AI News
AI news that matters. Updated daily.
No stories match your filters.
Inside Codex's System Prompt: No Goblins, Raccoons, or Trolls
"Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant."
Musk Testifies He Founded OpenAI to Prevent a 'Terminator Outcome'
Elon Musk told a San Francisco federal court on Monday that he co-founded OpenAI in 2015 specifically to prevent a "Terminator outcome" - a scenario where AI systems turn against humanity. The testimony came during the Musk v. Altman trial, where Musk is suing OpenAI and CEO Sam Altman, alleging the company abandoned the nonprofit mission he helped build.
Musk Takes the Stand Against OpenAI, Opens With His Origin Story
On the witness stand in San Francisco, Elon Musk opened his testimony against OpenAI and its CEO Sam Altman with an origin story: born in South Africa, arrived in Canada for college with $2,500 in traveler's checks, driven throughout by one goal - saving humanity.
AWS Adds OpenAI Models to Bedrock One Day After Microsoft Exclusivity Ends
One day. That's how long it took Amazon to move after Microsoft agreed to end its exclusive distribution rights to OpenAI's models. AWS announced on April 28 that it would offer a range of OpenAI products through Amazon Bedrock, including a new agent service.
Elon Musk Takes the Stand in OpenAI Lawsuit Against Sam Altman
Elon Musk took the stand Monday in his lawsuit against OpenAI CEO Sam Altman and company president Greg Brockman, moving a long-running legal dispute from court filings into live testimony for the first time.
Amazon Adds AI Audio Q&A to Product Pages with "Join the Chat"
Shopping on Amazon just got a voice layer. The company is rolling out a feature called "Join the chat" that lets shoppers ask questions about products and receive AI-generated audio responses directly on product pages, as reported by TechCrunch.
Claude.ai Goes Down in Widespread Outage
Claude.ai went down on April 28, 2026, leaving users unable to access the service. Anthropic acknowledged the incident on its status page, confirming that Claude was unavailable. No root cause was published at time of writing.
Mistral Medium Is Coming, Likely Lands at 128 Billion Parameters
Mistral's model naming conventions are giving away secrets. The internal designation for Mistral Small - appearing in model metadata as "Mistral-Small-4-119B-2603" - reveals a pattern that encodes version number, parameter count, and release date. Following that pattern, a Mistral Medium model appears to be in development at roughly 128 billion parameters ("parameters" being the numerical values that determine how a model interprets inputs and produces outputs).
Claude.ai Goes Down on April 28 - Anthropic Opens Incident Report
Claude.ai went offline on April 28, 2026 at around 5:41 PM UTC. Anthropic opened an official incident report and began posting updates at status.claude.com.
13B Model Trained Only on Pre-1931 Text Tests What LLMs Actually Learn
What happens when you train a language model on nothing but books, newspapers, and journals from before 1931?
Claude Gets Direct Connectors for Photoshop, Blender, Ableton, and Other Creative Apps
Anthropic has released a set of connectors that let Claude talk directly to creative software - Adobe Creative Cloud apps, Affinity, Blender, Ableton, and Autodesk among them. Instead of copying and pasting between a chat window and your tools, Claude can now reach into the software you're already working in.
Mistral Teases Something Called 'Vibe' for April 29
Mistral is announcing something called "Vibe" on April 29, 2026. The teaser came via the @mistralvibe account on X, with no technical details attached.
NVIDIA's Nemotron Omni 30B Reasoning Model Surfaces Before Official Announcement
NVIDIA's Nemotron-3-Nano-Omni-30B-A3B-Reasoning model has appeared in model repositories ahead of any official company announcement. The naming convention is dense but informative: 30 billion total parameters, with only 3 billion active at any given moment.
Lovable Launches iOS and Android Apps for Mobile Vibe Coding
Lovable just shipped native apps for iOS and Android, bringing its AI app builder to mobile. The tool falls into the "vibe coding" category - building web apps and sites by describing what you want in plain language, with AI writing the actual code, no programming knowledge required. Until now, doing this with Lovable meant being at a desktop or laptop.
Google Translate Turns 20: New Features and Two Decades of Language AI
Twenty years ago this month, Google launched a translation tool that most linguists assumed would stay limited forever. The 2006 version worked by pattern-matching text from existing translated documents - no actual language understanding involved. It was useful, and often memorably wrong.
NVIDIA Nemotron 3 Nano Omni Can Process 5+ Hours of Video, Audio, and Docs in One Pass
The name is a bit misleading. NVIDIA is calling this model "Nano," but it has 30 billion parameters - not exactly small. The naming makes sense once you understand the architecture: it uses a Mixture of Experts (MoE) design, meaning only 3 billion parameters are active on any given computation pass. The rest sit idle until needed for different task types. That's where the efficiency comes from, and why NVIDIA claims 9x higher throughput on multimodal workloads compared to alternatives.
Nvidia Executive: Running AI Systems Costs More Than the Employees It Replaces
The story of AI cutting costs is more complicated than the headlines suggest. An Nvidia executive told Fortune that AI infrastructure costs more than the human employees it supposedly replaces - a candid acknowledgment from the company that profits directly from selling the GPUs powering these systems.
Claude Gets Direct Blender Integration via MCP
Claude can now control Blender, the open-source 3D modeling and animation software used by designers, VFX artists, and indie game developers. The connection runs through Anthropic's MCP (Model Context Protocol) - a standard that lets AI assistants connect directly to external software and take actions inside it, not just describe what to do.
Nvidia's Nemotron 3 Nano Omni Is Built for Enterprise AI Agents
Nvidia is better known for the chips that run AI models than for building models itself. Nemotron 3 Nano Omni is another step toward changing that.
Musk vs. OpenAI Goes to Trial, and the Founding Story Gets Stress-Tested
The lawsuit Elon Musk filed against OpenAI heads to trial this week, and the proceedings are expected to pull back the curtain on the earliest and most contentious days of one of the most influential AI companies ever built.
YouTube Tests AI Search Answers for U.S. Premium Subscribers
YouTube is rolling out an AI search feature that generates guided answers directly inside search results, currently available as an opt-in test for Premium subscribers in the U.S., according to a TechCrunch report.
Reinforcement Learning Startup Raises Record $1.1B Seed Round
$1.1 billion. In a single seed round - traditionally the very first institutional funding a startup raises - a reinforcement learning company just closed what appears to be the largest seed funding in startup history. The company's stated goal is superintelligence.
AI's $20/Month Era Is Under Pressure as Compute Costs Hit Subscription Models
AI companies have been charging users well below their actual compute costs for years. Several signals suggest that math is getting harder to sustain.
FIDO Alliance, Google, and Mastercard Build Safety Rails for AI Shopping Agents
AI agents that can browse the web, fill out forms, and place orders on your behalf are coming fast. The security infrastructure to keep those agents from making unauthorized purchases? That's lagging behind.
Qwen 3.6 27B Quantization Tested: BF16 vs Q8_0 vs Q4_K_M
Running a 27-billion parameter AI model on your own hardware requires compression. A community evaluation of Qwen 3.6 27B - Alibaba's latest mid-size open model - tested three formats that represent the main options for local deployment.
Otter Launches Cross-App Search Across Gmail, Notion, and Salesforce
35 million users and $100 million in annual recurring revenue - Otter.ai has turned meeting transcription into a real business. Now the company wants to be the search layer across your entire work stack, not just your calls.
Microsoft Releases VibeVoice as Open-Source Voice AI on GitHub
Microsoft just dropped VibeVoice on GitHub, an open-source voice AI project the company is positioning as a frontier-grade system. It's available now for anyone to inspect, run, or build on.
Who Owns Code Written by Claude Code? The Legal Answer Is Complicated
The answer is messier than most developers realize. When you write a prompt and Claude Code generates the implementation, who holds copyright to that output? Your employer? You? Anthropic? Nobody?
DeepSeek Researcher Teases Upcoming Vision Model
A DeepSeek researcher just teased what comes next from the Chinese AI lab: a vision model.
ChatGPT's Image Generation Is Getting Noticeably Better at the Details That Mattered
Readable text inside AI-generated images was, until recently, a reliable way to identify AI output. Garbled letters, warped logos, signs that said nothing in particular - these were table stakes for any image generator. ChatGPT is getting noticeably better at this, and users sharing outputs are showing results that are hard to dismiss.
Bloomberg Terminal's AI Overhaul: Chatbot Controls Coming to Finance's Most Stubborn Platform
The Bloomberg Terminal hasn't changed much in 40 years. Traders learn its cryptic command codes the way surgeons learn anatomy — through repetition, until the shortcuts become muscle memory. Now Bloomberg's chief technology officer is telling WIRED that chatbot-style AI is coming to the platform, and the framing — "like it or not" — tells you everything about the internal tension.
Local LLMs for Coding Keep Failing the Same Test: Actual Work
Local LLMs for coding keep failing the same test: actual work.
Why Some Claude Code Users Burn Through Hundreds of Millions of Tokens a Month
Heavy users of Claude Code - Anthropic's AI coding tool - regularly report burning through hundreds of millions of tokens per month. For context: one token is roughly three-quarters of a word, so 100 million tokens represents around 75 million words. Developers who plan architecture before prompting, keep sessions focused, and work deliberately across a few codebases typically land at 15-25 million tokens per month even with intensive daily use. The gap between that baseline and the power-user ceiling - sometimes north of 500 million tokens per month - comes down to a few specific habits.
Musk v. Altman Trial Kicks Off With a Jury Pool That Already Dislikes Musk
Jury selection started Monday in San Francisco federal court for the lawsuit Elon Musk filed against Sam Altman and OpenAI over what he claims are broken promises about the organization's nonprofit mission. The legal battle, which has been grinding through pre-trial motions for more than a year, is finally headed to a jury - and seating one is proving difficult.
Local AI Models for Coding Cross the 'Good Enough' Threshold
A year ago, running an AI coding assistant on your own hardware meant accepting real quality tradeoffs compared to cloud-based tools. The models that fit on a consumer PC were noticeably weaker than what you'd get from Cursor or Aider connected to a frontier API. That gap has narrowed faster than most expected.
Anthropic Confirms Claude Opus Is Not Being Removed from Pro Plans
Anthropic's Claude Opus model is staying on Pro plans. After confusion spread among subscribers about a potential removal, it's been confirmed that the $20/month Pro tier still includes Opus access - no changes are coming.
Some Claude Pro Subscribers Seeing Claude Code as a 7-Day Trial
Some Claude Pro subscribers are seeing Claude Code framed as a 7-day free trial rather than a standard plan feature. The apparent A/B test - where a company shows different versions of their product to different user groups to measure responses - raises the question of whether Anthropic is moving toward charging separately for its coding agent.
Google Tests Conversational AI Search Inside YouTube
YouTube is testing a conversational search feature that lets you ask questions and get back a mix of longform videos, Shorts, and written text in a single response. Google confirmed the experiment is live now for a limited group of users.
Anthropic Connects Claude to Ableton, Adobe, Blender, and 5 More Creative Tools
Anthropic just connected Claude to eight professional creative software platforms in a single announcement. The new integrations - which Anthropic calls "connectors" - let Claude pull from official software documentation and APIs to answer questions, write scripts, and automate tasks inside tools creative professionals already use. The full announcement is on the Anthropic blog.
OpenAI Brings GPT Models, Codex, and Managed Agents to AWS
OpenAI's models, Codex (its code-generation tool), and Managed Agents are now available directly on AWS. The announcement means businesses already running on Amazon's cloud can build AI-powered products without routing data through OpenAI's own infrastructure.