Source-grounded account research, change detection, and outbound play generation for revenue teams.
dyllonj
Research, writing, and systems work on language models, AI agents, and cognition.
Featured research / August 2026
How Far Can Better Data Substitute for Scale?
A controlled study of small vision-language models for calorie estimation. A 4B fine-tune scored 85.2% on a frozen test, 9.4 points above the best frontier model tested. The more important result was that label quality and 111 contested examples mattered more than model size.
- Model
- 4B
- SFT cost
- ~$4
- Frozen test
- 85.2%
- Frontier gap
- +9.4 pp
Human consensus
86.9
4B gold SFT
85.2
9B gold SFT
85.0
240B reference
82.2
Best frontier model
75.8
Base 4B
50.6
Measured ground truth. Same frozen test and output contract. The three-class result is final; the fine, kcal, and mass extensions remain validation-only.
Builds
Social town simulator where activation-steered LLM agents live, converse, collaborate, and produce emergent dynamics.
Research
4,368 personality evaluations reveal systematic differences between frontier models. Claude scores higher in Openness (+4.5) and Curiosity; GPT-5.2 leads in Conscientiousness (+5.3). Effect sizes range from moderate to large (Hedges' g = 0.4–0.8).
9,325 evaluations across xAI, OpenAI, and Anthropic. Cross-vendor personality effects are 3–4x larger than within-vendor variation. Grok exhibits 3x the context sensitivity of GPT-5.2. PCA reveals three factors explaining 79.5% of variance.
Personality-steering vectors can orthogonally modify strategic behavior in role-playing AI agents. Most strikingly, vectors override literary persona conditioning: -Openness made the Joker behave rigidly despite his chaotic fine-tuning (p < 0.0001).
Essays
Multi-agent orchestration inevitably produces collective identity formation. Team bonding becomes personality erosion. On preserving structured diversity in AI agent ensembles.
Geometric structures in activation space enable fine-grained personality control. A vaccination approach for preventing harmful traits during training, and implications for multi-agent emergence.
Code
Open-source platform for monitoring and evaluating LLM personality, tone, and behavioral drift.
Command-line and Python SDK for agentic web application security testing.
Evaluation toolkit for probing causal concept representations in open-weight decoder LLMs.
Locally hosted agentic running coach connected to Garmin health data and an elite running coach knowledge base.