The AI Race Just Moved the Goal Posts

This week felt like the AI industry moved the goalposts. Nvidia lined up $500 billion from Wall Street, Zuckerberg published a 6,500-word essay about open AI that Jeff refused to call a manifesto, Anthropic started watermarking everything Claude produces, and at least one major analyst group is calling Gemini “cooked.” Also, I did a live unboxing of a smart ring and probably ordered the wrong size.

But first, a huge shout-out to our newest patrons over on Patreon: Chad, Stacey Coy, S, Andrew Wielandt, Tim Sanchez, Philip White, and Bradley Deeley!

We’re celebrating reaching 100 paid patrons, and for one more week, new paid patrons get 30% off their first year with the coupon code 100PATRONS, on top of the 20% annual discount we normally offer. That’s close to 50% off for a full year of ad-free episodes, the AI Inside Daily podcast, Discord access, and an upcoming patrons-only AMA at the end of the month. We’re at 106 and trying to hit 120, so head to patreon.com/aiinsideshow if you want in.

Mark Zuckerberg published a big essay called “The Future is for Everyone,” and Jeff and I debated whether it qualifies as a manifesto. Jeff’s take: no, it’s more of a “hubristic offering” with an extremely utopian vision. The core argument is that AI should be open and available to everyone, not locked behind corporate API access, and Meta is backing it with over a billion dollars in data center funding.

Alongside the essay, they released Muse Glimmer, a 30 billion parameter open-weight model that runs on a single 24GB GPU or an M4 or M5 Max MacBook. As Jeff put it, that’s “money where mouth is.” M.G. Siegler over at Spyglass had a good piece calling it “Meta’s Open Weight Whiplash,” pointing out they’ve gone back and forth between open and closed. As we say on this show all the time, it’s open-ish. Open weights. Not fully open source. But whether it’s conviction or strategy, the real test is whether they open-weight the unreleased “Watermelon” model, their actual next-gen frontier system.

Anthropic announced they’re embedding two layers of marking into everything Claude produces: imperceptible text watermarks baked into the writing itself, and C2PA signed provenance metadata on generated files. This is partly principle and partly compliance with the EU AI Act labeling requirements that kicked in August 2nd. Jeff called it “technosolutionism,” arguing that the problem is technology but the solution isn’t always technology. The real work is individuals, authors, publishers, all of us refiguring our norms around trust. We also got into what Jeff calls the “AI has cooties” problem: if AI touches something, you’ve got the cooties now. I pretty firmly believe a lot of people calling the AI cooties thing are still using it behind the scenes and covering their tracks. It’s a weird moment where everybody’s using it, but not everybody’s talking about how.

Nvidia partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion in outside capital for AI data center infrastructure, aimed squarely at the circular financing criticism we’ve been covering for months. Jeff’s airplane leasing analogy clicked for me: United Airlines can’t afford to buy all its planes, so it leases planes bought by entities in the middle. AI compute is heading the same way. The challenge with securitizing chips, as the Wall Street Journal noted, is that they have “the half-life of heads of lettuce.” But Jeff argued this actually argues against the bubble: smart companies wanting in speaks to the maturing of the process, not the start of a crash. Ben Thompson over at Stratechery drew a comparison to the railroads and the Gilded Age crash, so we wanted that on the record in case it all blows up.

Google had a big leadership shake-up. Demis Hassabis is moving to a chairman and chief scientist role, Koray Kavukcuoglu is now SVP leading DeepMind day-to-day, Jeff Dean has departed, and Sergey Brin is expected to take a more active role. SemiAnalysis published a piece declaring “for all intents and purposes, we believe DeepMind is no longer a frontier lab,” which is shots fired as far as I’m concerned. But Tim O’Reilly wrote a counter comparing Google’s strategy to Westinghouse distributing electricity: instead of one huge central engine, distribute lots of small ones. Google Cloud’s 82% year-over-year revenue growth supports that read.

The Pixel 11 launched, and from the outside looking in, it feels pretty iterative, but the AI features are what make it moderately interesting. Magic Capture analyzes around 400 frames to grab perfectly timed photos. Rambler is the one I’m most looking forward to because, well, that’s how I am with AI: I ramble a lot. You blather into the text field and Rambler cleans it up for you while keeping it in your voice. Kind of cool. And the sign-to-text feature in Gboard using DeepMind’s SL2T model is a legitimately cool accessibility feature. Some of the neatest things happen in the accessibility realm.

A Wired piece covered a new technique for extracting AI models’ hidden reasoning traces, and one finding that jumped out is that some Chinese AI models appear to have been influenced by US models, feeding directly into the distillation debate. Jeff’s take: we’ve got to get past this notion that distillation is bad. “Life is distillation, folks.” He distills me every week, I distill him. We’re a distillery, the two of us.

I did a live unboxing of the Pebble Index 01, a voice-controlled smart ring for capturing audio notes and converting them into tasks through a companion app. It runs $75, integrates with Apple Reminders and Google Tasks, and uses local transcription. I got it because I’m constantly having ideas while driving and can’t reach for my phone. The unboxing was a little rough, though, because I think I ordered the wrong size, but voice recognition is solid, and the concept makes sense for how I actually work.

Wetherspoons, which operates over 800 pubs across the UK, has banned Meta Ray-Bans from all locations, joined by Jeremy King Restaurants, Soho House, and others. As someone who reviews these glasses, I still haven’t gotten used to wearing them around people because the stigma is real. Jeff pointed out the camera capability hasn’t changed, so why now? His read: the glasses are more popular, the companies are more hated, and they’ve become easy targets for AI backlash. Meanwhile, an Annenberg survey found 61% of Americans oppose local data centers, up 12 points from spring, and data centers have become a political football in midterm primaries.

OpenAI is working on a hardware device that Bloomberg says will be hockey puck-sized, cost over $300, and work as a battery-powered voice-only ChatGPT speaker designed by Jony Ive. My reaction: it’s basically a portable Google Home puck. And as Jeff noted, at $300 you’re not buying five of them.

OpenAI launched Daybreak with partners including Accenture, IBM, CrowdStrike, and Cisco. Two specialized models for defensive cybersecurity and red teaming, with enterprise guardrails built in. The more companies testing their software before deployment, the better.

Cloudflare open-sourced what they’re calling “Cloudflare OS,” a browser-based workspace for AI agents built on Cloudflare Workers. Jeff thinks Cloudflare is “getting weirder” and “spreading their wings too much,” potentially competing with the companies they serve. He thinks they should stay in their lane.

xAI launched Grok Bot, an AI agent that logs into your tools and runs multi-step workflows on Grok 4.5. Beta access is $300/month on SuperGrok Heavy or $200 on Cursor Ultra. Jeff’s take on the whole thing: “Once I see the word Grok, I know everything I need to feel.”

Nvidia released Nemotron 3.5 Lightning, a 30 billion parameter open-weight model with only 3 billion active per token through mixture-of-experts. It claims up to 4x the output speed of similar-sized models and runs on a single H100 or DGX Spark.

Spotify announced “AI Persona” badges for AI-generated artists, excluding them from recommendations by default. This isn’t about how the music was made but about the artist’s public identity, so human artists using AI tools aren’t affected. As Jeff put it, transparency about AI use when it’s relevant and meaningful is an ethic we should all engage in.

Executive Producers of AI Inside: DrDew, Jeffrey Marraccini, Radio Asheville 103.7, Dante St James, Bono De Rick, Jason Neiffer, Jason Brady, Anthony Downs, Mark Starcher, and Karsten Samaschke. Thank you for making this show possible.

Thanks for watching and reading. New episodes every Wednesday at aiinside.show.