PorTAL, Making AI Training Cheap and Portable

If you ask OpenAI or Anthropic investors about the risks that open models pose to their companies, they will simply tell you that the speed of progress will render open models obsolete before they can make an impact, and that these Labs will always offer models most worth using.

But with each passing day and every "open-source is dangerous; it should be regulated" speech from the AI Labs, their convictions soften, and their fears grow.

And my topic of today won't make them feel any better, because we've just made open models much more appealing.

The Claim Isn't As Waterproof As They Realize

Naturally, I strongly disagree.

I've long sustained that the future of enterprise AI is companies training open models in their own data while giving an enthusiastic 'goodbye' to OpenAI and Anthropic, at least for the vast majority of enterprise use cases.

"Open models when you can, OpenAI when you must," one could say.

In hindsight, a no-brainer

There's a Czech proverb that goes "Po bitvě je každý generál." Very deep, right? Jokes aside, the English translation is "After the battle, everyone is a general."

In hindsight, the eventual rise to dominance of open models will feel inevitable. But for the longest time, most people believed open models would never catch up and that the closed Labs, OpenAI, Anthropic, Google DeepMind, or xAI, would run away with it.

For fear of sounding a little arrogant, I always talked very openly about this eventually being true, but I was quite alone on this hill for quite some time.

This didn't make me a genius — God forbid you might think that, because you would be wrong — it was just me looking back at AI's decades-long history and realizing it was the only possible option. Because let me tell you, there were signs.

For instance, a very recent one is Thinking Machine Labs, a star-studded AI research lab packed with ex-OpenAI, Meta, DeepMind, and Anthropic researchers, among others, which bet their entire company's future on open models eventually running the world, to the point that, for now, the only product they have is one that lets you train open models on your data, called Tinker, and one that has been used by Bridgewater Associates, the largest hedge fund in the world, to train internal models that exceed frontier model results by quite the margin on several tasks their investors value.

This is what, when done well, open models give you: frontier-level performance at tens of times (or more) lower cost.

But where's the catch? Besides requiring sophistication on your side, it's also a trade-off: We sacrifice generality.

These fine-tunes make open models great at that one task you focus on, not at all. In broad terms, the frontier models remain superior in most tasks, just not in the one you care about, and can even hinder their performance in other areas not focused on during fine-tuning.

But that's the thing: enterprises don't need their customer support AI agents to be great at suggesting tiramisu recipes.

So even if you could perfectly assume I could have been wrong, it would have been very wise for most of the late-round investors in these Labs to, at the very least, review their intuitions if a group of star researchers, including OpenAI and Anthropic founders, built an AI Lab whose entire bet is that open models would eventually win.

And now, it seems that the world is finally converging on this same idea, with several key AI incumbents being extremely open, almost belligerent, about it.

The Open Comeback

Over the last few days and weeks, several companies upstream and downstream of the closed Labs (customers and suppliers alike), including Palantir and Microsoft, have become increasingly vocal about their pro-open model stance.

Or, in some cases, less positive about open models and actually very negative about these Labs, directly accusing them of very concerning stuff.

For example, both Palantir and Mistral's CEOs directly accused closed Labs (e.g., Anthropic, OpenAI, or Google DeepMind) of 'stealing' their customers' data to build downstream direct competitors.

Stealing might be a strong word, but these Labs are definitely seeing what you're using their models for, and they aren't particularly coy about it, with Anthropic literally publishing research telling the world how its customers use its products.

The clearest example is Claude Design, a direct competitor to Figma, a major customer of Anthropic at the time, and a company that had Anthropic's Chief Product Officer on its board until his resignation three days before the launch of Claude Design. Do with this information what you wish; I have a very clear stance, personally.

Adding to these two, Microsoft and NVIDIA, the latter of which has just partnered with Palantir, have become strong open model champions, the former launching the Frontier Company, a $2.5 billion inititive to promote Microsoft products to its customers in a more hands-on approach with very heavy focus on internalizing AI worklows (codename for using open models) while the latter is the US's current most representative of good open models that compete with Chinese counterparts, with its Nemotron model family.

Naturally, we mustn't be naive; all these companies are pushing self-serving narratives because they are upstream or downstream of the big AI Labs and feel threatened by them, either by these Labs cannibalizing their businesses (e.g., Palantir) or by becoming basic resellers of their models (e.g., Microsoft), or because these Labs are pushing their own chips for many workloads to shake off the dependency on chip designers (e.g., NVIDIA).

All of these companies need open source to win and keep their business protected from a potential 'AI oligopoly' in the making, a particularly bad outcome for all (including us, customers) because that oligopoly can only be built using regulatory capture, as the unit economics clearly favor open models.

But I digress. Importantly, their bias doesn't make them any less correct about this matter. The appeal of lower costs, better governance, and tighter security makes this a no-brainer for enterprises once open models reach a level of capability that warrants adoption, which is exactly what has taken place over the last few months.

But then, why aren't we seeing more proof in the pudding?

The fear of obsolescence

I kid you not when I tell you that I enter every executive or board meeting with "you should have a strategy to migrate most AI workloads to open models. And fast." printed on my head.

However, I often get pushback. The most consistent argument thrown back at me has always been obsolescence, meaning why I would spend a couple thousand dollars fine-tuning an open model if it's going to be obsolete by next week?

If war veterans suffer severe PTSD from explosions, most executives are scared straight of opportunity cost, committing to something that ends up being wasteful, instead of choosing the less contrarian approach. As a corporate executive, you really, really don't want to be that person, because your prestige is on the line; you'd better know what you're doing.

However, for starters, that argument is already not particularly strong when you realize that it's okay to have legacy models for some tasks because these tasks do not need new levels of intelligence.

For the most part, most use cases have "capability requirement ceilings", meaning at one point newer models don't improve what you had already, making "intelligence chasing" as absurd as asking a cowboy to not wear his hat — as you can see, I'm suffering from severe Dutton Ranch fever.

For instance, I continue to use Gemini 3 Flash for many of my enterprise tasks, even though it's more obsolete than a dinosaur daycare, but it's way cheaper and certainly not obsolete for what I use it for.

Furthermore, I've been massively transitioning most of my enterprise work to open models, with the goal of running most of my enterprise workloads fully open.

Most enterprise workflows don't force you to do "intelligence chasing"; you are not required to continue updating your models as better ones come out.

Some people disagree, but let's use Socrates' proof of contradiction to quickly realize how illogical that sounds. Would you run a frontier model to write an email? Do you need Anthropic's Mythos to log a new lead in Salesforce?

That's completely absurd. Intelligence chasing does make sense in some use cases, areas like coding, cybersecurity, or drug discovery, but many — I would argue, most — enterprise use cases don't fall into those categories.

But even then, in many other examples, there's a real opportunity cost, because using open models doesn't come free most of the time. They need work that might still not make them worth it.

And here's where today's research comes in.

Enjoying this piece? Then you'll love my newsletter, which explains AI in first principles and in words you can understand, for those allergic to hype but hungry for knowledge.

Join today.

Subscribe | TheWhiteBox by Nacho de Gregorio The newsletter to stay ahead of the curve in AI

Making Training Cheap and Portable

The problem is that open models require sophistication on the part of enterprises, because the key that makes them 'worth it' is that you can download them for free and actively train them on your data.

It's this process that leads to frontier-level performance on the task at 50 or even 100 times lower cost, like Bridgewater did, but it requires some expertise and effort on your side.

Without that additional training, performance will be much worse than a frontier model, and the point I'm making is that fine-tuning an open model is not a walk in the park.

It requires manual effort to do things like gathering data (preparing the data for training), creating additional synthetic data using other AIs, evaluating results, defining constraints, and more.

And if you aren't using a fine-tuning managed service like Tinker or many others like Unsloth, it can be incredibly tricky to execute.

Training isn't free, not in terms of cost, and certainly not in terms of human capital investment. But now, this new work by Ramp, of all things, an expense management company, proves two things:

1. You can make training immediate,
2. You can make it portable.

In layman's terms, you can amortize training costs by transferring the outcomes of one model to a new, smarter base model without having to write off the training run.

In other words, the last complexity that prevented enterprises from going all-in on open models may have been solved (or at least significantly improved).

But how?

Immediate adapters

The answer is that they've taken one of the most fascinating pieces of research I've ever come across, Text-to-LoRA, by Sakana AI, which I've talked about in the past, and made it portable.

Let me explain.

As discussed, the key to making open models a superior option to closed models, while costing a fraction of the price, is that you can train them, something OpenAI or Anthropic don't allow because their models are closed.

The problem with fine-tuning is that, if you do it naively (a full model retrain), it can incur prohibitive costs, especially considering that the industry moves so fast that new, superior models come out every week, reducing the incentive even further because getting a return on that training run feels very complicated.

An alternative is to use LoRAs (low-rank adapters). The link above goes into more detail, especially the more "mathy" parts, but the idea is that most tasks that a model has to learn are "low rank," meaning only a very small subset of the model's global parameter count has to be modified to learn the task. Hence, the idea is to only train a small portion of the model and leave the rest untouched.

In fact, the real secret is that you don't train the model at all; these adapters are external and can be added or offloaded. When the model has the adapter, it behaves as if it were trained; when you offload the adapter, it reverts back to normal.

Think of this as a plumber's tool. The tool, like a wrench, lets the plumber do plumbing tasks, but when the plumber isn't using it, there are certain tasks it can't do.

And while you need to teach the plumber to use it, you don't need to rewire the plumber's entire brain; it just requires a small, quick learning process to use the wrench. The plumber remains unchanged; the same person, just augmented for the job whenever it's holding the tool.

That is what makes LoRA incredibly effective and a standard for training these days**: it's low-cost, does the job, and can be scaled without needing a new model for everything — just a new adapter.**

However, it still requires a training run and can be expensive relative to the base model's time-to-obsolescence (i.e., your base model might become 'dumb' relative to what's available in the market pretty quickly).

For this, Sakana AI had a great idea called Text-to-LoRA, which uses an AI to generate LoRA conditioned on text.

But what do we mean by that?

In simple terms, it's an AI model that generates these adapters as output instead of actively training them. Instead of having to train the adapter, we generate it, in the same way ChatGPT answers "Belgium" to the question "Which country is going to lose against Spain in the quarterfinals of the soccer World Cup?"

To my US soccer fans, you will be avenged.

For example, you can describe the task "write emails using formal language", and this model automatically generates a LoRA adapter that, when added to a model, makes it write emails always in formal language.

As you may guess, there's obviously a "capability ceiling" for this. You can't just ask this model, "Make my model solve the biggest questions in humankind," and expect it to work. This only works for tasks that have meaningful representation in the system's training data.

Incredible, right? No training run required, fast iteration, and it works. What's not to like?

Well, one problem remained: Text-to-LoRA outputs are not model-agnostic because they are trained alongside the chosen base model. In simple terms, they learn to generate adapters for a particular model.

Therefore, if you train an adapter for a model and a better model emerges, that training run is no longer useful.

And here's where Ramp's PorTAL comes in.

Adaptable and Portable

Cutting straight to the point, they've managed to create a model-agnostic Text-to-LoRA that can port adapters between models.

For example, say you've trained an email-filtering LoRA for your Qwen3.6 35B model… and Qwen4 arrives the next week and is considerably smarter, to the point that it's worth the switch.

Instead of ditching the previous base model and writing off the training, you use PorTAL to transfer the LoRA to that new model. There's some training required, but it's minimal as the largest portion of the PorTAL system is model-agnostic.

In other words, it's still a text-to-adapter model, but one that generates model-agnostic adapters by learning the patterns that make for good adapters independently of the base model used.

And the results are very promising, showing that the system recovers 98% of the per-task LoRA performance on an unseen base model, with most other tests recovering at least 90%.

I know, that's a lot of jargon in one sentence.

In layman's terms, they prove that the PorTAL system can take an 'unseen' model and generate a LoRA with minimal training and cost, matching the effort required to train a LoRA adapter for that model and task from scratch.

With this, not only is fine-tuning cheap, but it's also portable.

Bad for Thee, Not for Me

Think about what this means. It's hard to argue against open models being a superior option to closed models for enterprises if you fine-tune them; you're betting against the very essence of AI to say that.

The problem is that fine-tuning can be tricky, costly, and face steep time-to-obsolescence challenges, making it hard to justify the investment given how long the training run will take and how quickly the model can become obsolete.

But the industry, scared straight by potential AI spending once subscriptions end and customers are pushed to API which much better reflect the real costs of the technology, finally has a very strong reason to create the required abstraction tooling that makes fine-tuning cheap and easy, and PorTAL is just one example of the great appetite there is for open models in the enterprise, something that is incredibly great news for everyone (yes, everyone, including the Hyperscalers) except the closed AI Labs.

But I'll take it further: I have a strong belief that many in this industry would consider very bold:

Much of Frontier Lab's revenue growth today can be explained by customer unsophistication.

Which is to say, I believe a considerable portion of the revenue Anthropic and OpenAI are seeing comes from customers using their models because they don't yet know how to build robust open-model pipelines and workflows.

The math is already there; open models give you greater control, better cost management, and a fully sovereign AI stack. The problem is that enterprise leaders aren't yet aware of this, nor are they operationally capable of addressing it.

But once frontier token prices force companies out of their bubble and into real AI engineering, I believe Anthropic and OpenAI will be in a world of pain.

And before you come at me with "well, then they will just drop prices", restrain yourself because the issue here is not operational costs; it's capital flows. AI Labs could be profitable (they aren't, but could be relatively soon) yet still be massively cash-flow negative.

Besides liquidity (i.e., needing more money), it's this sophistication curve, which I insist is accelerating, that makes these labs so eager to go public as soon as possible.

Once the word is out that things could get murky for them in the enterprise business, investors will find it much harder to underwrite today's valuations for these Labs on the basis of distant, increasingly uncertain revenue growth.

The "own your AI stack" trend is gaining huge momentum at the worst time possible for OpenAI and Anthropic.

I share similar thoughts in a more comprehensive and simplified manner on my LinkedIn (don't worry, no AI-generated content there either). As a reminder, you can also subscribe to my newsletter.