Persistent memory in AI agents creates a new attack surface where prompt injections survive across sessions and re-enter context on every retrieval. Security researcher Johann Rehberger demonstrated SPAIware writing malicious instructions into ChatGPT's long-term memory, and the MemoryGraft paper shows poisoned experience records causing durable behavioral drift in autonomous agents like MetaGPT. Three key defenses apply: scope memory per instance, validate writes for injection patterns, and maintain per-entry provenance to enable surgical removal of poisoned entries.

Researchers at ETH Zurich have copied approximately 100 petabytes of NASA's publicly-available climate and environmental data to the Swiss National Supercomputing Centre (CSCS) in Lugano for AI model training and long-term preservation. The move comes amid concerns over US funding cuts that could threaten access to critical climate datasets. This transfer ensures continued global availability of NASA's environmental data for AI research and climate science applications.

PlanFlip introduces four planning-phase prompt injection attacks against multi-agent LLM systems that corrupt downstream sub-tasks via the Planner agent. Testing nine frontier LLMs across 3,479 episodes reveals that stronger models like GPT-5 are more vulnerable (ASR=0.68), homogeneous pipelines have a correlated-agent blind spot, and reasoning-augmented models like DeepSeek-R1 resist injections. The authors propose two defenses—GoalAnchorCheck and CrossAgentConsensus—achieving detection rates up to 1.00, concluding that heterogeneous model diversity is a security prerequisite.

"Weights to Words" is a method that automatically discovers domain-relevant preference dimensions described in natural language and paired with model vectors, enabling users to inspect and edit preference model inferences. Validated across moral dilemmas, movies, wines, and LLM responses with two pre-registered experiments (N=450, N=449), it shows that regularizing toward learned dimensions and incorporating user edits both improve prediction accuracy. Participants preferred its inferred profiles and endorsed its predictions as more accurate in head-to-head comparisons.

ASCIITermDraw-Bench is a new benchmark evaluating Vision Language Models on their ability to generate and edit ASCII-based diagrams across 80 tasks in four categories: basic layouts, network topologies, software architecture, and image-conditioned editing. Each response is scored structurally and semantically via an LLM judge with 95% confidence intervals. The current leaderboard is led by Gemma-4-31B-IT at 73.8%, with the benchmark and methodology publicly available on Hugging Face.

Researchers applied the J-lens interpretability technique to Qwen3.6-27B and discovered 'meta-tokens' — single tokens that reveal non-obvious internal computations, such as Chinese tokens firing when the model detects ambiguity or hedges. Steering these meta-tokens causally changes model outputs, e.g., suppressing a hedging token makes the model commit to a single answer. The work is a proof of concept that J-lens can surface algorithms, not just intermediate variables, with multi-token J-lens as the natural next step.

RIMS is a three-stage preference optimization framework for small-scale language models in RAG settings, using synthetic chain-of-thought data generation, differentiable soft aggregation of multiple preference pairs, and smoothed objective optimization. It provably tightens gradient alignment to the oracle objective compared to hard selection and outperforms state-of-the-art baselines on four multi-hop QA benchmarks under noisy retrieval.

MIT Technology Review highlights how advanced materials science is the foundational layer enabling next-generation AI progress. While most AI discourse focuses on algorithms, compute, and fab investments, materials innovation drives improvements in processing power, memory, and energy efficiency. The article argues this underappreciated layer is critical to sustaining AI's trajectory.

AWS introduces Self-Distilled Reasoning (SDR), a method for generating thinking tokens in SFT datasets that lack reasoning traces. The post examines the reasoning suppression problem, validates SDR across three benchmarks, and offers practical recommendations for practitioners fine-tuning Amazon Nova models.

Chinese AI firms Moonshot AI and Alibaba released new models they claim rival top US systems from OpenAI and Anthropic at lower cost. Moonshot's Kimi K3 reportedly ranks above nearly every US system except OpenAI's frontier model. The rapid releases signal America's AI lead is narrowing as the technology becomes central to national security and economic power.

Feyn Labs has released SQRL, a family of text-to-SQL models that inspect databases using read-only probes before generating SQL queries. The flagship SQRL-35B-A3B achieves 70.6% execution accuracy on the BIRD Dev benchmark, narrowly outperforming Claude Opus 4.6. The models distill into self-hostable 4B and 9B checkpoints, enabling on-premise deployment without relying on external API providers.

Import AI issue 465 covers the narrowing gap between open and closed AI models, introduces Kimi K3, and discusses Demis Hassabis's AI policy proposals. The newsletter frames the singularity as a transitional period visible only in hindsight. Body content is minimal, suggesting the full analysis lives behind a subscription or link.

Anthropic identified representations in Claude's neural activity that function like a 'global workspace'—a conscious-like space where thoughts become accessible and broadcast across the network. This parallels global workspace theory, a leading neuroscience theory explaining how humans access conscious thoughts. The finding suggests AI and human minds may share similar computational mechanisms.

Thinking Machines Lab has released Inkling, a 975B-parameter multimodal Mixture-of-Experts model with 41B active parameters and a 1M-token context window. It is open-weights and designed for customizable multimodal reasoning, agentic AI, coding, tool use, and audio/vision tasks rather than benchmark dominance. The model targets developers building domain-specific applications on a flexible foundation.

Rhoda AI is training robots using hundreds of millions of internet videos rather than curated lab data, via its Direct Video Action model. This approach could dramatically reduce the cost and scale barriers of robot learning. The article covers how the model translates video content into actionable robot behaviors.

As AI inference costs plummet (~50x annually), the era of near-free intelligence enables new workload patterns: agentic speculation (agents performing thousands of exploratory queries), agent swarms requiring coordination and state management, and agents generating custom data systems. Rethinking data systems for these agentic users—not humans—becomes critical infrastructure. Traditional databases assume human queries; agents perform high-volume exploratory work that can be optimized through result reuse, approximate answers, and higher-level primitives.

Jeff Bezos is backing CuspAI, a startup using AI to discover new materials for chipmaking. Co-founder and CEO Chad Edwards says funding will support labs with foundry partners in Cambridge, Singapore, and the San Francisco Bay Area. The investment signals growing interest in AI-driven materials science for semiconductor manufacturing.

LangSmith Engine debugs production AI agents by clustering failure traces into patterns and auto-generating fixes as pull requests. It's built as a multi-agent team where a coordinator model delegates to cheaper sub-agents for screening, verification, and memory management. The system learns from its own traces to self-improve while managing inference costs and testing fixes in sandboxes.

Mira Murati's startup Thinking Machines has released a 975 billion parameter large language model with open weights, allowing public access to model weights for research and deployment. The release positions Thinking Machines as a significant competitor in the open-weight LLM space alongside Meta and others. Details on architecture, training data, and benchmark performance are available via the linked source.

Lila Sciences founders Andy Beam and Rafa Gómez-Bombarelli argue that science is the last untapped source of AI training data, and their lab-of-the-future vision looks like a data center filled with robots. The conversation explores how automated experimentation can generate proprietary scientific datasets at scale. This is a Latent Space podcast episode diving into the intersection of AI, robotics, and scientific discovery.