AI News Daily

1.DeepSeek Slashes Prices Forever

2.AI Brings Back Dead Pilots' Voices

People used AI to reconstruct cockpit voice recordings from spectrogram images — visual snapshots of sound — and the results were good enough to force the US National Transportation Safety Board to temporarily shut down public access to its docket system. The NTSB uses recordings from crashed aircraft to investigate what went wrong. Those recordings are official safety records. The fact that someone could take a spectrogram — an image, not audio — and use AI to recover the original voices raises an immediate question: what else can AI reconstruct from data that was not meant to be audio?

3.xAI Abandoned Solar, Chose Gas

Elon Musk once promised a "solar-electric economy." Today, xAI's Memphis data center runs on natural gas generators, and SpaceX is focused on orbital data centers powered by solar in space. On Earth, the solar future Musk promoted has been quietly replaced by fossil fuels. This is not just an energy story. It reflects the brutal reality of what training large AI models actually requires: enormous amounts of power, available right now, no matter what the source is. Solar cannot meet that demand on the timelines AI labs need. Natural gas can.

4.AI Funding Numbers Are Fake

TechCrunch investigated how AI startups and their investors are inflating revenue numbers to attract attention and capital — and found that this is not accidental. Founders stretch what counts as "annualized recurring revenue," and their investors know it and allow it because higher-sounding numbers help everyone raise at better valuations. The report is specific: some startups count one-time revenue as recurring, count pilot deals as contracts, or use annualized projections of months-old growth as if they were current run rates. The numbers being reported publicly bear little resemblance to what the businesses actually earn today.

5.Antigravity Tops 3D AI Benchmark

Antigravity 2.0 has taken the top spot on the OpenSCAD Architectural 3D LLM Benchmark — a test that measures how well AI models can generate accurate three-dimensional code from natural language descriptions. OpenSCAD is a programming language for creating 3D models, and getting an AI to produce correct geometry from a text description is genuinely difficult. The result is generating real conversation on Hacker News today because 3D code generation has practical applications: product design, architecture, manufacturing, and any field where you need to go from an idea to a physical object. An AI that can reliably do that well is a working tool, not a research demo.

6.Your Code, Turned Into a Map

Understand-Anything is a tool that converts any codebase into an interactive knowledge graph — one that you can explore, search through, and ask questions about directly. It works alongside the AI coding tools most developers already use: Claude Code, Codex, Cursor, Copilot, and Gemini CLI. The idea is simple but practical. Large codebases are hard to navigate even for humans. For AI coding agents, they can be worse — the agent wastes time (and your money) searching for things it could find instantly with a structured map. Understand-Anything builds that map once and keeps it accessible.

7.Run One Company With AI Agents

Tycoon AI is a new platform that lets one person run an entire company using a team of AI agents. The system is built around an AI CEO called Astra, who manages ten or more specialized agents — a marketing agent, a coding agent, a social media agent, and more. You give Astra a goal, like "10x traffic this month" or "launch the onboarding flow," and she creates a plan, assigns agents, tracks what each one is doing, and asks for your approval only when a real decision is needed. No setup, no API keys, no custom configuration required. The agents come pre-built and ready to work. It connects directly with Claude Code and Hermes Agent for coding tasks.

8.A New IDE Built for Agents

Superset is an open-source IDE built from the ground up for running multiple AI coding agents at once. Three engineers — Avi, Kiet, and Satya, backed by Y Combinator's 2026 cohort — built it because they kept wanting to run several agents in parallel on different parts of their codebase simultaneously and had to hack together scripts to do it. Superset manages the underlying complexity: separate working copies of your code for each agent, development servers, terminal sessions, task tracking, and code review — all in one interface. It has also launched remote workspaces in beta, so agents can run on remote machines while you manage everything from the desktop app.

9.Spotify Is Becoming an AI Platform

Spotify made three AI announcements in the same week. First, a deal with Universal Music to let fans create AI-generated covers and remixes of songs, with artists getting paid. Second, a new desktop app for creating personal podcasts — feeding it sources, topics, and questions to generate custom audio briefs. Third, an ElevenLabs-powered tool for authors to turn their books into AI-generated audiobooks. Taken together, Spotify is making a move from music streaming platform to AI audio platform. The company that built its business on passive listening is now building tools for creating, not just consuming.

10.Faster AI, No Nvidia Required

General Compute launched today as an AI inference cloud built on ASICs — specialized chips that are purpose-built for running AI, not training it. Most AI infrastructure today runs on Nvidia GPUs, which were originally designed for graphics and adapted for AI. General Compute argues that running AI on hardware designed specifically for the job is both faster and cheaper. The company claims 5x faster responses and higher capacity per user compared to Nvidia-based clouds. The API is fully compatible with OpenAI's format, which means switching over requires changing one line of code — your base URL.