Attention has been the key component for most advances in LLMs, but it can’t scale to long context. Does this mean we need to find an alternative?
Presenting Titans: a new architecture with attention and a meta in-context memory that learns how to memorize at test time. Titans are more effective than Transformers and modern linear RNNs, and can effectively scale to larger than 2M context window, with better performance than ultra-large models (e.g., GPT4, Llama3-80B).
Latest books published by myself last month.
Navigating Inertia: Understanding Resistance to Change in Organisations https://t.co/2UGGeurEh0
Wardley Mapping Doctrine: Universal principles and best practices that guide strategic decision-making https://t.co/pcvpRkVFW4
AIconomics: The Business Value of Artificial Intelligence https://t.co/uMDiSjkbY3
#WardleyMaps
#Strategy
🚀 Exciting News for Wardley Map Developers! 🚀
1/ 🗺️ We're thrilled to announce the launch of our new API, designed specifically for developers creating SaaS or ChatGPT applications. Unlock the full potential of Wardley Mapping with our comprehensive suite of APIs!
https://t.co/PCrwBIVDWb
#WardleyMapping #API #SaaS
You can find the Wardley Map python package here, https://t.co/vdt6f7odTM
The wardleymap package designed for creating and visualising Wardley Maps.
Features:
Parse and interpret Wardley Map syntax.
Visualise maps using matplotlib for easy integration into Python workflows.
Export maps to SVG format for embedding in web applications or documents.
Supports use via OpenAI Actions and Anthropic Tools
Comes with a set of utilities to convert Wardley Map text into JSON, TOML, GRAPH and Cypher Text.
#WardleyMaps
🧵1/7 Exciting breakthrough in #WardleyMapping! We're now using #OpenAI to extract key mapping terms from web pages and documents, revolutionising how we approach strategic planning and analysis. Let's dive into this transformative method! 🌐 #WardleyMaps#Strategy#AI https://t.co/GdYtT9R4bs
🧵 1/8 Exciting news in the world of #DataSecurity! A recent article on https://t.co/5YV6Uw0B4i by Samuel K. Moore reveals a game-changer in computing: Chips designed to compute with encrypted data using Fully Homomorphic Encryption (FHE). https://t.co/z434hYnVyA #TechTrends#FHE #CyberSecurity
An example experiment on creating LangChain Tools which can be used by OpenAI Functions.
This experimental Python code provides a class WardleyMap to extract and process various pieces of information from these maps, specifically from the OnlineWardleyMaps tool. The WardleyMap class uses the APIs provided by OnlineWardleyMaps to fetch, parse, and manipulate map data.
https://t.co/WDr6d189fU
The best thing about the multi-modal models such as GPT4-V is that I can finally start having a conversation with the system, now that it can finally create and talk maps
The new technical debt. Data debt.
Like technical debt, data debt is the cost of avoiding or delaying investment in maintaining, updating, or managing data assets, leading to decreased efficiency, increased costs, and potential risks.
"The most amazing achievement of the computer software industry is its continuing cancellation of the steady and staggering gains made by the computer hardware industry." — Henry Petroski (1942-2023)
Fine-tuning a Llama 65B parameter model requires 780 GB of GPU memory.
This kind of compute is outside the purview of most individuals.
Thanks to parameter-efficient fine-tuning strategies, it is now possible to fine-tune a 7B parameter model on a single GPU, like the one offered by Google Colab for free.
A great example of such a method is LoRA which freezes the weights of the LLM and introduces new weights for training.
These new weights are based on the frozen ones but small enough to make fine-tuning a single GPU because the trainable parameters have been reduced.
What's even more interesting about this method is that it doesn't introduce inference latency.
The QLoRA method goes further and quantizes the model to 4-bit further reducing the computational requirements.
In the following article, I dive into these techniques, including:
• Why you should fine-tune an LLM
• Fine-tuning vs. prompting
• How LoRA and QLoRA work
• Fine-tuning a Llama 2 7 billion model with LoRA and QLoRA (with sample notebook)
Check out the article: https://t.co/kvXtpOhOMA
Launch of the First Global Report on Climate and SDGs Synergies https://t.co/XkmGlizABy
"In his priorities for 2023, the UN Secretary-General stated that “Climate action is the 21st century’s greatest opportunity to drive forward all the Sustainable Development Goals.” Yet, halfway to the deadline for the 2030 Agenda for Sustainable Development, the progress towards achieving the SDGs and addressing climate change is alarmingly insufficient."