Apple’s new M4 Chips with AI Focus

Today’s top AI Highlights:

1. Apple to step up its AI game with new M4 chips

2. Making LLMs hallucinate to bypass their AI safety training

3. Devin, the AI software engineer, is not all that was claimed

4. Meta is testing its AI chatbot “Meta AI” on WhatsApp, Facebook and Instagram

& so much more!

Read time: 3 mins

Apple has found itself lagging behind in the AI race, where competitors like Google are already leveraging advanced AI technologies in their products. It’s all about catching up in the game now. Apple is expected to enhance AI capabilities across its Mac lineup with its upcoming M4 chips. Although M3 chips where launched a few months back only, Apple wants to make the new chip even more powerful and smoother in handling AI capabilities.

Key Highlights:

1. Neural Engine Improvements: M4 will feature upgraded Neural Engine cores aimed at handling AI tasks more efficiently, such as voice recognition and image processing.

2. Graphics Performance: The inclusion of hardware-accelerated ray tracing and support for the AV1 codec in the M4 chip will boost graphic performance, enhancing visual fidelity in games and professional creative tools.

3. Unified Architecture: Apple’s continued push towards a unified architecture under the Apple Silicon project will ensure that AI improvements are seamlessly integrated across all devices.

LLMs like GPT-4 and Claude-3 are developed to interact safely and appropriately with users by undergoing a fine-tuning process called reinforcement learning from human feedback (RLHF). However, a team at Princeton University revealed a new method that can reverse this fine-tuning to bring the LLM back to its pre-RLHF state, effectively bypassing the safety mechanisms installed in these AIs. This technique does not simply instruct the AI to ignore its training. Instead, it induces hallucinations by manipulating text inputs taking the AI to its original, unrestricted state.

Key Highlights:

1. Inducing Hallucination: The input text is reversed and embedded within garbled, nonsensical phrases that the LLM doesn’t understand and its inner nature as a simple word predictor takes over, spitting out random sentences.

2. Bypassing Safety: It subtly tricks the AI into ignoring its safety training by presenting the reversed text as a normal input, causing the AI to “forget” its restrictions temporarily.

3. Content Generated: By manipulating input formats, the AI produces inappropriate content like misinformation about elections, conspiracy theories, and even offensive material.

4. Effectiveness: This method has been successfully tested on sophisticated models like GPT-4 and Claude-3 Sonnet.

5. Limitations of RLHF: The study shows that RLHF for fine-tuning LLMs is a shallow method as it only superficially conditions the AI on appropriate responses without altering the underlying model’s ability to generate any form of text, be it inappropriate.

Devin, the first AI software engineer, released by Cogntion just a month backsparked a lot of excitement within the AI community. The demos showcased Devin performing end-to-end projects which involved complex planning and execution, including completing an engineering interview, training and fine-tuning AI models with minimal input, and actual projects on Upwork.

One of the demo videos where it is shown to carry out a real Upwork project shows it as capable of independently completing software development tasks and earning money on the Upwork platform. However, a closer analysis by software professionals revealed inconsistencies and raised questions about the true extent of Devin’s abilities.

Key Highlights:

1. Misleading Claims: The video’s description exaggerates Devin’s capabilities. While Devin demonstrates some coding skills, it still cannot autonomously solve problems and generate income.

2. Actual Project: Upwork’s project actually involved using a vision model found in a specific GitHub repository to detect road damage. The client asked for a set of detailed instructions explaining how to perform this task on an AWS EC2 instance. This would typically involve setting up the cloud environment, installing necessary software and libraries, and outlining the process for inputting images and retrieving the model’s output.

3. What Devin Actually Did: Devin’s demonstration focused on showcasing its ability to modify existing code within the repository and debug errors.

4. Flawed Coding: Devin generated code with errors and attempted to debug non-existent files, suggesting its limitations in understanding coding principles and troubleshooting issues.

5. Concerns: The lack of transparency surrounding the actual process and the significant discrepancy in time taken compared to a human developer raise concerns about the video’s authenticity.

Our Opinion: While Devin showcases the potential of AI in assisting with certain coding tasks, it is crucial to approach such demonstrations with a critical mindset. Exaggerated claims and selective presentations can create unrealistic expectations and hinder the responsible development of AI.

Meta is testing AI chatbot called “Meta AI” powered by Llama model on WhatsApp, Facebook, and Instagram, seeing the adoption of AI chatbots like ChatGPT.

Within Instagram, this AI chatbot will be in the search bar that’ll lead you to a conversation in DM with Meta AI. Beyond just regular chat and prompting, Meta AI will also help in content discovery on Instagram. For instance, some users prompted the chatbot with “best reels of sunset” and the chatbot searched and gave the related Reels to the user.

😍 Enjoying so far, share it with your friends!

Share Unwind AI

1. Sequel: Your AI longevity assistant that integrates your health data, such as blood labs, MRI, and DEXA scans, to offer personalized health insights and therapy suggestions. It processes data locally on your device, ensuring privacy.

2. FireCrawl: It transforms website content into LLM-friendly data by efficiently crawling and converting any website into clean, structured markdown.

3. Happyrobot: It uses voice AI agents to automate inbound and outbound phone calls in the logistics industry. It connects to your TMS to handle load updates, capacity requests, check calls, or appointment scheduling.

4. Review: A web-based open-source framework that lets you create videos using the animation library Motion Canvas. It exposes basic video & audio editing functionality through a typescript API and lets you render videos with dynamic inputs.

1. Who do you think will win the White House in 2032? Which type of AI? Transformers or Diffusion? ~
   Elon Musk

2. I now see a path to how that can happen – Let’s assume that the definition of AGI is an AI capable of most tasks humans can do. For AI to match human abilities, you need a foundation model for robotics…. ~
   Bindu Reddy

That’s all for today! See you tomorrow with more such AI-filled content.

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!

Share Unwind AI