A comprehensive personal memory system, always available for additions, deletions and modifications.
You give it anything you want: notes, documents, conversations and more. It can consolidate and organize all of your information for you, or propose a way to organize it for your approval and proceed according to your instructions. Your information is always available, and you can chat with it: answers draw upon everything that you have approved. It can also surface insights and connections you might not see on your own.
Independent Research: Latent State Transfer Between Language Models
Aug 2026 – present
When can AI models efficiently and effectively share information without using text?
When AI tools work together, they pass information to each other as text. Researchers are now exploring whether models could skip the words and share their internal state directly. I tested this between two different models. It worked: the second model correctly answered new kinds of questions about information it never saw as text. However, it was not more efficient. Plain text did the same job at about 1/1,000 the size, so I stopped that line of research. I am now actively investigating other approaches. The research findings cover the whole investigation so far, including what failed.
Fig. 1 The transfer experiment. Both models stay frozen; only the small map is trained.
I designed and ran this project on my own: the question, the experiments, the evaluation rules, and the decision to stop. All models were open-weight, run and trained locally.
The question was simple to state. When AI agents work together, they hand off context as text: one writes a summary or JSON, the next reads it. Could a model hand over its internal state directly, to a different model or to a fresh copy of itself after a reset? And would that keep more information, or cost less?
What I found
It works, in a narrow sense. A small learned map moved a seven-field record from one model’s internals into a different model’s internals. The second model then answered kinds of questions the map had never been trained on: 0.90–0.92 on a held-out equality task and 0.78 on a held-out mapping task, matching what it achieves with its own representation of the real record.
It was not better than text. JSON produced the same decision accuracy (1.000 on 300 unseen cases) in 77 bytes. The internal route used 80,640 bytes and about a second of extra compute.
Memory across a context reset worked on synthetic data and did not work on my real coding-agent sessions.
Decisions that mattered
Knowing when to stop. In early October I gave the internal-state approach its best possible chance against plain text. Even then, it needed a message about 1,000 times larger to reach the same accuracy. On October 5, I stopped that line of work and wrote down a simple rule for which approach to use: plain text when the information is easy to write down, reuse of what the model has already read when it can go back to the original, and internal-state transfer only for information that is hard to put into words.
Auditing my own benchmark. The real-session memory results looked weak, so on October 29 I audited the benchmark itself and found it could be solved perfectly by exact string matching. The question already contained the key to its own answer. The honest conclusion is that this benchmark can’t show whether latent session memory works, and the write-up says so.
How I ran it
pass bars written into an experiment’s plan, and a miss stayed a miss
final test sets sealed and opened once
every transfer claim backed by a swap test plus zero and random controls
failures and corrections kept in the record, relabelled rather than deleted
Those rules overturned several of my own conclusions, including a leaked evaluation, a readout that wasn’t computing, and a result that turned out to be a lucky initialisation.
LoRA adapters; small one-layer transformer maps between models
Fig. 2 At equal decision accuracy, the hidden-state message was about 1,000 times larger than JSON. Log scale.
Athletics, Health & Human Performance
Elite junior cycling and the science of training, rehabilitation and performance.
My background in elite junior cycling sparked a deep interest in human physiology, biomechanics, and kinesiology. I was accepted to UCLA, but at the same time was offered a position at a clinic in Silicon Valley. I graduated early from high school and moved to Silicon Valley to pursue both my interests: the human body and technology. I became a Certified Personal Trainer (CPT) and bike fitter, and, at the physical therapy clinic, I’m developing apps to help clinicians work more efficiently with their patients and creating videos showing patients how to do their exercises properly and engage the right muscles. I give presentations at bike shops on exercises that help prevent injury and improve performance, and I run wind tunnel testing using AI-based tools. I also intern with a nutritional supplement company, reviewing clinical research to identify evidence-based ingredients for product formulation and helping build automated tools that guide customers to the right products.
Certified Personal Trainer (CPT) · bike fitter
4th · USA Cycling National Time Trial Championships (2024)
1st · Barentin Road Race, Normandy, France (2024)
Newbury Park HS MTB · VeloSport · Hot Tubes
Full detail
Cycling
I started racing mountain bikes and moved to regional and national road racing. Along the way I attended a USA Cycling talent identification camp, where I won the road-climb fitness test, raced in Europe, and was invited to join Hot Tubes, the junior development team of the INEOS Grenadiers.
Teams
Hot Tubes, INEOS Grenadiers junior development team
VeloSport Junior Elite Team
Newbury Park High School Mountain Bike Varsity Team
Racing taught me perseverance: how to keep working through things that are hard. In September 2025, bilateral hip arthroscopy took me off the bike, and the recovery is when I turned to software and AI.
Health and human performance
Certified Personal Trainer (CPT) and bike fitter.
Physical therapy clinic: personal training, rehabilitation coaching and bike fitting with clients.
Nutritional supplement internship: reviewing clinical research to identify evidence-based ingredients for product formulation.