Science

Today’s best in science and research, discover the breakthroughs making waves now.

Humanity's Last Exam reveals the gap between AI and human intelligence despite rapid progress

AI chatbots interpret 'likely' as 80% probability while humans assume 65%, new study reveals

Harvard Study Shows AI Can Predict 71% of Fund Manager Trades, Raising Questions About Fees

ChatGPT cracks decades-old gluon amplitude puzzle, marking AI's first major theoretical physics win

Humanity's Last Exam reveals the gap between AI and human intelligence despite rapid progress

Researchers created Humanity's Last Exam, a PhD-level AI benchmark with 2,500 questions designed to assess the limits of AI reasoning. Google's Gemini 3 Deep Think achieved the highest score at 48.4%, while human experts score around 90%. Despite progress, experts emphasize this doesn't signal the arrival of Artificial General Intelligence.

2 Sources

AI chatbots interpret 'likely' as 80% probability while humans assume 65%, new study reveals

A new study published in NPJ Complexity reveals that AI chatbots and humans interpret probability words differently, with language models assigning 'likely' to 80% while humans assume 65%. This probability misalignment poses risks in high-stakes fields like healthcare and government policy, where miscommunication about uncertainty could lead to flawed decisions.

2 Sources

Harvard Study Shows AI Can Predict 71% of Fund Manager Trades, Raising Questions About Fees

ChatGPT cracks decades-old gluon amplitude puzzle, marking AI's first major theoretical physics win

MIT Researchers Expose Hidden Biases and Personalities in Large Language Models

AI model predicts bone removal in cochlear implant surgery to enhance surgical planning

Explore Stories by Date

Follow fresh insights sorted by the date they surfaced.

27 Feb 2026

Humanity's Last Exam reveals the gap between AI and human intelligence despite rapid progress

Researchers created Humanity's Last Exam, a PhD-level AI benchmark with 2,500 questions designed to assess the limits of AI reasoning. Google's Gemini 3 Deep Think achieved the highest score at 48.4%, while human experts score around 90%. Despite progress, experts emphasize this doesn't signal the arrival of Artificial General Intelligence.

2 Sources

25 Feb 2026

Harvard Study Shows AI Can Predict 71% of Fund Manager Trades, Raising Questions About Fees

A Harvard-led study reveals that AI can predict 71% of mutual fund trading decisions, suggesting much of active management follows detectable patterns. The research, analyzing data from 1990 to 2023, found that the unpredictable 29% of trades is where genuine alpha lives. The findings challenge the justification for high active-management fees and highlight how automation could reshape the $54 trillion asset management industry.

3 Sources

AI chatbots interpret 'likely' as 80% probability while humans assume 65%, new study reveals

A new study published in NPJ Complexity reveals that AI chatbots and humans interpret probability words differently, with language models assigning 'likely' to 80% while humans assume 65%. This probability misalignment poses risks in high-stakes fields like healthcare and government policy, where miscommunication about uncertainty could lead to flawed decisions.

2 Sources

24 Feb 2026

Guide Labs open-sources Steerling-8B, a new interpretable LLM that traces every token to source

San Francisco startup Guide Labs has open-sourced Steerling-8B, an 8-billion parameter language model built with a novel architecture that makes every generated token traceable to its training data origins. Founded by Julius Adebayo and Aya Abdelsalam Ismail, the company aims to solve AI's black box problem by engineering interpretability directly into the model rather than analyzing it post-hoc.

2 Sources

AI model predicts bone removal in cochlear implant surgery to enhance surgical planning

Researchers from Vanderbilt University and partners developed an AI model that predicts bone removal during cochlear implant surgery. The system analyzes CT scans to forecast mastoidectomy shape, achieving a 0.72 Dice score across 751 cases. This medical imaging technology could support robotic surgical tools and improve patient outcomes.

2 Sources

20 Feb 2026

ChatGPT cracks decades-old gluon amplitude puzzle, marking AI's first major theoretical physics win

ChatGPT solved a complex gluon amplitude problem that had stumped theoretical physicists for decades, proving that certain particle interactions previously thought impossible can actually occur. Working alongside Andrew Strominger and a team of researchers, the AI simplified complex mathematical expressions and generated a proof in just 12 hours—what might have taken humans indefinitely to solve.

2 Sources

AI agents surge in capability but lack safety disclosures, MIT AI Agent Index reveals

MIT researchers analyzed 30 AI agents and found a troubling pattern: while developers eagerly showcase what their systems can do, only 19% disclose formal safety policies and fewer than 10% report external safety evaluations. The study exposes how agentic AI systems that can autonomously perform tasks like booking flights and managing workflows are advancing faster than their safety frameworks.

6 Sources

19 Feb 2026

MIT Researchers Expose Hidden Biases and Personalities in Large Language Models

Researchers from MIT and UC San Diego have developed a method to identify and manipulate over 500 hidden concepts within large language models, including biases, personalities, and moods. The technique uses a Recursive Feature Machine algorithm to steer AI responses, improving performance while also exposing potential vulnerabilities like jailbreaking and hallucinations.

3 Sources

13 Feb 2026

AI Prosthetic Arm Study Finds One-Second Movement Speed Feels Most Natural to Users

Researchers used virtual reality to discover that AI-powered prosthetic arms feel most like part of the body when they move at a moderate pace of about one second. The study tested six different speeds and found that both overly fast and slow movements reduced body ownership and usability, revealing critical design insights for future autonomous prosthetics.

3 Sources

10 Feb 2026

AI Decodes Rules of Mysterious Roman Board Game After 1,700 Years of Silence

A mysterious limestone tablet discovered in the Netherlands has puzzled researchers for over a century. Now, artificial intelligence has cracked the code, revealing it as a blocking game played by Romans—centuries earlier than such games were thought to exist in Europe. The breakthrough demonstrates how AI simulations can unlock secrets of ancient cultures.

7 Sources

09 Feb 2026

MIT's AI Algorithm Unlocks Brainstem Imaging, Revealing Signs of Parkinson's and Brain Injury

Researchers from MIT, Harvard, and Massachusetts General Hospital have developed the BrainStem Bundle Tool (BSBT), an AI-powered software that automatically segments eight distinct nerve bundles in the brainstem using diffusion MRI. The tool revealed distinct patterns in patients with Parkinson's disease, multiple sclerosis, traumatic brain injury, and Alzheimer's, and tracked a coma patient's 7-month recovery.

2 Sources

AI chatbots fail to improve medical advice for patients, Oxford study reveals

A University of Oxford study published in Nature Medicine found that AI chatbots offer no advantage over internet searches when patients seek medical advice. Despite large language models achieving 94.9% accuracy in controlled tests, real-world human-AI interaction in healthcare revealed a troubling gap between AI potential and performance, with patients struggling to provide complete information and receiving inconsistent advice.

14 Sources

06 Feb 2026

New AI method tackles century-old challenge of identifying which dinosaur made which footprints

Researchers developed an AI-powered approach that analyzes eight distinct traits in dinosaur footprints to determine which species left them. The algorithm examined 1,974 footprint silhouettes spanning 150 million years, offering paleontologists an objective tool to solve a challenge that has sparked debate for generations.

2 Sources

University of Michigan's Prima AI model reads brain MRI scans in seconds with 97.5% accuracy

Researchers at University of Michigan developed Prima, an AI model that analyzes brain MRI scans and delivers diagnoses in seconds with up to 97.5% accuracy. Trained on over 220,000 MRI studies, Prima can prioritize emergency cases like strokes and brain hemorrhages, automatically alerting specialists. The system aims to address growing radiology shortages and reduce diagnostic delays across health systems.

4 Sources

03 Feb 2026

Apple researchers unlock faster, more natural-sounding Siri through new speech technology

Apple researchers have published a breakthrough study proposing a new text-to-speech approach that could dramatically improve Siri's performance. The research introduces Acoustic Similarity Groups (ASGs) to accelerate speech token generation while maintaining quality, potentially addressing long-standing complaints about the voice assistant's sluggish response time.

3 Sources

Get all updates into your inbox 💌, Stay Up-to-date for Free 🥳

Your Daily Dose of Curated AI News

Don’t drown in AI news. We cut through the noise - filtering, ranking and summarizing the most important AI news, breakthroughs and research daily. Spend less time searching for the latest in AI and get straight to action.

The Outpost

News

About