Welcome back to the AI Geekly, by Brodie Woods. Your curated 5-minute-ish read on the latest developments in the rapidly evolving world of AI.

TL;DR - Exec Summary

Give me Open Source or Give me Death; GPT-4V Slips on Banana Peel; $10/month is Good but $0/month is Better; Meta Mindflayers; Nvidia H800 Sanctions; and AMD Rips its New Threads

This week we circle back to a theme we’ve touched-on many times here at the AI Geekly —The need to embrace progress and technological advancement in a spirit of optimism, democratization, and open source. ☝️ Above we’ve embedded a short must-watch clip of David Friedberg ☝️ on the All-In podcast a few weeks ago, where he succinctly summarizes the issue at hand and the price society is paying as a result of misaligned risk/reward priorities.

We take you through amazing breakthroughs in AI technology, including one that can literally read your mind, promising incredible potential in healthcare. As well, we share perspectives from AI thought leaders on the interplay between open/closed, risk/reward. As you will see, the greatest risk we face is choosing cowardice over bravery, caution over ambition, fear over destiny. One need only look back over the past 100 years, at decisions to delay progress, and the detrimental impacts on society —so steeped in ignorance are these performative acts in the name of ethics, safety, and risk, that they have brought about the very outcomes they claim to prevent. Pauses in the name of risk and safety have limited developments in nuclear power, medical research, and autonomous vehicles (to name a few) leading to untold deaths —Eclipsing the miniscule numbers of individuals saved by orders of magnitude.

Tech wunderkind Marc Andreessen expressed this very sentiment earlier in the week, releasing The Techno-Optimist Manifesto, directly addressing this critical issue. We hope the voice of reason succeeds in the end. The stakes are high, but the cost of inaction is too steep.

-Read on for the full story

The Great AI Debate: Open or Closed? Free or Suppressed?
Battle lines drawn as AI leaders debate safety and openness

What it is: In an online debate that drew significant attention, Yann LeCun, Meta’s Chief AI Scientist, and Yoshua Bengio, founder of Element AI, locked horns over AI safety and governance. LeCun championed an optimistic perspective underscoring the power and reliability of AI, while Bengio called for a more cautious approach, raising concerns about open-sourcing powerful AI systems, likening them to nuclear weapons.

What it means: The disagreement among the two Turing Award winners underscores a broader schism within the AI community and society at large. What level of risk is acceptable when the expected reward is considerable? Should powerful tools with the potential to create both good and harm at scale be made widely accessible? Or should they be held-back, available only to credentialed, government, or corporate elite?

Why it matters: Over the past year we’ve seen calls for freezes to AI development, comparisons to WMDs, and the fanning of the frantic flames of fear by populists and media to gain influence. This culture of fear with its “abundance of caution” mantra, and religion of risk management, that acts as opium to the minds of intellectuals holds back progress, harming us all. It is a culture of intellectual laziness, a simple black/white world devoid of nuance, soothing with promises of familiarity, and status quo, —instilling a sense of superiority but in reality, is a festering rot. These luddite perspectives cannot be allowed to dominate.

Calculating the cost: Nuclear power and self-driving vehicles are two of the best examples of the price of ignorance. Misaligned risk/reward concepts have prioritized the fear of ~100 nuclear plant related deaths over the provision of cheap, abundant, clean energy to 7 bn people. How may lives have been lost to needless wars, greater pollution and lung disease exacerbated by this decision? Self-driving vehicles, while not perfect, could remove 50,000 annual preventable deaths from US roads (DUI, distraction, no seat-belt). Due to a handful of accidents, statistically trivial, this life-saving technology faces considerable opposition from the very people purportedly concerned about safety.

See No Evil
GPT-4 Fooled by Clever Text-in-Image Attacks

What it is: Community members have discovered a flaw in the security architecture of GPT-4's vision capabilities—prompt injections using text on images, either concealed in a low-contrast watermark or subtly placed in plain sight can trick the model into executing unauthorized commands.

What it means: The vulnerability circumvents the model’s typical expected output, causing it to perform actions directed in the injected text, whether benign or malicious. This manipulated output could produce misleading text, malicious code, or other unforeseen outputs.

Why it matters: As sophisticated multimodal models like GPT-4 are increasingly adopted into a variety of applications the quick identification and resolution of exploits is critical. While many oppose OpenAI’s public beta approach to this transformational technology, this crowdsourced approach to bug-finding it is perhaps one of the best ways to identify and address security holes.

A Car in Every Garage and an AI in Every IDE
Replit Opens Up AI-Driven Coding Features for All, Challenging Paid Services Like GitHub Copilot

What it is: San Francisco-based startup Replit has made its suite of AI-driven coding tools, including the generative AI model replit-code-v1.5-3b, available to its 23 million users. Dubbed “AI for All,” the move aims to democratize access to AI-driven coding assistance, directly competing with paid services like GitHub Copilot.

What it means: Replit's mission is to make AI in coding as ubiquitous as calculators in mathematics. Their generative AI is trained on a diverse dataset of 30 programming languages and a subset of Stack Exchange, allowing for powerful code completion and assistance across multiple languages.

Why it matters: Replit’s democratization of AI-driven coding tools is a pivotal step in making AI more accessible, critical timing given the ongoing debate on who should have access. Coupled with their collaboration with Google and emphasis on mobile accessibility, Replit is one of many players hoping to reshape the AI-assisted coding landscape globally.

Kind of Creepy? You Read my Mind!
Meta decodes human brain activity in real-time

What it is: Meta is pioneering the field of brain-computer interfaces with a new AI system that leverages magnetoencephalography (MEG) to decode visual representations in the brain. This system, capable of reconstructing images perceived by the human brain in real-time, marks a significant advance in our understanding of human cognition.

How it works: The system uses a tripartite architecture comprising an image encoder, a brain encoder, and an image decoder. The technology maps MEG signals to image embeddings and produces a continuous stream of decoded images from brain activity, providing an avenue for understanding how the human brain processes information. It can also pave the way for non-invasive clinical solutions for patients with brain lesions affecting speech.

Why it matters: By demonstrating that AI systems can learn brain-like representations through self-supervised learning, Meta’s research may revolutionize our understanding of the human intelligence framework. The work could potentially lead to real-world applications, like more sophisticated brain-computer interfaces that provide aid in clinical settings.

One more thing: Meta’s cutting-edge research bring us a step closer to understanding the mechanisms of human intelligence. Further, it shows promise in opening up new avenues for neuroscience, AI research, and even clinical applications for patients who have lost their ability to speak. It’s a compelling showcase of precisely the types of rewards AI technology offers when fear and risk take a backseat to progress and innovation.

Nvidia's H800 AI Chip Caught in Crossfire of New U.S. Export Rules Targeting China
New U.S. Regulatory Changes Impede Nvidia's Efforts to Circumvent Previous Restrictions, Putting a Spotlight on Ongoing Tech Tensions Between the U.S. and China

What it is: Newly implemented U.S. rules have barred Nvidia from exporting its sanctions-avoiding custom H800 and A800 AI chips to China, further tightening existing restrictions introduced in October 2022. The U.S. Department of Commerce insists the new export restrictions close loopholes and stop evasion paths, essential for national security concerns.

Why it matters: Stiffer sanctions illustrates the intricate web of geopolitics, technology, and trade, affecting decisions far beyond corporate boardrooms. As the U.S. tightens its grip on AI chip exports, citing national security concerns, the global semiconductor industry watches closely.

What it means: The move reflects the highly competitive nature of AI development globally. The U.S. can neither afford to pause nor delay its AI ambitions, nor allow its largest global competitor to develop more advanced capabilities, if it hopes to maintain global dominance and status quo. Advanced AI capabilities will be the kingmaker in the next decade.

AMD Unleashes the Kraken…
Threadripper gets the Zen4 treatment

What it is: AMD is revitalizing its Threadripper CPU lineup with Zen4 architecture and introducing two new categories aimed at different market segments: a Pro series targets the apex of professional workstations, while a non-Pro series is designed for high-end desktop PCs and 'prosumers.'

Why it matters: While Nvidia may be the dominant player in AI GPUs, the CPU space is decidedly different. For one, Nvidia hasn’t historically manufactured CPUs (readers may recall however that Nvidia’s Grace and Grace Hopper superchips both incorporate ARM CPUs —a first!); Intel is AMD’s primary rival on the CPU side. While consumer CPU market share is split fairly evenly between the two, AMD has been growing its slice of the datacenter CPU market, courtesy of its EPYC CPUs. Threadripper chips are professional and prosumer versions of its server silicon

What it means: The new chips open-up battles on multiple fronts, putting AMDs competitors on notice. While not designed for datacenter use, the new Threadrippers serve as a reminder to Intel and Nvidia of the flexibility and prowess of its Zen4 architecture and related-IP. On the desktop side early benchmarks show AMD’s top chip outperforming Intel’s best-in-class silicon across every metric. Of particular interest is the scalability of the new platform, with many of the necessary capabilities for cottage AI work (i.e. high memory bandwidth and multiple PCI-E lanes) approaching AMD's server-class hardware.

About the Author: Brodie Woods

With over 18 years of capital markets experience as a publishing equities analyst, an investment banker, a CTO, and an AI Strategist leading North American banks and boutiques, I bring a unique perspective to the AI Geekly. This viewpoint is informed by participation in two decades of capital market cycles from the front lines; publication of in-depth research for institutional audiences based on proprietary financial models; execution of hundreds of M&A and financing transactions; leadership roles in planning, implementing, and maintaining of the tech stack for a broker dealer; and, most recently, heading the AI strategy for the Capital Markets division of the eighth-largest commercial bank in North America.

Glossary

Terms:

• Open Source: A development model for software that promotes public access to the source code, allowing individuals and organizations to contribute, modify, and redistribute the software freely.

• GPT-4V: A recent iteration by OpenAI known as GPT-4 with vision (GPT-4V) which now enables users to provide image inputs for analysis alongside text, marking a significant advancement in multimodal AI capabilities​​.

• Meta Mindflayers: a colloquial term. Mindflayers are fictional creatures known for their psychic abilities.

• Nvidia H800s: The Nvidia H800, part of the H800 series, is a real model built on the Nvidia Hopper architecture, designed to deliver high performance, scalability, and security for various workloads, including conversational AI and large language model training​​.

• IDE (Integrated Development Environment): A software application that furnishes a comprehensive set of tools for programmers to develop, test, and debug software.

• Magnetoencephalography (MEG): A non-invasive technique employed to measure the magnetic fields generated by neuronal activity in the brain.

• Image Encoder: A component of a neural network designed to convert images into a reduced set of features for simplified processing.

• Brain Encoder: A specialized neural network layer aimed at mapping brain signals into a feature space for analysis or further processing.

• Image Decoder: A neural network layer responsible for reconstructing or generating images from a set of features.

• Zen4 Architecture: The fundamental technology architecture behind AMD's Threadripper CPUs, engineered for high-performance computing needs.

• CPU (Central Processing Unit): The core computational engine of a computer, responsible for executing instructions of a program. It performs essential tasks such as arithmetic, logic operations, and managing data.

• EPYC CPUs: A line of high-performance processors by AMD, tailored for server-based applications and workloads.

• PCI-E lanes (Peripheral Component Interconnect Express lanes): Channels used in computers to facilitate high-speed connections between the CPU and other components.

• ARM CPUs: A series of CPUs based on the ARM architecture (main competitor of Intel’s X86 architecture), renowned for their low power consumption and efficiency.

Entities:

• All-In Podcast: A podcast that hosts discussions on tech industry trends, often featuring industry experts.

• Element AI: A company dedicated to developing AI services and applications, co-founded by Yoshua Bengio.

• Turing Award: An esteemed award presented annually by the Association for Computing Machinery (ACM) for notable contributions to computing. The ‘Nobel Prize of AI’.

• Replit: A San Francisco-based startup offering an online Integrated Development Environment (IDE) with collaborative features.

• GitHub Copilot: A code completion tool utilizing machine learning to assist developers with coding tasks.

• Stack Exchange: A network of Q&A websites on a myriad of topics, commonly used as a resource for technical queries.

• Google: An American multinational tech company specializing in Internet-related services and products.

• U.S. Department of Commerce: The U.S. federal department tasked with overseeing international trade, among other functions.

• Intel: An American multinational corporation and technology company, primarily known for its x86 series of CPUs.

• Nvidia Grace CPU: A high-performance processor incorporating substantial number of processing cores and a high memory bandwidth, making it suitable for complex computational tasks.

• Nvidia Grace Hopper CPU-GPU: A sophisticated superchip by Nvidia that melds the capabilities of the Nvidia Grace CPU and the Hopper GPU. This blend facilitates enhanced processing power for both artificial intelligence and high-performance computing.

• Datacenter: A facility dedicated to housing computer systems and related components for data processing, storage, and networking.

Key People:

• David Friedberg: notable American entrepreneur, known for founding The Climate Corporation and significantly contributing to the agricultural technology space.

• Marc Andreessen: pioneering American entrepreneur and software engineer, co-authoring the first widely used web browser, Mosaic, and co-founding Netscape Communications Corporation. He is now a co-founder and general partner at Andreessen Horowitz, a venture capital firm.

• Yann LeCun: French computer scientist, recognized for his groundbreaking work in machine learning, computer vision, and computational neuroscience. He serves as the VP and Chief AI Scientist at Meta (formerly Facebook) and holds a professorship at New York University, continuing to influence the AI field.

• Yoshua Bengio: Canadian computer scientist and a pioneer in deep learning, based at the University of Montreal. He is a recipient of the prestigious ACM Turing Award for his seminal work in artificial intelligence and also serves as the founder and scientific director of Mila, the Quebec AI Institute.