I like an interesting problem. I like it even more when solving it makes something faster, more dependable, less expensive, or more useful to someone. That is the direction I want Drantech Labs to take.
These are proposed directions, not announcements of products or completed research. I want to start small enough to measure the result, then decide whether the idea deserves more time.
The frontier is moving
On October 6, 2026, OpenAI released a collection of mathematical results from an internal model. When checked on October 11, its repository listed 719 manuscripts in 372 families of related work, with different levels of verification. Its October 7 revision history records three withdrawals and other corrections. “Over 700 problems solved” is not quite the same claim.
That is still a significant development. It changes how I would choose a problem. Before spending months on something, I want to check the current literature, inspect what has actually been established, and understand the assumptions. A claimed proof, a checked formal statement, and a practical implementation answer different questions.
Private AI that earns its place
Running a model locally is a starting point. I am more interested in whether an assistant can find the right information, respect who is allowed to see it, show where an answer came from, and stay useful after the first week.
A good first investigation would compare a narrowly scoped assistant with ordinary search on the same permitted documents. Wrong answers, unauthorized retrieval, response time, and the work required to maintain it all count. If that produces a dependable improvement, there may be a useful product or managed service behind it.
Efficiency that survives the whole measurement
I want to investigate how much useful work we can get from hardware people already own. vLLM documents quantization and hardware-dependent support; that is an existing baseline to learn from and test against.
Compression belongs here too. Zstandard already provides dictionary training for related small records. A new approach has to justify its extra complexity. I would count compressed data, dictionaries, model weights, decoding time, and memory, then test on data that was not used to tune the method. A smaller file alone does not tell the whole story.
Automation that handles a bad day
Home Assistant already supports fully local voice pipelines. I do not want to call that unsolved. The part worth investigating is how a complete installation behaves when the network fails, a device disappears, a sensor is stale, or a request is unclear.
A useful system should have a predictable fallback and a manual way out. That is something we can test on a bench before asking someone to depend on it at home or at work.
Evidence people can check
AI makes it easier to generate ideas, code, and mathematical arguments. I want the evidence to be easier to inspect too. Lean provides tools for formal proofs and verified software, while Sigstore provides artifact signing and transparency infrastructure.
The interesting work is connecting a claim to the exact inputs, code, environment, and checks behind it. A signature establishes something about an artifact and its signer; it does not make a scientific claim true. I would start with a small reproducible result, rather than promising a system that can verify any research.
Longer-term questions
NIST has published its first post-quantum cryptography standards and migration guidance. The practical opportunity I want to explore is helping identify dependencies and test changes in existing systems, using reviewed algorithms.
I also want room for cooling, quiet compute, sensors, and repairable physical products. Those connect to my experience with custom computers and water cooling. Materials research remains interesting, but it needs appropriate measurement equipment and collaborators. I would rather begin with a specific thermal problem we can measure properly.
Mathematics, symbolic methods, automata, and information theory remain part of the toolkit. Their value is in what they help explain or improve. I am not treating a new headline as proof that an entire field is finished.
What makes an idea worth continuing?
Someone should have a reason to care about the result. There should be a fair baseline, a test that could show the idea is wrong, and a practical way to repeat the work. Commercial potential needs customer conversations and a willingness to pay; technical novelty does not establish demand.
That leaves plenty to explore. The goal is to find a useful question, work on it carefully, and share what holds up when it is ready.
Source status checked October 11, 2026. These priorities are my proposed direction for the lab, not validated market forecasts or claims of novel results.