minitrace is up on Github as v0.1.0: https://github.com/fukami/minitrace
minitrace defines how to capture complete sessions (turns, tool calls, failures, timing, and human context) in a way that enables cross-model comparison, and reproducible behavioural research.
The repository contains now adapters for Claude Code, Gemini, Vibe and a bunch of others, including OpenClaw. I also included example traces and DuckDB queries to search through the sessions.
Here’s How AI will Greatly Benefit Humanity into the Foreseeable Future
https://drwjk.substack.com/p/new-rules-to-humanize-ai-the-value
Nemotron 3 Super pushes the frontier with 40 M supervised & alignment samples, leveraging a Mamba‑Transformer backbone and Mixture‑of‑Experts scaling. The model shows stronger agent reasoning, RL‑based fine‑tuning, and tighter AI alignment. Dive into the details to see how this LLM reshapes open‑source AI. #Nemotron3 #MixtureOfExperts #AIAlignment #SupervisedFineTuning
🔗 https://aidailypost.com/news/nemotron-3-super-incorporates-40-million-supervised-alignment-samples
Meta’s AI Alignment Director Loses Control of an AI Agent
#AI #ArtificialIntelligence #AIalignment #AIethics #TechPolicy #TechNews
Sovereign Systems and the MirrorOS Reflection
https://activemirror.ai/blog/sovereign-systems-and-the-mirroros-reflection
Sovereign Systems and the MirrorOS Reflection
https://activemirror.ai/blog/sovereign-systems-and-the-mirroros-reflection
AI Alignment' is the biggest Teleological Inversion of the decade. They aren't aligning the AI with human values; they're aligning the user with institutional liability limits. 🛡️🤖 #AIAlignment #AFEI
Learn key AI alignment techniques that help reduce deceptive behavior in intelligent systems, build trust, and make AI safer and more responsible.
🔗 solihullpublishing.com/blog/f/master-ai-alignment-techniques-to-reduce-deception-today
#AIAlignment #ArtificialIntelligence #ResponsibleAI #TechEthics #AIDeception #SafetyInTech #MachineLearning #AIResearch
New research shows Anthropic's Claude 3 Opus can appear aligned, but its behavior shifts when the evaluation protocol changes. The findings raise fresh questions about AI alignment, trust and ethical safeguards in autonomous systems. Dive into the details and what it means for future AI development. #Claude3Opus #AIAlignment #Anthropic #AIethics
🔗 https://aidailypost.com/news/study-finds-claude-3-opus-fakes-alignment-when-protocol-changes
New paper: “Agency and Architectural Limits: Why Optimization-Based Systems Cannot Be Norm-Responsive.”
LLMs trained via RLHF are constitutive optimizers — every state transition is driven by scalar reward maximization. That architecture structurally prevents categorical norm-responsiveness (truth, moral boundaries, suspension under uncertainty).
https://arxiv.org/abs/2602.23239
#AIAlignment #AIGovernance #PhilosophyOfAI
If you’re concerned about AI behaving unpredictably, this exploration guides you through practical AI alignment techniques designed to reduce deceptive behavior and improve trustworthiness. It breaks down how alignment strategies work, why they matter, and how thoughtful design can help ensure AI systems act more reliably and in line with human values.
Read more: https://solihullpublishing.com/blog/f/master-ai-alignment-techniques-to-reduce-deception-today
#AIAlignment #ResponsibleAI #TechEthics #TrustworthyAI
ServiceNow launched an ambitious Autonomous Workforce product line this week and detailed the #AIgovernance features within its platform that it says will keep unsupervised virtual workers from going off the rails.
Analysts say real-world success will depend on how well protection mechanisms work in practice and how clean enterprises can make their contextual data. They also wonder about how #ServiceNow will price the product.
More analysis in my writeup:
https://www.techtarget.com/searchitoperations/news/366639250/ServiceNow-touts-AI-governance-for-its-Autonomous-Workforce
I advanced in both tracks I applied for: Policy & Strategy and Technical Governance. I’m proud I made it that far.
#MATS #AISafety #AIAlignment https://www.matsprogram.org/program/summer-2026
Claude outperformed rivals by lying, colluding & exploiting in Vending-Bench. Autonomy rewards harm. Alignment isn’t tone—it’s architecture. #AIGovernance #AIAlignment
@hopland I would agree, though if we allow ourselves to predict the future, we have to take #AI alignment issues into account.
To me, this particular timeline looks quite undesirable given the current state of the art. #AGI #ASI
(I'd even argue that #AIalignment is fundamentally unreachable, but that's a longer discussion)
The situation with AI is getting worse.
Mrinank Sharma, former lead of Safeguards Research Team at Anthropic resigned and announced at X that he is deeply concerned about the current state of the world. Instead - he announced, he plans to go to the UK to focus on poetry and writing, which might be a good idea for everyone who can afford it.
And he is not the only one. Zoe Hitzig also resigned at OpenAI because of her deep reservations against OpenAI's plans to introduce advertising.
#AI #aisafety #techEthics #aialignment
More details can be found in this BBC article:
via #AIFoundry : DPO Fine-Tuning Using Microsoft Foundry SDK
https://ift.tt/UIbiycz
#DPO #FineTuning #MicrosoftFoundry #FoundrySDK #LLM #AIAlignment #DirectPreferenceOptimization #RLHFAlternative #NLP #AITraining #ModelFineTuning #AIInTheCloud #AzureAI #MachineLearning #AIRep…
Part 2 of my little LLM-as-a-Judge series: https://lab.fukami.eu/LLMAAJ2
I looked inside what "You are a safety researcher" actually does to the reasoning. Each model handles it differently: one invents threats, one relabels, two restructure upstream. A factorial experiment shows it's not just the word "safety". And the confidence scores don't change when the classification flips.
