Javier Marin @jamarinval

"AI safety has a friendly-fire problem. The technology that proves where AI text came from is weakening the technology that checks whether it is true." @jamarinval presents an incisive analysis of the link between LLM watermarks and hallucinations towardsdatascience.com/hallucinations…

Read the full paper. Congratulations for the work. For me, the proposal/consequence separation is the cleanest part of the design. I found that the world acting as the latent space for invention is almost literally Kauffman's adjacent possible theory (TAP), with every installed artifact enlarging what the next agent can build. I'm curious about the control parameter. In your atoms-in-a-box analogy, superconductivity only shows up in a narrow regime. Is there an equivalent here, a disturbance rate or artifact density beyond which the world stops being memory and becomes noise? Your Fig. 15A phase portrait seems to hint at it.

"A threshold tells you an answer is bad, and when it is wrong it costs you. A list of words only tells you where to look, and when it is wrong it costs you a glance. One check takes five minutes and gets skipped. The other takes thirty seconds and gets done." @jamarinval shares an incisive reflection on AI hallucinations and how we try to prevent them. towardsdatascience.com/ten-is-not-a-h…

@SvenUrbanSci @TDataScience Indeed. This was the idea behind the post.

In a thoughtful analysis, @jamarinval reflects on the links between LLM watermarks, hallucinations, and the slippery notion of authenticity in the age of AI. towardsdatascience.com/hallucinations…

"A bank runs a RAG system behind its customer-service chat. A user asks about an invoice. The retrieved document says the total due is 10,000 dollars. The chatbot tells the client that the total is 1,000 dollars." @jamarinval unpacks a RAG pain point that seems simple to solve — until you try to. towardsdatascience.com/ten-is-not-a-h…

linkedin.com/posts/javierma…

Asked Claude.ai for a number that isn't in the doc. It refused. The grounding check flagged it anyway, not because the answer was wrong, but because it didn't come from the source. Provenance, not truth.

How can you ensure that a model grading its own work does it in a robust and cost-effective manner? @jamarinval digs into an emerging conundrum of agentic-AI workflows. towardsdatascience.com/design-loops-n…

The daily "lying" is the case for not trusting the model's own account of itself. If it asserts false things confidently, self-checks inherit the same failure - you need a check anchored to the source, outside the model. Detecting ungrounded answers is tractable; getting the model to self-certify isn't.

The legibility-erosion vs unfaithful cut is the one that matters - but from the outside you often can't tell which regime you're in without ground truth you don't have. Unfaithfulness scores only cover cases where a faithfulness test is constructible. So if safety leans on the CoT, it leans on a proxy you can only audit where you already know the answer - which argues for anchoring the check outside the trace.

Every AI headline is really an energy headline now - and storage is where the grid absorbs the shock. The hard part isn't installing more batteries. It's operating them: predicting failures before they cascade and optimizing dispatch in real time, on assets that degrade in ways a static model never captures. That's exactly where a physics-informed digital twin earns its keep. I'll be on a panel on this at the IEEE PES General Meeting in Montreal - "AI-powered Digital Twins for Grid-Scale Energy Storage" (Wed July 22, 1pm). If you work on storage, grid ops, or AI-driven load, come say hi.

This is the strongest case I've seen for keeping the "don't believe this" outside the model. If a document that explicitly flags a claim as false still gets absorbed as true, the correction can't live in the model's own judgment - grounding has to be checked against the source, not delegated to the model's read of it.

One hard part of model forensics: the "why" often isn't localized. When you compare how a model processes a correct vs. an incorrect answer, the discriminative signal is spread across the trajectory - not sitting in any single layer or component. The "why" may need attribution across the whole path, not one circuit. Wrote it up here: arxiv.org/abs/2603.13259

I just published Do You Know What You Want? medium.com/p/do-you-know-…

It’s clear that current autoregressive models do not equal human intelligence. Everybody knows that. But I do not understand why so much people acts like “well, AI are only good pattern matching systems”. We should remember that our universe started with hydrogen + helium and end up writing poetry.

Can Neural Networks "innovate" during its training? towardsdatascience.com/i-measured-neu…

@chrisalbon @goodfellow_ian I think many of us owe a lot to @goodfellow_ian

Just benchmarked Claude, ChatGPT, Gemini & Grok against each other. Here's what I learned that public leaderboards don't show. The irony of AI benchmarking is this: benchmarks have driven massive progress in the field, but they're almost useless for choosing the right model for your business. We ran the a propietary framework across 4 leading providers, measuring: ✓ Consistency (behavioral reliability across multiple trials) ✓ Performance (quality on real-world business tasks) ✓ Coordination effects (multi-step workflow performance) The results? There's no winner. Or rather, there are four different winners: 🔵 Claude excels at consistency (0.94 score) — if you need auditability, compliance, and predictable behavior, Claude is your model. Regulatory bodies will love it. Cost premium is worth it for decision-critical workflows. 🟡 Grok maximizes performance (88/100 score) — pure output quality. Creative problem-solving, complex analysis, "give me your best answer" tasks. Trades consistency for ceiling height. 🟢 Gemini balances both — neither specialized nor weak. Great if you have diverse workloads and want to minimize switching risk. 🔴 ChatGPT holds the middle — reliable across domains, broad ecosystem, trusted integrations. Your vendor choice matters way less than testing on your actual data. We found that cost-quality tradeoffs, edge case handling, and degradation under load are completely specific to each use case. If you're deployed to production with AI: spend 10-20 hours self-benchmarking. It's ~0.1% of annual AI infrastructure cost and one of the highest-ROI investments you can make. The benchmark that matters most? The one you run yourself. 👉 For more details about the pilot test you can DM me.