Astra Programs AI, The Battle for TV, Longevity Goes Mainstream

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.

AI just learned how to edit AI, and the compounding from here is hard to overstate.

OpenAI’s GPT-6 Astra changed just 0.016% of a smaller model's weights and pushed its math score from 7.9% to 48.1%.

a16z published its State of Longevity report this week, interviewing 500 Americans and Andrew Huberman on where the science and the money are headed.

Instagram is paying YouTubers up to $500,000 to post long-form video while Netflix locks in Druski for a multi-year exclusive deal. TV attention is now a bidding war.

My bet is that within six months, we will see a publicly demonstrated case of a frontier model using weight-space programming to produce a meaningful capability leap. This will force the labs to disclose how much of their internal progress has already been driven by exactly this mechanism.

Below: SpaceX quietly making orbit a commercial utility, Greg Isenberg describing what it actually feels like to run a company in 2025, and Mark Cuban's argument for where the next wave of jobs is coming from.

Two research threads converged this week on the same finding. @rikiyatakehi showed that GPT-6 Astra directly edited the weights of a Qwen3.5-0.8B model, updating only 0.016% of parameters, and improved its math performance up to 6x and instruction following 3x with zero gradient-based training.

@kimmonismus picked up the same result from the MIT angle: math scores moving from 7.9% to 48.1% via surgical weight edits, no retraining pipeline required. The researchers are calling it Weight-Space Programming.

What this opens up is recursive. If a large model can read and write the weights of a smaller one, the question you should be sitting with is obvious: what stops it from doing the same to itself?

The labs have been tight about their internal training methodology. My bet above is that this is partly why. Watch for the first peer-reviewed paper that shows a model improving a same-scale peer, not a toy model.

That paper changes the trajectory of everything downstream of it. Something to keep an eye on.

Greg Isenberg runs a 50-person company across 12 countries and says he talks to AI agents more than his human team. The workflow: voice note to a bot, an agent interrogates the idea before a single line of code gets written, a spec drops automatically from the conversation.

What he describes as "kinda dystopian" is actually the first honest public account of what async-first AI-native management looks like in production.

The manager-to-agent relationship is not a productivity hack layered on top of the old org chart. It is a different org chart. If your team still treats AI as a tool that waits to be prompted rather than a collaborator that challenges the brief, you are leaving the most valuable part of these systems completely idle.

Mark Cuban is making the case, louder than most, that generic SaaS is functionally over. His argument via @guilleflorvs: AI will replace off-the-shelf software with custom tooling built to match how each individual business actually runs. There are 33 million small businesses in America alone. Someone has to build, deploy, and maintain the custom AI layer for all of them.

The opportunity is not building the next SaaS product. The opportunity is becoming the person who replaces SaaS for a specific vertical or customer type. If you have domain expertise in any industry and basic technical fluency, that combination is worth more right now than it has ever been. Start with one business you understand deeply and build the tool it actually needs.

He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.

They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.

Get started for free.

YC-backed Cyclon is building what they describe as an AI phone that operates itself: AI handles tasks in the background using your existing apps, builds new ones when something is missing, and responds to natural language across the board.

The pitch is that the app grid you have been swiping through for 15 years is a relic of a pre-agent world.

The graveyard of Rabbit R1 and similar devices is a real cautionary note here. The difference Cyclon is betting on is integration with existing apps rather than a parallel ecosystem.

Whether that holds up in practice is the only question that matters. Watch the retention numbers at six months, not the launch demo.

Anthropic released FlashProve this week, a cryptographic system that lets developers verify whether a given response actually came from a specific AI model. In a market filling up with model impersonation risk, fine-tuned imposters, and enterprise compliance requirements, provable inference is the trust layer the whole industry has been missing.

If you are building an enterprise product on top of any model, auditability is no longer a nice-to-have.

Regulators are going to require it and enterprise buyers are already asking for it. Get ahead of this before it becomes a procurement blocker.

SelendisAI launched Archon, a video foundation model co-built with Steve Aoki that takes raw footage and produces a finished edit in minutes. The more interesting part is not the automation.

Archon builds a personal taste profile from your past editing decisions, meaning the output gets closer to your creative voice over time rather than producing generic cuts.

Two moves happened in the same week that together define where the streaming wars have landed. Netflix signed Druski to a multi-year exclusive deal, with his first project "Collect and Praise" premiering December 1, built directly from a character he made viral through internet sketches.

Meanwhile, Instagram is paying top YouTubers up to $500,000 to post long-form video and algorithmically rewarding content that runs longer.

Even I find myself in front of the TV switching between Netflix and YouTube constantly because of my son’s demands. One of those is Danny Go on Netflix, a creator from YouTube. Instagram hasn’t made it into the rotation yet, but we’ll see.

SpaceX launched the Starfall website this week, formalizing a free-flying spacecraft that lets companies rent microgravity time for manufacturing, research, and testing.

Access is reusable and routine. You no longer need a seat on the ISS or a custom mission to run an orbital experiment.

I am genuinely bullish on the commercial space industry as the next hotbed of opportunity, and Starfall is the clearest proof yet that "space" is becoming infrastructure rather than spectacle.

Pharma, materials science, semiconductor manufacturing: the list of industries that benefit from microgravity is long. The barrier was always access and cost. Starfall removes both arguments.

SpaceX is reportedly exploring an acquisition of Boost Mobile or a network partnership that would transform it from a satellite provider into a full-stack mobile carrier. That means owning the satellites, the terrestrial network, and the rockets that launch and maintain both.

Every major carrier should be modeling what their customer retention looks like if SpaceX offers a combined Starlink and mobile plan at competitive pricing. That scenario is closer than it was 90 days ago.

Andreessen Horowitz published its State of Longevity report, built on a survey of 500 Americans and practitioner interviews including one with Andrew Huberman. When a16z publishes a category report, it is not just analysis. It is a market map of where they are about to write checks.

Longevity graduating from biohacker subculture to mainstream VC thesis is a meaningful inflection point. The capital that follows a16z attention tends to accelerate the entire category: more companies funded, more regulatory engagement, more public awareness.

If you are building anywhere near healthspan, aging biology, diagnostics, or longevity protocols, the tailwind just got materially stronger. The window to get in early is narrowing.

Will the whole AI builds better AI take off from here? It is entirely possible that this technique works beautifully on small models and breaks down completely at the scale that matters. That would make this week's result a fascinating footnote rather than a turning point.

What we do know for certain is that the gap between what is publicly available and what is running inside the labs has always been real. Weight-space programming is a credible mechanism for why that gap may be wider than anyone has acknowledged.

If GPT-6 Astra can do this in a public research context, the more likely scenario is that frontier models have been doing some version of this internally for longer than the published literature suggests.

The longevity thread connects here directly. The hardest constraint on biological research has always been iteration speed: hypothesis, experiment, result, revision. If AI systems can now improve other AI systems faster than humans can improve them, the same recursive acceleration becomes available to any research domain those systems touch.

Drug discovery, protein folding, aging pathway analysis. a16z publishing a longevity market map in the same week that recursive model improvement becomes a demonstrated reality is a coincidence in timing, not in direction. The two stories belong in the same sentence.

Greg Isenberg calling his working life "kinda dystopian" while running a 50-person global company through AI agents is the most honest thing anyone said in public this week. Most of the world is blissfully unaware that this is what operating a business looks like right now when using the latest tools.

Do you feel like you’re using the latest tools? Are you enjoying it? Sometimes I don’t even think I feel comfortable on the bleeding edge. Imagine being an insider at one of these labs. That would be an interesting perspective to see amidst this AI gold rush. Albeit stressful I’m sure..

Looking for more?

Are you ready? As revealed in our holiday playbook, Roku Ads Manager lets you easily set up, measure and optimize your self-serve holiday ad campaign. You can even create a shoppable ad unit that lets streamers buy directly from their remote.

Set up your campaign today.

Aligned shows you what buyers are doing between meetings inside your most important deals then surfaces the risks, openings, and next steps. No more guessing where a deal stands or finding out a champion went quiet too late. Just real visibility, so you can act before it's too late.

Start Using Aligned Free