LLMs Can Define Reward Params to Optimize for Robots
-LLMs define params to optimize for control policies to accomplish robotic tasks -Bridges gap from high-level language instruct & low-level robot actions
For the non AI-companies that want to train their own LLM similar to Bloomberg GPT:
the infrastructure and tooling is less of a constraint - it’s really only an option if you can hire sophisticated AI talent which is a subset of some of the roles on this slide
'The ‘Safeguarding the Future’ course at MIT tasked non-scientist students with investigating whether LLM chatbots could be prompted to assist non-experts in causing a pandemic.
'These results suggest that LLMs will make pandemic-class agents widely accessible as soon as they are credibly identified, even to people with little or no laboratory training.'
MIT tasked non-scientist students to investigate: can LLM chatbots be used by non-experts to cause a…
I love that pwning dumb incels can be now outsourced to LLMs. Here's they thought it will come to their rescue and end up getting smacked in the face with dry logic instead. This is a win.
Serving LLMs? My students found a way to accelerate serving by over an order-of-magnitude just by changing the way memory is managed (spoiler alert): gpu memory fragmentation = slow. Introducing vLLM with PagedAttention:
This great chart from BCV shows the limited number of companies that can construct their own LLM.
What’s interesting is the number of cos configuring LLMs for use cases specific to financial services, like risk modeling, fraud detection, policy underwriting or purchase order…
Had amusing exchange with someone defending LLMs as not memorizing content, particularly copyrighted content. So I jumped on chatGPT and tried a little experiment. Folks can decide for themselves. 12 prompts, each producing exact copy of a text snippet verbatim. There’s more…
Can large language models (LLMs) train themselves? The explosion of imitation-based open-source LLMs drew criticism due to cursory evaluation that covered up performance gaps. However, recent research shows powerful open-source LLMs can actually be created by imitating other…
Yesterday I submitted the first completed version of my thesis on streamwise vorticity currents and their effect on LLMs and tornadoes. Today I felt like I was rewarded for those efforts.
🚀 vLLM is an open-source LLM inference and serving library that accelerates HuggingFace Transformers by 24x and powers lmsys.org Vicuna and Chatbot Arena.
To my surprise, this paper is extremely detailed around training strategies of building such models and goes beyond “just prompting” ChatGPT and GPT-4.
In fact, I realised after reading-it doesn’t even mention these models in most of the…
🔧Thrilled to introduce #ToolQA , a new dataset to evaluate the capabilities of #LLMs in answering challenging questions with external tools. It offers two levels (easy/hard) across eight real-life scenarios. 🚀