Alexa+

Amazon has introduced Alexa+, its next-generation AI assistant powered by generative AI, which enables more conversational, personalized, and smarter interactions, allowing users to get things done, stay entertained, and manage their homes with ease. Alexa+ can understand and respond to natural language, orchestrate across tens of thousands of services and devices, and even navigate the internet to complete tasks on behalf of the user, making it a highly capable and intuitive assistant.

Read more

Y Combinator Supports AI Startup Dehumanizing Factory Workers

Y Combinator is supporting Optifye.ai, a startup that uses AI to monitor factory workers' performance, tracking their hand movements and output, which some critics see as dehumanizing surveillance. The company's launch video features a skit where a "boss" yells at a "worker" by calling them a number, highlighting the potential for the technology to be used to belittle and intimidate workers.

Read more

ForeverVM: Run AI-generated code in stateful sandboxes that run forever

ForeverVM is a code execution API that securely runs arbitrary Python code in a remote sandbox, allowing for stateful execution without session expiration. It uses memory snapshots to swap idle machines to disk, improving scalability and resource usage, and provides a REPL interface for interaction, as well as a CLI and API for integration with various applications and services.

Read more

Y Combinator Supports AI Startup Dehumanizing Factory Workers

Y Combinator is supporting Optifye.ai, a startup that uses AI to monitor factory workers' performance, tracking their hand movements and output to provide efficiency metrics to bosses. The company's pitch includes a video that demonstrates a dehumanizing approach to worker management, with a "boss" yelling at a "worker" and referring to them by a number, highlighting concerns about the potential for surveillance and exploitation of workers.

Read more

Show HN: Emdash – Slack/Zoom alternative for distributed team collaboration

Emdash is a platform designed to streamline teamwork and knowledge management for hybrid teams, offering features such as video and chat, AI-powered search, and integrated discussions to help teams stay aligned and informed. The platform aims to reduce distractions, minimize meetings, and maximize productivity by providing a centralized workspace for teams to collaborate, share information, and make decisions.

Read more

Diffusion LLM Has Arrived

LLaDA, a diffusion model, challenges the dominance of autoregressive models in large language models by demonstrating strong scalability and competitive performance with state-of-the-art models like LLaMA3 and GPT-4. Through extensive benchmarks, LLaDA shows impressive results in in-context learning, instruction-following, and reversal tasks, establishing diffusion models as a viable alternative to traditional autoregressive models.

ArXiv

The FFT Strikes Back: An Efficient Alternative to Self-Attention

FFTNet is a framework that uses the Fast Fourier Transform to achieve global token mixing in O(n log n) time, making it more scalable than conventional self-attention mechanisms. By transforming inputs into the frequency domain and using a learnable spectral filter, FFTNet efficiently captures long-range dependencies and outperforms traditional self-attention and fixed Fourier models in experiments.

ArXiv

A Comprehensive Formal Security Analysis of OAuth 2.0

The OAuth 2.0 protocol was formally analyzed in an expressive web model to establish strong authorization, authentication, and session integrity guarantees, covering all four OAuth grant types and considering various malicious scenarios. The analysis revealed four vulnerabilities that can be exploited in practice, also affecting OpenID Connect, and proposed fixes, ultimately proving the security of the fixed OAuth version with best practices in place.

ArXiv

Improving Consistency in Large Language Models Through Chain of Guidance

Consistency is a crucial aspect of trustworthiness in Large Language Models (LLMs), but currently, there is no known mechanism to control and guide LLMs to be more consistent at inference time. This paper introduces a novel alignment strategy called Chain of Guidance (CoG) that maximizes semantic consistency in LLM outputs, resulting in models that are more than twice as consistent as base models and show strong generalization capabilities.

ArXiv

MITRE's Offensive Security Evaluation Framework for LLMs

Researchers have developed a framework called OCCULT to assess the potential cyber security risks of large language models (LLMs) and AI used for offensive cyber operations, and their preliminary results show significant advancements in AI's ability to scale realistic cyber threats. The evaluation found that some models, such as DeepSeek-R1, can correctly answer over 90% of challenging offensive cyber knowledge tests, demonstrating the growing capabilities of AI in this area.

ArXiv

DeepGEMM: clean and efficient FP8 GEMM kernels with fine-grained scaling

DeepGEMM is a library designed for efficient FP8 General Matrix Multiplications (GEMMs) with fine-grained scaling, supporting both normal and Mix-of-Experts (MoE) grouped GEMMs, and its performance matches or exceeds expert-tuned libraries across various matrix shapes. The library is written in CUDA, has a simple design with only one core kernel function, and is optimized for NVIDIA Hopper tensor cores, making it a clean and accessible resource for learning Hopper FP8 matrix multiplication and optimization techniques.

Repo

Show HN: LLM plays Pokémon (open sourced)

The Fire Red Agent project aims to create an autonomous AI agent that can play Pokémon FireRed using a large language model, with capabilities to explore, battle, and respond to game events. The project uses a combination of emulator integration, game memory management, navigation, and LLM integration to enable the agent to make decisions and interact with the game, but development was paused due to technical hurdles with programmatic input control.

Repo

DeepSeek Open Source Optimized Parallelism Strategies, 3 repos

The DeepSeek Infra project is sharing profiling data from their training and inference framework to help the community understand their communication-computation overlap strategies and low-level implementation details. The profiling data, which can be visualized in Chrome or Edge browsers, includes training and inference profiles with various configurations, such as different numbers of expert layers, sequence lengths, and batch sizes, to demonstrate their overlapping strategy for computation and communication.

Repo

DualPipe: Bidirectional pipeline parallelism algorithm

DualPipe is a bidirectional pipeline parallelism algorithm that achieves full overlap of forward and backward computation-communication phases, reducing pipeline bubbles and improving efficiency. It is introduced in the DeepSeek-V3 Technical Report and can be used with PyTorch 2.0 and above, requiring a custom implementation of the overlapped_forward_backward method for real-world applications.

Repo

Datahawk – Text data browser for NLP, LLM researchers and developers

Datahawk is a lightweight app that allows for easy browsing and analysis of text data in various formats, including HuggingFace and JSONL, with features such as intuitive navigation, efficient browsing, and powerful analysis. The app can be installed with Python 3.8 or later using pip, and launched from anywhere, providing an intuitive interface with on-screen instructions for seamless data browsing and analysis.

Repo