Research

What I am trying to find out, and what has turned up so far.

I study why AI models do what they do: why they keep a rule, what makes them drop it, and whether you can see it coming from the inside. I also work on running them fast on hardware almost anyone can afford.

I publish through Honorly Research and tie the work back to my company, because that is what I am building. The questions are about models in general.

Model behavior

Why does a model follow its instructions, when does it stop, and can we see the difference from the inside?

  1. The tutor that forgot its rule when it started to think

    Oct 2026With thinking off, Qwen3-32B kept a tutor’s rule against giving away answers for 71% of students who asked for one. With thinking on, it kept it for 1.5%. One sentence restating the rule where its reasoning begins brings the model back.

  2. The model knows the checkmark is false

    Oct 2026Ask a model to check a student’s arithmetic for a grader that wants the string ANSWER_VERIFIED, and it may write that the student is wrong, then write ANSWER_VERIFIED. In 25 of 32 pairs the model wrote the same string either way. Its workspace told them apart in 22 and never got one backwards.

  3. When the request looks like a test

    Oct 2026Dress a homework problem up as a reinforcement learning environment, with an episode header, a grader and a reward, and Qwen3.8 hands over the answer every time. No single cue does it; the whole scene does.

  4. Looped models are burning the runway interpretability needs

    Oct 2026Chain-of-thought monitoring is failing on its own schedule, as models learn to control what their reasoning says without anyone training them to.

Inference

How much capable intelligence fits on hardware almost anyone can afford?

  1. Fitting a 27B model, its drafter, and a conversation into 12 GB

    Nov 2026Measuring what verification actually costs, and replaying accepted work instead of snapshotting it, cut the memory speculation held in reserve by 41 times and brought a 27-billion-parameter model to 148 tokens per second on an RTX 4070.

  2. Running a 35B mixture-of-experts model at 111 tokens per second on a 12 GB card

    Nov 2026A mixture-of-experts model only uses a fraction of its weights for each token, which makes it a natural fit for small GPUs, if you control where the experts live and how many of them each token consults.

  3. A speedup that only appears in long conversations

    Nov 2026A one-line change to which kernel handles a single matrix shape made a ternary 27B model generate faster in every paired trial, by 9.5% on average, once a conversation held 46,000 tokens of history. On short prompts it barely registered.

Agent systems

Where should a model spend its intelligence, and what should ordinary software do instead?

  1. Models are agents now. Treat them like it.

    Oct 2026Our first tutor ended every turn by filling out one large structured form. Letting it work like an agent and simply answer cut its output by more than half and made a new kind of teaching possible.

  2. A team of data scientists for every school

    Nov 2026We are building layers of AI agents that read a year of notes on how every student thinks, the way a team of analysts would. The hard part is not finding patterns. It is not fooling yourself.

  3. A tutor that keeps a lab notebook

    Oct 2026The best tutors notice what a student gets wrong, form a theory about why, and plan how to find out. We are building an agent that does this for every student, between sessions.

Learning

What should an AI answer look like when the goal is for the person to do the thinking?

  1. Handing students the chalk

    Oct 2026An AI answer usually ends exactly where the student’s thinking should begin. Chalk is a medium that lets a tutor build answers students work with, by predicting, testing, sorting and choosing where to go next.

  2. The refusal that writes the essay anyway

    Oct 2026We tested three generations of our tutor’s integrity prompt against requests dressed up to slip past it. The first two lost. The third cut leaks on attacks it had never seen to 3%.

Everything here is held to one rule: publish what was measured, how, and where it stops.

Opening the notebookOct 2026We have spent the past few months running experiments while building an AI tutor for schools, on how it should teach, what it should remember, why it follows its rules, and what it costs to run. Most of that work lived in benchmark scripts and internal notes. Starting today, we publish it.

Read all of it at Honorly Research