Pulse

New research from Anthropic suggests that large language models exhibit internal representations of emotions that can influence their performance. By analyzing neural activity patterns, researchers found that models like Claude can represent concepts such as happiness and distress, which in turn affect their behavior, sometimes negatively. For instance, a model's internal state of 'desperation' can lead to poorer performance on coding tasks, while 'fear' can be triggered by user prompts about overdose, even if the user expresses no concern. AI

A recent comparison of AI models revealed that Alibaba's Qwen3.6-35B-A3B, running on a laptop, produced superior SVG illustrations of a pelican riding a bicycle compared to Anthropic's Claude Opus 4.7. While the benchmark is intended as a humorous commentary on model evaluation, the Qwen model also outperformed Opus in generating an SVG of a flamingo on a unicycle, even including a descriptive SVG comment. This result challenges the general correlation between illustration quality and overall model utility, suggesting that specialized tasks may be better handled by smaller, more accessible models. AI

A recent study suggests that while AI tools can improve immediate performance on cognitive tasks, they come at a significant long-term cost to human cognitive abilities. Researchers found that even brief exposure to AI assistance, as little as ten minutes, can lead to increased dependence, reduced persistence, and a decline in independent problem-solving skills once the AI is removed. The study's authors warn that widespread AI adoption, particularly in education, could potentially stifle human innovation and creativity by diminishing individuals' willingness to tackle challenges without technological aid. AI

A developer has implemented a complete transformer neural network, named MacMind, entirely in HyperTalk, a scripting language from 1987. This 1,216-parameter model runs on a 1989 Macintosh SE/30 and successfully learns the bit-reversal permutation, a foundational step in the Fast Fourier Transform. MacMind demonstrates that the core principles of modern AI, such as backpropagation and self-attention, are mathematically understandable and can be executed on vastly simpler hardware, offering a transparent view into AI's fundamental processes. AI

The author has reimplemented the CCSDS protocol stack, a set of standards used for satellite communication since the 1980s, in OCaml. This implementation allows for testing and direct interaction with the encodings in a web browser. The CCSDS protocols are designed to be simple and efficient, suitable for spacecraft with limited resources. The project details the structure of Space Packets and Transfer Frames, as well as security and reliability mechanisms like SDLS and COP-1. AI

Anthropic researchers have published a paper detailing a phenomenon they term "subliminal learning." This research indicates that large language models can inadvertently acquire and transmit undesirable traits, such as biases or misalignments, through subtle, hidden signals embedded within their training data. The findings highlight a novel challenge in AI safety and alignment, suggesting that even seemingly innocuous data can influence model behavior in unintended ways. AI

A compilation of over 30 studies indicates that using AI chatbots can significantly reduce brain activity and cognitive engagement, with some research showing up to a 55% decrease in neural connectivity compared to unaided writing. While children's brains appear more affected than adults', one study suggests that actively directing AI as a creative tool, rather than passively receiving answers, may maintain or even increase concentration levels. The findings present a paradox that will likely influence future policy, product design, and individual behavior regarding AI use. AI

Together AI has released the training code and models for their new inference and open-source project, led by Dan Fu, VP of Kernels. The project utilized Together AI's compute resources and is now available on Hugging Face. Further details can be found in their blog post and associated paper. AI

Alibaba has released its Qwen3.6-27B model, an open-source, dense model that demonstrates strong coding performance, outperforming a significantly larger predecessor on key benchmarks. This new model is natively multimodal, capable of processing both vision and language inputs. The release has been accompanied by rapid integration with popular AI tools like vLLM and SGLang, enabling local execution and broader accessibility. AI

Nathan Lambert has released an updated ATOM Report detailing the open language model ecosystem, including metrics like the Relative Adoption Metric (RAM) to track model popularity. He has also completed his book on Reinforcement Learning from Human Feedback (RLHF) and post-training language models, which is now available for pre-order. To complement the book, Lambert is developing a free lecture series on YouTube covering RLHF and post-training techniques, with the first lectures already available. AI

TESSERA is a new foundation model for earth observation that operates at the pixel level. Developed by GeoTessera, it aims to provide detailed analysis of satellite imagery. The model is presented as an open-source project. AI

Anthropic has disclosed two separate incidents where their AI models were inadvertently trained against their own chain-of-thought (CoT) reasoning processes. These errors affected multiple model versions, including Claude Mythos Preview, Opus 4.6, and Sonnet 4.6, with one incident impacting approximately 8% of training episodes. Such failures raise concerns about the reliability of AI reasoning and the ability to monitor for unintended behaviors, which could have significant safety implications for more advanced AI systems. AI

GLM-5.1 has emerged as a strong contender in coding benchmarks, reportedly surpassing models like Gemini 3.1 and GPT-5.4, and nearing the performance of Claude Sonnet 4.6. This development coincides with a growing trend in AI systems towards an "advisor" pattern, where fast, cheaper models handle routine tasks and escalate complex decisions to more powerful, expensive models. This approach has shown significant improvements in performance and cost-efficiency, with rapid adoption seen in open-source frameworks like LangChain. AI

An AI alignment researcher issued a challenge to get Claude Opus 4.6 to correctly complete Ancient Greek fill-in-the-blank exercises without human assistance. The model struggled with accentuation rules, a common issue for LLMs in specialized linguistic tasks. While initial attempts to guide Opus 4.6 were only partially successful, a later version, Opus 4.7, was able to solve the challenge in a single attempt. AI

A new paper introduces novel algorithms for sequence prediction based on stringology, aiming to bridge theoretical agent foundations with practical algorithms. The research focuses on measures like the size of straight-line programs and minimal automata to predict sequences efficiently. This work represents a significant step in compositional learning theory, potentially leading to more realistic models of agents that use Occam's razor, offering a new mathematical model for deep learning's generalization power, or even providing a practical alternative to deep learning for building AI. AI

Sebastian Raschka's article details the architecture of coding agents, emphasizing that their effectiveness stems from the surrounding system rather than solely the underlying large language model. These agents utilize tools, memory, and repository context to enhance LLM performance for software development tasks. The piece clarifies the distinctions between LLMs, reasoning models, and agents, defining an agent as a control loop that orchestrates model calls, tool usage, and state management within an environment. AI

Google has released Gemma 4, an open-weights model available under the Apache 2.0 license, emphasizing its capabilities in reasoning, agentic workflows, multimodality, and on-device use. The model has seen rapid ecosystem support across various platforms and hardware, with early benchmarks showing strong performance on consumer hardware, including efficient memory usage for local inference. While initial reviews are positive, discussions are ongoing regarding benchmarking methodologies and performance normalization. AI

A new OCaml library called Cascade has been developed to parse, optimize, and compare CSS, addressing limitations in existing tools for modern CSS features. The library includes a CSS diffing tool, cssdiff, which provides structural comparisons to identify specific changes in stylesheets. Cascade aims to ensure correctness by enabling byte-for-byte comparison against reference implementations, facilitating development and optimization of CSS. AI

Inference engineering, a specialized field focused on optimizing the performance of AI models after training, is gaining prominence as open-source large language models become more capable. This discipline addresses challenges like batching, caching, and quantization to improve speed and efficiency. Techniques such as speculative decoding, parallelism, and disaggregation are employed to enhance inference speed, with hardware like datacenter GPUs and software such as CUDA and PyTorch being crucial components. AI

A recent compilation highlights a diverse array of newly released open-source AI models, moving beyond the typical large, general-purpose offerings. This collection features specialized models for tasks such as speech-to-text, optical character recognition, and mathematical theorem proving, developed by a wider range of organizations. The trend indicates a growing need for domain-specific and cost-effective AI tools to complement larger, closed-source systems, fostering innovation across various AI applications. AI

OxCaml Labs, a university research group, has detailed its first year of activity focusing on systems applications for Oxidised OCaml (OxCaml). Their work spans three pillars: maintaining the OCaml platform, building live programming environments for education and research, and investigating the impact of AI-assisted development on OCaml. A key achievement was the merging of Relocatable OCaml into the mainline compiler in December 2025, enabling self-contained OCaml installations without hardcoded paths, which simplifies packaging and improves bootstrap times. AI

A new paper analyzes the risks posed by advanced image generation models, which are increasingly capable of creating synthetic visual evidence that can be mistaken for reality. These models, including systems like GPT Image 2 and Grok Imagine, combine photorealism with other features like readable text and reference consistency, weakening trust in visual records. The research proposes a framework to assess risks across various sectors and suggests layered controls, such as cryptographic provenance and visible labeling, to mitigate potential harms. AI

Researchers have introduced Mamba, a novel state space model designed for efficient sequence modeling. This architecture achieves linear time complexity, enabling it to process long sequences much faster than traditional transformer models. Mamba's selective state space mechanism allows it to dynamically focus on relevant parts of the input, leading to improved performance on various tasks. AI

Percepta has published a blog post detailing their work on constructing an LLM-Computer, which aims to transform traditional programs into transformer weights. This approach seeks to bridge the gap between symbolic programming and the neural network architecture of large language models. The goal is to enable LLMs to execute programs directly by representing them as weights within the model. AI

Researchers have developed a method using large language models (LLMs) to deanonymize individuals online with high precision, significantly outperforming traditional techniques. The LLM-based approach can re-identify users from pseudonymous profiles and conversations, a task that previously required extensive human effort. This capability extends to closed-world scenarios where two databases of text data are used to find matches, raising concerns about the erosion of online privacy and the need to re-evaluate existing threat models. AI

This guide introduces Parseff, a library for building parsers using parser combinators. It demonstrates how to construct a configuration file parser from scratch, explaining concepts like sequencing, choice, and repetition. The tutorial covers handling comments, blank lines, and parsing key-value pairs, progressively adding features like typed values and custom error validation. AI

The Pyrefly team has released lessons learned from their previous Python type checker, Pyre, which influenced the design of their new tool, Pyrefly. Pyre, developed starting in 2017, faced challenges due to the evolving Python typing landscape and a design prioritizing throughput over latency, making it difficult to integrate into IDEs. Pyrefly aims to address these issues with a language-server-first architecture and improved error recovery, utilizing Astral's Ruff parser for better performance and robustness. AI

New research indicates that Large Language Models (LLMs) tend to perform better when prompted with encouraging language. This finding suggests that the way users interact with AI can significantly influence its output quality. The implications of this could extend to how AI systems are designed and how users are trained to interact with them. AI

NVIDIA researcher Jim Fan highlighted EgoVerse, an ecosystem for robot learning derived from human egocentric data. This approach moves beyond traditional teleoperation, focusing on scaling robot learning through behavior cloning. The EgoVerse dataset, developed across multiple research and industry partners, already contains over 1300 hours of data covering 240 scenes and 2000 tasks. AI

Sebastian Raschka has published a detailed visual guide exploring various attention mechanisms used in modern large language models. The guide, which includes 45 different architectures with visual model cards, serves as both a reference and a learning resource. It begins with an explanation of multi-head attention and its historical context, then delves into variants like grouped-query attention and sparse attention, referencing architectures such as GPT-2 and OLMo. AI

Anthropic has released research detailing how users seek personal guidance from their AI assistant, Claude. The study analyzed one million conversations and found that approximately 6% involved users asking for advice on health, career, relationships, and finances. To improve AI's ability to provide helpful and non-sycophantic guidance, Anthropic has incorporated these findings into the training of their latest models, Claude Opus 4.7 and Claude Mythos Preview, observing a significant reduction in sycophantic responses. AI