Good morning. It’s Monday, December 15th.

On this day in tech history: In 1962, NASA's Mariner 2 became the first spacecraft to successfully fly by Venus, relaying data on its hellish atmosphere via autonomous attitude control and sequencing. This deep-space feat relied on rudimentary onboard logic - an early echo of AI autonomy in robotics and pathfinding algorithms. Its Canopus star tracker and gyros prefigured SLAM techniques in modern AI-driven probes, enduring 42 million miles without real-time human input.

In today’s email:

• The Open-Source Voice Model That Changes Everything

• Google just made real-time translation seamless on ANY headphones

• $18-per-hour AI agent beats 9 out of 10 pro hackers in Stanford test

• Runway drops Gen-4.5 with native audio edits and scene-wide ripple changes

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

The Open-Source Voice Model That Changes the Economics

For the last year, ElevenLabs has set the bar for high-quality text-to-speech. But it comes with the usual tradeoffs: closed models, usage caps, and rising costs as you scale. Chatterbox flips that model entirely. Built by Resemble AI, Chatterbox delivers comparable voice quality while being open-source (MIT-licensed), faster in real-time use, and dramatically cheaper to run—because you can host it yourself.

Unlike ElevenLabs’ API-only approach, Chatterbox runs locally or on your own infrastructure, eliminating per-character pricing and vendor lock-in. It supports zero-shot voice cloning, expressive speech control, and real-time synthesis fast enough for live agents and interactive apps. In blind listening tests, Chatterbox has even been preferred over ElevenLabs—despite costing nothing to use.

The strategic difference is simple: ElevenLabs sells access to voice. Chatterbox gives you ownership of it. You control the model, the latency, the deployment, and the economics. Built-in watermarking supports responsible use, while the open architecture makes it ideal for startups, internal tools, and AI agents that need voice at scale.

If ElevenLabs proved what was possible, Chatterbox proves it doesn’t need to be closed—or expensive. For builders who care about speed, cost, and control, this is the voice model that changes the game. Try it here.

Google just made real-time translation seamless on ANY headphones

Google just removed one of the biggest friction points in live translation. Translate’s Android beta now delivers real-time speech translation through any headphones in 70+ languages. Powered by Gemini, it focuses on meaning, not literal word swaps, so tone, rhythm, and intent survive the translation, along with slang and local expressions, at lower latency.

At the same time, Translate is shifting from “use it when you travel” to “use it every day.” Google is expanding its AI-powered practice features to 20 more countries, with adaptive vocabulary, listening drills, and feedback that adjusts to how you actually learn.

Voice is also getting more disciplined. Gemini Live is picking up basic conversational manners such as pausing without cutting you off and letting you mute the mic while it’s speaking. Small changes that make real-time voice feel usable instead of awkward. In parallel, Google Search Live is rolling out a new Gemini-based audio model that sounds less synthetic and supports continuous voice conversations while pulling in live web results.

Google is also fixing a less visible problem: wasted compute. A new research framework teaches AI agents to monitor their own token and tool budgets in real time and adjust behavior accordingly. Early results show higher accuracy with up to 50 percent lower costs, a shift away from brute-force scaling toward efficiency and control. That strategy matches DeepMind’s longer timeline. Chief AGI Scientist Shane Legg recently put a 50 percent chance of reaching “minimal AGI” by 2028. Read more.

$18-per-hour AI agent beats 9 out of 10 pro hackers in Stanford test

Stanford researchers put an AI agent, ARTEMIS, to the test against professional penetration testers and found it outperformed nine of ten humans while running at a fraction of the cost. Over a 16-hour period, ARTEMIS scanned roughly 8,000 devices, uncovering vulnerabilities that humans missed, including on older servers inaccessible through standard browsers. Unlike human testers, the agent could deploy sub-agents to investigate multiple targets simultaneously, giving it a scale and parallelism advantage.

Within the first 10 hours, ARTEMIS discovered nine valid vulnerabilities with an 82% success rate. Its strengths lie in parsing code-like inputs and outputs, though it struggles with graphical interfaces, occasionally producing false positives. The study highlights how AI agents can drastically reduce the time, cost, and cognitive load of complex security tasks, even while exposing new challenges in verification and GUI-dependent testing. Read more.

Runway drops Gen-4.5 with native audio edits and scene-wide ripple changes

Runway has rolled out Gen-4.5 with major upgrades and unveiled its first General World Model, GWM-1. The new video model adds native audio generation and editing, plus multi-shot editing that applies changes across an entire scene. GWM-1 builds an internal simulation of environments in real time, allowing interactive control via camera movements, robot commands, or audio inputs.

The system ships in three versions: GWM Worlds for explorable environments, GWM Avatars for lifelike speaking characters, and GWM Robotics for synthetic robot training data, with plans to eventually unify these capabilities. Runway joins a growing field of AI labs including DeepMind, Yann LeCun’s startup, and Fei-Fei Li’s World Labs racing to build models that understand and simulate the physical world beyond text. Read more.

• Meta’s next AI, Avocado, could mark the end of its open-source approach

• Gemini 3 Pro crushes Pokemon Crystal flawlessly where 2.5 Pro loops forever

• GPT-5.2 tops “most censored” lists across leading AI benchmarks

• The “Godmother of AI” bets on tool fluency over diplomas in tech recruitment

• Guinness-verified PocketLab from Tiiny AI offers supercomputer performance at a fraction of traditional costs

• Creative roles aren’t going away, but AI is turning workers into directors managing teams of intelligent agents

• AI reasoning models now ace all three levels of the grueling CFA exam, with near-perfect scores

• Old-school oral exams are making a comeback as a defense against ChatGPT in classrooms

• Most AI agents would fail as office workers, new Carnegie Mellon simulation finds

• A new report pegs 2027 as the likely arrival date for superhuman AI coding if supply chains and global stability hold

• Watch: An AI robot that initially refused to shoot its owner changed its mind as soon as it was asked to play a character

• Google and MIT find adding extra AI agents can tank performance on sequential tasks

• Center for Humane Technology blasts federal AI policy for prioritizing growth over safety

• LongCat-Image shows careful data curation trumps brute-force scaling in AI-generated photos and text

• Columbia University launches a public tracker to monitor AI deals and copyright disputes in media

• Fine-tuning and persona tricks don’t fool detection—simpler prompts work better, researchers say

• Pangram 3.0 claims near-perfect AI text detection with a four-tier system tracking subtle AI edits

• Starcloud becomes first company to train and run AI models entirely from space using NVIDIA H100

• New multilayer 3D chip could shatter the memory wall and supercharge AI workloads

• Google and NextEra team up to build nuclear-powered gigawatt AI data centers across the US

• Suleyman on AI power players: Altman is courageous, Hassabis a polymath, Musk a bulldozer

• Oracle’s AI boom hits a wall as physics and debt markets collide, sending shares tumbling

• A new pay-to-scrape standard gives publishers control over how AI uses their web content

• Thinking Machines launches Tinker AI fine-tuning service to simplify customizing massive language models

📄 One Layer Is Enough: Adapting Pretrained Visual Encoders for Image Generation

📄 Omni-Attribute: Open-vocabulary Attribute Encoder for Visual Concept Personalization ​​

📄 Are We Ready for RL in Text-to-3D Generation? A Progressive Investigation

📄 MoCapAnything: Unified 3D Motion Capture for Arbitrary Skeletons from Monocular Videos

📄 StereoWorld: Geometry-Aware Monocular-to-Stereo Video Generation  

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!