Catching Up with the Recent Data & AI News
Let's Recap the Top Highlights from the Past Two Weeks
It’s been already two weeks since the last update on the data & AI space. We began to release Naas's new version (see Introducing Naas Alpha Workspace) and I did not have much time to write. So as usual, I will go lean and straight to the point. I made very short paragraphs and added a bit more than 5 items on your radar. Are you ready?
Anaconda Introduces Anaconda Assistant, and Open Interpreter Makes Its Debut
Last week, Anaconda made waves with the announcement of Anaconda Assistant, a tool designed to streamline your coding experience in Anaconda notebooks. In parallel, a new open-source project, Open Interpreter, was introduced. This innovative tool allows large language models to execute code locally on your computer. Open Interpreter has many features, from web browsing capabilities to adding subtitles to videos and even using a control net to animate images.
I need to spend more time testing these 2 projects, especially to highlight the difference or complementarity with what we do at Naas, so will probably make dedicated articles about them.
Read more on the Anaconda Assistant here and if you want to try open-interpreter, check out the website or the notebook shared in colab.
TII Releases Open Source Model - Falcon 180B
TII has released “the best open-access model currently available, and one of the best models overall.” This new open-source model, Falcon 180B has been trained on 3.5 trillion tokens, from refined web data. It is said to perform exceptionally well in various tasks like reasoning, coding, proficiency, and knowledge tests, even beating competitors like Meta's LLaMA 2.
You can try this model on Hugging Face and here.
Prompt2: An Open Source Project that Generates Deployable Models
Prompt2 is an impressive open-source project that generates deployable models from instructions. Why would you use it? Because thanks to it you can train a small special-purpose model that is ready for deployment, autonomously. It retrieves models, generates additional training and test data, fine-tunes the retrieved model, evaluates the model, and even creates an interface for interaction. One example states that the resulting model outperforms GPT 3.5 Turbo while being up to 700 times smaller.
Check out their GitHub to deep-dive
RLAIF Paper: Reinforcement Learning with AI Feedback
A new paper explores the performance comparison between reinforcement learning with AI feedback (RLAIF) and reinforcement learning with human feedback (RLHF). The study found that both RLHF and RLAIF produced higher-quality summaries than the supervised fine-tuned baseline (SFT). Moreover, RLAIF can reach human-level performance, offering a potential solution to the scalability limitations of RLHF.
I recommend this article from AssemblyAI to learn more before diving into the paper.
Speculative Execution for LLMs
Andrej Karpathy recently tweeted about speculative execution for LLMs, a concept aimed at optimizing inference time. The idea is to use a small and cheap draft model to generate a candidate sequence of tokens, which are then fed through the big model in a batch. In other words, an AI program takes a lot of time to read its instructions from its memory just like how you take time to read a book but with this concept, it takes about the same time.
Bonus: Time Magazine’s 1st-ever Top 100 AI List Honours Many Indian Talents
Time magazine’s first ‘TIME100 AI List’ highlighted the 100 most influential people in AI, including several Indians and Indian-origin techies. This list is a testament to the growing influence and contribution of Indian talent in the global AI landscape. I really applaud the work done by Time, they have created one article/ interview per personality. I found it very interesting to hear from my fellow French compatriots, including Clément Delangue from Hugging Face and Yann LeCun from Meta.
Warning from UK’s National Cyber Security Centre (NCSC)
Lastly, the UK’s National Cyber Security Centre (NCSC) has issued a warning about the increasing vulnerability of chatbots to manipulation by hackers, leading to potentially serious real-world consequences. The alert focuses on the practice of “prompt injection” attacks, where individuals deliberately create input or prompts inside web pages to manipulate the behavior of language models that underpin chatbots. I’m planning to do a special investigation and series of tests about it soon because it seems to be a problem that is inherent to the design of large language models.
Feel free to share your ideas!
That's it for this week. Stay tuned for more updates in the world of data and AI. See you next week!