- ThePrompt
- Posts
- LLama 2: Beyond ChatGPT and closed models
LLama 2: Beyond ChatGPT and closed models
PLUS: GPT-4 performance is declining, AI-generated South Park episode
Anita Kirkovska
July 19, 2023
Hi folks!👋🏻 This is The Prompt! We're your go-to source for all things AI.
Let's see what’s new
FEATURED
Llama 2 by Meta: Beyond ChatGPT and closed models🦙
Meta released Llama 2, an open-source and highly capable language model.
The base model is stronger than GPT3 and the fine-tuned chat models are on par with ChatGPT. This is a big step forward for open-source software, and a big blow to closed-source providers (OpenAI/Bard).
Meta did a great thing for the community by making this model open-source. It is estimated that they spent over $20 million to train it.
This will allow companies to customize and save tons of money.
In summary here’s what you need to know:
What was launched
Multiple models: Llama base and Llama Chat from 7, 13, 34, 70 billion sizes)
Performance
It outperforms every other open-source model out there
Probably the first time an open model (Llama Chat) is at the same level as ChatGPT (except for coding)
Still far behind GPT-4 and PaLM-2-L
Safety
They go above and beyond on safety issues (round of applause👏🏻 on this one)
Almost half of the paper talks about safety
They spelled out the whole recipe on the models (unlike OpenAI who didn’t share anything about GPT4)
Commercial use
You cannot use Llama-2's data or output to improve any other LLMs
You must request a special commercial license if your product's MAU exceeds 700M users in Jun. 2023
Other than that it’s royalty-free to use, copy and reproduce.
Links
TOGETHER WITH CHATPDF
Chat with your PDFs for free ✨
ChatPDF is the OG when it comes to chatting with your PDFs.
The tool is super simple and works really fast.
Plus, it’s free and doesn't require signing in.
You can talk with your books, research papers, manuals, essays, and legal contracts… you name it.
Try ChatPDF and get to chat with 3 PDFs per day for free👇🏻
WHAT ELSE IS GOING ON
🦙 New research shows GPT-4 performance is declining. They evaluated results between March and June. On math problems, specifically at identifying prime numbers dropped from 98% to 2.4%, on directly executable coding generations, 52% to 10%, and on visual reasoning, GPT-4 started to make more mistakes than before.
This might be happening due to OpenAI accepting worse quality in return for lower latency & cost savings… or due to fine-tuning the model from user input.
👀 AI-generated TV Shows and Showrunner Agents are coming. The authors of this paper generated a SouthPark episode using a prompt. Their model can write, animate, direct, voice, and edit for you.
Ok, this is mindblowing, right? Wait until you hear the rest of it.
Their goal is to build AIs that are truly alive, not just chatbots that answer questions, but AI people living real daily lives in simulations, growing over time. Their “showrunner agents” basically will run as a reality show, expanding the show to infinity. 🤯
RESOURCES
The best resources we came across lately that will help you become better at writing prompts & building AI apps.
📚 Mark Andreessen on the Lex Fridman podcast [ a must-listen]
👋🏻 AI (1981) [ interesting read ]
🎥 LLM Performance Drifts - And What To Do About Them [ interesting read ]
TOOLBOX
The latest AI tools to use or get inspiration from.
PROMPT OF THE DAY
TOOL
MidjourneyPROMPT
Grainy overexposed closeup photo of a man swimming against a rough current in a suburban flood, fighting for his life, climbing out of a car window, intense and scary dark night, electrical sparks popping, glowing in the water --ar 3:2 --w 250 --s 250 --v 5.2RESULT
What'd you think of today's edition? |