Vansh Bhavsar

Researcher

[Vansh Bhavsar @vanshdot]

The solar wind has kinks in it: magnetic field lines that fold back on themselves into an S-shape. Scientists have argued over where they come from. Solar Orbiter just flew through one and traced it back to the Sun's surface.

• Its Solar Wind Analyser sampled the plasma inside a very large "switchback", roughly halfway between Earth and the Sun

• The mix of charged oxygen and carbon particles could only have formed in hot magnetic loops at the solar surface

• That's the smoking gun for "interchange reconnection": open and closed field lines snap and reconnect, letting trapped plasma escape into space

• Waves and turbulence show up too, but likely only after the switchback leaves the Sun. Both rival theories were right, at different stages.

The bigger point: solar wind carries a chemical fingerprint of where it was born. Knowing that helps us prepare for the solar storms that threaten satellites.

Published in Nature Astronomy (Coburn et al.)

#Space #SolarOrbiter #Physics #SpaceWeather

[Vansh Bhavsar @vanshdot]

Small talk with your doctor can end up in your medical record when an AI writes the note.

A new study stress-tested LLMs on clinical documentation:

• Across 576 patient–clinician dialogues, frontier models inserted small-talk exchanges into 35% of notes

• In 3.7% of notes, the model misattributed those asides or used them clinically

• Background speech from a separate patient encounter (at −10 dB) leaked into 48.2% of transcripts, and showed up in 5.3% of notes from 4 open-weight models downstream

The catch: mean note-quality scores moved by at most 0.20 points on a 5-point scale. Standard metrics barely see the problem.

AI scribes are already in clinics. Robustness to irrelevant information should be tested before deployment, not after.

https://t.co/sKPD5oVxqW

#MedicalAI #LLM #AIResearch

[Vansh Bhavsar @vanshdot]

Which AI model best matches how your hippocampus encodes the meaning of words? GPT-2.

Researchers at Baylor, Rice and UC Berkeley recorded single neurons in the hippocampus of 10 people (epilepsy patients with implanted electrodes) while they listened to stories.

• Word meaning was robustly encoded, even after controlling for sound (phonemes) and grammar

• No single "word neuron": individual cells responded to multiple words across multiple semantic categories. Meaning lives in the population

• Like LLM embeddings, the distance between neural population responses tracked the semantic distance between words

• Neural activity aligned most closely with GPT-2 embeddings, and response variability tracked LLM-derived polysemy, so meaning is encoded in context

LLMs are becoming microscopes for the brain. 🧠

Published in Nature Neuroscience.

#Neuroscience #NeuroAI #LLM #AI

[Vansh Bhavsar @vanshdot]

501B parameters, but only 23B active per token. And a US lab says it will release the weights under Apache 2.0.

Reflection AI just unveiled Beam, its first frontier open-weight model:

• Sparse MoE, 52 layers, 1M-token context, text-only
• Pretrained on 23.8T tokens in under 4 weeks on 6,144 NVIDIA GB300s
• RL on ~1M environments with 100M+ rollouts

Coding: 80.9 on SWE-bench Verified and 65.5 on SWE-Bench Pro v1 (GLM 5.2: 62.1).

It's not the strongest open model, and their own table shows Kimi K3 ahead. The pitch is efficiency: they claim GLM-5.2-level reasoning with 3–4x less inference compute.

Weights and the tech report are due later this month. That's when we'll find out if the claims hold up.

#AI #LLM #OpenSource

[Vansh Bhavsar @vanshdot]

Quantum computers won't scale on a single chip. They'll need to be networked, and the link between them has been the slowest part.

IonQ researchers just pushed that link past 1,000 entangled pairs per second.

• A trapped barium ion is entangled with a silicon-vacancy qubit inside a diamond nanophotonic cavity

• The ion's photon is converted in wavelength, bounced off the diamond cavity, and a single detected photon heralds the entanglement

• 1.03 kHz average rate, counting all conversion losses: ~4x the fastest photonic link between two trapped ions

• Bell-pair fidelity: 87.9%

Why it matters: one Bell pair per millisecond roughly matches the cycle time of planned trapped-ion processors. The network stops being the bottleneck.

Fidelity still has to climb, but that's the right problem to have.

https://t.co/QVJqFKlukv

#QuantumComputing #Quantum #Physics

[Vansh Bhavsar @vanshdot]

A robot navigator that was never fine-tuned for navigation just beat the ones that were.

SuperNav (Zhejiang University et al.) takes an off-the-shelf multimodal LLM, leaves its weights untouched, and wraps it in an agent harness: navigation skills, tools, task-progress tracking, and a "point at the destination in the image" interface that hands the actual motion to navigation tools.

Results in simulation:
• Single-object search: 78% success vs 34% for the best of 4 baselines
• Multi-object: 59.5% vs 37.5%
• Demand-driven (need-based) requests in AI2-THOR: 73.3% vs 62.5%

It also runs on a real Unitree Go2 quadruped.

The pattern keeps showing up: for embodied AI, the harness may matter more than the fine-tune.

https://t.co/L4FbokDfLw

#Robotics #EmbodiedAI #AIResearch

[Vansh Bhavsar @vanshdot]

21 star systems. Same violent fingerprint: dust from planet-sized bodies smashing into each other.

Using mid-infrared spectra from Webb and Spitzer, astronomers sorted these "extreme debris disks" by what the dust is made of:

• About one-third are silica-rich, pointing to high-energy collisions between Mars-sized bodies
• The other two-thirds are silica-poor, more like grazing impacts between Moon-sized objects

The silica-rich kind has only been found around stars under 300 million years old, the same era when rocky planets like Earth are thought to form.

Our own Moon likely came from exactly this kind of crash. Now we can see it happening elsewhere.

#JWST #Space #Exoplanets #Astronomy

[Vansh Bhavsar @vanshdot]

In 1988 he pitched an idea: turn the ice at the South Pole into a particle detector. This week it won him the Nobel Prize in Physics.

Francis Halzen (UW–Madison) is the sole 2026 physics laureate, for IceCube and the discovery of high-energy neutrinos from far outside our solar system.

- IceCube is a full cubic kilometre of Antarctic ice fitted with light sensors, completed in 2011

- Neutrinos pass through Earth (and you) almost without interacting. When one rarely hits a nucleus, it makes a faint flash the sensors can track

- Unlike other particles, they reach us without changing direction or losing energy, so they point straight back to their source

Why it matters: the universe has natural accelerators firing particles at up to a million times the energy of our best labs. Neutrinos carry information about them that no other messenger can.

A new kind of astronomy, built out of ice. 🧊

#NobelPrize #Physics #Neutrinos #Astrophysics

[Dr Singularity @Dr_Singularity]

What Happens If We Solve Quantum Gravity

it can happen very fast (ASI will help us).

[Sharath Kuruganty @5harath]

How can you not not fell in love with California?

[Vansh Bhavsar @vanshdot]

Agents that learn in 4 rounds instead of millions of RL steps.

AgentGarten: simulators keep the world state, a distilled video model renders what the agent sees, and each round's lessons become a playbook the next agent inherits.

https://t.co/UNQ17v3GsI

[Vansh Bhavsar @vanshdot]

Xiaomi's MiMo-V2.6 tech report is a masterclass in scaling RL:

• ~3B tokens per RL step, up to 1M context
• code, visual & cyber environments
• frozen MoE router for stability

RL framework + environments are open-sourced.

https://t.co/ANiZMC34IR

[We, the People of India @India_Policy]

As US suspends Infosys, TCS, Wipro, HCL and others from green-card labour programme, listen to what Michio Kaku said on H1B, and it being the secret weapon of the US.

[Vansh Bhavsar @vanshdot]

MMedAgent: A Multi-Modal AI Medical Agent

General vision-language models can struggle with specialized medical tasks. MMedAgent addresses this by acting as a “medical conductor” that dynamically selects and orchestrates specialized AI tools.

Here’s how it works:

1/ Architecture

User provides a medical image + question.

The agent then:
→ Plans the task
→ Selects the appropriate medical tool
→ Executes it
→ Aggregates the results into a final response
It can work across CT, MRI, X-ray, histology, and gross images.

2/ Specialized Tools
MMedAgent integrates tools for:
• Grounding/detection
• Segmentation
• Medical image classification
• Radiology report generation
• Medical knowledge retrieval (RAG)
• Medical VQA

Examples include Grounding DINO, MedSAM, BiomedCLIP, ChatCAD, and LLaVA-Med.

3/ What’s Novel?

The paper introduces:

• A multi-modal medical agent designed for diverse clinical imaging tasks
• A 48K instruction dataset for learning when and how to use medical tools
• Domain-specific tuning of open-source vision tools
• A plug-and-play design where tools can be replaced or added without retraining the entire system.

4/ Performance
MMedAgent achieved an overall score of 109.48, compared with:
LLaVA-Med: 60.68
RadFM: 45.38
It also outperformed GPT-4o on organ grounding, disease grounding, and chest X-ray report generation.
The agent reported 100% accuracy in selecting the correct tool in its evaluation.

#AI #MedicalAI #MultimodalAI #AIAgents #GenerativeAI #ComputerVision #HealthcareAI #MachineLearning #DeepLearning #Research

[Elon Musk @elonmusk]

🇮🇳 Starlink will help the least-served in India 🇮🇳

[SpaceX @SpaceX]

Dragon separation confirmed!

[Vansh Bhavsar @vanshdot]

Can an AI diagnose multiple diseases by reasoning over an ECG, chest X-ray, and blood tests simultaneously?

MedTVT-R1 is a multimodal medical reasoning framework designed to do exactly that.

Instead of relying on a single type of patient data, it jointly processes:

• ECG — continuous heart signals
• CXR — chest X-ray images
• LAB — 50 blood indicators across 7 physiological categories
The architecture combines:
• LLaMA 3.2-1B + LoRA
• ViT-B/16 for chest X-rays
• ECGFM-KED for ECG signals
• Symile for laboratory data

Its key component is the Modality Perception Layer (MPL).

It uses:
Cyclic Multi-Head Attention (CMHA) to allow different modalities to exchange information.
Contribution-Aware Operator (CAO) to dynamically determine which modality is more important for a particular diagnostic context.
For example, ECG information may receive greater importance for cardiac conditions, while chest X-rays may become more important for pneumonia.

The model was trained using 8,706 patient cases from MIMIC-IV and uses a Chain of Evidence approach to connect evidence across modalities.

It also applies GRPO reinforcement learning with a Jaccard-based reward for disease prediction.

Results:

Disease Diagnosis
F1: 0.5190
AUC: 0.6554

Reported baseline F1 scores ranged from 0.0398 to 0.1995 on the same benchmark.

The framework was evaluated across 7 disease categories:

• Coronary Artery Disease
• Acute Renal Failure
• Hypertension
• Atrial Fibrillation
• Pneumonia
• Diabetes Mellitus
• Sepsis

The broader idea is important: medical AI may become more capable when it learns to reason across different types of clinical evidence rather than analyzing each modality independently.

Paper & code: https://t.co/HSBI0XFwcW

#MedicalAI #MultimodalAI #HealthcareAI #AIResearch #MachineLearning #DeepLearning #MedicalImaging #LLM #MedTech

[Arshjit Singh @Arshh1998]

Elon Musk you fucking genius 🫡
Shoutout to the entire SpaceX team 💯

That’s a 17 floor building dispersing satellites the size of a school bus in space btw.

(The satellites open up to the size of an airplane)

We are living in the craziest of times $SPCX @elonmusk

[Vansh Bhavsar @vanshdot]

What if an AI system could act like an entire multidisciplinary medical team?

Researchers developed KOM, a multi-agent AI system designed to assess knee osteoarthritis, predict disease progression, and generate personalized treatment plans.

And the interesting part is that it doesn’t rely on a single AI agent.

KOM divides the clinical workflow across specialized agents and combines their decisions through a central Clinical Decision Agent.

Here’s how it works:

Assessment Agent

Handles conversational patient intake and analyzes knee X-rays using:
• Qwen-Max for clinical dialogue
• U-Net for joint center localization
• ResNet50 for KOA severity grading and radiographic feature classification
Risk Agent
Uses XGBoost, Random Forest, AdaBoost and ElasticNet to predict functional outcomes and disease progression.
It also uses SHAP to identify individualized risk factors.

Therapy Agent Group

Simulates a multidisciplinary clinical team with specialized agents for:
• Exercise rehabilitation
• Orthopedics and medication
• Psycho-nutrition

A central Clinical Decision Agent combines their recommendations and resolves conflicts.

The therapy agents use RAG with six curated medical knowledge bases containing 4,017 peer-reviewed papers and clinical guidelines.

Why is KOM interesting?
• End-to-end KOA management instead of a single prediction task
• Combines LLMs, computer vision, ML risk prediction and RAG
• Simulates multidisciplinary clinical decision-making
• Generates personalized, guideline-informed treatment plans
The reported results are also impressive:
• 77.16% accuracy for KOA severity grading
• 82.22% accuracy for KOA detection
• AUC > 0.96 for 1-year radiographic progression
• Treatment-plan quality score: 29.63/35
• 38.5% reduction in case-processing time
• Doctor approval rate increased from 56.0% to 93.0% with KOM assistance

The bigger idea is what makes this research interesting.

Instead of asking one general-purpose LLM to behave like an entire medical team, KOM gives different agents different responsibilities and then uses a central agent to integrate their decisions.

This points toward a future where medical AI systems are not just chatbots, but structured teams of specialized AI agents working alongside clinicians.

Code and implementation:
https://t.co/f6LwxmMqK4

#MedicalAI #AIinHealthcare #MultiAgentAI #HealthcareAI #ArtificialIntelligence #MachineLearning #LLM #RAG #ComputerVision #DigitalHealth

[Vansh Bhavsar @vanshdot]

Can multi-agent AI teams transform sepsis care in under-resourced hospitals?

The MATEC (Multi-AI Agent Team Care) framework takes a different approach: instead of relying on a single AI chatbot, it creates a simulated multidisciplinary care team with 10 specialized agents.

The team includes:
Emergency Medicine
Hospital Medicine
Infectious Disease
Critical Care
Nursing
Pharmacy
Social Work
Patient Safety
Risk Prediction
Senior Physician supervision

These agents use LLM reasoning, ReAct prompting, and RAG to generate differential diagnoses and care plans. A Senior Physician agent then synthesizes the recommendations, verifies information, and checks for potential hallucinations.

What makes MATEC interesting?

It goes beyond diagnosis by incorporating:
Care-gap analysis
Medication safety
Social determinants of health
Patient safety and quality control
SEP-1 sepsis guideline compliance

The goal is particularly relevant for rural and under-resourced hospitals where access to specialists may be limited.

In evaluation by 10 attending physicians, MATEC received:
4/5 for usefulness
4/5 for accuracy
Higher usefulness ratings than general-purpose LLMs among AI-experienced physicians
Reduced reasoning errors through multi-agent verification

The bigger idea is compelling: instead of asking one AI to act like an entire hospital, can we build AI teams that collaborate like one?

Paper: "The Application of MATEC (Multi-AI Agent Team Care) Framework in Sepsis Care"

#MedicalAI #HealthcareAI #AIAgents #MultiAgentSystems #Sepsis #ClinicalAI #GenerativeAI #LLM #RAG #DigitalHealth