Ask yourself these questions, assuming you have a serious medical condition and your doctors are going to be using GenAI to scan your images and test results to discover the breadth and depth of your condition and to recommend treatment.
First, do you want that AI based on a closed source, proprietary model from OpenAI or Anthropic, an open weight model from Meta Platforms, Moonshot AI, Alibaba, DeepSeek, OpenAI, Google, Mistral, or the countless models being offered up on Hugging Face? Or do you want a fully open source model like the Nemotron 3 models from Nvidia or the now ancient Llama 3 and 4 models from Meta Platforms? Do you want a model made by a US company only?
Second, do you want that model to have the largest number of input parameters and output weights as possible or do you want a streamlined model with less data and fewer weights?
These are good questions, but you will not be making these choices. Medical software suppliers and healthcare companies will be doing that for you. But what are the right answers?
Having grown up in the proprietary software era of the 1980s and the Unix, Linux, and Windows Server revolutions of the 1990s, I see the value of both proprietary system and application software and the rise of open source. I also know the value of creating custom and inhouse application software, which dominated the first several decades of corporate computing and which is still done by hundreds of thousands of large corporations around the globe. Control your code, control your corporate fate.
But the bias I have is towards commercially supported or, in the case of the hyperscalers and cloud builders, self-supported open source software once that open source software rises to the same level of functionality as the closed source, proprietary alternatives.
This is what has happened in operating systems and their related middleware stacks over the decades. Back when I was a cub reporter there were probably on the order of 25 different datacenter hardware platforms and maybe twice that number of operating systems, and over the years this has collapsed down to X86 and now Arm iron with a smattering of IBM z mainframe and Power systems machinery, and IBM’s platforms run the last supported Unix (AIX) and last supported proprietary platform (IBM i) in midrange and enterprise systems, and its mainframes run proprietary z/OS and sometimes older stuff in partitions as well as Linux. (Power Systems run Linux too, obviously, and not just Red Hat Enterprise Linux.)
And while Windows Server footprints are shrinking out there, I do not think usage is for the core back office and front office applications upon which that platform jumped from our desktops to our datacenters two and a half decades ago. This is more a statement about how many cores you can cram in a server chassis these days and the rise of data analytics and now AI, which runs exclusively on Linux. But, that said, even Microsoft is embracing Linux, having used Linux underneath its SONiC Linux distro for the past decade and having created the Azure Linux distro for its cloud this May.
Application software is a more varied story. Back in the early days of computing in the 1960s through the 1980s, most large enterprises controlled their own back office application code, starting with basic accounting and evolving into MRP systems then full-blown ERP systems.
But about the time data analytics on that commercial data started taking hold, the first wave of SaaS hot at the same time commercial ERP suites got good enough for companies with a huge amount of technical debt in their applications – and no easy way to deal with the Y2K bug as well as move to modern Internet technologies – said to hell with it and tossed out their code and modified third party applications from the likes of SAP, Oracle, and Microsoft and changed their businesses to meet that code as much as changed that code (where possible) to meet their business.
They traded control for speed, and are betting the application software providers can add features and provide security that they would have trouble doing in house.
Those deploying net-new AI applications or augmenting their existing applications with AI are facing the same choices, and those running third party applications seem inclined to wait for their software partners to add the AI features. There are those, however, who want to own their AI stack and wrap around these existing applications rather than gut them – and spend more money for more functionality – much as they did with adding web technologies to their ERP, SCM, and CRM suites. Either way, you pay. You either have to invest in software developers and spend time or wait for the software houses to do the work and pay them. The choice and the math is rarely crisply clear.
And for the record, we pay for both OpenAI GPT (Nicole) and Anthropic Claude (Tim) in our household, just so we can see how these things work. An API works just fine for our limited use cases, but I do not believe most enterprises want to pipe their data out to an external model unless they live in the cloud already. They will want to have their AI models close to the software suites they either created and maintain themselves or they bought from third parties.
Meta Platforms flip-flopping on whether to be open source with its models and Nvidia’s aggressively generous terms for its most recent and open sourced Nemotron models as well as the workflow manager for stitching together agentic AI applications are bringing this dilemma top of mind this week.
As we have pointed out before, Nvidia is probably the only company that can afford to be generous and give away the source code, the weights, and the datasets used to train its Nemotron 3 models. The way we see it, at the price you are paying for Nvidia hardware, the hardware damned well should be free – and not coincidentally, just as application software programming was with the IBM System/360 mainframes six decades ago. IBM sent in teams of programmers to teach companies how to do applications, and the code was open source to the users. No one thought to share application code across companies back then because this was a strategic advantage.
Meta Platforms has to be commended for both the hardware it opened up through the Open Compute Project a decade and a half ago as well as for open source systems software such as the React JavaScript generator and the PyTorch AI framework. These were opened for very self-serving reasons – to get programmers to do interfaces and AI the Meta Platforms way – and there is absolutely nothing wrong with enlightened self-interest.
The company kept its initial AI models closed and proprietary, but was open weight (but not open source) with its Llama 3 models launched in February 2023 and expanded in April 2024. With Llama 4, the models got bigger, but they were also not open sourced but rather open weighted. With Musk Spark 1.1, its latest and very different GenAI model launched in April this year, the software was closed, but with the upcoming 1.2 release, Meta Platforms says it will go open weight. The smaller Muse Glimmer model announced this week, with 30 billion parameters and designed to run on a PC with a single GPU, is open weight from the get-go.
Just for fun, I put together a giant table of the open weight and open models. The feeds and speeds of the proprietary models from Anthropic, Google, OpenAI, and Amazon Web Services are not provided to the public. This table shows any model with 30 billion parameters or more. Anything smaller than that is really only suitable for personal devices and limited use as far as I am concerned. I think of 30 billion parameters are the bare number of synapses for thinking, and I like as many parameters in a model as I can get for the hardware that I am restricted to running it on.
One last thing: Many of the smaller models are created by distilling down a larger model, whether it is a dense one or a mixture of experts design. In his super-chummy AI manifesto this week, Mark Zuckerberg, co-founder and chief executive officer of Meta Platforms, said a lot of things about how altruistic the company was going to be as it advanced its ideals of “personal superintelligence,” and how open weight models would be part of that effort for sure, but also that he thought that distillation across models was something we needed to permit here in the US. (I would say that China has already been distilling from US models to train its own.)
“For the US to lead in open source, we will need to rethink our policies in several areas, including distillation and data use in training,” Zuckerberg said. “The ability for models to learn from other models is an important principle of how the open source ecosystem works. All AI models are derived from human knowledge. Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe. This is how the world works, and the US will not be able to lead if we restrict ourselves on this front.”
I wonder what happens to models when they are all averaging against each other. And if that can happen, how will anyone be able to charge for anything other than the underlying hardware supporting inference for a GenAI model?
Advantage, Nvidia.
Then again, if the quality of the closed models can stay better than the open models, they can hang in there and sell API access to anyone and everyone and still license their models for on-premises use cases.
Advantage, better models.