Grok's journey to the top!

Author

Jason - The AI Digest
August 28, 2024

Welcome, AI Enthousiasts!

In today’s AI newsletter:
→ Quick AI Newsflash
→ Grok 2 improves speed, accuracy and transparancy
→ AI learns to plan better without humans
→ ChatGPT teams up with ASU
→ Why you should always include examples in your prompt
→ 5 Best tools for academic/professional writing
→ Quick Newshits

Reading time: 12 minutes

Apple is rumored to be working on a new AI personality that will be distinct from Siri, designed specifically for future robotics devices. While the company’s first ventures into robotics aren’t expected to roll out until 2026 or 2027, Apple is also planning to introduce an upgraded, AI-enhanced version of Siri with the release of iOS 18.1 later this fall. That said, the most impressive new features of this Siri upgrade won’t make their appearance until sometime in 2025.

Anthropic has revealed the system prompts used for all their models and refreshed the dates in the knowledge database to ensure up-to-date information.

Salesforce has introduced xGen-VideoSyn-1, a text-to-video model that turns written descriptions into realistic video scenes. The model uses a video variational autoencoder (VidVAE) to compress video data, making it less demanding on resources, and a Diffusion Transformer (DiT) to ensure smoother transitions and better adaptability over time.

The Rundown: xAI’s Grok-2 and Grok-2 mini have just seen major upgrades - the mini version now runs twice as fast, and both models boast improved accuracy, all within days of their beta launch.

The details:
→ Grok-2 mini is now twice as fast, thanks to a rewritten inference stack using SGLang.

→ Both Grok-2 and its mini version have seen a slight boost in accuracy due to reduced quantization error, as noted by an xAI employee.

→ The Grok-2 models have also joined the LMSYS Chatbot Arena leaderboard for greater transparency, with the larger Grok-2 model securing the #2 spot, surpassing Claude 3.5 Sonnet.

→ Grok-2 stands out in math, ranking #1, and excels in hard prompts, coding, and instruction-following, performing at a state-of-the-art level.

Why it matters: In just around 18 months since its founding, xAI has managed to create an LLM ranked third globally, leaving the entire AI community amazed. Grok-2 is now a serious contender in the AI race, and its success is likely to ramp up competition, possibly speeding up advancements across the industry.

The Rundown: IBM Research and Cornell University recently created
AutoToS, a system that teaches AI to solve complex planning problems at 100% accuracy - without needing a human to check its work.

The details:
→ AutoToS is like a smart tutor for AI, helping it learn how to break down and solve tricky problems step-by-step.

→ The system uses clever tests to check the AI’s work, pointing out mistakes and showing examples of how to do better without human interferance.

→ This approach seems to work equally as well for smaller and larger models.

→ AutoToS succeeded in teaching AI to solve complex puzzles, including classic problems like arranging blocks and solving Sokoban, a box-pushing game.

Why it matters: Right now, it’s difficult to trust AI agents to completely autonomously perform actions on your behalf, but AutoToS is solving complex tasks at a 100% accuracy. If this system works in the real world, it’s the next big step in creating more reliable AI assistants.

Image source: Midjourney

The Rundown: OpenAI’s ChatGPT is headed to Arizona State University (ASU), where the university is integrating the AI assistant into over 200 projects across teaching, research, and operations.

The details:
→ ASU is using ChatGPT Edu, a version designed for universities with enhanced privacy and security features.

→ The university also launched an ‘AI Innovation Challenge’ for faculty and staff, receiving an overwhelming demand for using ChatGPT to maximize teaching, research, and ops.

→ Key projects include an AI writing companion for scholarly work, 'Sam' (a chatbot for med students to practice patient interactions), and AI-assisted research recruitment.

→ The partnership has inspired other institutions like Oxford and Wharton to pursue similar collaborations.

Why it matters: While some schools are attempting to resist AI, ASU is embracing ChatGPT to make learning more personalized and to prepare students for an increasingly AI-driven job market. As education continues to change in the age of AI, case studies like this will be instrumental in shaping the future of academia.

New research reveals why you should always include examples in your prompts. 

You know how most prompting advice tells you to include a few example outputs with your request? Well, a group of researchers from Amazon and UCLA just found out why that works as well as it does. 

It turns out, large language models excel at learning patterns, even with only a few examples. What they don’t excel at? Applying rules—especially in unfamiliar scenarios.

Here’s what happened: 
→ The paper looked at two types of reasoning in AI: “deductive” (applying rules) and “inductive” (figuring out patterns).

→ The researchers tested large language models like GPT-4 on a variety of tasks, from math to language to spatial reasoning. 

→ They discovered that AI are actually really good at inductive reasoning, often getting perfect scores, but struggle with deductive reasoning, especially on “counterfactual” tasks (things they weren't trained on).

To test their theories, the researchers created a new method, called SolverLearner, which uses a two-step process: 

→ Function proposal (the learning step): The AI is given a few examples and asked to figure out the underlying pattern or rule.

→ Function execution (the application step): Once the AI proposes a rule, it's tested using an external “verification” system (like a code interpreter), not the AI itself.

By using an external system to apply the rule, the researchers could focus solely on how well the AI learned the pattern without mixing in its ability to apply the rule.

This means future AI models will need to prove their deductive skills have specifically improved before they can claim anything close to AGI…

Here’s how to apply this in your own work with AI: 
→ When asking an AI to solve a problem, play to its strengths and always give it a few examples. 

→ Don't assume the AI can perfectly follow complex instructions or rules, especially with unique situations.

→ When possible, break down complex tasks into smaller pattern-recognition problems rather than long sets of rules.

In general, this paper is a good reminder to be aware that while an AI might understand a concept, it can still make mistakes when trying to apply it.

5 Best AI Tools for Academic Writing

Jenni AI generates content and improves writing style, providing suggestions for structuring essays and papers.

Afforai finds academic papers, summarizes research, and generates writing ideas for quick information gathering.

AcademicHelp AI Essay Writer assists in writing academic essays with templates, suggestions, and citation support.

StudyX aids in note-taking, summarizing textbooks, and creating study guides to enhance study sessions.

QuillBot rephrases sentences to improve clarity, avoid plagiarism, and enhance readability.

Good to know:

→ Asana's marketing head reveals how AI can fix top 4 productivity killers.

→ Amazon could launch AI subscriptions for Alexa in October.

→ Creating calendar entries from an image using Claude 3.5 Sonnet.

→ India’s startups are betting on AI voice products instead of text.

→ Google AI Studio released a native prompt gallery featuring long context, multi-model inputs, and structured outputs for enhanced AI development.

That’s a wrap!

We had a lot to talk about, so let’s wrap it up. If you have any questions, feel free to shoot over an e-mail and we wil get you a response within 24 hours.

If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.