Science

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

Large language models achieve under 1% accuracy at basic multiplication, new study reveals

DeepSeek unveils mHC architecture to slash AI model training costs and boost efficiency

Consciousness May Require Physical Brain Computation, Not Just Abstract Code

AI-Generated Faces Fool Even Super Recognizers, But Five-Minute Training Boosts Detection

Large language models achieve under 1% accuracy at basic multiplication, new study reveals

Despite handling complex coding tasks, large language models fail at four-digit multiplication, achieving less than 1% accuracy with standard training. Researchers from University of Chicago, MIT, Harvard, and Google DeepMind discovered the culprit: models can't store and retrieve intermediate computations. But a specialized Implicit Chain of Thought method achieved 100% accuracy by teaching models to internalize reasoning processes.

2 Sources

DeepSeek unveils mHC architecture to slash AI model training costs and boost efficiency

Chinese AI startup DeepSeek has released a research paper detailing Manifold-Constrained Hyper-Connections (mHC), a new training architecture that could make advanced AI development accessible to smaller developers. The framework addresses signal degradation in neural networks while reducing memory overhead by over 90% compared to existing methods, potentially reshaping how the industry approaches model scalability.

6 Sources

Consciousness May Require Physical Brain Computation, Not Just Abstract Code

AI-Generated Faces Fool Even Super Recognizers, But Five-Minute Training Boosts Detection

AI Tool Predicts Underfeeding Risk in ICU Patients on Ventilators, Enabling Early Intervention

New Open-Source Tool RNACOREX Maps Gene Regulation Networks to Predict Cancer Survival

Explore Stories by Date

Follow fresh insights sorted by the date they surfaced.

02 Jan 2026

DeepSeek unveils mHC architecture to slash AI model training costs and boost efficiency

Chinese AI startup DeepSeek has released a research paper detailing Manifold-Constrained Hyper-Connections (mHC), a new training architecture that could make advanced AI development accessible to smaller developers. The framework addresses signal degradation in neural networks while reducing memory overhead by over 90% compared to existing methods, potentially reshaping how the industry approaches model scalability.

6 Sources

30 Dec 2025

Large language models achieve under 1% accuracy at basic multiplication, new study reveals

Despite handling complex coding tasks, large language models fail at four-digit multiplication, achieving less than 1% accuracy with standard training. Researchers from University of Chicago, MIT, Harvard, and Google DeepMind discovered the culprit: models can't store and retrieve intermediate computations. But a specialized Implicit Chain of Thought method achieved 100% accuracy by teaching models to internalize reasoning processes.

2 Sources

27 Dec 2025

AI-Generated Faces Fool Even Super Recognizers, But Five-Minute Training Boosts Detection

Artificial intelligence has advanced to the point where even elite facial recognition experts struggle to identify AI-generated faces. Research published in Royal Society Open Science reveals that super recognizers perform no better than chance at spotting fakes, while typical recognizers mistake AI faces for real ones 70% of the time. However, a brief five-minute training session on common rendering errors dramatically improves detection accuracy.

3 Sources

24 Dec 2025

Consciousness May Require Physical Brain Computation, Not Just Abstract Code

A new theoretical framework proposes biological computationalism as a third path between computational functionalism and biological naturalism. Researchers argue that consciousness cannot be reduced to abstract information processing because brain computation is inseparable from its physical, hybrid, and energy-constrained dynamics. This challenges assumptions about whether digital AI can truly recreate conscious experience.

2 Sources

23 Dec 2025

AI Tool Predicts Underfeeding Risk in ICU Patients on Ventilators, Enabling Early Intervention

Researchers at Mount Sinai developed NutriSighT, an AI tool that predicts which critically ill ICU patients on ventilators face underfeeding risk during their first week of care. The study, published in Nature Communications, found that 41-53% of patients were underfed by day three, with the system updating predictions every four hours to help clinicians intervene earlier.

2 Sources

20 Dec 2025

AI image generators converge on just 12 visual clichés, study reveals creative limitations

A study published in Patterns journal reveals AI image generators repeatedly default to the same 12 generic visual motifs when generating images autonomously. Researchers paired AI models in a visual telephone experiment, finding that despite diverse starting prompts, systems like Stable Diffusion XL consistently produced Eurocentric themes like Gothic cathedrals and Parisian nightscapes—what they call 'visual elevator music.'

3 Sources

18 Dec 2025

US Department of Energy enlists 24 tech giants for Genesis Mission to accelerate AI research

The US Department of Energy has signed agreements with 24 major tech companies including OpenAI, Google DeepMind, Nvidia, Microsoft, and Anthropic for the Genesis Mission. This national AI initiative aims to create an integrated platform that combines federal scientific data, national laboratory supercomputers, and private cloud capacity to accelerate scientific discovery across energy, security, and research domains.

7 Sources

16 Dec 2025

AI identifies single protein interaction that blocks virus entry into cells before infection

Washington State University researchers used artificial intelligence and molecular simulations to identify a critical amino acid interaction that stops herpes viruses from entering cells. By introducing a targeted mutation to this single interaction among thousands, they successfully blocked viral fusion, pointing toward a new direction for antiviral therapies that intervene before infection begins.

3 Sources

14 Dec 2025

Stanford team builds first commercial 3D chip at US foundry with 4x AI performance gains

A Stanford-led research team has manufactured the first monolithic 3D integrated circuit at a commercial US foundry, showing four-fold performance improvements over conventional flat chips. The prototype, built at SkyWater Technology, vertically stacks memory and logic using carbon nanotube transistors and could eventually deliver up to 1,000-fold gains in energy efficiency for future AI systems.

2 Sources

12 Dec 2025

AI Water Consumption Sparks Debate as New Study Reveals Environmental Impact Estimates

A new study estimates AI could consume up to 765 billion liters of water in 2025, matching global bottled water consumption. But a lack of transparency from tech companies makes accurate measurement difficult, sparking debate between researchers and industry advocates over AI's true environmental impact.

5 Sources

09 Dec 2025

AI Co-Pilot for Bionic Hands Transforms How Amputees Control Prosthetics with Intuitive Grasping

University of Utah researchers developed an AI co-pilot for bionic hands that dramatically improves control for amputees. The system uses proximity and pressure sensors combined with artificial intelligence to enable natural, intuitive grasping. Success rates jumped from 10-20% to 80-90% in tests involving fragile objects, while significantly reducing the cognitive burden on users.

4 Sources

26,000 AI researchers admit they can't explain how models work as interpretability crisis deepens

At this year's NeurIPS conference, a record 26,000 AI researchers confronted an unsettling reality: no one fully understands how today's most advanced AI systems actually work. As companies like Google and OpenAI pursue diverging approaches to AI interpretability, the field grapples with the black box problem and an evaluation crisis that threatens trust in increasingly powerful systems.

2 Sources

08 Dec 2025

AI models achieve breakthrough accuracy in diagnosing brain tumors from MRI scans

Two groundbreaking studies reveal AI's potential to transform brain tumor diagnosis. York University researchers developed an AI model achieving over 85% accuracy in distinguishing tumor progression from radiation necrosis, while Thomas Jefferson University's automated machine learning system reached 97% accuracy in differentiating pituitary macroadenomas from parasellar meningiomas using preoperative MRI scans.

2 Sources

04 Dec 2025

AI Chatbots Sway Voters More Effectively Than Traditional Political Ads, New Studies Reveal

Two landmark studies published in Nature and Science reveal that AI chatbots can shift voter opinions by up to 15 percentage points—far exceeding the impact of conventional political advertising. Researchers tested nearly 80,000 participants across multiple countries and found that chatbots persuade through factual information, but often spread misinformation, particularly when advocating for right-leaning candidates.

18 Sources

03 Dec 2025

OpenAI trains AI models to confess when they lie, cheat, or hallucinate in new safety experiment

OpenAI is testing a novel approach to AI safety by training models to produce 'confessions'—secondary outputs where they admit to misbehavior like hallucination or rule-breaking. The experimental technique rewards models solely for honesty, not performance, and has reduced undetected failures to 4.4% in controlled tests. While promising, researchers caution that confessions don't prevent bad behavior—they only flag it after the fact.

8 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