Frontier Research
I like to understand how things work and use that to build cool things. Sometimes, that means treehouses for my daughters. And sometimes, that means new AI capabilities.
In modern AI (over the past ~10 years), I’ve mostly focused on four things: reliability, world models, human-AI augmentation and agents. Given the nature of my work (e.g. adversarial market-making at $100B/yr scale at Opendoor), I couldn’t publish most of it. It’s been fun publishing again as an indie researcher over the past several months — 3 published, 2 under review, 1 accepted (TMLR).
Reliability
Neural Diversity Regularizes Hallucinations in Language Models
Basic Question: What if LLM hallucinations are caused by noise (not lack of knowledge)?
Status: Accepted, Transactions in Machine Learning Research (2026)
Link: https://arxiv.org/abs/2510.20690
Although LLMs have demonstrated remarkable progress in gaining capability, LLM hallucination (especially without tool-calling) has been quite stubborn. Sometimes, if a problem seems surprisingly hard despite really smart people working on it, it’s because we’re asking the wrong question.
Borrowing from portfolio theory (tl;dr: diversifying risks reduces total risk), we re-interpret a model’s parallel reasoning streams as a portfolio and find them badly under-diversified, so their errors line up instead of canceling. De-correlating them reduces hallucination up to 25.6% at no cost to capability and produces the first provable bounds on hallucination in ensembled LMs, which predicts 94.3% of the real-world variance we see.
Upshot: Neural diversity is a third axis to improve AI reliability, alongside data & parameters.
Multi-Head Attention is a Multi-Player Game
Basic Question: What’s the price of ignoring intra-model collaboration?
Status: Under Review
Link: https://arxiv.org/abs/2602.00861
Modern transformers are actually internally multi-agent — attention heads compete and coordinate — yet we train them as if they were monolithic entities.
Borrowing from algorithmic game theory (tl;dr: selfish players can settle into equilibria arbitrarily worse than the global optimum), we show cross-entropy training induces an implicit game among heads and gradient descent lands on Nash equilibria with unpriced externalities. We (i) prove that the price of this anarchy manifests as head redundancy and hallucination and (ii) demonstrate that explicitly managing this anarchy reduces hallucination by up to 18% without knowledge cost.
Upshot: Attention is all you need, but coordination is what you lack.
Reliability Scales Inversely: Bigger Models Compound Mistakes Faster
Basic Question: How does reliability scale?
Status: Under Review
Link: <forthcoming>
We keep assuming bigger models will be trustworthier. This paper shows the opposite: reliability scales inversely.
As models get bigger, they get more fragile. In long-form responses, their answers start truer but degrade faster — capability and reliability are separate axes and scale drives them in opposite directions. In particular, we show as models scale, the knowledge gap (capability) improves by 2-7x, yet knowledge degradation (reliability) worsens by 3–39x. The culprit is a hidden, auto-regressive risk regime: the model blurts out one low-odds word, treats its guess as settled fact, and builds on it. If you squeeze down that risk, hallucinations drop by 35-74%.
Upshot: Bigger models are more fragile.
World Models
Spatiotemporal World Models (2023)
Basic Question: What do world models buy that predictions can’t?
Status: Deployed at Opendoor (NASDAQ: OPEN) — $100B/yr scale
Link: <proprietary>
How do you make $100B in offers on O(million) homes in <1 minute? Naively, it’s just prediction, right? Estimate the home’s value, then offer that. But, a binding cash offer isn’t a point estimate, it’s a standing option the seller exercises against you. (Zillow launched a prediction engine, lost O($1B) at steady-state and exited iBuying.)
We built an adversary-robust world model: physics-informed temporal dynamics, a hierarchical causal world model over space, active learning on a multi-million-dollar budget, fusing five modalities into one spatiotemporal field enabling counterfactual reasoning under different rates, spreads, ops capacities and risk constraints. It beat Zillow’s AVM and 10+ SOTA models by >10% MAPE on a 30K holdout and helped lead to Opendoor’s first post-COVID profitable quarter in Q2 2025.
Upshot: Market-making at $100B/yr scale requires simulation-capable world models.
Human-AI Augmentation
Cost Centers → Capabilities (2024)
Basic Question: What if your compliance cost is actually your training signal?
Status: Deployed at Opendoor (NASDAQ: OPEN) — regulated, $15B/yr scale
Link: <proprietary>
Compliance required every final offer to be independently human-underwritten — a 100-person team that looks like pure cost. But the disagreement everyone throws away as “QA” is the most valuable signal in the building.
Borrowing from active learning, we made the UX the instrument: triage by uncertainty so humans take the hard, high-stakes, adversely-selected cases, and capture every override as a labeled preference delta — without anchoring them. Overrides become labels, confidence rises, escalation and cost/offer fall, and the audit trail makes compliance stronger. A single budget can buy training signal and regulatory defensibility at once.
Upshot: Augmentation » automation.
Agents
TalentWorks (2017)
Basic Question: What if the bottleneck to trustworthy agents isn’t capability but testability?
Status: Operational at TalentWorks to 2019 — autonomous, LM-grounded agents in production
Link: <proprietary>
TalentWorks developed and operated autonomous agents at production scale that applied to real jobs for real people — years before “agentic AI” had a name.
Borrowing from simulation-based verification (tl;dr: for a stochastic system you don’t check a single run, you check the distribution of behaviors across many counterfactual worlds), we made evaluation a hardened product: a harness that replayed each agent across synthetic job applications and scored its tail reliability, not its median or average. Done right, even primitive (BERT, spaCy, etc.)-based agents worked wonders.
Upshot: An agent is only as trustworthy as its tail.