Data Machina #194
GPT- Plugins. MS GPT-4 Sparks of AGI paper. Databrick's Dolly LLM. Dalai, Alpaca in any computer. Modelscope txt2vid diffusion model. Faster long-range transformers.
On The New Open AI GPT Plugins, and GPT-4 Sparks of AGI. Open AI is no longer an “open research” company, but effectively a product company. A few days ago they launched GPT Plugins (recommended read.) A game-changer perhaps?
I guess this is the Apple store moment for Open AI. GPT-4 & GPTChat now power a fully-fledged, intermediation platform for AI apps and AI startups.
For starters, GPT Plugins shows that Open AI has amazing capabilities in fast-productising AI innovation at scale. Back in February, Meta AI published Toolformer: Language Models Can Teach Themselves to Use Tools. afaik GPT Plugins is the fastest ever commercial productising of the Toolformer concept.
The release of Open AI GPT Plugins also completely nukes those naysayers who have criticised the inabilities of LLMs regarding aspects like: web search, facts retrieval, knowledge mapping, neuro-symbolic computation, or math calculations.
Here is some interesting stuff on 3 of the main GPT Plugins:
The Wolfram plugin- Add neuro-symbolic capabilities to ChatGPT. Stephen Wolfram posted a blog on how ChatGPT Gets Its “Wolfram Superpowers”!
The Retrieval plugin - enable semantic search, “memory,” and knowledge retrieval in ChatGPT. Checkout the repo: ChatGPT Retrieval Plugin which is now top trending. Also worth reading how to enhance ChatGPT retrieval with LangChain
The Code Interpreter plugin: super powerful… a ChatGPT model with a Python interpreter in a sandbox!... @AndrewMayne has been playing with Code Interpreter, and wrote about his first impressions in ChatGPT + Code Interpreter = Magic
As a side note, I think that LangChain is doing a clever job riding the GPT wave while extending the AI apps ecosystem beyond Open AI’s plugin store. See: How to use ChatGPT Plugins within LangChain abstractions.
In the meantime, four days ago Google enabled access to Bard chatbot under controlled demolition supervision :-) in the US & UK. I got access to Bard and I’m really not impressed. Google’s world class leadership in AI research is obvious. Perhaps its AI products are becoming boring or GPT-4 has set the expectations quite high.
I read that Meta AI is aggressively going after LLaMA repos with DMCA's (Digital Millennium Copyright Act), and taking down many derivative repos of LLaMA and Alpaca. Talk about productising LLMs innovation not!
Instead of sending the legal suits, why not capitalising on and enabling the LLaMA/ Alpaca movement, and promoting awesome projects like:
MS announced GitHub Copilot X, which is AI-Pair-Programming on steroids. And yes, it’s very impressive too. The fact that MS has full control of Copilot, Github and Open AI is a bit too much centralisation. The AI community is building alternatives:
Sparks of AGI in GPT-4. MS Research published a new, massive, 154-page paper titled: Sparks of Artificial General Intelligence: Early experiments with GPT-4. The paper is remarkable and you should read it.
The paper has ignited a further polarisation of views between those AI/ ML researchers who believe that GPT-4 sparks of AGI are real or those who don’t buy it. The great Ferenc reflected about why many DL/ ML researchers didn’t believe in LLMs. This is a great read: We May be Surprised Again: Why I take LLMs seriously.
If you don’t have time to read the Sparks of AGI paper, you can watch this great video review that goes through the main 15 capability examples of GPT-4 detailed in the paper.
@leopold also wrote a nice summary with some of the highlights of the paper in this thread: GPT-4: A Qualitative Jump in AI Capabilities.
@zvi is not fully convinced about those sparks of AGI. He wrote a critical review of the paper here: MS Research Paper Claims Sparks of AGI in GPT-4. Worth reading too.
If this is a lazy Sunday for you, you can always play games:
Every Possible Wordle Solution Visualized. Is this cheating?
Try to buy the rug from the Turkish merchant for as cheap as possible, a GPT game
Have a nice week.
10 Link-o-Troned
Vesuvius ML Challenge: Decode the Herculaneum Papyri, Win $1M
How I Built Lex-GPT with ChatGPT, LangChain, Whisper, Pinecone DB
the ML Pythonista
the ML codeR
Deep & Other Learning Bits
AI/ DL ResearchDocs
Diffusion over Diffusion for eXtremely Long Video Generation
Learning Real-world Conversation from Large-Scale Web Videos
data v-i-s-i-o-n-s
MLOps Untangled
AI startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.