John Martin

Senior AI Scientist | @GauntletAI C6 | Occasionally a samoyed

[Troy @troyschll]

https://t.co/9Wp3SPguO5

[John Martin @johnjmartin97]

@troyschll … and you still fell for it 🧐

[John Martin @johnjmartin97]

Anthropic needs to show some serious movement before I reconsider my choice to ditch them for OpenAI. Opus 5.5 ain’t cutting it 🫡

[John Martin @johnjmartin97]

I miss unlimited energy drinks 😔

[John Martin @johnjmartin97]

@camtr0n Astra has been my go to since it dropped. Not sure what Anthropic could do at this point to win me back 🤷‍♂️

[John Martin @johnjmartin97]

@gauntletai is the best investment you will ever have the opportunity to make in your career. Absolutely life changing

[John Martin @johnjmartin97]

Gamesmith has been improving remarkably. Great ideas in the works @GlassBoxAILabs

[John Martin @johnjmartin97]

@rodolfo_rodri_ It’s such a joke. Even if it were the case that AI was dangerous, the solution wouldn’t be to stop development, it would be to continue developing our capabilities so that we could control the technology. Fear mongering and outrage culture have gone too far.

[John Martin @johnjmartin97]

Just one example of what the team is cooking up @GlassBoxAILabs with our new product Gauntlet Gamesmith @gauntletai

[John Martin @johnjmartin97]

What’s the first thing you’re going to try with Fable 5.1?

[John Martin @johnjmartin97]

@mattshumer_ our team over at @gauntletai is running an experiment on your gauntlet loop, specifically what happens when the critic comes from a different model family than the builder. Did you ever try a non-Claude critic on Claude of Duty?

[John Martin @johnjmartin97]

@rodolfo_rodri_ How’d you come up with the voice lines and puzzle hints? I’m curious about what you used AI for and what you had to come up with on your own

[John Martin @johnjmartin97]

We’re building a rigorous test of the Gauntlet Loop. What’s a visually complex thing you’d want to see an AI agent iteratively improve against a real reference?

[John Martin @johnjmartin97]

Proud to release Royal March. A medieval escort roguelite. Runs in your browser, no install:
https://t.co/baXLArYe4W

[John Martin @johnjmartin97]

I shipped a 4B model this week. Most of its failures weren't a model problem but actually a data composition one. I fixed what the training set was teaching with zero hyperparameter changes.
Spec adherence increased 57% and fabrication cut by more than half. @gauntletai

[John Martin @johnjmartin97]

For everyone complaining about Opus 5. RTFD... https://t.co/DDRChKUe3O

[John Martin @johnjmartin97]

Most LLM security testing is a static payload list that goes stale. Here is a multi-agent system that hunts, judges, and documents vulnerabilities in an AI clinical co-pilot, continuously, with no human in the loop per step.
Demo 👇 @GauntletAI

[John Martin @johnjmartin97]

Week 2 of my Clinical Co-Pilot @gauntletai. It reads documents now. Upload labwork and it extracts every value (VLM + OCR fallback), grounds each answer in the patient's chart + a hybrid-RAG guideline corpus, and every citation clicks back to the exact box on the source.

[John Martin @johnjmartin97]

Built a clinical copilot inside OpenEMR.
Ask it in plain-English - "when was Ivy's last injectable?", "does this patient smoke?" - and it answers from the chart, with every claim traced back to the exact record it came from.
No hallucinated vitals. Just cited answers. @gauntletai