiness and fulfillment, while others believe it is to serve a higher purpose or to follow a certain set of values or beliefs. Ultimately, the meaning of life is a personal matter that each individual must determine for themselves based on their own experiences, beliefs, and desires.
bamboo:is it 42?
vicuna: The answer “42” is a reference to the famous book “The Hitchhiker’s Guide to the Galaxy” by Douglas Adams. In the book, the character Deep Thought is asked to determine the meaning of life, the universe, and everything. After thinking for a long time, it gives the answer “42.” However, this answer is meant as a joke and doesn’t provide a real answer to the question of the meaning of life.
I used the following prompt: ./main -m ./models/vicuna_13B/ggml-vicuna-13b-4bit.bin -t 4 -c 2048 -n 2048 --color -i --reverse-prompt ‘### Human:’ -p ‘### Human: What is the meaning of life? ### Assistant:’
I think the answers are actually quite close to what ChatGPT produces.
Your answer from GPT 3
Is the comprehensible input approach a good way to learn a language?
Yes, the comprehensible input approach is considered a very effective way to learn a language. This approach emphasizes the importance of exposing language learners to input that is just slightly above their current level of understanding, so that they are challenged but not overwhelmed. The idea is that by gradually increasing the complexity of the input that learners receive, they will be able to internalize the language more effectively and eventually become proficient speakers.
Research has shown that this approach can be particularly effective for developing speaking and listening skills in a second language, and it has been widely used in language teaching programs around the world. However, it is important to note that different learners may respond differently to this approach, and that individual differences in learning style, motivation, and prior language experience can all play a role in determining how effective it is for a given person.
Tell me more.
The comprehensible input approach is based on the idea that language learners need to be exposed to input that is just slightly beyond their current level of understanding in order to make progress. This means that the language input should be challenging enough to stretch the learners’ abilities, but not so difficult that they become frustrated and unable to comprehend it.
One of the most famous proponents of this approach is the linguist Stephen Krashen, who developed the Input Hypothesis. According to this hypothesis, language acquisition occurs when learners are exposed to input that is both comprehensible and slightly beyond their current level of understanding. This means that learners should be able to understand most of what they hear or read, but also encounter new words and structures that they can gradually internalize over time.
The comprehensible input approach can be applied in a variety of ways, depending on the needs and goals of the learners. For example, teachers might use authentic materials like news articles or podcasts to expose learners to natural language use, while also providing support and scaffolding to help them understand the content. They might also use activities like role plays, conversations, or games to create opportunities for learners to use the language in meaningful contexts.
Research has shown that the comprehensible input approach can be particularly effective for developing speaking and listening skills in a second language. However, it is important to note that different learners may respond differently to this approach, and that individual differences in learning style, motivation, and prior language experience can all play a role in determining how effective it is for a given person.
@xxdb @bamboozled @S.I.
Here just an experiment for you if you want to understand it more. Running your own version on the web or locally.
This is the link: https://windowai.io
Here’s where he explains what he’s doing: https://twitter.com/xanderatallah/status/1643356106670981122
@xxdb: I have another strategy you can try. Chatbots have the capabilities to read links outside them. They can create stories with the content from other pages including list of words. So, if you have a database or list of words in a public blog post, for example, you can ask the chatbot to create a story based on that list.