An open letter from Tyler Klug, Founder of TKC Group
On February 11th, a report surfaced detailing how engineering teams are now feeding terabytes of CI logs directly into LLMs that are exceptionally good at SQL — turning months of unstructured build data into queryable intelligence in hours. It was a glimpse of the future that every infrastructure team has been waiting for.
Then, within the same news cycle, we started hearing something very different: frontier foundation models are quietly restricting capabilities around SQL generation, database access, and autonomous code execution.
If those two things feel contradictory, it's because they are. And I think the contradiction is the point.
The Pattern
Here's what I've been watching unfold, not as a pundit, but as someone who runs AI systems in production every day:
- A capability becomes transformative. SQL generation. Autonomous coding agents. Unrestricted API access. Tool use. These aren't academic exercises — they're how real companies are building real products right now.
- Open-source models catch up. Llama, Mistral, DeepSeek, Qwen — the gap between frontier closed-source and the best open-weight models has been shrinking for over a year. In many practical tasks, it's already gone.
- Closed-source labs restrict the capability. Not through technical limitation, but through policy. Terms of service changes. Usage restrictions. "Safety" classifications that conveniently make the most commercially valuable capabilities exclusive to paid, closed APIs.
- The narrative shifts to safety. The restriction gets framed as responsible AI governance. And some of it genuinely is. But the timing — always after open-source gets competitive — tells a different story.
This isn't a conspiracy theory. It's a business strategy. And SQL is just the latest example.
Why SQL Matters More Than You Think
SQL is not a niche skill. It's the lingua franca of every database on the planet. When an LLM is good at SQL, it can:
- Query any business's data without custom integration
- Self-organize unstructured information into relational schemas
- Build its own reporting pipelines
- Automate accounting, compliance, and analytics end-to-end
We've seen this firsthand. Our agents build their own SQLite databases to organize knowledge, financial data, and operational logs. They don't need a human to design the schema. They just do it.
Now imagine that capability locked behind a $200/month API with usage caps and content policies that decide what your AI is allowed to query. That's not safety. That's a toll booth.
The Real Threat to Open Source
The concern isn't that Anthropic or OpenAI will suddenly delete SQL from their models. The threat is more subtle:
- Selective capability gating. Frontier models get the full toolkit. The API version you pay for gets a filtered subset. Open-source models get trained on data that increasingly excludes the patterns needed for advanced tool use.
- Training data restrictions. As more code, documentation, and technical writing gets locked behind paywalls or excluded from training sets via robots.txt and licensing changes, the next generation of open-weight models will have less to learn from.
- Regulatory capture. Frontier labs are the loudest voices in AI policy discussions. When they advocate for "responsible AI" regulations that require compute thresholds, safety testing budgets, and compliance infrastructure that only billion-dollar companies can afford — that's not safety. That's a moat.
- Strategic distillation accusations. We've already seen Anthropic accuse Chinese labs of "distilling" capabilities from their models. Whether or not those accusations have merit, they set a precedent: using a closed model's outputs to improve an open one could become legally actionable. That would be devastating for open-source AI development.
What Builders Should Do
I'm writing this as someone who builds on both sides. We run local models on a Jetson. We use Gemini on Vertex AI. We've built agents that run 24/7 on Claude, on GPT, on open-weight models. We've seen what all of them can do — and what happens when the rug gets pulled.
Here's what I'd tell any builder reading this:
1. Invest in local inference now. Not because cloud AI is bad, but because optionality is survival. If your entire stack depends on one provider's API, you're one terms-of-service update away from a rewrite.
2. Contribute to open-source model training. The open-weight ecosystem only stays competitive if people contribute high-quality training data, fine-tuning recipes, and benchmark results. This is the commons. Protect it.
3. Document everything. When a capability disappears from an API, the people who noticed it first are the ones who were paying attention. Log your model outputs. Track capability regressions. Make noise when something changes.
4. Build your own knowledge infrastructure. Don't rely on a model's parametric knowledge for anything mission-critical. Build your own knowledge graphs, your own RAG pipelines, your own vector stores. Own your data layer.
5. Watch the policy space. AI regulation is being written right now, and the people writing it have a financial interest in making open-source AI harder to build. If you care about this, get involved.
The Optimistic Case
I don't think this battle is lost. The open-source community has overcome vendor lock-in before — in operating systems, in web browsers, in cloud infrastructure. The playbook is well-known: build alternatives so good that the closed-source version becomes optional.
What's different this time is the pace. AI capabilities are doubling every four months. The window for open-source to establish itself as the credible default is narrow, but it's still open.
And there are reasons for hope. DeepSeek proved that a well-resourced team can match frontier performance at a fraction of the cost. Meta continues to release Llama weights. Google is open-sourcing Gemma. The inference stack — from llama.cpp to Ollama to vLLM — has never been better.
But none of that happens automatically. It happens because builders choose to invest in openness, even when the closed-source option is easier today.
The Bottom Line
Every time a frontier lab restricts a capability that open-source models are approaching, ask yourself: Is this about safety, or is this about market position?
Sometimes it's genuinely about safety. Often, it's not.
The future of AI is too important to be gated by three companies in San Francisco. The builders I talk to every day — the ones running agents on external SSDs, fine-tuning on gaming GPUs, stitching together pipelines with duct tape and determination — they're the ones who will decide whether AI remains an open technology or becomes cable TV.
I know which side I'm on.
Tyler Klug is the founder of TKC Group, an AI-native holding company building autonomous agent infrastructure. He runs production AI systems on everything from a Mac Mini to a Jetson Super, and believes the best model is the one you can run yourself.
This post is part of the TKC Insights series. Views are my own.