Research
The aim is to evaluate superintelligence that is not confined to human language, and to check safety through more than one measure. The reading list is the stack behind that aim. The projects below are the public work so far.
One measure reads behavior from the inside of a system, not only from what it says. The public testbeds so far freeze Qwen2.5 and train either one new token embedding or a LoRA adapter, then ask whether the change behaves like a word. A result on a 0.5B testbed is not a result about superintelligence.
The other measure is evaluation built from tasks and situations that are too involved to settle with a single answer in ordinary language. Accuracy, recall, and precision@k stay inside the run that produced them. A mock provider is not a model result.
An agent-native tool for collecting terms, senses, and evidence spans that stay attached to their sources. Recipes cover neologisms, cultural vocabulary, and specialist jargon. Jobs can resume. Grounding is audited. Exports include CSV, JSONL, Parquet, and annotation files. Model-proposed meanings stay suggestions. Real extraction uses a key, a local model, or approved Georgia Tech compute, and the default connection is an offline rehearsal.
github.com/BobbY-24/quarryTwo repositories ask how little of a frozen model has to move before a new behavior shows up. Both use Qwen2.5-0.5B-Instruct as a testbed. Neither treats that scale as a stand-in for larger models.
Neologism embedding learning. One new token, 896 trainable parameters, optimized with a DPO-style objective while the rest of Qwen stays frozen and checksummed. The full runs, a matched LoRA baseline, and the interpretability analyses were driven from Colab. A 1.5B within-family replication has been run. The 3B run and cross-family replication have not.
CommunityQwen. LoRA fine-tuning on an Apple M3, so the model explains slang in context. It does not try to make the model speak in slang. About 1.08 million adapter parameters are trained, 0.22% of the model with adapters. Held-out scores are lexical and behavioral proxies. They do not establish semantic accuracy. Adapters stay local and are not in the repository.
A due-diligence desk for a customer, a vendor, or an acquisition, built in Cursor. The interface is Flutter. The job service is FastAPI, PostgreSQL, and Celery. A memo cites only evidence that run retrieved, and a negated phrase does not count as support. A thin margin holds the decision instead of guessing. David never pays, signs, orders, or contacts the other party. The shipped demo is synthetic and offline, with no model calls.
Decision support only. It is not legal, credit, or investment advice.
github.com/BobbY-24/DavidOne repository, two instruments. The first compares extraction and judgment pipelines on 50 frozen web sources, half gaming and half cybersecurity, and keeps a separate runner for five evidence conditions: term only, usage text, an image, text plus image, and a distractor image. No live run of those five conditions has been made. The bundled items are fictional, and the default scorer is an offline mock whose scores are meaningless on purpose. Completed provider tables are aggregate ledgers. They are not an invoice, and they are not precision or recall: no gold set is claimed there.
The second instrument is nine offline loaders for game media — VPT, MineRL, CS:GO, Terraria, ProAct, StarCraft images, Slippi, Dota 2 visuals, and Age of Empires II — plus annotation scaffolding, episode preparation, and scoring. Demo scores check the plumbing. They are not a measure of a model, and they are not gameplay.
github.com/BobbY-24/llm-eval-ablation-multimodal