Deep Dives collects longer analysis of AI products, model behavior, technical changes, and strategic shifts. The hub is designed for readers who want more context than a news brief and need durable explanations for AI decisions.
20 articlesDeep Dives collects longer analysis of AI products, model behavior, technical changes, and strategic shifts. The hub is designed for readers who want more context than a news brief and need durable explanations for AI decisions.
Use this hub to follow the latest related articles, then open individual posts for source-grounded details, context, and practical implications.
MIT researchers have developed PlasmoSniff, a portable chip that detects pneumonia and other lung conditions from exhaled breath after inhaling engineered nanoparticles. The device…
Deep research agents leak private info through web queries. MosaicLeaks exposes the problem and offers a fix that cuts leakage from 34% to under…
After decades of searching, WIMP detectors are hitting the neutrino fog. But physicists aren't giving up—they're getting creative with quantum sensors, Jupiter probes, and…
Researchers used an OpenAI reasoning model to help diagnose rare genetic diseases in children, identifying 18 new diagnoses in cases that had stumped clinicians…
OpenAI and Molecule.one let GPT-5.4 loose on a tough medicinal chemistry reaction. The AI didn't just suggest ideas — it ran experiments, iterated, and…
Google's AMIE conversational AI went head-to-head with primary care physicians in a Nature study on disease management. The results are genuinely impressive, but there's…
Google Research releases a vectorized dataset of fine-scale ecological features like hedgerows and copses across England, using a deep learning framework to turn satellite…
OpenMed built an end-to-end protein-to-mRNA pipeline covering structure prediction, sequence design, and codon optimization. CodonRoBERTa-large-v2 won with perplexity 4.10 and CAI 0.40. They scaled…
Most Arabic LLM benchmarks have quality issues nobody talks about. QIMMA fixes that with a validation pipeline that filters out garbage before it skews…
IBM's VAKRA benchmark exposes how poorly current AI agents handle real-world tool use across 8,000+ APIs. A breakdown of the four task types and…
Google Research and Beth Israel Deaconess tested AMIE, a conversational diagnostic AI, with real patients in a primary care clinic. The results show promise…
Google Research's TurboQuant, QJL, and PolarQuant algorithms promise extreme vector compression for LLMs with zero accuracy loss. Here's how they work and why they…