Doesn’t This Thing Go Any Faster?

Welcome back to the AI Geekly, by Brodie Woods.

The EU has agreed to a binding AI Act, set to take effect in 2025. We believe that traditional political instruments move too slowly to address AI’s opportunities and potential risks, this being a prime example, however there are promising alternatives... Researchers have been dropping models like crazy this past week, each of which raises the bar in its respective way. Robotics news from Tesla and a team of NYU researchers, respectively, also caught our eye, presenting an interesting case study where wildly different budgets deliver comparable outcomes.

Read on for the full story

We Might Need To Try Something Different
Attempts to regulate AI demonstrate exactly why we can’t

What it is: EU lawmakers recently agreed to landmark regulation introducing restrictions on AI development in the EU beginning in 2025. Applying a risk framework, the regs classify AI systems and introducing increasing levels of oversight and restriction depending on risk severity, up to and including banning certain “Unacceptable” uses.

What it means: Striking a balance between fostering innovation and safety is a delicate art, particularly with the dull and heavy hand of transcontinental regulation. Given the pace of legislation (‘21-’25), this hardly seems feasible with available political tools. Further compounding the ineffectiveness of the slow pace is the sizeable-and-growing knowledge gap between lawmakers and the rapidly evolving field.

Why it matters: Ensuring the impacts of AI are widely beneficial is infeasible with traditional means of control via high-latency government. New concepts and tools present our most realistic path to societal prosperity: Democratization of AI (via Open Source), and AI Superalignment are the two emerging solutions with the most promise. The former empowers both the individual and the many, giving ubiquitous access to safe, transparent, and powerful AI tools to all, while the latter endeavor to align superintelligence with humanities goals and values.

It’s Raining Models!
Flurry of models released during supposed quiet December

What it is: The last week has seen a flurry of new AI models, including several new highly performant LLMs, and a few new image generators. The models include:

• LLM360’s Amber 7B, and CrystalCoder 7B) - Open Source

  • Fully open (training data, model weights, everything)

• DeciLM-7B - Open Source

  • Top performing 7B parameter model

• Microsoft’s Phi-2 (2.7B) - Open Source

  • Incredibly small, highly performant model at only 2.7 Bn parameters vs. GPT 3.5 at 175Bn

• Mixtral 8×7B (Mix of Experts) - Open Source

  • Open model weights, optimizes performance and cost

• Google DeepMind Imagen-2 - Closed Source

  • Highly realistic image generation excelling at accurate anatomy and text display

• Stable Zero123

  • Considerably outperforms prior version in extrapolating 3D models from 2D images

Why it matters: We highlight these models to showcase the breadth and depth of developments (limited solely to LLMs and image generators) in just a single week during a holiday slowdown period. This underscores our comments in the preceding notes on the mismatch between the pace of government and development. This moves very fast.

DANGER WILL ROBINSON! YOUR CLOTHES NEED MORE STARCH
Better AI cores for more robots in 2024

What it is: Two robot announcements caught our eyes this week, at the intersect between Tech and AI: Tesla’s Optimus 2.0 and Open-Source chore bot Dobb·E.

What it means: We’ve noted the ambitious pace of development in the AI space and hope by now this is widely accepted. Following on this then is the application of this same AI to robotics in novel ways. While Tesla’s Optimus 2.0 robot gets its flashy dance moves from the company’s Dojo supercomputer, Dobb·E was built by seven NYU researchers and can perform 109 household tasks, taking only 20 minutes to learn a new task.

Why it matters: Tesla’s Optimus 2.0 is starting to approach the level of Boston Dynamics’ bipedal robots in a fraction of the time it took BD to get to the same point —they’ll have a solid robot soon. Dobb·E is an inspiring application of AI and robotics paired in a very simple, yet performant package. It’s not as splashy as Tesla’s robot, but it only cost about $24k to make. That’s about the price of one of the 10,000 H100 Nvidia GPUs that Tesla uses to train Optimus but providing 80% of the utility for 0.000000001% of the price.

“I Understood That Inference”
AI-specific Neural Processors bring inference to local hardware

What it is: Chip designs are increasingly incorporating Neural Processing Units (NPUs) to offload local AI-related tasks to specialized components, freeing-up the CPU and GPU for other tasks. NPUs deliver significant improvements in AI model performance and present a way to dramatically reduce the energy needs.

What it means: AMD has incorporated NPUs into its mobile Ryzen 7000 and 8000 series chips (rebadged 7000s with a 60% overclock to the NPU) while Intel is incorporating NPUs into its next generation Meteor Lake (14th Gen) mobile chips, while mobile-focused Qualcomm also has found a coveted place for an NPU on its die. We need to give credit where credit is due though —Apple has included NPUs on its iPhones (aka its shadow R&D arm) since 2017 with its A11 Bionic chip, and more recently with the introduction of its proprietary M1 chip in 2020.

Why it matters: First, it’s worth pausing for a moment and reflecting on how quickly hardware is adapting to the AI Age. ChatGPT wasn’t introduced that long ago and hardware is already converging towards AI development (including at the datacenter level). The three ingredients/accelerants of AI are: compute, data, and algorithms and this will contribute materially to the first.

One more thing: The proliferation of dedicated consumer AI hardware will have a multiplying effect on the reach and development of AI as millions of developers get access. This widespread availability is critical to ensuring the democratization of AI.

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

• AI Act: a European Union regulation that categorizes and regulates AI applications based on their risk to harm, aiming to ensure AI development aligns with safety and ethical standards​​.

• Democratization of AI: making AI development accessible to a wider audience, including non-experts, facilitated by large companies through user-friendly resources​​.

• AI Superalignment: the process of ensuring that superintelligent AI systems align with human values and goals to prevent potential catastrophic risks​​​​​​.

• LLMs (Large Language Models): advanced AI models capable of processing and generating human language on a vast scale.

• Tesla’s Optimus 2.0: Tesla's Optimus 2.0 is an advanced robot that utilizes Tesla's AI technology for various functionalities.

• Dobb·E: an open-source chore robot developed by NYU, designed to learn and perform household tasks efficiently.

• Neural Processing Units (NPUs): specialized hardware components in computer chips designed to optimize AI-related tasks.