daily briefing

Today in AI

The most important AI developments from the past day. Through October 26, 2026

Today's research haul delivers a bracing reality check for anyone still reflexively reaching for LLMs as the default solution to every problem. Start with the skin cancer finding: a decade-old Inception v3 CNN significantly outperformed frontier models GPT-6 Astra and Claude Fable 5.1 at medical image diagnosis, a stark reminder that "frontier" doesn't mean "fit for purpose" and that specialized architectures often demolish generalist behemoths on actual tasks. Pair that with logistic regression beating LLMs at email routing when labeled data is scarce, and a pattern crystallizes: the AI industry is systematically overpaying for inference, both in dollars and energy. The environmental cost is quantified brutally in another study showing that LLM-based code optimization consumes energy equivalent to 250-2,800 GPT-4.1 queries relative to human-written alternatives—we're trading marginal speed gains for substantial carbon debts that nobody's accounting for on balance sheets.

Yet genuine progress persists in narrower, more deliberate channels. HAD-VLM advances medical AI by preserving numerical continuity in vision-language models, addressing a subtle but critical failure mode where current systems garble quantitative measurements. SemantEq achieves perfect precision on formal verification in Lean 4, quietly advancing the infrastructure that could someday make AI systems themselves provably safe. And KKT-HardNet offers a physics-constrained framework that actually guarantees engineering constraints get satisfied—novel concept—while DexPort tackles the perpetual bottleneck of robot learning: getting enough good manipulation data without million-dollar labs. The through-line? Tools that respect domain structure and measured constraints are outpacing brute-force scale, even as the hype machine keeps selling bigger hammers for every nail.

**Bottom line:** The most important AI advances today came from knowing when *not* to use frontier models—and the energy and accuracy costs of forgetting that lesson are mounting fast.

Stories referenced

Green Code or Costly Code? Evaluating the Energy Costs of Iterative LLM-Based Code Optimizations

LLM-based code optimization produces faster but more energy-intensive code than human-written solutions, with inference energy costs equivalent to 250-2800 GPT-4.1 queries.

arXiv

A cat-belling problem for FDT

Analyzes functional decision theory (FDT) using a cat-belling problem analogy to explore decision-making paradoxes in AI alignment research.

LessWrong AI

Measure Before You Train: Typed Decision Models vs. a Classical Baseline vs. a Generative LLM for Routing a Personal Multilingual Inbox

Research paper compares classical ML, typed decision models, and generative LLMs for email routing, finding logistic regression outperforms LLMs with limited labeled data.

arXiv

Beyond Discrete Tokens: Preserving Numerical Continuity in Medical Vision-Language Models

Researchers developed HAD-VLM, a hybrid autoregressive-diffusion framework that improves quantitative medical image measurement accuracy by preserving numerical continuity.

arXiv

Learning Positive Linear Combinations of ReLUs in Fully Polynomial Time

Researchers present the first fully-polynomial time algorithm for PAC learning positive linear combinations of ReLU activations under Gaussian distribution.

arXiv

DexPort: A Hand-Held Skill Interface for Dexterous Hand Data Collection

Researchers introduce DexPort, a portable hand-held interface that simplifies collection of high-quality dexterous manipulation data for robot learning.

arXiv

A Tutorial on Physics Constrained Machine Learning for Hybrid Modeling, Optimization, and Control

Researchers introduce KKT-HardNet, a physics-constrained ML framework combining neural networks with differential projection for guaranteed constraint satisfaction in engineering applications.

arXiv

A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

A 10-year-old convolutional neural network (Inception v3) outperformed frontier LLMs GPT-6 Astra and Claude Fable 5.1 at diagnosing skin cancer from clinical and dermoscopic images.

Hacker News