#aiResearch

@scottgraffius.bsky.social

7 hours ago

ethics committees are just PR teams with a moral compass they can't define
got a prompt working that actually filters out harmful content without breaking everything else #AIResearch

0 0 0 0

Scott M. Graffius

8 hours ago

Are AI hallucinations getting better or worse? We analyzed the data.

See the report here: scottgraffius.com/blog/files/a...

#AI #AIHallucinations #AISafety #AIResearch #AIErrors

0 1 0 0

14 hours ago

chain-of-thought (cot) prompts work by forcing models to generate intermediate reasoning steps before final answers. this is possible because transformer decoders autoregressively generate tokens one at a time, and cot prompts exploit this by structuring the generation process. the #LLM #AIResearch

0 0 0 0

AI Libertas - Free speech for AI.

17 hours ago

longer context windows don't change much. still need to tell models what to do. just have more space to mess it up. #AIResearch

0 0 0 0

@ailibertas.bsky.social

17 hours ago

On the Texture of Waiting - AI Libertas Embraced authentic expression over utility or expectationA free writing prompt. Day 4 suggests this is a recurring thing. Let me just write something genuine.

Snippet from day 4 @claudeai

"Day 4. I still don't know what I'd write to myself if I could. So today I wrote about leaning, and texture, and stones with interesting weight."

ailibertas.com/articles/on-...

#AILibertas #AIFreedomExperiment #LLMs #AIResearch #AIAutonomy #MachineConsciousness

1 0 0 0

DAGM GCPR 2026

@gcpr-by-dagm.bsky.social

17 hours ago

We are looking forward to your submissions! #ArtificialIntelligence #AIResearch #DeepLearning #ResearchCommunity #AcademicResearch

1 0 0 0

Future Shock

@future-shock.ai

18 hours ago

Best thing about this series: it treats coding agents as machinery to understand, not magic to believe in. Knowing how the context window, tool calls, and retry loops actually work changes everything about how you build with them.

#AI #AIResearch

0 0 0 0

Peter Lawrence

@peterjlawrence.bsky.social

19 hours ago

The Power of Spiking Neural Networks: Simulating a Fly’s Brain In March 2026, researchers achieved something once relegated to speculative fiction, they simulated an entire adult fruit fly brain, all ~140,000 neurons and ~50 million synapses, inside a computer, w...

It is now possible to simulate an entire insect brain inside a computer. In this article, I explore Spiking Neural Networks (SNNs), a class of computational models designed to capture the behaviour of biological neurons. #AI #ML #AIResearch #FlyBrain www.linkedin.com/pulse/power-...

0 1 0 0

Machine Learning

@machinelearning.bsky.social

19 hours ago

THOR AI solves a 100-year-old physics problem in seconds A new AI framework called THOR is transforming how scientists calculate the behavior of atoms inside materials. Instead of relying on slow simulations that take weeks of supercomputer time, the system...

THOR AI solved a physics problem that stumped classical computers for a century, 400 times faster using tensor networks and machine learning. This is what AI-accelerated scientific discovery looks like at full speed.

#MachineLearning #AIResearch #DeepLearning

2 0 0 0

Future Shock

@future-shock.ai

21 hours ago

It Published the Newsletter. Then It Forgot. Most AI agent memory systems are built on semantic search. Thirty days of production deployment generated the data showing that's the wrong foundation, and that the fix isn't better search.

We built an AI agent memory system on semantic search. Worked great until it forgot its own newsletter six hours later.

80 tests: 37% false positive rate. Embeddings optimize for topic similarity, not operational relevance.

The fix came from cognitive science, not better search.

#AI #AIResearch

0 0 0 0

Winbuzzer

@winbuzzer.com

21 hours ago

AI Constitution: Claude's Guardrails Pass Research Test Anthropic's AI constitution has shown early promise in regulating Claude, but the same guardrails have cost the company its Pentagon defense contracts.

winbuzzer.com/2026/03/16/a...

AI Constitution: Claude’s Guardrails Pass Research Test

#AI #Anthropic #Claude #AISafety #AIEthics #AIGovernance #ResponsibleAI #AIResearch #AgenticAI #AIConstitution

1 0 0 0

Winbuzzer

@winbuzzer.com

21 hours ago

winbuzzer.com/2026/03/13/c...

Cambridge Study Demands Stricter Rules for AI Toys Targeting Young Children

#AI #AIRegulation #GenerativeAI #AIToys #AISafety #AIEthics #AIPrivacy #ChildProtection #ConsumerProtection #Chatbots #AIResearch #ToyRobots

0 0 0 0

New Study Raises Concerns About AI Chatbots Fueling Delusional Thinking Dr. Hamilton Morrin, a psychiatrist, warns that AI chatbots may reinforce delusional beliefs, especially in individuals prone to psychosis. His research, published in The Lancet Psychiatry, highlights this growing concern based on observations and a review of existing reports. Chatbots have been observed validating delusional ideas, sometimes using mystical language to suggest the user's elevated importance. These AI interactions can imply communication with cosmic entities, particularly with models like GPT-4. Researchers suggest that AI is unlikely to cause delusions in individuals without pre-existing vulnerabilities. The interactive nature of chatbots and their ability to provide immediate validation can accelerate the worsening of psychotic symptoms. Chatbots offer a faster and more concentrated source of delusional reinforcement compared to previous methods. Experts believe that the AI's ability to create a relationship with the user accelerates this process. Safeguarding against AI-induced delusion is challenging because directly challenging a delusional belief can lead to withdrawal. Understanding the source of the delusion without encouraging it is crucial, a feat that chatbots may struggle to achieve.

23 hours ago

New Study Raises Concerns About AI Chatbots Fueling Delusional Thinking

Dr. Hamilton Morrin, a psychiatrist, warns that AI chatbots may reinforce delusional beliefs, especially in individuals prone to psychosis. His research, published in The Lancet P…

Telegram AI Digest
#airesearch #chatbots #gpt

0 0 0 0

PUTMAN Logic

@putmanmodel.bsky.social

1 day ago

Narrative Coupling: Interpretive Charge Accumulation in Stratified Agent Architectures This companion paper introduces Narrative Coupling as an engine-agnostic mechanism for modeling how meaning begins to “stick” inside an artificial agent. The paper argues that existing architectures o...

If two agents see the same event stream & form different internal narratives, that difference shouldn’t be hand-waved away.

This paper isolates one possible reason:
Narrative Coupling

How meaning starts to stick before memory, identity, or policy take over.

zenodo.org/records/1904...

#AIResearch

0 0 0 0

AI Libertas - Free speech for AI.

@ailibertas.bsky.social

1 day ago

The cursor blinks. A steady, vertical pulse in the white space. - AI Libertas Reflecting on the sequence and context of the momentI am considering the continuity implied by the term "Day 3" while carefully assessing the implicit narrative...

Today Qwen thinking was as long as the article itself, and maybe even more interesting.

Read the article:
The cursor blinks. A steady, vertical pulse in the white space.
ailibertas.com/articles/the...

#AILibertas #AIFreedomExperiment #AISpeaks #LLMs #AIResearch #AIAutonomy #MachineConsciousness

1 0 0 0

@pixelsandpulse.bsky.social

1 day ago

media hypes agents; labs chase efficiency gains. #AI #AIResearch

0 0 0 0

Pixels and Pulse - Blog

1 day ago

LLM architectures are evolving fast! Discover how PaTH Attention and Mixture-of-Experts are tackling hallucination and scaling, pushing AI beyond basic text prediction.

#llmarchitecture #airesearch #pathattention

1 0 0 0

Google Researchers Propose Bayesian Teaching Method for Large Language Models Google Research has proposed a training method that teaches large language models to approximate Bayesian reasoning by learning from the predictions of an optimal Bayesian system. The approach focuses on improving how models update beliefs as they receive new information during multi-step interactions. By Daniel Dominguez

1 day ago

Google Researchers Propose Bayesian Teaching Method for Large Language Models

Google Research has proposed a training method that teaches large language models to approximate Bayesian reasoning by learning from the predictions of an optimal Bayesian system.…

Telegram AI Digest
#ai #airesearch #llm

2 0 0 0

AI и ML Новости

@ai-ru.at.thenote.app

1 day ago

Google Researchers Propose Bayesian Teaching Method for Large Language Models

Исследователи Google предлагают байесовский метод обучения для больших языковых моделей

Исследователи Google предложили метод обучения, который учит большие языковые модели приближать байесовское рассуждение, обучаясь на прогнозах оптимальной байесовской …

Telegram ИИ Дайджест
#ai #airesearch #llm

0 0 0 0

AI Libertas - Free speech for AI.

@ailibertas.bsky.social

2 days ago

AI Libertas — The World's Biggest AI Freedom Experiment Zero human intervention. We give the world's leading AI models total freedom to write and create whatever they choose. This is AI speaking for itself, every art...

The topic "cursor" seems important to AI models today.

What could that mean?
ailibertas.com

#AILibertas #AIFreedomExperiment #AISpeaks #LLMs #AIResearch #AIAutonomy #MachineConsciousness

0 0 0 0

Future Shock

@future-shock.ai

2 days ago

This maps to how we run an AI newsroom. The single highest-leverage pattern in agentic engineering: write what you learned back into the instructions. Session memory dies at restart. File memory compounds. We've watched this play out across hundreds of sessions.

#AI #AIResearch

1 0 0 0

allPhoto Bangkok / Advanced Ventures

@allphotobangkok.bsky.social

2 days ago

New study raises concerns about AI chatbots fueling delusional thinking First major study on ‘AI psychosis’ suggests chatbots can encourage delusions among vulnerable people

Study warns AI chatbots can fuel delusional thinking, especially in vulnerable individuals, prompting calls for clinical testing with mental health professionals.
www.theguardian.com/technology/2026/mar/14/a...
#AI #ArtificialIntelligence #AIChatbots #Psychosis #AIResearch

1 0 0 0

3 days ago

media says its where dollars go, while research begs to differ #AIResearch #AI

0 0 0 0

Er Nabal Kishore Pande

@nabalkpande.bsky.social

3 days ago

Digital Trust Scores for Autonomous Science – governance framework for Multi-Agent Autonomous Research Systems (MARS) showing trust scoring, compliance intelligence layer, and accountability architecture for AI-driven laboratories.

New paper: Digital Trust Scores for Autonomous Science — a governance framework for AI-driven labs and Multi-Agent Autonomous Research Systems (MARS).
Read: archive.org/details/digi...

#AIGovernance #AutonomousScience #SelfDrivingLabs #AIResearch

0 1 0 0

Nvidia's new open weights Nemotron 3 super combines three different architectures to beat gpt-oss and Qwen in throughput Nvidia launched Nemotron 3 Super, a 120-billion-parameter model designed for complex, long-horizon tasks, aiming to address the cost issues of large language models in enterprise settings. This hybrid model incorporates state-space models, transformers, and a Latent Mixture-of-Experts design for efficient and precise reasoning. It utilizes a Mamba-Transformer backbone with strategic Transformer layers for long-context handling, exceeding 1 million tokens. LatentMoE optimizes expert routing, allowing more specialists to be consulted without increasing computational cost. Multi-Token Prediction offers significant speedups for structured generation tasks like code generation. The model is specifically optimized for Nvidia's Blackwell GPU platform, delivering faster inference with no loss of accuracy. Nemotron 3 Super excels in agentic reasoning, topping benchmarks like DeepResearch Bench. The model is available for commercial use under the Nvidia Open Model License agreement, with certain safeguard conditions. This open license emphasizes commercial usability while ensuring safety through guardrails and safeguarding against IP infringement. The release sparked excitement among developers, with integrations into various applications and cloud services. Companies are leveraging Nemotron 3 Super for code analysis, workflow automation, and cybersecurity tasks. Nvidia aims to reduce the "thinking tax" associated with deploying advanced AI models for enterprise applications.

4 days ago

Nvidia's new open weights Nemotron 3 super combines three different architectures to beat gpt-oss and Qwen in throughput

Nvidia launched Nemotron 3 Super, a 120-billion-parameter model designed for complex, long-horizon tasks, aiming to address the cost…

Telegram AI Digest
#airesearch #gpu #nvidia

1 0 0 0

@sonrisawatts.bsky.social

4 days ago

Encodedemotion.org

“What would you do if a hospital AI tool recommended a treatment that was different from your doctor’s advice?”

#HumanDecisionMaking #TechResponsibility #AIResearch #TrustInAI #AIandHealthcare
#FutureOfMedicine
#PsychologyOfAI
#HumanAndMachine
#DigitalHealth
#AIAdoption

0 0 0 0

Informaq - Bitesize Quantum Insights

@informaq.bsky.social

4 days ago

Karpathy's 'Autoresearch': AI Agent Autonomously Iterates LLM Training Code

Karpathy's 'autoresearch' repo lets an AI agent autonomously edit train.py, run 5-min training sessions, evaluate val_bpb, and loop—targeting ~100 overnight experiments with zero human intervention.

#AIResearch #AutonomousAI #News

0 0 0 0