80% Error Rate in Medical Transcripts 🚨, Apple's AI Debut (finally)🍏, Google's Browser Agent Jarvis 🤖
Plus: Meta's New AI Search Engine 🔍, OpenAI's Custom Chips 💽, and More
Welcome to this week in AI.
This week's news: Apple launches Apple Intelligence, bringing privacy-focused AI integration to devices; Google's Project Jarvis introduces agentic AI for seamless web automation; and Meta develops its own AI-powered search engine to rival Google and Bing.
Plus, explore OpenAI's move towards self-reliance with custom AI chips, the speculation and denials around OpenAI's 'Orion' model, and the risks of using AI transcription in healthcare due to hallucinations in transcripts.
Let’s get caught up!
🎵 Don’t feel like reading? Listen to two synthetic podcast hosts talk about it instead
Step into the Future of Design
Equip yourself with the tools and knowledge to design for AI.
Apple Intelligence: Key Features and Privacy-Focused AI Integration Finally Arriving
After delays, Apple has launched Apple Intelligence with iOS 18.1 and macOS Sequoia 15.1 updates, introducing tools for productivity, creativity, and privacy. Key features include:
Systemwide Writing Tools: Features for rewriting, proofreading, and summarising text across Mail, Messages, Notes, and other apps.
Enhanced Siri: Added support for typing, improved context understanding, and deeper system integration.
Photo Search and Clean Up: Enhanced search capabilities using natural language and tools to remove unwanted elements from photos.
Device Compatibility: Available for devices with M1/A17 Pro chips or later. Some users may face a waitlist.
Future Features (December update): Planned additions include ChatGPT integration, Image Playground, and Genmoji.
Why It Matters
Apple Intelligence introduces integrated AI tools aimed at enhancing user productivity while maintaining a focus on privacy.
Key features include on-device processing and Private Cloud Compute, intended to keep user data secure while providing AI functionality.
By maintaining a privacy-oriented approach, Apple offers these AI features without compromising user data security.
The upcoming December update will add further functionality to support creative and communication-related tasks.
Google's Project Jarvis: Agentic AI for Seamless Web Automation
Google's Project Jarvis is an agentic AI designed to autonomously navigate web browsers, executing tasks through visual interpretation and interaction.
By analysing screenshots to perform actions such as clicking elements, inputting data, and navigating pages, Jarvis aims to automate routine digital operations like online shopping, research, and travel reservations.
Set for a limited preview in December, alongside Google's new Gemini model, Jarvis represents a strategic step towards sophisticated AI that minimises user involvement in everyday web-based activities.
Why It Matters
Project Jarvis exemplifies the trend towards agentic AI—autonomous systems that execute complex tasks for users.
In a competitive landscape featuring similar agents from Anthropic and Microsoft, Google's Jarvis, coupled with the Gemini model, could lead in AI-driven digital assistance.
By acting as an autonomous browser agent, Jarvis transforms browsers into dynamic productivity platforms, allowing users to delegate routine tasks and enhancing efficiency and convenience in everyday digital interactions.
Meta Joins the AI Search Battle: A New Rival for Google and Bing
Meta is developing its own AI-powered search engine to reduce reliance on Google and Bing, employing proprietary web-crawling technology.
This infrastructure has been in development since early 2024, aiming to allow Meta AI to provide real-time information independently.
Recently, Meta also partnered with Reuters, integrating its news articles into the Meta AI chatbot, indicating a broader effort to control information sources and improve AI capabilities.
Meta AI now reaches 185 million weekly active users across platforms like Facebook, Instagram, and WhatsApp.
Why It Matters
Meta’s initiative to build an AI-powered search engine represents a strategic shift towards reducing dependence on rival technologies and establishing self-sufficiency in real-time information delivery.
By developing proprietary web-crawling and partnering with reputable news sources like Reuters, Meta aims to directly control the quality and scope of information provided by its AI.
This move reflects an increasing trend among tech giants to create independent ecosystems, which could fuel competition in AI search capabilities, with other companies like OpenAI and Apple also moving in this direction.
OpenAI's Path to Self-Reliance: Custom AI Chips in Collaboration with Broadcom and TSMC
OpenAI is developing its first custom AI chip in collaboration with Broadcom for design and TSMC for fabrication, targeting deployment in 2026.
This move aims to reduce dependence on Nvidia GPUs. Simultaneously, OpenAI is incorporating AMD's MI300X processors to diversify its hardware.
These efforts are part of a broader cost-optimisation strategy, as OpenAI projects a $5 billion deficit against $3.7 billion in revenue for 2024.
A team of 20 engineers, including former Google TPU designers, leads this initiative.
Why It Matters
Control over compute resources is crucial in AI. OpenAI's proprietary chip development and partnerships with Broadcom, TSMC, and AMD reflect a strategy towards hardware self-sufficiency and cost mitigation.
By diversifying its chip supply and building in-house capabilities, OpenAI aims to remain competitive in an environment where hardware autonomy is as vital as advances in AI models.
OpenAI's 'Orion': Speculation, Denials, and Industry Implications
Recent hype regarding OpenAI's "Orion" model escalated when CEO Sam Altman dismissed the claims as "fake news." Reports suggested a December release with early access to partners like Microsoft.
Despite Altman's denial, earlier cryptic comments about "winter constellations" have fuelled speculation.
OpenAI asserts there are no immediate plans for Orion but hints at other technological advancements.
Orion is described as a potential successor to GPT-4, allegedly 100 times more powerful.
The development aligns with OpenAI's $6.6 billion funding round, involving a transition to a for-profit structure and significant organisational changes, including the departure of CTO Mira Murati.
Why It Matters
The speculation around Orion underscores the uncertainty of OpenAI's next moves.
Altman's denial contrasts with insider rumours, keeping industry interest alive.
The restricted early release to partners like Microsoft highlights OpenAI's focus on strategic collaborations to maximise product integration and showcase capabilities.
Risks of Using AI Transcription in Healthcare: Hallucinations in up to 80% of Analysed Transcripts
OpenAI's Whisper transcription tool has come under scrutiny for frequent hallucinations—fabricated content not present in original recordings.
These hallucinations, including invented medical treatments and racial commentary, were found in up to 80% of analysed transcripts.
Despite these risks, Whisper is used by over 30,000 healthcare professionals to transcribe doctor-patient consultations, raising serious concerns about its reliability.
Whisper remains the most downloaded open-source speech model, with over 4.2 million downloads last month.
Its integration into healthcare workflows, despite OpenAI's warnings against use in high-risk domains, reveals a troubling gap between technological readiness and appropriate application.
Why It Matters
Using Whisper in healthcare poses significant risks due to its propensity for hallucinations in critical environments.
Errors in transcription can lead to incorrect diagnoses or treatments.
Despite OpenAI's warnings, its adoption by over 30,000 clinicians highlights the pressure to deploy AI prematurely, without adequate safeguards.
This points to an urgent need for stricter regulations and a higher standard of reliability for AI tools in healthcare.
