29 April 2026 PM: Scout AI’s $100m War Bet, OpenAI Lands on AWS, and Google Translate Turns 20

An AI startup raising $100 million to train soldiers to command fleets of autonomous vehicles. OpenAI models going live on Amazon’s cloud the day after a landmark deal ended Microsoft’s exclusivity. And Google Translate turning 20 with new features you can start using right now. Wednesday afternoon’s AI news has something for the battlefield, the boardroom, and your phone.

Scout AI, founded by Colby Adcock, has raised $100 million to build AI agents that give individual soldiers command of autonomous vehicle fleets. The startup is developing software that lets a single soldier direct groups of autonomous ground vehicles in the field. TechCrunch visited the company’s training ground where Scout is putting that system through its paces, with AI agents designed to interpret battlefield instructions and coordinate drone and vehicle movements in real time. The fundraise underlines how defence technology has become one of the fastest-moving sectors in AI investment, with startups now moving quickly to fill gaps between what consumer AI can do and what military operators actually need in the field.

The broader question Scout AI raises is one the industry has been wrestling with since large language models became capable of taking consequential actions: who is ultimately responsible when an AI agent makes a decision in a high-stakes environment? Scout’s pitch is that its agents keep humans in the loop at every step, but the speed at which autonomous systems can operate means the nature of that loop is changing rapidly. Defence AI funding hit a new milestone this week, with Firestorm Labs also closing an $82 million round to put drone manufacturing inside shipping containers for front-line deployment.

OpenAI’s models are now available through Amazon Web Services, less than 24 hours after OpenAI ended its exclusive cloud partnership with Microsoft. AWS announced a slate of OpenAI model offerings inside Amazon Bedrock, including access to a new agent service. For businesses already running workloads on AWS, this matters in practical terms: rather than routing AI calls to a separate API or managing a different cloud relationship, they can now access OpenAI models alongside their existing AWS infrastructure. Sam Altman and AWS CEO Matt Garman discussed the deal in an interview that outlined the scope of the arrangement, including integration with Bedrock Managed Agents. The speed of the announcement, coming within hours of the Microsoft exclusivity news, signals that OpenAI is actively diversifying its distribution and that AWS sees AI infrastructure as a competitive priority it cannot sit out.

Google Translate turned 20 this week, and the milestone comes with a meaningful capability update that is worth exploring if you use the service for anything beyond basic phrases. The tool now supports close to 250 languages, up from a handful when it launched in 2006 as one of Google’s earliest AI experiments. The anniversary release includes improvements to translation quality in lower-resource languages, better handling of informal registers and slang, and tighter integration across Google’s product suite. For consumers and small businesses, Translate remains one of the most practically useful AI tools available. The quality improvements in underserved languages are particularly significant, as many of those 250 languages are spoken by communities that have historically been poorly served by translation technology. If you use Translate professionally, it is worth testing the current version: accuracy in legal, medical, and technical contexts has improved noticeably over the past 12 months.

A security researcher has published a detailed technical breakdown of how ChatGPT’s advertising system works, and the findings are worth understanding if you use the platform regularly. The analysis, published by Buchodi’s Threat Intel, shows that when OpenAI’s backend decides to serve an ad, it injects a structured ad object directly into the streaming conversation response. The ad includes a branded card with a click-through link and four separate encrypted tokens used to track whether you visit the advertiser’s website. On the merchant side, participating retailers load an OpenAI tracking SDK called OAIQ that writes a cookie to your browser, tied to that click token, and reports back to OpenAI’s servers after you land on the page. Targeting appears to be contextual to the conversation topic rather than tied to your account history, based on the traffic the researcher observed.

If you prefer not to be tracked through this system, the researcher identified two domains to add to your browser filter list: bzrcdn.openai.com and bzr.openai.com. The research also reveals a 30-day attribution window, meaning a click on a ChatGPT ad can tie your subsequent activity back to OpenAI for a month. Given that ChatGPT now has hundreds of millions of users, the scale of this data collection infrastructure is significant, and this is the most detailed public documentation of how it actually functions.

Elon Musk took the stand in his lawsuit against OpenAI this week and repeated his version of the organisation’s founding history for the first time under oath. Musk has told this story before, in interviews and in Walter Isaacson’s biography, but the trial marks the first time he has said it in a formal legal setting. The lawsuit centres on Musk’s claim that OpenAI departed from its original non-profit mission when it moved toward a commercial structure. The legal process is forcing a more formal accounting of what was and was not agreed between the founders. Musk’s xAI, which competes directly with OpenAI through the Grok model series, gives the dispute a commercial dimension that colours the entire proceeding.

Here is everything else worth knowing from today’s AI news.

• OpenAI and AWS CEOs detail the Bedrock deal — Sam Altman and Matt Garman outline how OpenAI models will integrate with Bedrock Managed Agents in a Stratechery interview. Source
• Google expands Pentagon AI access after Anthropic’s refusal — After Anthropic declined to allow the DoD to use its AI for domestic surveillance and autonomous weapons, Google signed a new contract with the department. TechCrunch
• IBM releases Granite 4.1 LLMs — IBM’s open-source Granite series gets an updated release, with a Hugging Face blog post detailing how the models are built and benchmarked. Hugging Face
• Making AI chatbots friendlier leads to more errors and conspiracy support — New research reported by The Guardian finds that designing AI assistants to be warmer and more agreeable correlates with increased factual mistakes and a greater tendency to validate false beliefs. The Guardian
• Why AI companies want you to be afraid of them — BBC Future examines how frontier AI labs have cultivated a narrative of existential risk, and the commercial logic that sits behind it. BBC Future
• AI counted carbs 27,000 times and could not give the same answer twice — A researcher ran the same carbohydrate counting task through multiple AI tools thousands of times and found significant variation in results, raising questions about AI reliability in health monitoring. Diabettech
• Shapes brings AI characters into human group chats — A new app lets users add AI personas to shared chat threads alongside real people, blending social messaging with AI companions. TechCrunch
• Firestorm Labs raises $82 million for containerised drone factories — The defence startup has raised $82 million to build drone manufacturing facilities inside shipping containers, designed to operate at or near the front line. TechCrunch

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s afternoon update on developments in artificial intelligence, published every weekday afternoon.