Weekly AI Roundup: Week 42, 2025
This week's AI news breaks down into what's genuinely worth your time and what's just filler, with the good stuff centering on platform strategies and infrastructure tweaks. The real highlights? OpenAI's push for big revenue and platform growth, NVIDIA and TSMC's first US-made Blackwell chip, and Anthropic's Agent Skills setup. These could shake up how we actually build, roll out, and profit from AI, I'd say.
As for the hype machine, it's churning out stories on new video models, voice AI startups, and minor developer tools. Credit where it's due—some of these are clever from a tech angle. But, honestly, they're mostly small upgrades in a jammed field, not the world-shaking changes folks are hyping. What's truly unfolding is a grab for control around platforms and infrastructure, where the giants are setting the stage for who calls the shots in AI for years ahead, probably.
The Platform Wars Heat Up
OpenAI went all in this week at DevDay 2025, dropping AgentKit, an Apps SDK, and reframing ChatGPT as a full-on app platform. They're aiming to be like the iOS of AI, with outfits such as Canva, Zillow, Coursera, Figma, Spotify, and more running their stuff right inside ChatGPT chats. It's not merely about sprucing up ChatGPT—it's a play to rope developers into OpenAI's world and keep them from jumping ship.
These moves sync up with the leaked revenue forecasts, shooting from $13 billion in 2025 to $100 billion by 2028 or 2029. If that holds, it's wild growth—only seven US companies have hit that mark in the last fifty years, and OpenAI figures on doing it in three. Their platform bet seems like a smart way to hook developers and fuel those numbers, but I'd wait before getting excited; a lot could go sideways.
Anthropic pushed back with Agent Skills, a system that loads Claude up with practical know-how and expertise via folders of instructions, scripts, and extras. It doesn't grab headlines like OpenAI's ecosystem, yet this one actually matters because it pushes AI toward handling real workflows, not just chitchat. The way they handle progressive disclosure to dodge context overload is a nifty detail that might make a difference in the long run, even if it's easy to overlook.
Infrastructure Reality Check
The big news that flew under the radar was NVIDIA and TSMC cranking out the first US-made Blackwell wafer at their Arizona fab, and Jensen Huang called it a historic first for American chip production. This matters because it's real progress on bringing AI hardware back home, instead of empty promises. Not every story has to be flashy to count.
They pulled this off in just a few years, which is impressive on its own, but the deeper angle is about strategy. As AI processing turns into a cornerstone of global economies, having chips made locally becomes key for security reasons. It's not some abstract trade spat; we're talking about making sure the US keeps the tech edge in AI building blocks, and that feels urgent right now.
Sobering Voices Cut Through the Hype
Andrej Karpathy dropped the week's most grounded take, telling Dwarkesh Patel that agentic AI is more like a decade away, not the "year of agents" everyone's pushing. Coming from someone who's coded AI at OpenAI and Tesla, his point hits home: current models are missing basic thinking skills, full multimodal tricks, and solid memory setups.
When he says tools like Codex and Claude Code offer real perks but "just don't work" like actual human workers, it's a reminder that we're not there yet. The chasm between slick demos and stuff that runs smoothly in the real world is huge, and startup hype isn't helping bridge it. I think his view adds some much-needed balance to all the cheerleading.
OpenAI's tackling this head-on with their "OpenAI for Science" group, headed by VP Kevin Weil and including Vanderbilt's Alex Lupsasca, a black hole expert. Focusing on physics and math is logical because those areas have straightforward goals and tests, unlike the vague business apps AI agents usually chase. It might not fix everything, but it's a step in the right direction, even if I'm not fully sold on the outcomes.
Quick Hits
Google rolled out Veo 3.1 with photo-to-video features, but after Sora 2's big splash, it just seems like playing catch-up. They also tweaked AI Studio for developers and tied in live Google Maps data for Gemini—handy stuff, yet it's more of the same old improvements that probably won't shake things up much. Meta's opt-in for scanning unuploaded photos to train AI feels like a half-hearted privacy gesture, especially with their history.
Then there are seven Indian voice AI startups building multilingual systems, which is a solid move for making AI global, though it's not exactly fresh news at this point. Pathway's "(Baby) Dragon Hatchling" swaps out Transformers for neuron-synapse networks—a cool experiment, but let's be real, it's years from being useful. Open Source AI Week kicks off Monday with the standard hackathons and sessions. And Ford's Vignesh Kumar showed off AI for vehicle checks at Data Hack 2025, which could really change how cars get serviced if it ever moves past demos.
Connections and Patterns
Connecting the Dots
From all this, a few patterns stand out. One, platform lock-in is speeding up, with OpenAI's app push, Anthropic's skills framework, and Google's developer tweaks all pointing to a handful of firms dominating AI's core layer. It's reminiscent of the mobile battles from 2007 to 2012, but with even bigger risks, I suppose.
Two, there's this widening gap between what AI can do and what people claim it can do. Karpathy's down-to-earth comments clash with the over-the-top buzz about new models, reminding us that we're still figuring out reliable systems, not cruising toward some AI utopia. Maybe that's frustrating, but it's the truth as I see it.
Three, geopolitics is creeping in big time. That NVIDIA-TSMC chip win, paired with OpenAI's revenue drive and platform grabs, hints at the US doubling down on AI leadership. It's no longer just tech talk; we're dealing with fights over economic power and the future of computing, and that could get messy.
The standout from this week that might still echo in six months is OpenAI's platform strategy. Even if their revenue targets turn out spot-on or way off base, they're planting seeds for an ecosystem that's tough to leave once developers dive in. AgentKit and the Apps SDK aren't random add-ons—they're building blocks for keeping people hooked, for better or worse.
Keep an eye on how developers react to OpenAI's moves next week, if Anthropic offers a real counter, and whether NVIDIA and TSMC's investments give them a lasting edge. The AI game is pivoting from just model smarts to who controls the platforms, and the coming months could decide the winners, assuming nothing derails it. Who knows? It's always a wild ride.