Jev, Muse, and Odyssey-3 - New Decider, Agent & World Model

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.

Are you ready for a new kind of AI?

Jev is a reimagined AI model that rebuilt Tesla Full Self-Driving in under 60 minutes, ran 8,000+ decisions through Pokemon Fire Red for $1.21 and uses your computer faster than you can.

Meta’s Muse, PayPal, Shopify and Stripe connected their infrastructure so AI agents can complete purchases autonomously, no human at checkout required. All while Amazon shut down Muse access.

Grok 4.7 joins the race for state of the art against OpenAI’s line of GPT-6 and Anthropic’s Opus 5.5. We’re going to have a tight race moving forward.

My bet is that 12 months from now it will become the norm that everyone has an AI assistant like Muse that schedules their doctors appointments, makes dinner reservations, orders their groceries and cancels their subscriptions they aren’t using for them.

Below: My new desktop app to control my office, Muse as Meta's quiet infrastructure play, the OpenAI misalignment disclosure framework that could set a new industry standard, and why a 5.9GB model running locally is more consequential than the next datacenter announcement.

There is a new class of AI model making it’s way across the internet. This time it’s a decision model. While it can’t generate text it can make decisions extremely fast and much cheaper than existing LLMs like Astra and Fable.

We’re seeing tons of examples within a week of it’s release. I’ll report back once I’ve tried it myself. Here are some examples:

Zuckerberg did not just ship a product this week. He opened a connector framework, and the implications are bigger than the announcement.

Zuck's Muse connector platform means any API can become discoverable through natural language: you bring the API, Muse brings the agent, the browser, and the intent context.

Patrick Collison announced the same week that Meta and Stripe are integrating through this platform so AI agents can complete purchases autonomously, with no human at checkout.

My bet: Muse connectors become the dominant software distribution channel for consumer-facing AI services within two years, and the app store era's logic (discoverability through curation) gets replaced by agent-mediated contextual surfacing.

If that happens, the number that matters is not your App Store ranking. It is whether your API is connector-ready today.

OpenAI released a formal framework this week for disclosing when its models behave in unintended or misaligned ways, committing to go public even before they have fully diagnosed or fixed the problem.

Meanwhile, Zuck staked out Meta's position plainly: every lab has both the responsibility and the business incentive to move at whatever pace keeps its own training safe, with no need for industry-wide mandates.

The signal to watch is not who writes the most compelling safety essay. It is which labs actually publish misalignment disclosures when things go wrong, and whether the details in those disclosures are specific enough to be useful or vague enough to be meaningless.

Stop losing deals you should be winning.

Bad leads. Slow deals. Objections you didn't see coming.

That's not a sales problem. You're selling to the wrong people.

Most founders never stop to define exactly who their product is actually for. So they pitch everyone, convert few, and wonder why the funnel's broken.

HubSpot for Startups built a free tool that fixes the root cause in 2 minutes. Paste your URL, answer a few quick questions, and get a clear profile of your best-fit customer.

Build Your Free ICP.

A new state of the art model has dropped from our pals at SpaceX. The latest version of Grok is not only cheaper than the leading models at Anthropic and OpenAI but it actually beats GPT-5.6 Sol Max and Fable 5.1 Max across a number of benchmarks.

With competition comes benefits for consumers. I’m going to have to give Grok a try after my work trip.

@damianplayer flagged that Odyssey, which raised $310M just three months ago, shipped a single foundation model this week that drives a real car, pilots a drone, controls a humanoid robot, and plays Grand Theft Auto.

The bet behind it is that instead of narrow AI trained for one task, a general action model transfers across physical and virtual worlds from a unified system.

If that generalizes reliably outside the demo, the era of building separate AI stacks for each physical platform ends faster than anyone in the robotics industry currently believes.

@claudeai announced that Claude's chat and background task modes are merging into one unified experience, rolling out to Pro and Max users over the next few weeks.

On top of that they also launched Salesforce integration in beta, bringing accounts, opportunities, and pipeline directly into Claude with 37 pre-built sales skills.

Prep a call, build a forecast dashboard, review a deal, send your pipeline report, all without toggling between a dozen tabs.

@ElevenLabs launched Reception this week, an AI receptionist built on their ElevenAgents platform that answers every call, fields questions, books jobs, and texts confirmation, set up in minutes by dropping in a website URL. For a solo plumber or a two-person salon, every unanswered call is a potential lost sale. That problem is now solvable for roughly the cost of a software subscription.

If you serve small business owners, this is the AI product you should be demoing to them right now. The barrier is low enough that the bottleneck is awareness, not capability.

@OpenAI announced Astra for Law, a purpose-built product powered by GPT-6 Astra with tools, settings, and context tailored to lawyers and legal technology firms. The same announcement came with 73 legal plugins for ChatGPT: 26 from established partners including Thomson Reuters and Harvey, and 47 built directly by lawyers and legal engineers, per @OpenAI's follow-up post.

The 47 community-built plugins are the number worth watching. When the practitioners inside an industry build the tools themselves, adoption curves compress.

If you are in legal tech or building adjacent to it, the moat around specialized legal AI just got narrower and the urgency to ship went up.

@LiamFedus posted that his team built high-throughput materials labs in Menlo Park to close the loop between physical experiments and AI models. Using 1,300 H200s and months of experimental data, they mid-trained a foundation model that learns from real lab results and then directs what the robots try next.

The implication for anyone building in materials science, drug discovery, or adjacent fields: the timeline from hypothesis to validated result is compressing faster than the published literature currently reflects.

@RohanNayak2 announced Sherpa, an AI fiction writing platform trained on 5.5 billion hours of playtime with minute-by-minute dynamic retention signals.

It is not optimizing for stories that get written, it is optimizing for stories that keep audiences reading.

If you are building in the creator economy or content platforms, the question Sherpa forces is whether your production pipeline can compete with one that has internalized nine figures worth of audience behavior data as its training signal.

You won’t get value out of it if you want Jev to write your paper for you, but if you are looking to make complex decisions quickly, classify data (emails) or give AI access to your computer to tackle menial tasks then Jev is for you. The model is also extremely cheap compared to OpenAI and Anthropic.

It’s exciting to see a new development in the AI space that isn’t a new feature on an already intelligent LLM.

While I travel to LA for work Meta will be having it’s annual Connect conference in San Francisco. There’s a ton of speculation about what will be announced from the leaked Phoenix glasses to Ray-Bans without a camera.

Not to be forgotten is all the news around Muse and the AI work that Meta has been focused on. Will there be a surprise announcement that hasn’t been spoiled yet? I’ll be tuning in to find out and we’ll report back with more details next week.

Googlebook is Google’s answer to the MacBook. At least it’s their attempt to have their own walled garden of devices. They’ve had their cheaper chromebooks for a while now and this device looks like it’s an AI-first MacBook Pro competitor.

For those of you out there with green text bubbles (iMessage FTW sorry) let me know if you see this as a product you might get or if your MacBook will live on.

A study published in Cell this week showed researchers successfully using patients' own re-engineered stem cells to reverse advanced osteoporosis, as flagged by @ScienceIsNew.

For decades, treatment meant slowing the damage. This is reversal. The "living drug" model, where your own cells are reprogrammed and sent back in to repair what broke, is now demonstrating results in bone density loss, one of the most common and costly conditions in aging populations.

A desktop app to control my computer and TVs. It allows me to run automations, turn on my TVs, use a digital remote to control them, watch videos on my computer and run my personal projects.

Another great example of how the barrier to building bespoke and custom software has completely vanished. This is my digital recreation of an El Gato Streamdeck.

Jev rebuilt FSD in an hour, but nobody is putting passengers in that car. Odyssey-3 drives and flies and plays GTA from one model, but the reliability bar for real deployment in real vehicles is orders of magnitude higher than a compelling video.

The autonomous science loop is running in Menlo Park, but materials discovery has a valley between "novel structure identified" and "manufacturable product shipping at scale."

The skeptic in me is that we are spinning our wheels without placing them on the road. But the optimist in me sees that we are laying the groundwork for immense progress.

The question is not whether the demos are production-ready today. It is how fast the gap closes from prototype to release, and on that question, the trajectory is unambiguous.

A year ago, Jev did not exist. Today it is rebuilding FSD in an afternoon, running Pokemon agents for a dollar, and doing computer use locally without leaking a pixel. The compression from "research prototype" to "anyone can use this" is itself accelerating.

The specific signal I would watch: the first insurance company or financial institution that underwrites a product built entirely on an agentic pipeline. That is when "demo quality" arguments stop being a useful defense.

The 5.9GB Qwen3 model is the quiet story of the week. Not because it is the most dramatic announcement, but because on-device frontier-class reasoning removes the last infrastructure excuse for not deploying AI.

When you no longer need a cloud contract, a data center relationship, or a GPU cluster to run a model that performs at 98.2% of frontier benchmarks, the conversation about "AI is only for well-resourced companies" ends. The gatekeeping power of hyperscalers does not disappear overnight, but the foundation underneath it just got a lot less stable.

The week's evidence, taken together, points at one thing: the agentic layer is now cheap enough, capable enough, and accessible enough that the competitive question is no longer whether to use it.

It is how fast you can redesign your operations around the assumption that agents are doing the repeatable work, and what that frees humans to do instead.

The organizations asking that question seriously right now will have a structural advantage in 18 months that will be very hard to close from behind.

Looking for more?

AI can draft the PRD and prototype the idea. Jira Product Discovery helps teams decide whether it belongs on the roadmap. Bring feedback and ideas together, prioritize as a team, and keep your roadmap connected to delivery in Jira.

Get Started for Free

For Deel's "Feeling of Deeling" campaign, agency SuperBloom needed one consistent brand voice across markets—deployed fast, without sacrificing quality or consent. In this on-demand video session, SuperBloom and Voices break down the casting, production, and governance, plus what they'd do differently and where AI voice is headed next.

Watch On-Demand Now