AI Geekly - Huff and Puff

Welcome back to the AI Geekly, by Brodie Woods, bringing you yet another week of fast-paced AI developments packaged neatly in a 5 minute(ish) read.

TL;DR Let my model weights goooo; Bezos bucks; Sleepwalking LLMs

This week we take a quick peek at Databricks’ new open-source (wooo!) model, why it outperforms, and how companies should be thinking of using these types of models to accelerate their AI efforts. Next, we take a look at Amazon’s latest AI investment and how it stacks up relative to its peers. We finish with thoughts on a new potential risk factor with LLMs: sleeper agents that lie dormant in a model and come-out to wreak havoc when least expected.

Leading the Way to the Open Source Promised Land
Databricks’ Mosaic GenAI team wows with new model

What it is: The Databricks team via it’s Mosaic Research Team this week released a new state of the art large language model (LLM) called DBRX, an open-source decoder-only model which surpasses all previous open-source models including Meta’s Llama 2 and French company Mistral’s Mixtral 8x7B, by a wide margin in coding (38% and 16% better, respectively) and math (15% and 5% better, again). Like Mixtral 8x7B, DBRX is an MoE model (Mixture of Experts) meaning that rather than a single LLM, there a several smaller models that work together to respond to user inputs. DBRX uses 16 experts and narrows responses down to four of them to respond (based on the topic) whereas Mixtral 8x7B uses only eight experts and chooses two to respond. Benchmarks favor DBRX.

What it means: There’s a new top dog in open-source model land, and it’s DBRX. It wasn’t too long ago that Mixtral 8x7b held that title for three months and prior thereto Meta’s Llama 2 four months (note it took one month less to usurp the top model this iteration). Normally one would say that the space is very competitive but given that these are open models one might say instead the space is very additive, with various groups of researchers contributing their work to the global open source community.

Why it matters: As big proponents of the democratization of AI and the widespread availability of open-source AI tools as a means to ensure equitable outcomes, we are very pleased to see models like this. Aside from the societal benefits of open-source AI, they offer a compelling value proposition for corporate use cases (enterprise through to small and medium-sized businesses): open-source models are cheaper to experiment with, easier to fine-tune, and most importantly for certain highly regulated industries, like finance and healthcare, available for on-prem use instead of relying-upon external cloud.

Spinning a faster flywheel: Not only do open-source models provide an affordable and rapidly iterative model solution for enterprise and small business customers (often better than models specifically designed for enterprise) but they accelerate global development of AI, sharing learning and insights publicly and making those available to all. Unlike the espoused “AI for the benefit of all humanity” we see from increasingly closed OpenAI, these players and researchers are doing more than simply talking the talk —they walk it.

Free Prime Two-Day Shipping for Purchases over $4 Bn
Amazon increases its investment in Anthropic

What it is: In the AI space we like to toss around absurd numbers. Here’s a great one: $2.75 Bn. It’s big. Almost cartoonishly large. Using what one can only assume is a giant checkbook of novelty-sized checks, Amazon has just announced a $2.75 Bn investment in OpenAI rival Anthropic. Readers will recall from earlier Geeklies that Anthropic recently released its highly performant Claude 3 family of models (Sonnet, Haiku, Opus) to much fanfare, with Claude 3 Opus generally recognized as America’s Next Top Model (a less impressive feat when you consider that OpenAI’s GPT-4 is almost a year and a half old).

What it means: The AI pissing-matches are in full… uh… spra… no, never mind. Well, the competition is heating up anyways. Here’s a summary of spend-to-date by the major tech companies (external investment).

Microsoft:

• OpenAI - $13 Bn

• Mistral AI - 15 mm Euro

• Inflection.ai - $650 mm

• Adept.ai Labs -part of $350 mm Series B round

Alphabet/Google:

• Anthropic - $2 Bn

• AI21 Labs - part of $155 mm Series C round

• Runway ML - part of $100 mm investment round

• Hugging Face - part of $235 mm investment round

Amazon/AWS:

• $4 Bn invested in Anthropic to date

• Hugging Face - part of $235 mm investment round

Nvidia:

• Adept.ai Labs -part of $350 mm Series B round

• Inflection AI: $650 mm

• Hugging Face - part of $235 mm investment round

AMD:

• Hugging Face - part of $235 mm investment round

Intel:

• Hugging Face - part of $235 mm investment round

• Stability AI - $50 mm

Why it matters: This list is not exhaustive (well, it is to write…) but it illustrates the reality that a lot of cash is being thrown around in pursuit of GenAI value. Consider the trillion-dollar rocket ships that names like MSFT and NVDA have become over the past year and the sheer volume of dollars becomes more apparent. Investors and corporates are deploying capital in AI across structures, markets, and value chains. We haven’t yet separated the companies with real long-term AI potential from the also-rans and there’s still going to be a lot of tears ahead. The bigger they are…

Alignment is All you Need
Sleeper agents introduce new threat vector

What it is: a new whitepaper is making the rounds, focused on the risk behind so-called “sleeper agents” in models, essentially models that have been trained to be malicious (or have trained on data that imbues them with such traits coincidentally). These negative behaviors lie dormant inside an LLM and, when activated, lead to corrupted outputs.

What it means: These attacks are reminiscent of phishing and social engineering attacks. There are limitless ways to trick both humans and AI models into undesirable behaviors. The best defense in both cases is better training, for the humans and the AIs.

Why it matters: Enterprise players know well that there are infinite threat vectors affecting their businesses, therefore it should come as no surprise that a promising tool like AI brings with it additional risk. Finance players know well that risk is a necessary part of the risk/reward equation —there is no reward without risk -the trick is finding ways to hedge or mitigate that risk.

Hedging bets: So how do you hedge your bets in AI? Talent and partners —a firm hand on the rudder is needed to avoid rough waters. Most companies don’t have this talent internally; management teams would be wise to look around and snap-up talent and partner bandwidth now. The longer they wait the more costly it becomes. Choose wisely.

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.