Welcome to Daily Inference, your daily dose of the most important developments shaping the world of artificial intelligence. I'm your host, and today we've got a packed episode covering everything from AI agents that rewrite their own code, to a massive surge in AI-generated abuse material, to Jensen Huang dropping one of the biggest claims in tech history. Let's dive in.

But first, a quick word from our sponsor. If you've been putting off building a website, 60sec.site is the tool you need. It uses AI to get you from zero to a fully designed website in under a minute. No coding, no hassle. Check it out at 60sec.site.

Alright, let's start with what might be the single most provocative statement made by any tech executive this week. Nvidia CEO Jensen Huang appeared on the Lex Fridman podcast and declared, quote, I think we've achieved AGI. Now, AGI — artificial general intelligence — is one of those terms that the industry has been both chasing and running away from simultaneously. In recent months, many tech leaders have tried to rebrand around the concept, creating new terminology to avoid the hype. But Huang went the other direction entirely. What makes this interesting is the timing. We're seeing AI systems right now that are doing things that genuinely push the boundaries of what we thought machines could do — and that brings us directly to our next story.

Meta AI has unveiled something called the Darwin Gödel Machine, and it represents a genuinely different kind of AI architecture. The idea is called recursive self-improvement — meaning the system doesn't just get better at a task, it gets better at how it learns. Think of it like an employee who doesn't just improve their skills, but rewrites their own job description and training manual as they go. This concept has existed theoretically for decades under the name the Gödel Machine, but it's always been considered too impractical to build in the real world. Meta's new hyperagent framework appears to have actually made this work. When you combine this with Huang's AGI claim, you start to see why the debate around machine intelligence is heating up so fast — the capabilities really are advancing in ways that are hard to categorize.

Now, on the research side, there's some fascinating work coming out of Yann LeCun's lab. LeCun, who famously disagrees with the current large language model approach to AI, has been pushing a framework called JEPA — Joint Embedding Predictive Architecture — as a path toward more human-like machine intelligence. The new research, called LeWorldModel, tackles a tricky problem called representation collapse. Here's the simple version: when you try to train an AI to predict what the world will look like next based on raw pixel data — like a video feed — the model tends to cheat. It learns to produce vague, redundant outputs that technically satisfy the goal without actually understanding anything meaningful. LeCun's team is working on architectural fixes that prevent this shortcut behavior, which is a crucial step toward AI that can genuinely reason about and plan in the physical world.

Speaking of reasoning before acting, Luma Labs just released a new image generation model called Uni-1, and it takes a notably different approach from the standard diffusion models that power most AI image generators today. Instead of jumping straight into generating pixels, Uni-1 first reasons about what it's actually trying to create — processing the intent behind a prompt before producing the output. The team is calling this closing the intent gap, and it's part of a broader shift in generative AI away from pure statistical pattern matching toward something more structured and deliberate. This connects directly to what LeCun is pursuing and what Meta's hyperagents represent — AI that thinks before it acts.

Now let's talk about some serious, darker territory. The Internet Watch Foundation released sobering data showing that AI-generated child sexual abuse material found online surged by fourteen percent in 2025. The watchdog verified over eight thousand pieces of realistic AI-generated content, and what's particularly alarming is a more than two hundred and sixty fold increase in videos compared to prior years, with sixty-five percent of those videos falling into the most extreme category. This is a direct consequence of increasingly accessible and realistic generative AI tools, and it represents one of the most urgent harms the technology is enabling right now. It's a stark reminder that as we celebrate the capabilities of these systems, the same power is being weaponized in devastating ways.

Anthropics relationship with the government is meanwhile getting complicated. Senator Elizabeth Warren has sent a letter to Defense Secretary Pete Hegseth, calling the Pentagon's decision to label Anthropic a quote supply-chain risk an act of retaliation rather than a legitimate security concern. Warren argued that if the Department of Defense had genuine issues with Anthropic's contract, it could have simply terminated the agreement rather than blacklisting the company. This comes as Anthropic is also rolling out new remote control capabilities for Claude — essentially allowing the AI to navigate the web and take actions on your computer on your behalf. So Anthropic is expanding what Claude can do in the real world at the exact same moment its government relationships are under strain. Quite a tension to manage.

On the infrastructure side, two stories point to where the real money is flowing. OpenAI CEO Sam Altman is stepping down as board chair of fusion energy startup Helion, as the two companies are reportedly in talks for Helion to sell roughly twelve and a half percent of its power output to OpenAI. The energy demands of AI are enormous and growing, and fusion would be a transformative solution if it ever comes online at scale. Meanwhile, a startup called Gimlet Labs just raised eighty million dollars for technology that lets AI inference run simultaneously across chips from Nvidia, AMD, Intel, ARM, Cerebras, and d-Matrix. That's the infrastructure problem nobody talks about — different hardware ecosystems creating bottlenecks. Gimlet's approach could be a significant unlock for deploying AI at scale.

Finally, let's zoom out to the economic picture. BlackRock CEO Larry Fink used his annual letter to investors to warn that the AI boom risks widening inequality. The boss of a fourteen trillion dollar asset management firm pointed out that the financial rewards of AI are likely to concentrate among a relatively small number of companies and investors, even as the technology becomes central to geopolitical competition between the US and China. When the person managing more money than most countries' entire GDPs says the wealth gap might get worse, it's worth taking seriously. This connects to everything we've discussed today — the governance tensions around Anthropic, the safety failures enabling abuse, and the scramble to control AI infrastructure. The technology is racing forward. The question of who benefits, and who gets protected, remains very much open.

That's your Daily Inference for today. If you want to go deeper on any of these stories, head over to dailyinference.com to subscribe to our daily AI newsletter — we break down the most important developments every single day. Thanks again to 60sec.site for sponsoring today's episode. Build your website in sixty seconds at 60sec.site. We'll see you tomorrow.