Coffee Shop Thoughts - 11/25/24
It’s been a minute since I’ve dumped some thoughts down… so here we are! Straight from the top of the dome, here are the latest unvarnished thoughts that I’ve been having lately. Enjoy a windy trip Beyond the Yellow Woods!
AI / LLMs
We continue to see incredible pace and innovation in this field. However, it is marred with some challenges and headwinds. For starters, companies that are focusing on wrappers are getting their asses handed to them. So many vendors are trying to wrap a basic UI around an LLM and charge an insane price for it. The challenge? Someone else does the same thing and offers it at a cheaper price, and thus a pricing war occurs, creating a deflationary effect that pushes these products to zero.
I think the ultimate end state here is that companies are going to be forced to do 1 of 2 things: focus on the foundational layer or focus on a truly novel application experience.
At the foundational layer, the mainstream LLM still lacks a crucial element: knowledge graphs. We are hitting a plateau in capabilities from the training that can occur on the public corpus of human knowledge. The challenge? 80% of the data is privately owned in silos within businesses. On top of that, there is the challenge of mapping semantic reasoning to inform a foundational LLM model on how to reason based on a business’s context and perspective.
This argument ultimately comes down to the industry needing to build an easy on-ramp to employing both a GraphRAG as well as a Knowledge Graph in order to get a much higher yield on intelligence from LLMs. I don’t see many folks working on this yet so… the hunt continues.
From the application layer, we must evolve from the basic chat interface. This is cute and reminds me a lot of the early internet days with forums. However, it’s an absurdly inefficient way of communicating knowledge. I’m not suggesting that we move to voice or anything, although I think that would be a nice benefit if we could figure out the lag/latency. What I’m suggesting is that we need industry-specific applications that truly leverage the nomenclature and concepts of that industry. Think Jarvis from Iron Man. I should be able to talk to a system and have it understand a large corpus of instructions, tasks, and collaboration keys to create things. We don’t have that today and it’s a large gap. We don’t have a UI that builds things in real-time, visually, for the end user to collaborate on.
So, alas, we continue the hunt here as well.
Geopolitics
Hoooo boy, did we all have fun the last few weeks or what?? Well, Trump is in office. The world continues to heat up. It’s going to be a volatile 4 years.
First off, I think we are going to experience some brutally volatile movements across the USA. Between the DOGE initiative to nuke large portions of the wasteful government to solving the border crisis, I’m fully expecting we see a large immune response from the “deep state” that pushes back hard against this. Like cancer, we must give a heavy dose of chemotherapy. It’s necessary and needed for us to get back on track.
On the geopolitics front, we see continued escalation on the Eastern front. Biden sent one last fuck you to the world by allowing Ukraine to launch USA-produced long-range missiles into Russia. Russia has responded by launching intermediate-range ballistic missiles into Ukraine and threatening to start bombing other countries.
We need sanity back in the world because the East and West theaters are getting hot. We need Europe to get their shit together and prepare for war. They need to be helping stabilize the Middle East but the broader EU is still hell-bent on claiming Israel is the bad guys. We need a bastion of hope to stabilize everything and I believe that Trump will provide some of that for the East but we will see further cracks in the West/APAC region.
My eye is on Australia and the Philippines. If you look at the stock market positioning, Europe is getting sold off while Australia is getting purchased. We have been developing strategic arms relations as well as large amphibious joint training exercises with those countries in what I’m expecting is our push for deterrence.
My full expectation is that Trump focuses on the global situation while JD Vance focuses on the domestic situation. This will be a violent fight but we either die fighting on our feet or die on our knees. My take is that we must lead the world to security and prosperity through incredibly strong deterrence - and back what we mean.
People may disagree or call me callous for saying this, but we must back up our threats. This means that we are going to have to break our own rules and violate sovereign nations with operations, just like we did with Pakistan and ISIS. If we don’t, we will have substantially more needless suffering.
One last thing that I feel will happen extremely fast is that we are going to very aggressively retool our supply chains (more than we have) to focus on the Americas. I fully expect a much more robust NAFTA-style trade agreement and to develop 1:1 hyper-strategic relationships with key emerging countries, such as El Salvador and Argentina. We will fortify our bloc and, by doing so, can help support and fund the East and West blocs. We are operating on a containment strategy whereby we contain China via the APAC region via military, energy, and economics while we contain Russia via economics and NATO/EU proxy through de-escalation. We do not want that region to escalate further because the EU is woefully unprepared to fight right now.
I pray for a safer world but I’m not sure we’ll get it.
Mars
I fully believe we’re going to Mars in the next decade. In fact, it dawned on me that Elon is playing the ultimate chess game right now. Like a game of Starcraft, we are going to grow robots here on Earth (Optimus) and then go explore the “unexplored” parts of the map via rockets to Mars, sending droves of Optimus robots to plant and grow a new colony on Mars. Why send humans initially and risk the death when we can teleoperate these Robots from afar?
When you think about unleashing economies, the ability to expand into new territories is critical. One only needs to look at the USA history for this with events such as the Louisiana Purchase and the Westward Expansion. We didn’t really know what lay beyond the veil but we ventured into the unknown regardless, only to be rewarded with incredible suffering and unimaginable wealth (for that time). It’s the ultimate destiny for mankind and we’ve saturated what can be accomplished on this planet from a wayfaring perspective.
Ad Astra, to the max. Effective Acceleration.
Novel Knowledge Engineering
I got into a tiff on Twitter/X this past weekend with someone who clearly doesn’t understand the limitations of LLMs. Their contention was that we already have everything we need from the corpus of public knowledge to create AGI.
This couldn’t be further from the truth.
In my opinion, we need 3 critical abilities to achieve AGI:
Ability to generate and smith novel knowledge
Ability to transfer context and knowledge across disciplines (ie. cross-disciplinary proofs)
Ability to generate semantic reasoning based on the prior 2 points
When you teach a child new concepts, their connectome/neural pathways best develop when the training material is formulated in memetic packaging. The “memes” provide surface-level context about a topic while creating a sort of “tracer bullet” of knowledge and content behind it, creating a lineage of information that informs the learning from the meme.
All of this, however, is predicated on having access to the data. Over 80% of the world’s data is proprietary and locked away in corporate vaults that public foundational models like Claude3 and ChatGPT 4o don’t have access to. These LLM model weights have no insight into this data and thus have no way of incorporating the memetic packaging of that data to influence their model weighting.
As I’ve written about many times before, you need a multi-modal architectural approach to this problem that supersedes just “LLMs”. In my opinion, this comes down to a few critical elements:
Knowledge Graphs
GraphRAGs
Foundational Large Language Models
There is a superb paper that was made public recently by Markus J. Buehler of MIT that explore this exact topic with absolutely incredible findings.
Titled “Accelerating Scientific Discovery with Generative Knowledge Extraction, Graph-based Representation, and Multimodal Intelligent Graph Reasoning”, it dives into a hypothetical system that covers the 3 architectural concepts I bulleted out with findings of commonalities from radically different communities of data.
It’s well worth the read even though the paper is long. At the end of the day, our reality is based on a globally shared language of mathematics that constructs everything we experience. We are on the cusp of developing systems that are capable of reverse engineering critical common structural beams from the substrate of different memetic corpora of data, giving us Humans the tooling needed to accelerate novel discovery radically.
That’s it for now! I’ve been bad about posting/writing lately so I'm going to make a more concerted effort to do this more frequently.
Cheers!
- Ryan