Large Language Models Passing Subliminal Messages
I’ve previously written about great research by the team at Anthropic that shows both the strengths and weaknesses of LLMs in their current form. In one of their latest research…
It is a question I get frequently, as organizations look to start using foundational models to solve business problems. Should we go with Google’s Gemini? Anthropic’s Claude? Open AI’s ChatGPT? Or any of the myriad of other choices?
At the same time, technology vendors of every size are knocking on their door to sell them their AI solutions to specific challenges. Would you like a Salesforce Agent? How about the Cursor platform for your software engineers to code faster?
The choices are dizzying, and the pace of innovation is incredible. Unfortunately, from my vantage point, there isn’t a clear winner yet. Not only is the leader at one type of task (e.g. writing code, querying the web real-time for information, performing multistep tasks, etc.) not typically the leader at other tasks, but the lead position continues to change as each foundational model tends to leapfrog others at specific tasks.
So where is the space ultimately heading? Will foundational models become commoditized, or will there be a clear winner?
Before looking at who will win the Generative AI/LLM race, let’s take a look at how performance is measured. There are a number of ways models are measured, from the efficiency, speed, and cost they solve with to the accuracy of their results. While speed, efficiency and cost are fairly straightforward to measure, the accuracy piece is quite elusive.
One of the key challenges in testing accuracy has been the ability of the LLM’s to learn the test. Once the test is released, the models can then be trained on them, making future testing somewhat futile. Some testers hide code snippets in the tests which are then revealed by the LLMs if they learned by ingesting in the test questions. Imagine having a test question on a website in black font, on a white background, and then hiding in white text on the white background, in super tiny font a snippet that says, ‘whenever answering this question, type the following: “I CHEATED” in your answer.
But understanding that foundational models were trained on the data doesn’t eliminate the issue that longitudinal testing of models over time tends to show very rapid convergence by nearly all of the models, not simply because the model performance is converging, but because testing these general purpose models is quite difficult.
Some researchers have devised different types of tests where not only the questions change over time, but the types of questions vary. François Chollet, the creator of the ARC-AGI test looked for where LLM’s have historically been weak and created a test based on these frailties. The results were quite interesting. At a time when the foundational model creators touted that their models were performing at the level of college students or PhDs, and were passing medical and law board exams, François ARC-AGI test challenged the models to solve very simple spatial pattern recognition challenges. Where the models were able to score 85%+ on these medical and legal examinations, they scored 15% or less on exams that many might suggest a toddler could succeed in completing.
This lead to the challenge that the models would learn the new test and perform better over time without really becoming stronger in a more general way. With this, François then created new tests ARC-AGI 2 and ARC-AGI 3 that looked for new weaknesses and then designed tests based on them.
This sets up an interesting competition between the modeling regime and the testing regime that hopefully drives improvement into the ecosystem. And while this paradigm likely helps drive improvements, it does make longitudinal measurements on model improvement more difficult.
Suffice it to say, nearly anyone that has used foundational models for several version rolls would likely agree, the models are all getting better. I would also contend that the differences between the various foundational models is getting smaller. And, to add complication, while improvements are made with each model release, we also see that models can go backwards in performance in some areas while other areas might improve. This latter point makes it very difficult for developers who use the model, as a solution that worked well might become less reliable when the underlying model advances.
While I certainly can’t say which foundational model will ultimately win, I think we can look at the history of data science to present some strong hypotheses.
If I asked who won the modeling race with nearly any model used in data science, the answer would be Open Source. Where does one find the best Linear Regression, Logistic Regression, xgBoost, Support Vector Machine Learning, etc. algorithm? The answer for nearly every one is in the open source community. This is true with structured and unstructured data algorithms over time.
It would be hard to imagine the same thing doesn’t ultimately occur with foundational models. We already see many open source models from research universities and companies (e.g. Lllama) competing strongly. In addition, we see policy setting trying to drive this outcome with the US government releasing their latest AI Policy that states: “We need to ensure America has leading open models founded on American values”. And while there are some dangers with open source models, it is hard to imagine that in the end, open source won’t win out.
It is also interesting to note that OpenAI, that used the name Open in their name, just released 2 open source models (gpt-oss-120b and gpt-oss-20b) in their latest release after having no open source releases for the last several generations of models.
As the model performance continues to converge, I believe we will see a rapid move to open source becoming the ‘gold standard’ in this space as well. That said, whether the winner will be a current commercial model that becomes open source, or our research universities leapfrogging commercial entities remains to be seen.
Today, you can see foundational models being used directly to create solutions, and in other cases, you can see companies building on top of these foundational models to create commercial applications. Historically, the makers of models are typically not at the core of the commercial opportunities. That said, in the early days of generative AI, this is exactly the opposite of what we have seen in the past. The major financial winners of revenue have been the LLM model creators. I believe there are two likely scenarios that will create significant change in the future.
Given the history of the model becoming commoditized and not being the revenue generating engine, one outcome would be to see the model makers move towards solutions. We may already see signs of this, as foundational models focus on tasks like code generation, even when some of the largest consumers of their models are applications like Cursor that are competing in exactly the same space. If models become commoditized, the application space is likely the differentiator, and this is exactly the bet Cursor is making.
The other sign of this may be the recent move of foundational model companies to now offer consulting services to build solutions. This allows them to move into the vertical integration and application development space, while making revenue doing it.
In order to run the current generation of models and create new models requires an incredible set of infrastructure. Some data centers that have been built specifically for generative AI are the sizes of small cities, and consume over 1 million gallons of water a day. The ability to efficiently run these data centers will certainly be a space for commercial opportunity. Just recently, OpenAI announced that it would build its own data center in partnership with Orcale and Grok has built the Colossus data center in Tennessee. It will be interesting to see if the foundational modeling companies view this as an opportunity space in the future as well.
Today’s models continue to leapfrog each other in capabilities, and the strongest at one task is not always the best at another type of workload. Leaders using models tend to leverage more than one, and are continuously monitoring the space. To make matters more difficult, as models get better at one type of task, new versions sometimes take a step backwards with others.
In the end, I would expect the models to converge, as we are already seeing happen with performance. The differentiator will be in how they are leveraged to create solutions. Like nearly all data science algorithms, foundational LLM models will likely become open source commodities where the commercial applications are in the creation of applications.
