AI coding tools turn to ads, and it’s actually working

The AI race is shifting from text-predicting chatbots to world models that can actually grasp how the physical world works. These systems learn from video, sensor data and simulations to understand motion, cause and effect, gravity and how objects interact over time. The goal is to build AI that can navigate real environments, not just generate fluent text.

Some of the biggest names in the field are betting on this shift. Fei-Fei Li’s World Labs just released its first commercial model, Marble. Yann LeCun plans to launch a world model startup after leaving Meta. Google, Meta, and several Chinese labs are racing to build versions that help power robotics and more realistic video systems. Even OpenAI sees better video models as a path toward a full world model.

The challenge is data. Unlike language models that feed on the open internet, world models need vast amounts of high-quality multimodal data that is scattered across formats and industries. Encord, which released a billion-item dataset spanning video, 3D point clouds and audio, says it is still only a starting point.

World models are essential for robotics, autonomous systems and next-gen simulation tech, but it is not yet clear whether they can advance as quickly as today’s language models. The investment surge suggests many in the industry believe they will.

Supported by AssemblyAI

Building Voice AI is complex—manual workflows, multiple vendors, and high costs. AssemblyAI makes it simple with a single, developer-first API.

With one platform, you’ll get:

Transcribe audio in 99+ languages with low latency and best-in-class accuracy. Scale production-ready apps without commitments, and connect seamlessly to leading LLMs like Claude, ChatGPT, and Gemini.

Trusted by top teams (Granola, Dovetail, Cluely), AssemblyAI lets you focus on building smarter Voice AI apps instead of managing infrastructure.

AI coding assistants like Cursor, Replit and Devin have been growing fast on paid plans, but Sourcegraph is trying something different. Last month the company launched a free, ad-supported tier for its coding assistant Amp, and early results suggest the model might catch on.

Ads sit just above the input box and are targeted to the work a developer is doing. If you are adding authentication, you might see an ad for WorkOS. Around 25 companies are already buying these placements, including Vanta, Baseten and Turbopuffer. CEO Quinn Slack says advertisers are willing to pay up to $1,000 for a qualified action, and the new ad business is already generating between $5 million and $10 million a year.

Slack admits he assumed developers would hate it, but feedback has been surprisingly positive. Strong demand also allowed Sourcegraph to drop its initial requirement that free users share code for model training.

There is a catch. Ads alone cannot cover the cost of running top tier models, so Sourcegraph’s free plan relies on cheaper ones like Claude Haiku and Kimi K2. Still, the move signals where the market is heading. If even a premium tool like Amp can make ads work, bigger players may not be far behind.

Oracle’s 30-year bond has slid to 65 cents on the dollar, an 8 percent drop since October, and investors are getting nervous about how much Big Tech is borrowing to fuel the AI boom.

Bank of America says Oracle’s credit risk is widening faster than the broader investment-grade market. Five-year credit default swaps have jumped to their highest level in two years, a sign that bond buyers are starting to question whether tech companies can keep spending at this pace.

The shift comes at an awkward moment. Financial conditions are easing, risk assets are rallying, and just two weeks ago Meta’s latest bond sale was four times oversubscribed. Now investors are pulling back, worried that the AI capex race may be outpacing actual returns.

The concern is simple. Companies are pouring billions into chips and data centers with no clear timeline for when the payoff arrives. If demand softens or the economics do not improve quickly enough, the cost of financing this AI buildout could become a drag on even the biggest balance sheets.

Jeff Bezos is stepping back into an operating role for the first time since leaving Amazon in 2021. According to the New York Times, he will serve as co-CEO of Project Prometheus, a new AI startup focused on building advanced systems for engineering and manufacturing across computers, cars, and spacecraft.

The startup has already raised $6.2 billion, partly from Bezos himself, making it one of the most heavily funded early-stage AI companies to date. He will lead the company alongside Vik Bajaj, a physicist and chemist who previously worked with Sergey Brin at Google X.

Project Prometheus has hired nearly 100 employees, pulling talent from OpenAI, DeepMind, and Meta, as it prepares to compete with the biggest players in the AI race, including OpenAI, Google, and Meta. Reuters has not independently confirmed the report.