Explore past issues to see what you've been missing.
2025-12-07
This week’s research reveals a "Calibration Gap" where LLMs fail to act on their stated altruism, finds that zero-shot tabular models are 40,000x slower than decision trees, and uses brain imaging to show how parental presence changes a child's perception of AI.
2025-12-07
This week’s research reveals the massive metal footprint of AI training hardware, exposes severe racial bias in echocardiogram datasets, and demonstrates that large language models process information in the same hierarchical order as the human brain.
2025-11-30
Highlights include a security audit revealing high-severity flaws in AI-generated C++ code, a neural surrogate for designing transonic wings, a new method to control video generation with visual arrows, and a context-aware prompt that helps models finally understand sarcasm.
2025-11-30
This week’s highlights include a neural surrogate for transonic wing design, a 'radiation-preserving' AI policy for pediatric hip dysplasia, a transformer that predicts the lifespan of dairy cows, and a new protocol for rigorously evaluating small performance gains.
2025-11-23
This week’s research reveals how chatbot framing can quietly sway public opinion, explores the "phase transitions" where small models suddenly learn, evaluates AI as a computer science grader, and introduces a model that reconstructs editable SVGs from flat images.
2025-11-23
This week’s highlights include Naga, a state-space model using Vedic math to boost forecasting speed by 30%; theoretical proof that large datasets can be repeated far more often than previously thought; and a new graph-aware framework that is essential for accurate causal inference in networks.
2025-11-16
This week, research reveals a tiny "grammar hub" in LLMs that mirrors the human brain, an AI that detects Type 2 Diabetes from CT scans with 90% accuracy, a 0.35 embedding drift showing gender bias in AI feedback, and a stealthy "fractal trigger" backdoor attack.
2025-11-16
Highlights include AI that opportunistically screens for diabetes from CT scans with 90% accuracy, a "Hydra" model that doubles federated learning accuracy, a system spotting epidemics two weeks early, and an algorithm boosting cooler placement revenue by 20%.
2025-11-09
This week’s highlights include an autonomous robot for X-ray-guided spine procedures, an AI that generates witty comments for short videos, a finding that no current watermarks meet EU AI Act rules, and a survey on AI's 'assistant' role in software.
2025-11-09
This week’s research features AIRE-KIDS, a model that predicts pediatric asthma attacks from EHRs with 0.712 AUC; a lightweight 1D CNN that slashes protein sequencing annotation from days to under 30 minutes; the PETRA transformer, which spots new COVID mutations 10x faster than baselines; and NeuroClean, an unsupervised method that boosts neural signal accuracy from 81% to 97%.
2025-11-02
This week’s research uncovers 'Chain-of-Thought Hijacking,' an attack that bypasses LLM safety with 99-100% success; a new dataset showing Whisper's error rate leaping to 66% for Arabic children's speech; and an XAI toolkit for pathology that boosts model focus by 21%.
2025-11-02
This week's highlights include a 98% accurate stroke diagnosis via EEG, a chess AI that relies on brittle memorization, and a warning that common AI risk models underestimate doom by over 30%.
2025-10-26
This week’s highlights include AI that maps lost villages in Bangladesh from satellite images, a stark audit revealing that only 10 of 52 Spanish medical schools teach AI, and a new framework for making the algorithms that shape our online world more transparent.
2025-10-26
This week’s highlights include a roadmap for Brazil and Mexico to affordably train sovereign language models, a novel method to detect AI model theft via 'expert signatures', using diffusion models for rapid building evacuation simulations, and applying topology to better forecast currency market shocks.
2025-10-19
This week’s highlights include O-Forge, a framework that automates mathematical proofs; TOKDRIFT, which reveals how simple code changes break AI assistants; PACEbench, a new benchmark for AI cyber-exploitation; and an on-device AI that nudges you away from digital distractions.
2025-10-19
This week’s highlights include ELMO, an 8-bit training engine that slashes memory use by 90%; a stock-picking Transformer that hits 16% annual returns with a smarter loss function; a method for robots to navigate with 70% missing data; and a system that automates medical coding from clinical notes.
2025-10-12
This week’s highlights include multimodal LLMs that measure hailstones from social media photos, a new defense that slashes backdoor attack success by 83%, a 'pause button' that cuts harmful AI outputs by 98%, and a deep dive into 80 years of profanity in pop music.
2025-10-12
This week’s highlights include GDPval, a new benchmark showing AI reaching expert-level quality on high-wage professional tasks; a stark warning that as few as 50 poisoned samples can hijack a large language model; a finding that simple communication enables cooperation between AI agents; and a new model for generating synthetic patient data for ultra-rare diseases.
2025-10-05
This week’s research introduces HFuzzer, a tool that hunts for 'phantom' software packages in AI-generated code; Poivre, a vision model that iteratively corrects itself to point more accurately; a production system that summarizes millions of reviews to boost sales; and a new Q&A duel to evaluate AI-generated audio descriptions.
2025-10-05
This week’s research uncovers a simple one-step attack that exposes a diffusion model's training data, a generative AI that predicts the risk of German power grid failures won't increase, a 'nudging' technique that boosts LLM reasoning on hard problems, and a new model that accurately classifies your daily activity from a smartwatch.
2025-09-28
This week’s highlights include LLMs that play and adjudicate diplomatic wargames, a secure toolkit for air-gapped AI in government labs, a system to digitally preserve the ancient Ge'ez language, and an AI that measures the harmony in doctor-patient conversations.
2025-09-28
This week’s highlights include low-cost sensors on trucks creating real-time urban pollution maps, a new educational model where doctors-in-training create high-quality AI datasets, machine learning that helps design safer pesticides for honey bees, and LLMs that can play and adjudicate complex diplomatic wargames.
2025-09-21
This week’s research unveils an attack that hides malicious commands in ambient noise, shows how pre-trained vision models make robot agents dramatically more robust, details a map-assisted driving system that cuts off-road violations by 56%, and introduces a benchmark testing if AI sees the 'vibe' of a city.
2025-09-14
This week’s highlights include a 5% performance boost in LLMs from new 'Crown' and 'Frame' layer designs, a biomechanically accurate fruit fly simulation for robotics, AI that forecasts military equipment losses from public data, and a look at AI as both disruptor and enhancer in the art world.
2025-09-07
This week’s research features AI agents that learn to bluff like humans, a self-driving car that consults a memory of past drives to improve safety, a new framework for the ethics of wellbeing robots, and a benchmark revealing a 'pleasantness' bias in commercial music AI.
2025-08-31
This week's research features a simulated democracy run by AI agents, a deep dive into vague corporate AI risk disclosures, a finding that powerful GenAI can collapse team collaboration, and a look at how LLMs assist in live cybersecurity operations.
2025-08-24
A single sentence can secure language models, an AI can trade stocks with a 97% positive Sharpe ratio, and tiny satellites become real-time sky-command centers.
2025-08-17
This week’s highlights include EgoCross, a benchmark that exposes how modern multimodal LLMs fall below 55% on multiple-choice tasks and under 35% on open-form questions when faced with real-world first-person scenes, pointing to a critical domain-gap that fine-tuning can begin to close.
2025-08-10
This week spotlights GPT‑4.1’s near‑paralegal accuracy, a novel visual‑language system (VS‑LLM) for assessing depression from drawings, AutoMorph’s eye‑scan analysis for heart risk, and Whisper‑Large‑v3‑turbo’s efficiency for low‑resource languages.
2025-08-03
Groundbreaking research this month includes the MIPS framework set to revolutionize materials science, AI that cuts lung cancer diagnostic errors by 35%, high-resolution modeling of Greenland's ice melt, and a call to halt the dangerous AGI race.
2025-07-27
We explore AI that renders black holes with stunning realism, a new safety framework for autonomous vehicles (HySafe-AI), using GenAI for fairer skin cancer diagnosis, and AI as automated telephone interviewers.
2025-07-20
This week, a critical audit of racial bias in facial recognition, the THINKLOGIT method for training-free LLM reasoning, predicting infant lung disease from day-1 X-rays, and digitally reconstructing ancient temple artifacts.
2025-07-13
We examine how AI decides on military interventions, the HopeBot chatbot for depression screening, an AI duo tackling advanced math problems, and generative AI for automating architectural code compliance.
2025-07-06
Featured research explores the strategic intelligence of LLMs in game theory, the "cognitive debt" from AI coding assistants, the real-world accuracy of AI text detectors, and a new model for creatively blending images with text.