Apple's AI Moat – Transforming Hardware and Setting the Stage for Industry Leadership

Several AI publications have expressed doubts that Apple has any moat. The lack of notable LLM moves, absence of AI mentions during their WWDC presentation in June, and the lack of eye-popping AI investments or partnerships have raised eyebrows.

Well, Apple is playing its own chess game, probably maneuvering its queen in a strategic play that could checkmate the industry. Apple has been here for a while, it hasn't merely dipped its toes into the bubbly waters of AI; in fact, it's plunged in headfirst, making AI not just a feature but a core component of its hardware offerings. But what does this trend signal? In a nutshell, Apple is setting the stage for leading in the hardware sector transformed by AI.

The signs are clear when you look closer:

Apple is far ahead in AI acquisitions compared to other behemoths (according to CBInsights):

The signs are clear when you look closer:

1. Apple is far ahead in AI acquisitions that other behemoths (according to CBInsights):

2. Apple has multiple teams focusing on AI (according to The Information): The Foundational Models team led by Ruoming Pang and overseen by Apple's AI chief, John Giannandrea, stands out as a significant player. This team specializes in LLMs and is primarily comprised of engineers who previously worked at Google. Their budget has rocketed to millions of dollars per day. They are working on advanced models like Ajax GPT, which reportedly outperforms OpenAI's GPT-3.5. There are also other teams within Apple that work on Visual Intelligence and multimodal AI, which focus on generating images, videos, or 3D scenes and understanding both text and images respectively.

3. Apple aims at hardware transformation (a trend we regard as the most significant): Shifting its focus from solely software capabilities to hardware integration, Apple is extending beyond mere chatbot experimentation. They are embedding advanced language models into Siri, transforming the iPhone into a more autonomous, intelligent device. The funny-looking VR glasses they recently unveiled could evolve into a potent tool, enriched with multimodal AI.

4. Apple is hiring massively.

The Apple Moat

What does all this mean for Apple's business strategy? First, by integrating AI deeply into its hardware, Apple is creating a unique selling proposition that is hard to replicate. This is how the edge becomes a moat. The company's preference for running software on devices rather than in the cloud gives it an advantage in terms of both privacy and performance. However, fitting a model as big as Ajax GPT into an iPhone is a technical challenge – one that could potentially become Apple's 'AI moon landing' moment if they pull it off.

In a landscape where companies often stretch themselves thin trying to be everything to everyone, Apple's targeted approach to integrating AI into its hardware is a definite bid for success.

Other examples from this week how AI transforms hardware

• AI Pin by Humane, that looks like apple-watch remade to brooch, is a wearable AI assistant with projectors and cameras, aimed at blending AI-powered optical recognition with privacy-centric features.

• AI Sunglasses by Meta and Ray-Ban’s smart glasses offer upgraded audio, a 12 MP camera, hands-free live streaming, and voice-controlled interaction via Meta AI.

• Mysterious AI-infused intuitive physical device from OpenAI and formal Apple designer Sir Jony Ive. Allegedly, nothing is decided, but “many different ideas on the table” (if they connect the table to ChatGPT, it might help with the decisions!)

We recommend

Deriving Business Value from LLMs and RAG

Date: October 17th, 10 am PDT / 7 pm CEST

We are excited to support an upcoming webinar with Databricks and SuperAnnotate where we'll learn how to derive business value from Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG). In this webinar, Leo and Quinn will delve into these capabilities to help you gain tangible insights into assessing these models for optimal alignment with your objectives.

Join us for a knowledge-packed session that offers actionable insights for companies, big or small, looking to leverage the might of LLMs and RAG. Equip yourself with the information to drive strategic AI decisions. Secure your spot today. It’s free (of course).

News from The Usual Suspects ©

NVIDIA’s Reign Questioned

• In the midst of regulatory scrutiny in Europe over alleged anti-competitive practices, NVIDIA, the dominant player in the AI chip market, faces a potential challenge from an unexpected quarter. AI startup Lamini reveals a year-long successful run employing AMD GPUs, which have traditionally been overshadowed by NVIDIA's hardware in AI applications. This move, alongside their claims of achieving software parity with NVIDIA's CUDA via AMD's 'ROCm', hints at a possible break in NVIDIA's near-monopoly. As Lamini ventures forth with its AMD-powered "LLM Superstation", it echoes a broader industry aspiration for more competition, potentially nudging open a door for AMD and others to challenge NVIDIA’s predominance."

In the world of AI unicorns, there's unrest

• Jasper AI has slashed its internal share value by 20% amid potential growth slowdown since its $1.5 billion valuation, now facing stiff competition from OpenAI's ChatGPT, upon which it heavily relied for powering its marketing-centric writing tool. Well, we had doubts who to cover next, them or The Runway – they both had the same valuation, but now it’s clear. Stay tuned with our Unicorn Chronicles.

Meta places its bet on visuals

• Meta has published a detailed tech stack behind Codec Avatars, which made the conversation between Lex Fridman and Mark Zuckerberg in the Metaverse possible.

• Oh, and they’ve also launched AI stickers ¯\_(ツ)_/¯

Amazon and its Bedrock

• AWS unveiled Amazon Bedrock to turbocharge generative AI endeavors, aiding in crafting AI apps and boosting productivity. It packs a roster of foundation models for customization, ensuring your data stays as private as your diary. Amazon Titan Embeddings is on board for better text analysis, making Retrieval-Augmented Generation (RAG) less of a snag. Other features include Meta's Llama 2 on Amazon Bedrock, Amazon CodeWhisperer's new customization capability for bolstering developer productivity, and generative Business Intelligence (BI) authoring in Amazon QuickSight for rapid visualization creation.

RAG again, now with Cohere

• Cohere launches a public beta of its Chat API with RAG, aiming to spice up product experiences by making AI conversations more grounded and verifiable. This move tackles the ghostly hallucinations in AI, where it spews out plausible yet incorrect info. With RAG, Cohere’s Chat API can now fetch real-time data from external sources, making AI responses not just smarter but also verifiable. It's like giving your AI a reality check before it chats away.

Mistral has its second big moment now with the smallest but powerful LLM

• Mistral AI, that made headlines with a modest $113 million seed (!) round, has unveiled Mistral 7B, spearheading the open model voyage against the typical AI behemoths. This debut is a petite powerhouse, employing novel attention architectures, outclassing larger models in performance. The project has been warmly supported by fellow French, like Yann Le Cun and Clement Delangue.

Twitter Library

Fresh Research Papers, categorized for your convenience

(all links lead to the original papers)

Multimodal Learning & Interfaces

• GPT-4 with Vision (GPT-4V) from OpenAI marks a significant shift by integrating vision capabilities into GPT-4. It's particularly noteworthy for its focus on safety measures for image inputs →read more

• AutoGen from Microsoft Research dives into the architecture behind LLMs, focusing particularly on multi-agent setups. It discusses components like communication and safety guardrails →read more

• Any-Modality Augmented Language Model (AnyMAL) takes multimodality a step further by introducing encoders that can handle various data types, enhancing the reasoning capabilities of LLMs →read more

Algorithmic Innovations

• Making PPO even better: Value-Guided Monte-Carlo Tree Search (MCTS) decoding bridges the gap between training and inference in text generation by combining PPO with MCTS, exploiting the often-neglected value network from PPO →read more

• Finite Scalar Quantization (FSQ) offers an alternative to Vector Quantization (VQ), aiming to simplify the architecture without sacrificing performance →read more

Data & Transparency

• Demystifying CLIP data aims to bring transparency into the data curation processes, specifically in the case of CLIP, through the MetaCLIP algorithm →read more

Efficiency & Scalability

• Effective Long-Context Scaling of Foundation Models showcases the Llama 2 Long model that excels at handling long sequences, outperforming even GPT-3.5 Turbo. The paper highlights the use of techniques like Rotary Positional Embedding and reinforcement learning →read more

  Additional Read:

  • 7 months ago Meta launched Llama, here they describe The Llama ecosystem: its past, present and future.

  • Non-engineers guide to train a LLama 2 chatbot

Synthetic Media & Generation

• LAVIE: High-Quality Video Generation with Cascaded Latent Diffusion Models emphasizes the transition from Text-to-Image (T2I) to Text-to-Video (T2V) synthesis. It introduces LaVie and a new dataset, Vimeo25M, to improve performance →read more

• RealFill: Reference-Driven Generation for Authentic Image Completion delves into completing missing parts of images. It introduces RealFill and the RealBench dataset to demonstrate its effectiveness →read more

Thank you for reading, please feel free to share with your friends and colleagues. We are launching our referral program soon! Start today :) 🤍

Another week with fascinating innovations! We call this overview “Froth on the Daydream" - or simply, FOD. It’s a reference to the surrealistic and experimental novel by Boris Vian – after all, AI is experimental and feels quite surrealistic, and a lot of writing on this topic is just a froth on the daydream.