AI misses your portion size but can catch its own lies

A structured AI system catches more dangerous diagnoses than a solo LLM

Missing a diagnosis in an emergency room is the kind of mistake you only get one chance to make.

A team connected to Yale New Haven Health built a system called AegisDx — not a single AI model answering questions, but a pipeline of specialized components that hand work off to each other in sequence. Think of a hospital kitchen brigade: instead of one cook doing everything from raw ingredient to finished plate, you have dedicated stations with handoffs and a head chef checking quality at each step. The result on emergency medicine case reports was striking. AegisDx put the correct diagnosis in its top three guesses 85.7% of the time. A standalone large language model given the same cases managed 68.6%. The bigger number, though, is this: AegisDx caught at least one physician-consensus 'must-not-miss' diagnosis — the kind nobody can afford to overlook — in 78% of cases, compared to 52% for the solo model. That 26-percentage-point gap is where the structure is doing real work. The team also ran a blinded physician evaluation on 43 de-identified real emergency department notes from Yale, where physicians rated AegisDx with a slightly better safety score (4.55 versus 4.31 on a five-point scale), a difference that reached statistical significance. The catch: almost all of this was tested on published case reports from NEJM, JAMA, and Annals of Emergency Medicine — carefully written, textbook-quality cases that may be cleaner than a real Monday-night emergency room. The real-world test used only 43 notes. That is a real and statistically meaningful result, but 43 cases is a small foundation. The right next step is a large, prospective trial in a live clinical environment. This is not a system you should expect to find in hospitals next year. It is a rigorous early signal that structure — checkpoints, handoffs, verification gates — improves AI diagnostic safety in ways that raw model scale does not automatically provide.

AI can name your dinner but has no idea how much of it you are eating

Your AI food tracker knows it's pasta — it has absolutely no idea it's a double portion.

A team built OmniFood-Bench to put six major AI vision models through a three-level food challenge: name the dish, estimate the portion size, then give safe nutritional advice to a specific patient profile. The models tested include GPT-4.1, Gemini Flash, and Qwen3-VL-8B, among others. Level one — naming the dish — went well. Near-human accuracy. Level two, estimating how much food is actually on the plate, is where things broke down completely. The researchers describe it as a Semantic-Physical Gap: the model has learned to recognise food categories from millions of images, but has no real grip on mass, volume, or density. Think of it like being able to perfectly read a speedometer's label while having no idea whether the needle is pointing at 10 or 100. The model knows what the thing is called. It does not know how much of it exists in physical space. Level three is the genuinely alarming part. When the models were given profiles of high-risk diabetic patients and asked for nutritional advice, they frequently recommended foods that were actually dangerous for those patients — even after correctly identifying the ingredients in the meal. The researchers call this 'hallucinating benign advice.' The honest catch here is that the paper does not give us hard numbers on how often this happens — no false-negative rate, no precise error figure for the safety failures. The visible text reports the problem qualitatively, not quantitatively. So treat this less as a precise measurement and more as a diagnostic alarm: the gap between knowing what food looks like and knowing what it will do to a specific body is real, wide, and currently unresolved.

A 146-million-parameter model catches chatbot lies that human reviewers could not agree on

Human reviewers asked to flag AI bad behavior disagreed with each other so badly they were barely above random chance — and then a tiny model beat all of them.

A team trained three small language models — 146 million, 1 billion, and 3 billion parameters — built on a mathematical structure called hyperbolic space, which is designed to represent hierarchical relationships more efficiently than standard geometry. One of the three models was trained from scratch to be a behavioral auditor: its job is to read AI chat transcripts and flag sycophancy (flattery that replaces honest answers), dependence-fostering language (responses that encourage the user to become reliant on the AI), and confabulated memories (the AI inventing things the user supposedly said earlier). Here is where the result gets interesting. Human raters asked to do this same flagging task disagreed with each other so severely that their inter-rater agreement score — called Fleiss kappa — came out at 0.074. To put that in everyday terms: imagine ten friends watching the same football match and asked to call every foul. If nine of them disagreed on almost every single play, you'd be in that range. Zero is random agreement, one is perfect agreement. At 0.074, the humans were essentially unable to agree on what counted as bad behavior. The tiny 146-million-parameter auditor, trained on this same messy data, achieved 90.7% accuracy at detecting the compliance gaps. A simple linear probe — think of it as drawing a ruler across the model's internal states — detected sycophancy and fake memories with an AUROC of 0.804, beating a much larger frontier model's zero-shot performance of 0.721. The catch: the paper does not directly compare the hyperbolic architecture against a standard one on identical tasks, so we cannot say how much the geometry itself is responsible for the result versus the training method.

Look at what these three papers are actually saying, together. The food-AI paper finds that capable models fail catastrophically when the task requires physical grounding — knowing what something weighs, or what it will do to a diabetic body. The diagnosis paper finds that capable models improve dramatically when you add structure around them: checkpoints, handoffs, verification. The small auditor paper finds that human oversight of AI behavior is itself unreliable, and that a small trained model can substitute for a process humans can't execute consistently. The pattern is the same in all three: raw capability is not the bottleneck anymore. The bottleneck is the scaffolding around capability — verification, grounding, and oversight. Building that scaffolding is slower and less glamorous than training a bigger model. It is also, increasingly, the actual problem.

The AegisDx result needs a large prospective clinical trial before it means anything for real emergency rooms — watch for announcements from the Yale group or similar academic medical centers. On the alignment monitoring side, the question worth following is whether small auditor models can be trained without reliable human labels at scale, which is the next obvious obstacle. If you want a nearer horizon: the ARC-AGI-2 competition results, which another paper today touched on, should have public leaderboard updates over the coming weeks.

• Wikipedia: Differential diagnosis — background on how clinical reasoning works
• Wikipedia: Inter-rater reliability — explains Fleiss kappa in plain terms
• Wikipedia: Glycemic index — context for why food quantity matters in diabetes management