Gemini 4 Argon, DoorDash Air, Griffin's Video Turing Test

For those of you that are new here. Welcome to the party, pal.

The format I’m working on: one prediction at the top, the stories that earned their spot and the vision to close it out.

The gap between "AI as a tool you use" and "AI as an entity you interact with" is almost gone.

Google's Gemini 4 Argon is a frontier model built for complex coding pipelines, legal analysis, and cybersecurity defense, rolling out to trusted testers through the Fairwind Program.

DoorDash launched both a drone delivery service and a text-to-order AI agent living natively inside iMessage.

Tavus's Griffin model just passed the video Turing test: 48% of live human conversation partners thought they were talking to a real person. Previous systems had a pass rate under 3%.

My bet is that within 18 months, at least one Fortune 500 company will publicly acknowledge using AI video agents for a customer-facing workflow where the AI's nature was not disclosed upfront, and it will become a legal and regulatory milestone.

Below: how Cloudflare is quietly becoming the backend for the agentic era, why Figure killing its own robot brought back the Terminator, and what OpenAI just did in mathematics that should recalibrate your timeline for AI-driven scientific discovery.

If you know someone building something, forward this to them.

Google DeepMind announced Gemini 4 Argon this week, a frontier model built specifically for sustained, multi-step reasoning across complex coding pipelines, enterprise knowledge work including legal and financial analysis, and cybersecurity defense.

It is not rolling out broadly. It is starting with a curated group of trusted testers through something called the Fairwind Program, which tells you Google is being deliberate about this one.

Tavus just released Griffin, the first model to pass the video Turing test: 48% of people who held a live video conversation with it believed they were talking to a real human. The previous best for any system was under 3%.

If you are building anything that involves identity verification, video interviews, customer service calls, or any interaction where the humanity of the other party is assumed, your threat model just changed.

Watch for regulation targeting AI disclosure in video communication within the next 12 months. That is the signal that will tell you how quickly this reshapes legal and enterprise norms.

DoorDash launched a text-to-order AI agent that lives inside iMessage, letting you order food without opening an app. Useful on its own. But the more important signal is the enterprise MCP waitlist, which positions DoorDash as infrastructure for AI agents handling corporate meal ordering.

The consumer feature gets the press. The MCP layer is the actual business move.

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free

A team of Claude Opus agents identified two room-temperature magnetic semiconductor candidates for next-generation spintronic memory: one entirely novel material the agents designed from scratch, and one buried in a 1999 paper that nobody had connected to this application.

Materials discovery has historically taken decades of trial and error. An autonomous AI research loop just compressed that.

@ClaudeDevs announced that Claude Code is now open to community modification: developers can change how it behaves, customize the UI, and ship their own features using a few lines of TypeScript, with plugins installable via /plugin in the CLI or desktop app.

This is the exact playbook that turned VS Code from a good editor into the dominant editor.

@claudeai announced that Claude now lives as a sidebar inside Google Docs, Sheets, and Slides, reading your open file and editing it in place, with an approval step before each edit commits.

Most standalone AI writing tools fail not because the output is bad but because inserting it into a real document is friction. Not anymore.

TechCrunch reported that OpenAI is rolling out shopping features letting users upload their own photos to virtually try on clothing and save favorites to a library, a direct move against Google Shopping and Amazon at the purchase-decision moment.

One of my recent projects has been in this space. I built a Chrome extension which lets you put yourself in an outfit in a model's pose on retail sites.

@AravSrinivas shared the run stats: Perplexity's Decisions API cleared Pokémon FireRed's Elite Four and Champion in a single shot with 592ms median response time, 96.4% of responses under one second, and a total inference cost of $0.028 across 137 live API calls.

The number that should stay with you is not the Pokémon result. It is $0.028 for 137 sequential decisions in a complex, rule-based environment. AI is becoming too cheap to meter. Well some of it.

The Department of the Air Force's Chief Technology Office unveiled Project Meridian via @DoWCTO, framing it as the initiative to secure American dominance across every future warfighting domain through disruptive emerging technologies.

The military AI race is no longer being fought quietly in procurement documents. Watch which commercial AI infrastructure contracts follow this announcement over the next two quarters.

@hark_labs launched Hark Pro on web, iOS, and Android, pitching it as an operating system for getting things done across your existing tools. Separately, serial founder @adcock_brett, the person behind Figure AI and Archer Aviation, announced that his new venture Hark is launching this week with the first 100,000 signups getting a paid plan free.

The AI productivity layer is getting crowded fast, with Hark competing against Muse, Dots, and Grok for the same real estate in your workflow.

@stanleytang, a pilot, made the key point in his own announcement: you would not build a plane and then ask if it can carry cargo. DoorDash Air was designed the other direction, around 13 years of real delivery data covering what people order, how much it weighs, and how far it needs to travel.

@Figure_robot announced the decommissioning of F.02, its current humanoid platform, framing it as a deliberate generational leap rather than a failure. Think of the T-800 being lowered into molten steel: a deliberate, eyes-open ending that makes room for something that cannot yet be shown.

In the humanoid race where Tesla, Agility, and others are iterating incrementally, Figure is signaling it will jump a generation.

@natfriedman announced Muse Gadgets: an open source ESP32 firmware and Linux SDK letting developers build custom hardware that connects directly to Muse's AI capabilities. Grab an API token, point a coding agent at the GitHub repo, and you can have a physical device querying your second brain before lunch.

I have been using AI to design an Elgato Stream Deck-style device for a few weeks now, so this landing in the same week felt timed for me personally.

Two products landed this week making the same fundamental argument. @zainmfj introduced Core, pitching itself as the first true personal AI computer where intelligence lives on the device rather than someone else's server.

Days later, @Luofei_Chen announced Oki Home, a personal AI computer shipping mid-December targeting consumers who want their data to stay local.

Both products are making a bet that frustration with Big Tech data dependency will drive real purchasing behavior.

@CallofDuty announced Server Culling for Modern Warfare 4: a server-side technique that simply never sends players information about enemies they cannot legitimately see. If the data never reaches the client, the data cannot be exploited. No cheat software can read what was never transmitted.

@modretro is dropping a new colorway of the Chromatic on October 8, the premium Game Boy Color revival that Palmer Luckey's team built for people who grew up with the original and never really moved on.

@davidellison and Ynon Kreiz introduced the Skydance leadership team this week, with @paramountco separately announcing the team that will lead Skydance following the close of its acquisition of Warner Bros. Discovery.

The people now in these seats will decide what happens to HBO, DC, CNN, Paramount Plus, and Max as streaming economics continue to punish everyone who got there too late.

Progress in distinct domains, models, drones, video AI, materials science, hardware, does not automatically compound. History is full of capable technologies that coexisted for decades without creating the acceleration everyone expected.

The transistor and the laser both existed in 1960. It took thirty years and an entirely separate set of infrastructure investments before they rewired commerce. The reasonable skeptical case is that Griffin passing the video Turing test, AI agents discovering spintronic materials, DoorDash Air launching, and Gemini 4 Argon targeting enterprise workflows are four separate stories that happen to be in the same week's news cycle, not evidence of a single compounding wave.

For that skeptical case to be right, you would need the infrastructure layer to stay fragmented, you would need model capability gains to slow significantly, and you would need hardware costs to stop falling.

None of those conditions currently hold. Cloudflare is assembling the agentic backend. Inference costs just demonstrated $0.028 for 137 live sequential decisions. The ESP32 costs less than a lunch. And materials discovery, the thing that historically bottlenecked hardware progress for decades, is now running on autonomous AI research loops.

What makes this moment different from previous technology booms is that the feedback loops are themselves being automated. AI agents discovering room-temperature magnetic semiconductor candidates are accelerating the hardware that will run better AI agents. DoorDash's 13 years of delivery data trains the logistics intelligence that makes its drone viable, which will generate new data that trains better logistics intelligence.

The compounding is not metaphorical. It is structural, and it is happening at every layer of society simultaneously.

That is why I write this newsletter. Not to chronicle what shipped, but to map the structure underneath it so you can orient toward what comes next.

The pace of progress this week was not a headline. It was a baseline. Every edition from here is going to feel like that. The question is whether the thing you are building is pointed in the right direction before the next wave of announcements lands.

Think about the last thing that got you fired up about your job this year. Odds are it didn't come from a report. It came from someone who does the job, saying what happened when they tried it.

So we've filled a day with people like that. 

Maya Shaff, ŌURA. Leah Walker, Adobe. Georgia Humphries, Stanley 1913. Enara Roy, Halfday. Tyler Vaught, Edelman. Josh Rangel, Ogilvy. Sarah Whittle, ex-Crocs and Duolingo. 

And that's just the first wave of speakers. 

Return on Influence Festival '26 is a free, virtual day dedicated to influencer marketing. Like any good festival, there are stages. But no mud, no queues, and absolutely no porta-potties. 

Sign up today, and the rest of the lineup lands in your inbox first.

Save your spot

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more