Things are going to get forking ridiculous. I think there’s already enough results in that I can make a prediction. And that’s this…
These new tools we have are going to get a lot better and also somewhat worse at the same time. Like a lot of you, I’m having great fun using some of these new tools. Sometimes as a consumer. Sometimes for digital product craft. And sometimes for personal experiments. Insofar as those of us working in Product should be paying attention to trends and adjusting our efforts, we should be watching this emerging area of a patchwork modelverse of LLM dark matter. (Apologies for mixing metaphors so early, but I’m just really an LLM myself with drastically limited training data.)
Why should we care from a Digital Product Management perspective? As we look at some of the To Do items below, it’s not necessarily a product person’s job to handle the specific task, but we should understand the risks and make sure at the very least that table stakes items are part of the work plan. Because we need to ensure what we build is not only performant to its fit-for-purpose build, but also be confident that go-to-market claims are trustworthy. And I think it’s possible a lot of our products may end up sitting on shelves next to dain bramaged garbage. Trust is still going to matter. Actually, more so than ever. But we’re going to have to prove quality and branding will still matter.
Just What the Fork is Happening Now?
Here’s the short answer: There’s so many of these things now, both larger and smaller models, and they’re getting forked, fine-tuned and RAGged, along with more folks choosing to train their own plus an increasing disparity of training data for various foundation models. (Not to mention being bolted onto, into or through anything that even sounds like the word “agentic.”) Add to this that these things are not deterministic anyway. Meanwhile, there’s a parade of new claims of how just yesterday Model Gen Y beat Model Gen X at some new LLM math quality criteria. But then, this morning, Model Gen Whatever beat Y at crafting the best elevator music ever. To say that AI projects are popping up like weeds may be a perfectly parallel metaphor, given how so often removing them is poorly managed.
How is this Happening?
The short answer is poor model stewardship in some cases and just some animals escaping from the zoo into the wild. In the best controlled conditions right now it’s challenging to keep models usable, explainable, and improvable, not just now, but over time. Top shops are trying to do that, along with a growing industry of providers claiming to help with such things.
But here’s the thing… models will fork. And drift. And since they’ll be all over the place, that drift will come in a wide variety of forms, and it’s likely going to speed up partly due to forking. That is, the folks who fork won’t always go back to update from the source. Those who integrate with major models maybe won’t face this particular issue. But everyone who calculates that it may cost less to run their own models is another story. Part of the sell for some foundational models is that they’re open source. Great. Even relatively low-skilled tech folks can bolt these on to products. And yet, similar to what I’d written about Agent Rot for agentic flows, how many will properly maintain their creations? I’ve been guilty of this myself. At least, briefly. When DeepSeek came out, I installed it on an old Linux box and tested using it for a hobby site I’ve got. I figured why spend tokens when I can get this free model? Well, the answer ended up being I didn’t like DeepSeek output, so I went back to using ChatGPT.
Why Will Models Drift: High Level
Here’s the high level areas for model drift and stewardship challenges.
- Base model usage (from vendors or open-source)
- Fine-tuning (your own domain/data)
- RAG (Retrieval-Augmented Generation pipelines)
- Forks (open-source model divergence)
These areas are across contexts; from well-known providers like OpenAI/Anthropic APIs to small LLMs run locally with custom training.
Why Will Models Drift: Details
Data Drift (Inputs Changing Over Time)
- User behavior evolves; query language or tone shifts.
- Content formats, APIs, or sensors feeding the system change.
- RAG pipelines may pull from outdated, irrelevant, or incomplete sources.
- Fine-tuned models may be trained on data that becomes obsolete.
- Example: Developers reported that GitHub Copilot suggests deprecated syntax or API methods: See Why GitHub Copilot is Not Updated.
Label Drift (Definitions of Correctness Shift)
- What’s considered “relevant,” “appropriate,” or “correct” changes over time (e.g., social norms, spam detection criteria).
- Fine-tuned models can hardcode outdated or biased labeling assumptions.
- RAG ground truth may also evolve, requiring updated gold standards.
- Example: Content moderation and misinformation detection models for COVID-19. See: As humans go home, Facebook and YouTube face a coronavirus crisis
Concept Drift (Real-World Target Shifts)
- The underlying meaning of what you’re trying to model changes, e.g. fraud, threats, or customer intents evolve.
- Fine-tuning can lock in assumptions from an earlier era.
- Without live RAG content refresh, the model answers stale problems.
- Example: Joint Detection of Fraud and Concept Drift in Online Conversations with LLM-Assisted Judgment
User Population Drift
- Different users start using the system (new languages, locations, expertise levels).
- Fine-tuned models may not generalize well outside the original user cohort.
- RAG retrieval may bias toward dominant or majority queries if not carefully balanced.
- Examples: On the generalization of language models from in-context learning and finetuning: a controlled study, Understanding the effects of language-specific class imbalance in multilingual fine-tuning
Model Staleness
- Base models become outdated as newer, smarter versions appear.
- Fine-tuned weights do not “age gracefully” as the base model evolves.
- Forked models diverge and miss performance, safety, and capability improvements.
- Embedding models used in RAG degrade in effectiveness if not updated.
- Examples: Fighting Redundancy and Model Decay with Embeddings, MUSCLE: A Model Update Strategy for Compatible LLM Evolution
Retrieval Drift (RAG-specific)
- Indexed documents in your vector store may grow stale, incorrect, or untrustworthy.
- Embedding models may no longer reflect the LLM’s semantic space.
- Retrieval quality may degrade if ranking or filtering logic is not regularly reviewed.
- Content duplication, decay, or bias can skew context windows.
- Examples: Understanding the Retrieval-Augmented Generation (RAG) Pipeline, Measuring Embedding Drift, Knowledge Drift — The Silent AI Killer in RAG models, Long-Term Maintenance Challenges for RAG Systems
Infrastructure & Toolchain Drift
- Tokenizers, libraries, or attention mechanisms change over time.
- Your fork or custom model may become incompatible with standard RAG pipelines or adapters.
- Hardware (e.g., GPUs) or deployment environments affect reproducibility and performance.
- LLM APIs may change behavior silently over time unless pinned to a specific version.
- Example: (Why) Is My Prompt Getting Worse?
Evaluation & Testing Drift
- Benchmarks used to evaluate model performance become outdated.
- Newer risks or failure cases aren’t covered in legacy test sets.
- Evaluation scripts may no longer reflect the actual task setup (e.g., prompt structure or system messages).
Safety & Governance Drift
- Regulatory frameworks evolve (e.g., EU AI Act, privacy laws).
- What was safe/legal/responsible 12 months ago may no longer be acceptable.
- Forked or fine-tuned models may bypass safety filters unless intentionally maintained.
- RAG documents may introduce unmoderated or harmful content.
Feedback Loop Breakdown
- Fine-tuned models and RAG systems that don’t incorporate user feedback drift silently.
- Lack of reinforcement learning or logging means you don’t know when things go wrong.
- Inconsistent or unstructured feedback prevents targeted improvements.
Integration & Contextual Drift
- The surrounding product (chatbot, search UI, assistant) evolves, but the model doesn’t adapt.
- Fine-tuning or RAG may rely on assumptions about inputs/prompts that no longer hold true.
- Prompting strategies or formatting conventions change across time or teams.
What Can We Do as Consumers?
As consumers? Not much. Wait for some of the big kabooms. Let the lawyers sort it out. In many ways it comes down to the same as things ever were. You have to decide what brands to trust. And just like any advice, continue to think critically. There’s a reason people get second opinions on medical issues, even if their doctor is top notch. We still need to use some judgment. Now, there will be times this may be challenging to impossible in some scenarios. There will be increasing instances of AI use that are agentic or automated. That is, there are things we may leave in the hands of AI for when we’re not looking. For those we’ll just have to pay attention as best we can and make sure we’re getting the outcomes we want and that there’s limits for anything safety critical. E.g., there will be a difference between letting your streaming service pick the next show to autoplay and relying on an AI to recommend a treatment plan after reviewing your medical history. As well, there are some things we simply will not be able to directly control as some of these systems become more deeply embedded; essentially becoming ambient parts of the technological fabric of society. And of course, we’re subject to others’ use of such tools from credit worthiness scores to medical diagnosis and many more areas to quickly come.
What Can We Do as Product People and Marketers?
Take more care with the whole Minimal Viable Product (MVP) thing. Or Minimum Viable Test, or whatever the latest Flavor of the Minimum Viable Whatever might be. Quality will be challenging to judge with a lot of these tools and how they’re integrated, at least in some environments. But you can probably sort out some test rubrics if you start with the Why in the first place. If you took some time to determine what the values are you’re trying to add for your customers, what it is you need to produce and test for should flow out of that.
Traditional machine learning models, like those detecting tumors or fraud, rely on clear testing frameworks. Model cards document their purpose and performance, enabling consistent accuracy scores through standardized tests. For example, medical scan tests can yield reliable percent-correct metrics. But generative LLMs, with their open-ended probabilistic outputs, are harder to pin down, risking inconsistent results in chatbots or content tools. Unreliable LLMs can erode customer trust or invite regulatory scrutiny, making robust testing a strategic priority. Product leaders should invest in custom evaluation frameworks to ensure AI delivers value and protects brands from drift-driven failures.
Rubrics for GPTs are harder than for more static, deterministic products. Both to develop and to test. And they’ll likely drift over time. Here’s a checklist for what to strive for:
- Version Control: Base models, fine-tuning code, training data and labels, hyperparameters. If you’re building your own, tag the releases the same as the majors.
- Reproducibility: Can you retrain or audit the model? What are the dependencies? Save your training logs and evaluation metrics.
- Evaluation & Testing: Do continuous evaluation on task-specific benchmarks. Try to test for adversarial and edge cases. There’s performance metrics for accuracy, precision, recall, F1 and more. Use monitoring tools like Prometheus and Grafana or Evidently, or the many others that are increasingly available. (See: A Journey Into Machine Learning Observability with Prometheus and Grafana, Part I)
- Monitor for Drift: Is the model maintaining some baseline accuracy? What about latency and cost? Most importantly, feedback from real-world usage.
- Documentation: Make sure you’re clear on the model’s purpose and limitations, where the datasets came from, (both provenance and any licenses), and if there’s ethical considerations.
- Access Control & Auditability: Who can re-train or fine-tune the model? Who can deploy it?
- Compliance & Governance: Align with any risk management frameworks. Keep up with changes to these. Use model cards and evaluation cards. Document choices affecting fairness, privacy or safety.
- Lifecycle Management: Consider when to retire outdated models, or retrain due to data, domain or regulatory shifts. Make sure you continue to have reliable data pipelines. You probably took great care to set them up in the first place. Are they still OK?
Wrapping Up
So yeah. Things are going to get forked up. Not everywhere, not all at once, but enough to matter. The age of neat, centralized language models is already giving way to a fragmented ecosystem of semi-maintained clones, fine-tuned Frankenmodels, and RAG-patched guess engines duct-taped into products with minimal oversight. (This last part is just an opinion. But it feels that way from what I see everywhere from job postings to product and AI forum discussions.)
As product people and builders, our job isn’t just to bolt the latest model into the stack and call it done. It’s to ask the hard questions: Is this reliable? Is it traceable? Will it make us look like idiots six months from now? Will it put us in court? Will it just fail a customer or hurt a customer?
We won’t stop the drift. But we can design for it. Build with versioning. Monitor for failure. Document what the thing is supposed to do. And don’t outsource trust to a model you can’t explain.
Because when the model gets forked, your product will still get judged and you’ll be on the hook for any bad outcomes.