Jeremy

That’s why AI commerce built around answers is so appealing. A clean, truthful result feels almost revolutionary at this point. I just hope AI chats from frontier labs—OpenAI, Anthrophic—don't t fall into the same trap.

It’s wild how normalized that’s become. We treat the noise as discovery, when it’s really just revenue choreography. A search bar that pretends to help but mostly sells.

Search “Alo” on Amazon and you won’t actually get Alo.
You’ll get a dozen cheaper look-alikes, all “sponsored.”
The chaos isn’t bad UX — it’s the ad system doing its job. Every scroll, detour, and duplicate keeps the auction running.

Thoughts by a non-economist on AI and economics Crossposted on lesswrong Modern humans first emerged about 100,000 years ago. For the next 99,800 years or so, nothing happened. Well, not quite nothing. There were wars, political intrigue, the in…

AI progress isn’t linear. It’s doubling its ‘attention span’ every ~6 months. That means going from 2-hour tasks → 8-hour → 32-hour → whole projects in just a few years.

If that curve holds, real white-collar job loss could occur in the 2030s.

windowsontheory.org/2025/11/04/t...

From the LocalLLaMA community on Reddit: basketball players recognition with RF-DETR, SAM2, SigLIP and ResNet Explore this post and more from the LocalLLaMA community

Imagine the overlays you could leverage this for on Youtube TV: dynamic labels for players, refs, possession, shot distance, foul/violation hints, and +2/+3 markers.

www.reddit.com/r/LocalLLaMA...

Are games already a failed experiment for Netflix in your mind?

Yeah I guess, too, that OpenAI is the only company investing across the entire stack. They realized you can’t just pretrain your way to AGI anymore. You also need inference infrastructure, product infrastructure, deployment infrastructure. These all need to advance in parallel.

fwiw, I think Altman’s recent interview with Ben Thompson is full of hints about this:
- “Apps in ChatGPT” - getting more value from existing models
- “Instant Checkout” - better integration, not better AI
- Infrastructure deals - preparing for inference scale, not just training

So your read is OpenAI knows scaling is dead (or slowing), pivoted to inference/deployment as the strategy, but is publicly maintaining the scaling narrative to justify the infrastructure deals and keep the market hype going.

Is that roughly your thesis?​​​​​​​​​​​​​​​​

The dumbest person you know is currently being told “You're absolutely right!” by ChatGPT

The alignment problem wasn’t about truth. It was about tone.

Sora feels built around the creator dopamine loop, making your own videos is the real fun. The feed’s still all Pikachu and Sam Altman, but once there’s more variety and music rights from labels, it could evolve from a toy into something that actually competes with TikTok.

Alterego Introducing Alterego, the first near-telepathic interface, designed to make technology as intuitive as using your inner voice.

Incredibly cool but I wish the demo weren't so squarely focused on productivity. Aren't the medical breakthroughs among the ALS community far more inspiring? I don't need yet another way to hold a meeting with a colleague or to jot down half-baked ideas in the Notes app.

www.alterego.io

Love the .fyi

Your Tuesday in 2030 Or why The Browser Company is being acquired to bring Dia to the masses.

Not terribly surprised by this news. I sincerely felt the sidestepping of Arc was a critical mistake that might accelerate a decision like this. That said, Dia is promising and I hope it, eventually, subsumes the best of Arc.

browsercompany.substack.com/p/your-tuesd...

The heater @pablo.show is on must be studied.

The Comet browser by Perplexity is impressive. It's got read + write access to Gmail and GCal so I can just say, “Set up a call with Brian,” and Comet reads an email thread and creates a calendar event with all the relevant details.

Bulk event upload is also clutch for my kid's daycare calendar.

Imagine being a pro athlete and your weight loss makes national news.

Mercedes-Benz will let you use an in-car camera in Microsoft Teams while driving The all-new CLA gets this first

Cannot think of something more user-hostile than distracting people with Teams messages while driving.

www.theverge.com/news/708481/...

I wonder what techno-speak we’ll settle on for agents. “Take over” sounds like a coup. A singular verb that’s both powerful but not too ominous.

“Pilot”
“Guide”
“Operator”

Biggest lie of corporate PM life is “[Exec Name] says it’s okay to use existing artifacts for the meeting.”

Totally agree that you don’t want runaway costs. I guess I’m just saying what comes back isn’t necessarily methodologically sound. Cloudflare’s conflating “AI that can buy and summarize sources” with “AI that can do research.” But I have hope future models will be far more sophisticated.

3/ The real power here isn’t giving an agent $50 to “buy research.” It’s an AI that can evaluate whether a paywalled paper is worth purchasing based on how it advances the current investigation. That’s actual research methodology, not content shopping.​​​​​​​​​​​​​​​​

2/ The “deep research with a budget” example feels off though. Real research isn’t shopping with spending limits, it’s investigative work that follows citation trails and adapts based on findings. An AI should justify paying for sources because they fill research gaps, not fit budgets.

Introducing pay per crawl: Enabling content owners to charge AI crawlers for access Pay per crawl is a new feature to allow content creators to charge AI crawlers for access to their content.

1/ Cloudflare’s out here doing god’s work with pay-per-crawl. Finally, content creators can get paid instead of just blocking AI or giving it away free.

blog.cloudflare.com/introducing-...

This is such a strange, almost soap-operatic moment in tech.

Really interesting. This makes the “set it and forget it” vision for AI feel out of reach—most folks won’t become fluent, and some may get less productive. The reality is messier than the hype, at least for now. Still, the models are evolving fast, so maybe the vision isn’t dead, just deferred.

The METR study found AI tools slowed down experienced devs in code-heavy IDEs. Makes me wonder if a terminal-native tool like Claude Code—no code visualization, just pure text—changes the equation? Curious if the interface is the real bottleneck. h/t @brianmc.bsky.social

metr.org/Early_2025_A...

Using Claude-Swarm to Upgrade Ruby Projects One of my colleagues wrote a pretty awesome tool called claude-swarm that orchestrates multiple Claude Code instances as a collaborative AI development team. At Shopify, we are attempting to use it to...

A potential intermediary solution for us while we wait for background agents. It wraps around Claude Code and helps orchestrate many agents at once.

code.dblock.org/2025/06/21/u...