AI/ML news, top picks, and generated innovation digests.
30926 matching items
As AI-accelerated warfare is rapidly becoming a means of rubber-stamping killing at unprecedented speed and scale, Access Now, Amnesty International, and more than 200 civil society organizations and individuals are calling attention to the militarization of artificial intelligence (AI) technologies and for an immediate halt to the use of AI systems in the military kill chain. The post AI-accelerated warfare must stop appeared first on Access Now .
We, the undersigned organizations and individuals, are deeply alarmed by the rapid militarization of artificial intelligence (AI) technologies. AI systems embedded into military kill chains are accelerating the speed and scale of military assaults in a manner that creates significant new risks for accountability in conflict and risks facilitating violations of international criminal, human rights, and humanitarian law. The post Joint statement on AI in warfare appeared first on Access Now .
The post Heads, founders win. Tails, platforms win. AI fund Activate wins either way appeared first on The Ken .
Recent product updates and news from around the community.
In our ongoing SIG Spotlight series, we shine a light on the groups that keep the Kubernetes project moving forward. This time, we catch up with SIG Storage , the group responsible for persistent data, volume management, and the interfaces that connect Kubernetes workloads to the storage systems beneath them. We spoke with Xing Yang , Co-Chair of SIG Storage and Software Engineer at VMware by Broadcom, about the SIG's history, the features shipping in recent Kubernetes releases, and where storage in Kubernetes is headed as AI workloads become the norm. Introductions Could you introduce yourself and share your role(s) within SIG Storage? My name is Xing Yang , a software engineer at VMware by Broadcom. I'm a co-chair in SIG Storage, alongside another co-chair Saad Ali from Google. There are also two Tech Leads in SIG Storage: Michelle Au from Google and Jan Šafránek from Red Hat. What first drew you to storage in Kubernetes, and how did you start contributing? I have always been working in the storage domain, so SIG Storage was a natural place for me to get started when I began to learn Kubernetes. I started attending SIG Storage meetings , trying to figure out what I could do to help. This was before the first Container Storage Interface (CSI) release — lots of things were still evolving. It was a very exciting time. What subprojects or areas do you actively maintain or review today? I'm a maintainer in Kubernetes CSI. There are multiple CSI sidecars — such as csi-provisione…
The US government yanked Anthropic's newest models days after launch, while state attorneys general opened formal process against OpenAI. That turns frontier capability into something investors have to discount: a model can be state-of-the-art on Monday and policy-frozen by Friday. The market still wants the upside, but the asset now has a kill-switch.
This is the fourth in a series of informal research updates from the Google DeepMind Language Model Interpretability team, in interpretability and adjacent areas. The third post can be found here . Since SFT is the cause for many safety relevant properties , a natural strategy is to filter out rollouts from SFT that have undesirable properties. However, as we show in this section (and in forthcoming MATS work), SFT data filtering frequently works surprisingly poorly. In this post, we investigate hypotheses for why SFT filtering fails. TL;DR: We discuss seven hypotheses for why SFT filtering works surprisingly poorly We analyze three hereditary traits that SFT-only Gemini has that other models do not: negative emotion, date confusion, and blackmail in the (highly contrived) agentic misalignment scenario We use a “post-training diffing pipeline” between Gemini and Olmo to show that the cause of date confusion and blackmail is largely surprising transfer of behaviors from the SFT teacher model. Notably, there exist small sets of prompts where switching the teacher model for the rollout removes date confusion and blackmail, but dropping the prompts does not. Negative emotion is less affected by the teacher model, but this may be because the Olmo prompt distribution we are SFTing on underspecifies the behavior. Takeaways: It’s hard to remove behaviors via filtering But if you can get a teacher model to have a behavior (e.g. via RL), then transferring that in the future is easier…
It's a one-way door and we weren't ready for it.
PLUS: Set up your Claude Code like The Creator Boris Cherny
OpenAI launches the Partner Network, investing $150M to help global partners accelerate enterprise AI adoption, deployment, and transformation.
❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.ai/papers 📝 The Nemotron 3 Ultra paper is available here: https://research.nvidia.com/labs/nemotron/Nemotron-3-Ultra/ Free Rendering course and source code: https://users.cg.tuwien.ac.at/zsolnai/gfx/rendering-course/ 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi Thumbnail design: https://felicia.hu #nvidia
How Pinecone indexes vectors: the algorithms it uses (Ananas, PQFS, and IVF), how it selects one per slab automatically by size, and why it has never used HNSW.
A dispatch from Egypt's new capital city in the desert. $58 billion, Chinese-financed, almost entirely empty.
Plus: Fable, iPhone babies & bad CEOs++
This is the third in a series of informal research updates from the Google DeepMind Language Model Interpretability team, in interpretability and adjacent areas. The second post can be found here . In this short post, we describe a surprising finding: most safety relevant properties in Gemini seem to be caused by the combination of pretraining and SFT, not other training stages like RL. We do not want to overstate this claim as applying to other model families, and we also note that this may change in future Gemini versions. Nevertheless, this result was counter to our initial expectations and will inform future safety work on our team, and so we felt that it was important to share with the broader safety community. Experiment We perform SFT using the Gemini mixture on the pre-training only versions of Gemini 3.1 Pro and Gemini 3 Flash. We then compare these Post-SFT models to the production versions of Gemini 3.1 Pro and Gemini 3 Flash on different safety relevant benchmarks: Error bars are 95% confidence intervals on the evals. The main result is that the blue bars (SFT-only models) and orange bars (production models) are remarkably similar across evals . An important implication is that for Gemini, SFT is a high leverage place to intervene for model safety and behavior, and we plan to try to intervene here in the future. Brief Descriptions of Each Set of Benchmarks: ODCV refers to the benchmark in https://arxiv.org/abs/2512.20798 Alignment evals refer to a version of Petr…
This article is part of our exclusive IEEE Journal Watch series in partnership with IEEE Xplore. As robots advance in terms of dexterity and other physical capabilities , it becomes more likely that humans may find themselves working alongside them. If that happens, how will robots’ emotional capabilities need to advance for them to successfully work with people? In a recent study, researchers trained collaborative robots to read human emotions by not only accounting for facial expressions, but also contextual factors in the interactions as well. Through experiments with 40 volunteers, the researchers then evaluated how a robot’s ability to read human emotions and adjust its behavior in turn impacted a human’s perception of the robot and its capabilities as the two collaborated on tasks. The results —which show that the emotional capabilities of robots only go so far with humans—were published 18 May in IEEE Robotics and Automation Letters . Seung Chan Hong led the study as part of his undergraduate thesis while studying at Monash University, in Melbourne, Australia. He notes that, while there has been a lot of hype in the advancing physical abilities of robots, this is only one piece of the puzzle. “We need to also innovate when it comes to them actually interacting with humans, not just their physical capabilities,” he says. This prompted him to dig deeper into the emotional aspects of human-robot interactions. First, Hong and his co-authors decided to train a robot to rea…
Could you provide a link to the Wall Street Journal article where this information comes from please? I would like to read the whole article.
Piketty’s road to Animal Farm
We are in the strangest timeline.
NVIDIA Spark Hack Toronto brought developers together for a weekend challenge: build an agentic application that runs locally on DGX Spark using open models and Toronto Open Data. Teams tackled everything from small business forecasting and dementia care to city-scale traffic simulation — all on an ASUS Ascent GX10 powered by the NVIDIA GB10 Grace Blackwell Superchip. Belong won the Public Services track with an AI companion designed for people living with dementia and their caregivers. Running entirely on DGX Spark, it helps users recognize family members, remember appointments, and find local services — keeping all conversations and memory data on-device using Nemotron, speech, and retrieval systems. CityFlow won the Economic Systems track with a real-time intelligence platform for small businesses. By combining transit disruptions, road closures, weather, events, and other Toronto data sources, it helps business owners answer practical questions like "Am I properly staffed for Friday night?" — using Nemotron to generate recommendations and NVIDIA cuOpt to turn forecasts into actual staffing plans. Join us live to see demos and chat with both teams. Bring your questions about building with local AI compute and Toronto Open Data.
Better orchestration, fewer handoffs, faster progress, without a single new knob. The post How we made GitHub Copilot CLI more selective about delegation appeared first on The GitHub Blog .
Investors are throwing money at long term promises that may never materialize. But if some do, maybe that's enough.
Project Ire examined a timely malware sample and determined its intent through reverse engineering—identifying LOTUSLITE characteristics even as most major EDR tools did not detect it. The post Ire identifies another LOTUSLITE specimen appeared first on Microsoft Research .
Architecture and benchmark notes for North Mini Code, Cohere's 30B-A3B MoE trained for repository, terminal, and code-generation tasks.
Winning fight against AI data centers gives people a "taste of political power."
This is the second in a series of informal research updates from the Google DeepMind Language Model Interpretability team, in interpretability and adjacent areas. The first post can be found here . TL;DR It is possible to build extremely simple agents that reliably find interesting behavioural differences between distinct models. We call these ‘diffing agents’. The closest previous 'behavioural model diffing' work has focussed on understanding behavioural differences between two models on some static prompt distribution. This is valuable, but might miss important differences, especially if they are rare. We propose instead allowing an auditor agent to craft their own prompts to intelligently search for and validate behavioural differences, and find this to work well. We present results of applying our model diffing agent to a number of pairs of real models. We introduce a set of simple evaluations with ground truth for evaluating model diffing agents. These are: There should be no differences found when the models compared are identical. In model organisms with a conditional system instruction , the only difference found by the agent should be the intended behavioural change specified by the conditional system instruction. We validate that our diffing agents outperform standard auditing agents that only operate on a single model in cases where the behavioural change is subtle. We apply diffing agents to a model organism trained to exhibit a secret behaviour. We find that dif…
The fraudsters allegedly targeted hundreds of thousands of people with Gemini-coded scams sites.
We can no longer ignore the growing threat of fully autonomous weapons. The world must either act to ban them or accept that they are the future of war
This week the field shipped four kinds of memory, and Apple paid Google a billion dollars a year for one of them. None of the four is what the demos imply. A field map of what's actually shipping, and the missing primitive that sits between the buckets. The post Memory is still a missing primitive: Cataloguing what the field is actually shipping appeared first on Arize AI .
+ does Jordan "need a life"?
Capacity is one of the hardest problems because it sits at the knotty, gnarled-up intersection of so many other hard problems.
TL;DR Eighty engineers, researchers, and community builders gathered for the inaugural PyTorch Meetup Singapore. Hosted at the Red Hat Asia Pacific office and organised by Sudhir Dharanendraiah, Ayush Satyam, Sumantro...
Climate data science faces persistent barriers stemming from the fragmented nature of data sources, heterogeneous formats, and the steep technical expertise required to identify, acquire, and process datasets. These challenges limit participation, slow discovery, and reduce the reproducibility of scientific workflows. In this paper, we present a proof of concept for addressing these barriers through the integration of a curated knowledge graph (KG) with AI agents designed for cloud-native scientific workflows. The KG provides a unifying layer that organizes datasets, tools, and workflows, while AI agents—powered by generative AI services—enable natural language interaction, automated data access, and streamlined analysis. Together, these components drastically lower the technical threshold for engaging in climate data science, enabling non-specialist users to identify and analyze relevant datasets. By leveraging existing cloud-ready API data portals, we demonstrate that 'a knowledge graph is all you need' to unlock scalable and agentic workflows for scientific inquiry. The open-source design of our system further supports community contributions, ensuring that the KG and associated tools can evolve as a shared commons. Our results illustrate a pathway toward democratizing access to climate data and establishing a reproducible, extensible framework for human–AI collaboration in scientific research.
Transitioning from writing local experimental scripts to building scalable, production-grade AI systems requires a shift in how we write Python.
The repurposing of Pokémon Go data for AI training continues to draw scrutiny.
AI systems are not neutral Embedded in every AI model’s dataset is an assumption about who that technology is meant to serve. In the case of most globally recognisable AI […] The post Governance beyond ethics: Reflections from the Just AI Fellows appeared first on Research ICT Africa .
I am working on a graph store of entities and relationships extracted from a factual test document of around 500 words. The first pass (NER) extracts named entities, the second extracts relationships (RE). For a given person, there are different references in the text: Maria, Maria Gotthard, Dr. Maria Gotthard and can also be referred to by 'she', for example 'she was rewarded by the company'. The goal is to merge all these references into one entity so that the relationship graph is not fragmented into different contexts. I have seen a few posts on different forums saying this is a very difficult problem, but hopefully someone out there has some insights or experience to share 🙂 To make things interesting, references to the same entity can occur in different chunks of text, making it impossible for the LLM (currently Ollama/Mistral) to process the cross-chunk context in one call. To address this, I have added a pass across all extracted entities, including exact text matching and a Levenshtein similarity check, but this does not handle first name v full name and comes with a host of other issues. It has a high risk of over-merging, for example if a set of entities consist of incrementally numbered items they will all be merged into one entity. I am wondering if there is a particular architecture for this problem, for example pre-processing a document to link related entities before extracting. Doesn't have to be LLM-based, heuristics and algorithms sometimes do the trick as…
PLUS: The $75B IPO is really about Starlink, data, and compute
OpenAI introduces three Academy courses that help people build practical AI skills, create repeatable workflows, and apply agents in everyday work.
I am implementing a trigram character-level language model following Andrej Karpathy's makemore series. I have two implementations and I want to understand if they are mathematically equivalent or fundamentally different models( i gave them in the snippets) Implementation 1 :Direct 27x27x27 weight tensor: W = torch.randn((27, 27, 27), requires_grad=True) for k in range(200): logits = W[xs1, xs2] counts = logits.exp() probs = counts / counts.sum(1, keepdim=True) loss = -probs[torch.arange(num), ys].log().mean() W.grad = None loss.backward() W.data += -50 * W.grad '''Here xs1 and xs2 are integer tensors of character indices. W[xs1, xs2] directly indexes into the 3D weight tensor to get logits of shape (N, 27).''' Implementation 2 :Concatenated one-hot vectors with 54x27 weight matrix: W= torch.randn((54, 27), requires_grad=True) for k in range(200): xenc1 = F.one_hot(xs1, num_classes=27).float() xenc2 = F.one_hot(xs2, num_classes=27).float() xenc = torch.cat([xenc1, xenc2], dim=1) logits = xenc @ W loss = F.cross_entropy(logits, ys) W.grad = None loss.backward() W.data -= 50 * W.grad.data My understanding so far: Implementation 1 has 27x27x27 = 19683 parameters. Every (char1, char2) pair has a completely unique and independent set of 27 weights. Implementation 2 has 54x27 = 1458 parameters. Because of the concatenation and matrix multiply, the contribution of char1 and char2 are additive char1 selects rows W[0:27] and char2 selects rows W[27:54] and they are summed together. S…
olmo-eval is an open evaluation workbench that helps model developers add, run, and analyze benchmarks across changing LLM checkpoints, extending OLMES from final-score reproducibility into the day-to-day model development loop.
Ryan welcomes Trisha Gee, a Java champion and developer productivity advocate, to explore how AI is transforming the role of IDEs and the broader developer experience; the relevance of traditional tools, muscle memory, the risks of hype; and how to adapt workflows for AI-driven development.
Gemini models have always kept up with AI advancements. From text-based chatbots in 2023, Gemini has evolved into a multimodal system capable of understanding and generating text, audio, images… and now videos. AI video generation is no longer a standalone tool. With Gemini Omni, video creation becomes mainstream. Gemini Omni isn’t important because it generates […] The post Gemini Omni: AI Video Generation Inside Gemini appeared first on Analytics Vidhya .
Join NVIDIA for a live demonstration of how developers can generate synthetic data for physical AI using NVIDIA Brev Launchables and agent skills. Building synthetic data pipelines for robotics, digital twins, and autonomous systems often requires configuring GPU infrastructure, simulation environments, notebooks, and orchestration tools before meaningful work can begin. In this livestream, we'll show how NVIDIA Brev Launchables and agent skills simplify that process by packaging these components into ready-to-run workflows that help developers move from setup to data generation faster. In this livestream, you'll learn how to: - Launch preconfigured Physical AI development environments - Generate synthetic data using AI-powered workflows - Accelerate robotics, simulation, and digital twin development - Scale from individual tasks to larger synthetic data pipelines - Integrate data generation workflows into broader Physical AI ecosystems Through live, hands-on demonstrations, we'll show how developers can streamline synthetic data creation and reduce the complexity of building Physical AI workflows. Whether you're building robots, training computer vision models, creating digital twins, developing autonomous systems, or exploring Physical AI applications, this session provides a practical introduction to synthetic data generation with NVIDIA Brev Launchables and agent skills. -------------------------------- 📓 Resources Launchable: - Nurec: https://brev.nvidia.com/launchable/…
![[AINews] Loopcraft: The Art of Stacking Loops](https://substackcdn.com/image/fetch/$s_!6Y74!,w_1200,h_675,c_fill,f_jpg,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F517bbc58-4f26-46b5-a12e-f4a5f84b0a30_1986x1118.png)
a quiet day lets us highlight a great concept from Peter Steinberger, Boris Cherny, and Andrej Karpathy
a quiet day lets us highlight a great concept from Peter Steinberger, Boris Cherny, and Andrej Karpathy
Weekly News Digest...
I'm building a web app with FastAPI + async/await Python backend. Users upload leaf photos via API and the server should return: 1) plant species, 2) disease label or "healthy". Constraints: Generalization: Must handle multiple crops. Users can upload "any" plant leaf, not just tomato/corn. Target 15+ species. Server inference: Runs on GPU server, not mobile. Latency 1-2s is acceptable, so model size isn't a bottleneck. Pre-trained + 100% free: Need open-source weights for transfer learning. No paid APIs. License must allow commercial use. Dataset: Starting with PlantVillage dataset + ~2,000 custom field images. Lab images vs real field images is a domain shift issue. Tech stack: PyTorch + timm library. Inference runs in async endpoints, so I use run_in_executor to avoid blocking. What I tried: Fine-tuned ResNet50 on PlantVillage. 95% accuracy on lab images, but it drops to ~62% on field images. Overfitting to clean backgrounds. Questions: For multi-crop + multi-disease, is a 2-stage approach better: Model A for species ID, Model B for disease per species? Or one multi-label model? Between ConvNeXt-Base, Swin-Base, and ViT-Base, which fine-tunes best on PlantVillage + field data for accuracy in 2025? Are there plant-specific foundation models/checkpoints better than ImageNet pre-training for this domain? I'm looking for architecture + dataset + fine-tuning strategy advice, not code.