A series of Ukrainian drone strikes has damaged three Yandex data centers in Russia, knocking out 82 Yandex Cloud services including its YandexGPT AI chatbot and smart home platforms. The severe disruption has compromised Russia's cloud infrastructure and critical backup zones, leading Yandex Cloud to advise customers to migrate their hosting elsewhere.
Sources: HuggingNews, Reddit AI
Multiple AI developments have emerged, highlighted by startup Reflection AI releasing Beam, a 501-billion parameter open-weight model, and Mistral announcing Mistral Large 4 for cybersecurity tasks. Additionally, research labs introduced new benchmarks like Humanity's Sixth Sense and TasteVal to evaluate visual reasoning and research capabilities, while Anthropic updated Claude policies and released Claude Haiku 5.5.
Sources: Digg AI, Import AI
OpenBMB has introduced MiniCPM-V 4.7, featuring the 35B-A3B variant which is the first Mixture of Experts model in the MiniCPM-V series. It introduces Lookahead Expert Scheduling to enable efficient on-device inference on resource-constrained hardware like smartphones.
Sources: Reddit AI
The White House lifted a ban restricting the UK's AI Safety Institute from reviewing frontier models from Google, OpenAI, and Anthropic. This policy reversal allows pre-release safety audits of advanced AI systems following successful lobbying by British officials.
Sources: HuggingNews
Quantum computing company IonQ has agreed to buy cybersecurity firm Elisity in an all-stock deal valued at up to $1.3 billion. The acquisition integrates Elisity's network security capabilities to add an identity and security layer to IonQ's offerings and expands the company into enterprise security.
Sources: HuggingNews
NVIDIA is rumored to be preparing an RTX 5080 with 24GB of GDDR7 memory for release in Q1 2027, which would replace the existing 16GB version. This hardware update is significant for high-performance AI workloads and gaming desktop setups.
Sources: Reddit AI
Microsoft's high-end Surface Laptop Ultra configuration featuring the NVIDIA RTX Spark and 128GB of unified memory has quickly sold out, driven largely by demand for NVIDIA's CUDA software ecosystem in local AI workloads.
Sources: Reddit AI
A new Harvard study analyzing data from over 700 firms reveals that while AI coding tools increase code generation by 30 percent, the gains are absorbed by a 49 percent increase in code review time, resulting in little net change in overall software output.
Sources: Hacker News
A new method called DiffusionART has been developed to fine-tune diffusion models using online reinforcement learning efficiently and stably without needing a differentiable reward model or per-task tuning. It achieves significantly higher training rewards with a fraction of the GPU hours compared to existing methods like DiffusionNFT and Flow-GRPO.
Sources: arXiv AI
A new research paper details Elastic Latent Recurrence (ELR), a training recipe for looped language models that matches non-recurrent model accuracy at a single loop while allowing inference compute to be traded for higher accuracy through additional loops. This approach yields the Koi-4B model checkpoint, which scales effectively across different inference budgets.
Sources: arXiv AI
Researchers introduced SynAct, an active learning framework that strategically allocates scarce expert feedback at the level of individual decisions for LLM agents. The approach improves held-out accuracy across multiple agent benchmarks by prioritizing consequential decisions and distilling corrections into reusable rules.
Sources: arXiv AI
Researchers introduced DistressBench, a benchmark comprising 718 clinician-authored conversations to evaluate how well language models handle suicide and self-harm crises. The study found that while frontier models excel at refusal, many fail to appropriately route users to human help or perform active de-escalation.
Sources: arXiv AI
Users can now integrate GrokBot into Claude Code using the grokbot-webhook skill, allowing Claude to delegate X research tasks and continue workflows with the retrieved information.
Sources: Vault Bookmarks
An independent creator used xAI's Grok Bot to build, launch, and autonomously operate a new AI agent use cases directory, handling features, pull requests, SEO, indexing, and sponsor acquisition.
Sources: Vault Bookmarks
AWS introduced multi-node Ray LLM Deep Learning Containers to support prefill and decode disaggregation across multiple GPU nodes on Amazon EKS. This enables separating compute-heavy prompt processing from latency-sensitive token generation for large language models.
Sources: AWS ML Blog
At AI Engineer New York, Anthropic drew on lessons from Unix and Linux to explain how agents can tackle bigger projects without overwhelming developers.
Tomorrow and Tuesday. If you see this you’re in!
Inside Standard Bots’ AI stack, pretrained models learn factory tasks from demonstrations and improve through corrections from real deployments.
Goodfire researcher Eric Bigelow joins the show to investigate how large language models arrive at decisions at the level of mechanistic interpretability. Drawing on his research into forking paths,…
A lot of people in AI treat evolution as a dumb fallback, basically random search for when you can't take a gradient. Akarsh Kumar thinks that is wrong. Selection hangs on to partial solutions, so…
From the Bitter Lesson of AI scaling to the unsolved mysteries of protein folding, Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido are rethinking what it takes to build AI that truly…
Is the future of frontier AI open or closed? ReflectionAI co-founder and CEO Misha Laskin joins Sarah Guo and Elad Gil to discuss the launch of Beam, the company’s 500 billion parameter open-weight…
A special Science pod and Engineering pod crossover.. with Forward Deployed Engineering kicker!
Nathan Labenz reports back from The Curve conference, sharing off-the-record insights from frontier lab leaders on shortened timelines and whether humanity is approaching an intelligence threshold it…
OutSystems CEO Woodson Martin joins Nathan to discuss how enterprise software can evolve at AI speed without breaking critical operations. Martin explains how an intermediate layer of abstract…