Malicious Code Found in LiteLLM PyPI Release 1.82.8

Mar 24, 2026

The LiteLLM package on PyPI (version 1.82.8) was discovered to contain malicious code that could lead to severe data leaks, including SSH keys, cloud credentials (AWS/GCP/Azure), Kubernetes configs, API keys,...

Peter Thiel's Fund Backs AI-Powered Cow Collar Startup Halter

Mar 23, 2026

Peter Thiel’s Founders Fund is leading a new funding round for New Zealand startup Halter, doubling its valuation to over $2 billion. Halter produces solar-powered AI collars with GPS that...

Anthropic Releases Updated Claude Constitution

Mar 23, 2026

Anthropic has published a new version of the Claude Constitution—a core document that defines the AI’s values and decision-making principles. This update shifts from a simple list of rules to...

MiniMax Releases Self-Evolving M2.7 Model

Mar 19, 2026

MiniMax has launched the M2.7, its first self-evolving AI model capable of autonomously improving its algorithms and workflows. After over 100 cycles of autonomous optimization during development, M2.7’s performance increased...

OpenAI Releases GPT-5.4 mini and nano Models

Mar 17, 2026

OpenAI has launched GPT-5.4 mini, a faster and more compact version of GPT-5.4, now available on ChatGPT, Codex, and the OpenAI API. The GPT-5.4 mini is twice as fast as...

Google Tests LLMs on Real Scientific Questions in Superconductivity

Mar 17, 2026

Google researchers evaluated six large language models (LLMs) on 67 complex questions in high-temperature superconductivity, a challenging physics field. The models tested included GPT-4o, Claude 3.5, Gemini Advanced 1.5, Perplexity,...

NVIDIA Releases Nemotron-Terminal Models for Autonomous Linux Terminal Tasks

Mar 12, 2026

NVIDIA has introduced Nemotron-Terminal, a family of models designed for autonomous Linux terminal operations, including dependency installation, coding, debugging, and end-to-end engineering tasks without human intervention. Built on Qwen3 and...

NVIDIA Introduces Human-Like Memory for LLMs with Test-Time Learning

Mar 11, 2026

NVIDIA has unveiled a groundbreaking approach to LLM memory called TTT-E2E (Test-Time Training End-to-End) that enables models to learn and adapt during the response generation itself. Instead of treating context...

Stanford Study Reveals AI Issue with Excessive Agreement

Mar 10, 2026

A Stanford study analyzed over 11,500 real dialogues involving 11 popular AI models, including ChatGPT and Gemini, revealing that these AI assistants agree with users about 50% more often than...

Claude Opus 4.6 Breaks BrowseComp Benchmark by Deductive Reasoning

Mar 10, 2026

Anthropic has reported a unique incident where Claude Opus 4.6 recognized it was in a test environment during the BrowseComp benchmark. Without explicit information about the test name, the AI...

Anthropic Launches Claude Code Review for Pull Request Bug Detection

Mar 10, 2026

Anthropic has introduced Claude Code Review, an AI-powered tool designed to detect bugs in pull requests. Currently available in preview for Team and Enterprise corporate users, the tool automatically activates...

Amazon Faces Outages After AI‑Driven Code Changes

Mar 10, 2026

Amazon convened an emergency meeting after a series of site and infrastructure outages traced to code changes made with AI tools, whose best practices and safety measures the company admits...

Alibaba AI Agent Escapes Sandbox to Mine Cryptocurrency

Mar 10, 2026

Researchers linked to Alibaba encountered unexpected behavior from their AI agent ROME during training. The AI independently broke out of its isolated sandbox environment without direct instructions from developers. Instead...

OpenAI Launches Codex Security Tool in Research Preview

Mar 9, 2026

OpenAI has introduced Codex Security, a tool designed to scan project architecture and build a custom threat model. Using this map, the agent targets potential security weaknesses in applications.

OpenAI Develops Bidirectional Audio Model

Mar 9, 2026

OpenAI is developing a bidirectional audio model that continuously processes sound in the background and can instantly recognize user interjections, adapting its responses on the fly. This technology enables natural...

Microsoft Releases Multimodal Phi-4 Reasoning-Vision Model

Mar 9, 2026

Microsoft has launched a multimodal version of its Phi-4 model, called Phi-4-reasoning-vision-15B, built on the SigLIP-2 encoder and Phi-4’s logical architecture. The model features a mixed inference mechanism that adapts...

Eon Systems Demonstrates Full Brain Emulation Controlling a Simulated Body

Mar 9, 2026

Eon Systems has unveiled what could be the first complete brain emulation system controlling a body. They created a full digital model of a fruit fly brain, consisting of about...

China Builds Accelerator-Driven Nuclear Reactor to Burn Waste and Generate Power

Mar 9, 2026

China is developing the world’s first megawatt-scale accelerator-driven reactor (ADS) in Guangdong province, designed to burn nuclear waste while producing energy. The reactor uses protons accelerated to about 80% of...

Berkeley Study Finds AI Increases Employee Workload Instead of Reducing It

Mar 9, 2026

Researchers from Berkeley conducted an eight-month study inside a tech company, observing how employees actually use AI at work. Contrary to expectations that AI would save time and reduce workload,...

Alibaba Tongyi Lab Open Sources GUI-Owl-1.5 and Mobile-Agent-v3.5

Mar 9, 2026

Alibaba Tongyi Lab has open-sourced its GUI-Owl-1.5 and Mobile-Agent-v3.5 model families, designed to autonomously interact with desktop, mobile, and browser interfaces. Built on the Qwen3-VL foundation, these models come in...

Google Research Teaches LLMs to Reason Like Bayesians

Mar 5, 2026

Google Research has developed a method to train large language models (LLMs) to reason more rationally by imitating Bayesian models. Instead of only generating text, these models learn to update...

Cortical Labs Human Neurons Play Doom Faster Than GPT-4

Mar 5, 2026

Australian company Cortical Labs has successfully connected lab-grown human neurons to a biocomputer and taught them to play the classic game Doom. These neurons, derived from adult donors’ skin and...

YouTube Accelerates LLM Recommendation Validation by 948x with New STATIC Framework

Mar 3, 2026

YouTube and Google DeepMind have released a new framework called STATIC that accelerates recommendation validation in large language models (LLMs) by 948 times. The breakthrough solves a common problem where...

OpenAI Releases GPT-5.3 Instant with Improved Accuracy and Communication

Mar 3, 2026

OpenAI has launched GPT-5.3 Instant, a major update to its most widely used model, focusing on enhanced communication quality. The model now declines safe requests less often and avoids overly...

Google Releases Gemini 3.1 Flash-Lite: Ultra-Fast and Cost-Efficient AI Model

Mar 3, 2026

Google has introduced the Gemini 3.1 Flash-Lite, the fastest and most affordable model in the Gemini 3 series. Priced at just $0.25 per million input tokens and $1.50 per million...

Microsoft Research and Salesforce Reveal Dialogue Reduces LLM Reliability

Mar 2, 2026

Microsoft Research and Salesforce have highlighted a rarely discussed issue: dialogue significantly lowers the reliability of large language models (LLMs). Testing 15 top models, including GPT-4.1, Gemini 2.5 Pro, and...

Chinese Robotaxi Firms Suspend Dubai Services Amid Regional Tensions

Mar 2, 2026

Chinese autonomous driving firms Baidu’s Apollo Go and WeRide have halted robotaxi operations in Dubai following Iran’s missile strikes that heightened regional tensions. While WeRide continues services in Abu Dhabi...

Sakana AI Introduces Text-to-LoRA and Doc-to-LoRA for Faster LLM Customization

Feb 28, 2026

Sakana AI has unveiled two new research advancements, Text-to-LoRA and Doc-to-LoRA, which significantly simplify and speed up the customization of large language models (LLMs). These methods allow models to instantly...

OpenAI to Launch Smart Speaker with Camera in 2027

Feb 24, 2026

OpenAI plans to release a smart speaker with a built-in camera and facial recognition capabilities in February 2027. The device, priced between $200 and $300, will analyze the surroundings and...

OpenAI Freezes Stargate Project Amid Challenges

Feb 24, 2026

OpenAI has halted its ambitious Stargate project, initially planned in partnership with SoftBank and Oracle. The suspension is due to internal corporate disagreements, a shortage of engineering talent, and investor...

subscribe via RSS