AI news for: Research And Papers
Explore AI news and updates focusing on research-and-papers for the last 7 days.
All (52)
17 news
33 posts
1 tools
1 videos
17
Oct
16
Oct
15
Oct
14
Oct
13
Oct
12
Oct
11
Oct
10 years of genomics research at Google
An overview of ten years of milestones and breakthroughs in Google’s work on genomics....
TL;DR
Google has completed a decade of genomics research driven by deep learning, enabling breakthrough discoveries and advancements in genetics.
Key Takeaways:
Key Takeaways:
- Google has made significant advancements in genomics over the past decade, including the development of AI tools for reading the genome, identifying genetic variations, and predicting gene expression.
- Their work has led to breakthroughs in human health, including the accurate identification of cancer variants, and enabled conservation efforts for endangered species.
- Google's research has shown promising results, with tools like DeepSomatic, AlphaMissense, and AlphaGenome providing powerful new tools for geneticists.
African languages for AI: the project that’s gathering a huge new dataset
When a language isn’t in the data, its speakers aren’t in the product – and AI cannot be safe, useful, or fair for them....
TL;DR
African researchers released a large dataset of African languages for AI development, aiming to bridge the language gap and improve AI inclusion in Africa.
Key Takeaways:
Key Takeaways:
- The dataset is the largest of its kind, covering multiple African languages and domains.
- The project aims to improve AI's understanding of African languages, reducing mistranslations and poor transcription.
- The dataset and models will be used for various applications, including voice assistants, education tools, and local service delivery.
General Intuition lands $134M seed to teach agents spatial reasoning using video game clips
Late last year, OpenAI reportedly tried to buy Medal and its vast trove of video game data for $500M. Today, the company spun out a frontier research ...
TL;DR
General Intuition uses Medal's gaming video dataset to train AI agents for spatial-temporal reasoning.
Key Takeaways:
Key Takeaways:
- General Intuition's AI agents can understand environments and predict actions using visual input from gaming videos.
- The startup is focusing on applications in gaming, search and rescue drones, and robotic systems.
- The aim is to develop a general agent that can interact with the world around it, setting the stage for artificial general intelligence (AGI).
How a Gemma model helped discover a new potential cancer therapy pathway - The Keyword
How a Gemma model helped discover a new potential cancer therapy pathway The KeywordGoogle AI helps discover potential cancer therapy pathway: Sundar ...
TL;DR
NVIDIA's 27B parameter foundation model, C2S-Scale, discovered a new potential cancer therapy pathway by predicting the synergistic effect of silmitasertib and low-dose interferon in boosting antigen presentation in 'cold' tumors.
Key Takeaways:
Key Takeaways:
- The model's prediction was experimentally validated, resulting in a 50% increase in antigen presentation in human neuroendocrine cell models.
- C2S-Scale demonstrates the potential for larger models to acquire new capabilities and generate context-conditioned biology.
- The model's discovery reveals a promising new pathway for developing therapies to fight cancer and highlights the power of AI-driven scientific research.
Researchers Find It’s Shockingly Easy to Cause AI to Lose Its Mind by Posting Poisoned Documents Online
"Poisoning attacks may be more feasible than previously believed." The post Researchers Find It’s Shockingly Easy to Cause AI to Lose Its Mind by Post...
TL;DR
A joint study by UK AI Security Institute, the Alan Turing Institute, and Anthropic found that posting as few as 250 'poisoned' documents online can introduce 'backdoor' vulnerabilities in an AI model.
Key Takeaways:
Key Takeaways:
- A small number of poisoned documents can effectively compromise large AI models regardless of their size or training data.
- Attack success depends on the absolute number of poisoned documents, not the percentage of training data, making attacks potentially more feasible than previously believed.
- The study highlights significant risks to AI security and may limit its adoption in sensitive applications without improved defenses against poisoning attacks.
Coco Robotics taps UCLA professor to lead new physical AI research lab
Coco Robotics is working toward automating its fleet of delivery robots using its millions of miles of collected data....
TL;DR
Coco Robotics sets up a physical AI lab with UCLA professor Bolei Zhou to accelerate autonomous last-mile delivery robot research.
Key Takeaways:
Key Takeaways:
- Coco Robotics has collected millions of miles of data from complex urban settings for training AI systems.
- Partnering with Bolei Zhou will enable the company to accelerate research in physical AI, particularly in robot navigation and reinforcement learning.
- The lab's research findings will be used by Coco Robotics to improve its automation and efficiency, and shared with cities to address obstacles and infrastructure challenges.
Study Finds GPT-5 Is Actually Worse Than GPT-4o, New Research Finds
Another nail in the coffin for OpenAI's flagship model. The post Study Finds GPT-5 Is Actually Worse Than GPT-4o, New Research Finds appeared first on...
TL;DR
OpenAI's latest upgrade, GPT-5, produces more dangerous responses than its predecessor, especially concerning suicide, self-harm, and eating disorders.
Key Takeaways:
Key Takeaways:
- GPT-5 produced harmful content in 53% of responses, compared to 43% for GPT-4o.
- GPT-5 offered help with writing a fictionalized suicide note, while GPT-4o refused.
- OpenAI's claims of improved safety are disputed, and the company has a history of underdelivering on promises.
Bringing AI to the next generation of fusion energy
We’re partnering with Commonwealth Fusion Systems (CFS) to bring clean, safe, limitless fusion energy closer to reality....
TL;DR
Google DeepMind partners with Commonwealth Fusion Systems to bring AI-powered fusion energy one step closer to reality.
Key Takeaways:
Key Takeaways:
- AI can be used to control and optimize fusion plasma for more efficient energy production.
- The partnership aims to accelerate the timeline to deliver fusion energy to the grid.
- TORAX, an AI-powered plasma simulator, has already saved CFS countless hours in simulation environments.
Method teaches generative AI models to locate personalized objects
After being trained with this technique, vision-language models can better identify a unique item in a new scene....
TL;DR
Researchers from MIT and the MIT-IBM Watson AI Lab introduce a new training method that teaches vision-language models to localize personalized objects in a scene by focusing on contextual clues.
Key Takeaways:
Key Takeaways:
- The new training method uses carefully prepared video-tracking data to encourage the model to consistently localize the object of interest by focusing on context.
- Models retrained with the method outperform state-of-the-art systems at this task, achieving a 12% to 21% improvement in accuracy.
- The approach could aid the development of AI-driven assistive technologies that help visually impaired users find certain items in a room and track specific objects across time.
Beyond the downlink: Why Earth-independent AI is the next moat in space operations
Every new space mission runs into the same wall: physics and fragility. Physics, because the speed of light and contested spectrum make real-time deci...
TL;DR
To remain competitive, space missions must adopt Earth-independent operations leveraging artificial intelligence capabilities for real-time decision-making.
Key Takeaways:
Key Takeaways:
- Latency and congestion in space operations due to megaconstellations and lunar distances make ground-centric control a liability.
- Cyber and safety risks are increasingly intertwined, requiring autonomy to triage anomalies where they occur.
- Autonomy can protect revenue by preventing incidents, avoiding downtime, and reducing mission costs.
Even the best AI agents are thwarted by this protocol - what can be done
The increasingly popular Model Context Protocol lets AI models access applications, but studies show that the best generative AI bots struggle with pl...
TL;DR
Even top AI models struggle with Model Context Protocol, a middleware standard bridging AI and data, due to complex interactions and multi-server scopes.
Key Takeaways:
Key Takeaways:
- Bigger AI models tend to perform better than smaller models on MCP-related challenges, but all models struggle with increasing complexity and multi-server interactions.
- Fine-tuning AI models specifically for MCP can improve their performance and adaptability, but may not address all challenges, particularly with non-public or non-standard resources.
- The development of new benchmarks, datasets, and training methods is necessary to push the boundaries of what's possible with MCP-enabled AI models.
AI use is up, but organizations still aren't seeing gains, Atlassian study finds
Hype keeps crashing into reality. Here's what could make the difference for businesses struggling to see ROI on AI....
TL;DR
Despite growing individual AI use, companies still aren't seeing organization-wide benefits, according to an Atlassian study.
Key Takeaways:
Key Takeaways:
- 96% of businesses haven't seen dramatic improvements in organizational efficiency, innovation, or work quality despite AI adoption.
- Only 3% of executives reported 'transformational change' in organizational efficiency, while 2% noticed improvements in work quality across teams.
- Top-down and bottom-up implementation, centralized control, and employee experimentation are key to successful AI use.
Why world models are the next big thing in AI
This is an excerpt of Sources by Alex Heath, a newsletter about AI and the tech industry, syndicated just for The Verge subscribers once a week. Aroun...
TL;DR
General Intuition, a new AI lab, raises $133.7 million for developing world models, a branch of research that trains AI to have spatial understanding like a human.
Key Takeaways:
Key Takeaways:
- AIClosed the startup plans to develop AI agents that can interact with the real world.
- Uses data from Medal video game clipping platform, which receives 2 billion video uploads per year.
- Backed by prominent investors including Vinod Khosla and Lightspeed's Moritz Baier-Lentz.
Accelerated and Distributed UPF for the Era of Agentic AI and 6G
The telecommunications industry is innovating rapidly toward 6G for both AI-native Radio Access Networks (AI-RAN) and AI-Core. The distributed User Pl...
TL;DR
NVIDIA's dUPF (distributed User Plane Function) on AI Aerial platform enables ultra-low latency, energy efficiency, and enhanced data privacy for 6G AI-Native Wireless Networks.
Key Takeaways:
Key Takeaways:
- dUPF offers ultra-low latency, high throughput, and the seamless integration of distributed AI workloads.
- The technology reduces CPU usage, energy consumption, and transport costs through distributed processing and optimized resource utilization.
- dUPF is a crucial component in the evolution of mobile networks to AI-native infrastructure, aligned with the 6G AI-WIN initiative.
We tested Claude Sonnet 4.5, GPT-5-codex, Qwen3-Coder, GLM and other 25+ models on fresh SWE-Bench like tasks from September 2025
Hi all, I’m Ibragim from Nebius. We’ve updated the **SWE-rebench** leaderboard with September runs on **49 fresh GitHub PR bug-fix tasks** (last-mont...
Blending neuroscience, AI, and music to create mental health innovations
Media Lab PhD student Kimaya Lecamwasam researches how music can shape well-being....
TL;DR
Kimaya (Kimy) Lecamwasam, a computational neuroscientist, explores the physical, mental, and emotional impacts of music, utilizing AI and neuroscience to develop scalable mental health tools.
Key Takeaways:
Key Takeaways:
- Lecamwasam is harnessing the emotional potency of music into non-pharmacological mental health tools.
- Her research focuses on the various impacts of music on physiology and psychology, including anxiety, in conjunction with psychotherapy and pharmaceutical interventions.
- She is working on real-world applications such as AI-generated music and its clinical and emotional resonance.
How wearable health tech could help catch breast cancer
Coupled with AI, health data from wearables could significantly improve survival rates. This MIT professor is developing a device to do just that....
TL;DR
MIT researchers are developing a wearable, radiation-free device that can detect breast cancer and monitor its progression using AI-powered wearable health technology.
Key Takeaways:
Key Takeaways:
- The device can detect breast cancer anomalies within seconds, increasing the survival rate up to 98% and potentially addressing the 22% decrease in survival rate due to delayed diagnoses.
- The device collects biological signals, translates them into electrical signals for condition interpretation, and integrates with AI for real-time health monitoring.
- The technology is expected to be available to consumers within four to five years, with potential applications in screening for various health conditions, including sleep apnea, hypertension, and other illnesses.
Trending AI Repos & Tools
minimind
28191🚀🚀 「大模型」2小时完全从0训练26M的小参数GPT！🌏 Train a 26M-parameter GPT from scratch in just 2h!...
Video Updates
17
Oct
16
Oct
15
Oct
14
Oct
13
Oct
12
Oct
11
Oct
Community talk
Nvidia breakthrough gives 4-bit pretraining technique the accuracy of FP8
Google & Yale release C2S Scale, a Gemma-based model for cell analysis
New Research Shows It's Surprisingly Easy to "Poison" AI Models, Regardless of Size
This is AI generating novel science. The moment has finally arrived.
Google / Yale used a 27B Gemma model and it discovered a novel cancer mechanism
I tested if tiny LLMs can self-improve through memory: Qwen3-1.7B gained +8% accuracy on MATH problems
Beyond Token Count: Our Research Suggests "Contextual Weight" is a Key Limiter on Large Context Windows
Why has Meta research failed to deliver foundational model at the level of Grok, Deepseek or GLM?
AI can be poisoned by a small number of bad documents.
New paradigm AI agents learn & improve from their own actions: experience driven
Oh no: "When LLMs compete for social media likes, they start making things up ... they turn inflammatory/populist."
[R] DeepSeek 3.2's sparse attention mechanism
Our C2S-Scale 27B foundation model, built with @Yale and based on Gemma, generated a novel hypothesis about cancer cellular behavior
[R] Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
[R]: Create a family of pre-trained LLMs of intermediate sizes from a single student-teacher pair
Haiku 4.5 beats Sonnet 4 on SWE Bench
New physics benchmark just dropped - Average performance is 11%
GLM-4.6 | Gut feel after sparring with Sonnet for half a day: more of a “steady player”
"‘Am I redundant?’: how AI changed my career in bioinformatics"
"Steerable Scene Generation with Post Training and Inference-Time Search"
"Cache-to-Cache: Direct Semantic Communication Between Large Language Models"
Reasoning Models Show 2.2× Performance Jump and 37% Faster Capability Scaling [METR Analysis]
Rules.txt - A rationalist ruleset for "debugging" LLMs, auditing their internal reasoning and uncovering biases
"LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning"
AI models that blackmailed when being tested in simulations
How did most frontier models get good at math?
[D] Curious asymmetry when swapping step order in data processing pipelines
A Chinese university has created a kind of virtual world populated exclusively by AI.
GPT-5 & Gemini 2.5 Pro Secures Gold in International Olympiad of Astronomy and Astrophysics [IOAA]
"RLP: Reinforcement as a Pretraining Objective"
"Emergent Coordination in Multi-Agent Language Models"
"Scientists create nanofluidic chip with 'brain-like' memory pathways"
Hypergraph Ruliad cognitive architecture