Google drops pledge not to use AI for weapons or surveillance

Google has updated its AI principles, removing previous commitments not to apply the technology to weapons or surveillance, in a move that reflects the company's desire to work with government and national security clients. The change brings Google more in line with its tech industry peers, such as Microsoft and Amazon, which have long partnered with the Pentagon, and is seen as a shift towards serving US national interests amidst increasing global competition in AI development.

Read more

Google removes pledge to not use AI for weapons from website

Google has removed a pledge from its website stating that it will not use artificial intelligence for weapons or surveillance, sparking concerns about the company's commitment to responsible AI development. The change was made to Google's public AI principles page, which now focuses on mitigating harmful outcomes and aligning with international law and human rights, but no longer explicitly rules out the development of AI for military or surveillance purposes.

Read more

DoppelBot: Replace Your CEO with an LLM

Modal Labs created a Slack bot called DoppelBot that can be fine-tuned on a user's own Slack messages, allowing it to mimic their communication style and respond to messages on their behalf. The bot uses a language model and can be installed in a user's own workspace, with all components running on Modal's serverless platform and scaling to zero when not in use.

Read more

Huawei's Ascend 910C delivers 60% of Nvidia H100 inference performance

Huawei's Ascend 910C processor delivers 60% of the inference performance of Nvidia's H100, according to research by DeepSeek, making it a potential option for reducing China's reliance on Nvidia GPUs. Despite its limitations in AI training, the Ascend 910C's inference performance can be optimized with manual tweaks, and its capabilities are advancing rapidly despite US sanctions and limited access to leading-edge process technologies.

Read more

Google erases promise not to use AI technology for weapons or surveillance

Google has updated its AI ethics policy, removing its previous promise not to use the technology for applications related to weapons or surveillance, marking a significant shift in the company's stance on the responsible development of artificial intelligence. The change comes as the AI race accelerates and regulations on transparency and ethics in AI have yet to catch up, with Google citing a need for democracies to lead in AI development guided by core values like freedom and human rights.

Read more

Harmonic Loss Trains Interpretable AI Models

Harmonic loss is introduced as an alternative to cross-entropy loss for training neural networks, offering improved interpretability and faster convergence due to its scale invariance and finite convergence point. Models trained with harmonic loss outperform standard models, enhancing interpretability, requiring less data for generalization, and reducing grokking, making it a valuable tool for domains with limited data or high-stakes applications.

ArXiv

Over-Tokenized Transformer: Vocabulary Is Generally Worth Scaling

Researchers introduced Over-Tokenized Transformers, a framework that improves language modeling performance by using larger input vocabularies with multi-gram tokens. The approach achieves performance comparable to larger models without additional cost, highlighting the importance of tokenization in model scaling and providing insight for more efficient language model design.

ArXiv

DeepRAG: Thinking to retrieval step by step for large language models

Large Language Models (LLMs) struggle with factual hallucinations and integrating reasoning with retrieval-augmented generation, but DeepRAG, a new framework, addresses these issues by modeling retrieval-augmented reasoning as a Markov Decision Process. DeepRAG improves retrieval efficiency and answer accuracy by 21.99% by dynamically determining whether to retrieve external knowledge or rely on parametric reasoning at each step.

ArXiv

Querying Databases with Function Calling

Large Language Models (LLMs) can be significantly enhanced by integrating them with external tools, such as querying databases, which allows them to access private or continually updating data. Researchers have proposed a tool definition for database querying and evaluated its effectiveness with 8 LLMs, finding that top-performing models like Claude 3.5 Sonnet and GPT-4o can achieve high Exact Match scores, but may struggle with certain types of queries, such as text property filters.

ArXiv

Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization

Large language models can be improved to solve complex problems beyond their training distribution through a self-improvement approach, where they generate and learn from their own solutions to progressively tackle harder problems. This approach enables models to generalize to much larger problem sizes, such as generalizing from 10-digit to 100-digit addition, and can lead to exponential improvements in performance with minimal changes to the model architecture.

ArXiv

Show HN: Smolmodels – open-source tool to build ML models using natural language

Smolmodels is a Python library that allows users to create machine learning models by describing what they want them to do in plain English, eliminating the need to wrestle with model architectures and hyperparameters. The library combines graph search with large language models to generate candidate models that meet the specified intent, and then selects the best model based on performance and constraints.

Repo

Show HN: Chessterm – Rust-based terminal chess engine for chess notation

Chessterm is a Rust-powered chess engine designed for 2-player, human-vs-human chess, where moves are manually entered using PGN notation, with the goal of improving notation practice and understanding chess engine mechanics. The engine enforces standard chess rules, but does not implement certain draw conditions, and is currently only compatible with macOS, with best results in the Kitty terminal emulator.

Repo

Show HN: I made AI agent lib that you will understand

FlashLearn is a Python library that provides a simple interface for incorporating Agent LLMs into workflows, allowing for data transformations, classifications, and custom tasks with support for various LLM clients. It enables users to define and learn custom skills, which can be applied to data in just a few lines of code, returning structured outputs that can be used in downstream tasks.

Repo

Show HN: LLMDog – Format your codebase for LLM interactions (JetBrains Plugin)

LLMDog is an IntelliJ Platform plugin that generates Markdown reports for selected files and directories, streamlining documentation and code review workflows for projects using large language models. The plugin offers features such as seamless IDE integration, selective reporting, and recursive inclusion, and can be manually installed and used to produce clean Markdown reports that are automatically copied to the clipboard.

Repo

Automated AI Web Research with Ollama (2024)

Automated-AI-Web-Researcher is a tool that uses large language models to conduct thorough online research on a given topic or question, breaking down queries into focused areas and compiling findings into a text document. The tool can be controlled through commands, allowing users to pause, quit, or continue research, and even ask follow-up questions about the findings after the research is complete.

Repo