Announcement: PocketLLM was featured on ProductHunt! | Check out the latest AWS blog on ThirdAI benchmarks 

AI-Assisted Exploration and Discovery of All ICML 2023 Papers with PocketLLM

Dive into more than 1800 ICML Papers Instantly on Your Personal Device via RAG (Retrieval Augmented Generation)

The International Conference on Machine Learning (ICML) 2023, the leading event in the machine learning world this year, is starting this week in Hawaii. AI is currently the hottest trend, and enthusiasts worldwide are eager to stay updated on the rapidly evolving field by exploring its latest and greatest advancements. However, absorbing the vast amount of information presented in over 1800 dense and technical full papers won’t be an easy feat. Simply skimming titles and abstracts won’t suffice to grasp the full insights.

Google Search and ChatGPT wont cut it, as the information in the papers are too specific as well as too recent. Good news, there is now a third way, the ThirdAI way of specialized LLMs for this task.

Simplify your ICML 2023 Exploration: Instant Semantic Discovery on your Device. Even works Offline without Internet.

Get Started with PocketLLM: Download the 250MB application on your Mac or Windows device from this link. You will need the most recent application files as we constantly upgrade our capabilities. In addition, download the Neural Database containing 1827 accepted ICML papers (2.9GB file, which includes 2.7GB of papers in PDFs for references) from this link. After downloading, unzip the file. Now, simply double-click PocketLLM, go to ‘load existing models,’ and point to the extracted folder. Wait a few seconds, and you’re all set to dive in!
After loading the Neural Database, you’ll find a search bar. Enter your query or the paper idea you’re seeking, whether it’s something you heard from a colleague or a friend whose paper made it to ICML this year. In the example query below, we explore the advancement in the popular ideas of aligning memory access patterns with parameter access patterns to leverage cache hierarchies (a cheap promotion of one of my ICML papers this year!). The figure below displays retrieved paragraphs from various papers and the response generated by ChatGPT (Use the magic wand to enable RAG with ChatGPT). Clicking the small pop-out button at the bottom of the retrieved paragraphs will directly take you to the relevant text in the paper.

Refer to our PocketLLM Beta user documentation to unlock other capabilities like reinforcement learning and many more.

Next Level Exploration: With semantic search, we elevate information discovery to the next level. Imagine being intrigued by a paragraph in one paper during your exploration and wanting to find similar ideas in other ICML papers. Just type the whole paragraph, and you’ll uncover related concepts. See the figure below for an example of how it works.

The power of Neural Search: Drill down on any paragraph to find similar or related ideas from other papers.
(We are only showing NeuralDB Retrieval by switching off Generation in PocketLLM)

Personalized Exploration with Real-Time Incremental Teaching

As a researcher, I’m eager to explore the topic of efficient inference, which, in my opinion, encompasses various ideas like pruning, quantization, and more. Different researchers may have different associations. Now, with PocketLLM, you have the power to personalize your search in real-time by teaching associations to the AI. Just inform the tool that Efficient Inference is semantically related to a list of other ideas and click the “Teach Associate” button. The next time you search for “Ideas in Efficient Inference,” the NeuralDB will refine your results based on your provided preferences. Feel free to create as many associations as you wish.
In addition, you can also update the model with ranking preferences using the tick button.
Teach the AI to Understand your Jargon using the Teach Association Feature in PocketLLM. Don’t forget to save the model in the end.

Know the Limitations

We’ve all experienced “chat with your document” demos with small PDFs, but the scale we’re dealing with here is vastly different. We’re looking at 1827 full ICML papers, 175,000 lengthy paragraphs, and over 11.5 Million words (around 16 million tokens). Foundational LLM Models cannot memorize such vast information and hence, their role is often reduced to answering simple queries like “give an overview of all this information” whereas their performance over complicated queries deteriorate. However, with RAG (Retrieval Augmented Generation), we can tap into these papers to gain deeper insights into the latest ideas by querying specific pieces of information with longer queries, as shown above.

Behind the Scenes: ThirdAI’s NeuralDB with Just 20 minutes of Pre-training on an AMD Desktop!

The complete NeuralDB model was built from scratch on an AMD Milan Desktop in under 20 minutes, without using any base or foundational models. We pre-trained a 200 million parameter model from scratch on 1827 ICML texts using this simple NeuralDB script. The Neural DB index, is less than 1GB in size. Additionally, the end-to-end retrieval latency on a standard laptop is less than 10ms. That’s the power of ThirdAI’s NeuralDB. No GPUs in the loop, all local computations.

Instead, if we opt for embedding models and vector databases with 16 million tokens and approximately 1M sentences, the scenario changes significantly. The sentence embedding alone (excluding ANN index), assuming 1500 dimensions, would occupy about 6GB of memory to store 1.5 billion numbers. You should be prepared to spend thousands of dollars every month on cloud services to maintain the embedding models and vector database, as someone has to cover the cost of GPUs. Even with all this expenditure, the search latency would still be quite high due to slow embedding model inference followed by a vector database micro-service query.

The advantage of ThirdAI’s technology is that you can build the ICML exploration system completely free of cost and make it accessible even without constant internet service. All you need is a computing device with PocketLLM, a completely free app, to unlock it. Now millions of users can explore ICML papers, in any part of the world, without anyone spending a single penny!

Level Up: From AI User to AI Builder with PocketLLM!

Conference proceedings are just one illustration. With PocketLLM, anyone can build and share similar capabilities on any custom corpus using the user-friendly UI or simple NeuralDB scripts provided here. If you’re familiar with Windows or Mac, you have the power to create your own AI and make it instantly accessible. Get started with PocketLLM today!