State of AI · Public evidence through July 22, 2026
Hedge funds are building the research machine before the finance model.
Man AHL leads the disclosed agentic workflow. CFM owns the clearest narrow fine-tuning result. AQR-affiliated researchers have published the deepest language-model training program. Most competitors are still exposing the build layer through hiring.
01 · Competitive position
The public AI journey has two independent axes.
Moving right means more of the investment-research loop is disclosed. Moving up means stronger evidence that finance-specific language-model weights changed. Neither axis measures returns.
Position is a State of AI evidence rubric, not a vendor score. Sources: first-party firm disclosures, technical papers, job listings, named practitioner interviews, and a dated public-profile audit; retrieved through July 25, 2026.
02 · The category leaders
There is no single leader because the market is solving four different problems.
Man AHL
AlphaGPT and AlphaTrend disclose the fullest hypothesis-to-code-to-evaluation loop.
No Man-trained finance LM disclosed.CFM
Compact financial NER fine-tuning improved reported F1 from 87.0% to 93.4%.
Extraction task, not investment reasoning.AQR-affiliated
Point-in-time models trained from scratch, with monthly checkpoints and LoRA instruction tuning.
Portfolio deployment is not public.Citadel
AI Assistant is described as used by nearly all equities investors.
“Trained on” does not establish changed weights.03 · Evidence stack
The firms cluster by what they have chosen to disclose.
A filled cell means public evidence exists. It does not mean a competitor lacks an undisclosed capability.
| Firm | Talent | AI runtime | Agentic research | Named deployment | LM weights changed |
|---|
04 · People and models
Public recruiting data reveals direction, not team size.
No audited firm publishes a complete count of AI employees. The defensible measures are a public minimum of named practitioners and the number of distinct AI role families observed. Current-role checks also prevent alumni evidence from being presented as current ownership.
Public AI talent signal
Named practitioners plus distinct AI/ML role families in the evidence set. The worker audit now covers research engineering, MLOps/HPC, investment engineering, GenAI integration, and quant software. This is a disclosure footprint, not headcount.
Verified finance-LM weight modification
Public minimum. “0 disclosed” is not evidence that no private model exists.
The strategic read
The moat is the testable research environment.
The strongest public precedent combines frontier reasoning, proprietary data and tools, code execution, institutional methodology, and hard evaluation gates. Smaller adapted models enter where a repeated task produces a measurable quality or cost advantage. The evidence does not support beginning with a giant proprietary “finance GPT.”
Build the research harness and trace every hypothesis, code artifact, data lineage record, rejection, and human override. Fine-tune only after those traces reveal a stable repeated task.
Open the full audit, source boundaries, and verification ledger →