We read both vendors' documentation and the recent work on looped transformers with three questions: what does a model's reasoning cost, what can you steer, and what can anyone read afterwards. The answers sort into three places.

On the page, as visible text. You pay the output rate, you steer it with the prompt, and you can read it, though studies show these visible traces are often unreliable.

Off the page, as hidden tokens. The example in OpenAI's own guide shows 1,186 output tokens charged, 1,024 of them reasoning; you pay for every one and the text stays with the vendor. Anthropic's docs show the same: you're billed for full thinking tokens whether shown or not. The rate is the same, the steering is an effort knob in place of a prompt, and what you can read is at most a summary. The exact count is in the usage object, and on the newer models earlier turns' hidden reasoning stays in context and is billed again as input.

Inside the network, as extra passes through the same layers, which is what recurrent depth means. Compute rises per token, the token count does not move, and there is no text for anyone to read, the vendor included.

For those who don't know:

• The EU has proposed a law intended to protect children from online abuse.

• But the draft includes a highly controversial measure: mandatory scanning (client-side scanning) of private communications, including encrypted ones, for child sexual abuse material.

• Experts and authorities (such as the EDPB and EDPS) warn that this kind of mandatory scanning could lead to mass surveillance through automated tools, posing a serious threat to fundamental rights like privacy and data protection.

• According to reports, 15 member states support the proposal, 6 are against it, and 6 remain undecided. The law would not apply to political officials, but it would affect the reading of communication of ordinary citizens.

With all the recent data leaks and surveillance laws, it feels like the vibe is shifting toward 100% local, on-device compute, even if it means settling for smaller models. Running things locally used to feel like a niche hobby, but with the massive leaps in open-source efficiency, it is becoming a viable daily setup for privacy-first projects.

Granted, dialing in the right hardware and dealing with speed tradeoffs is still a massive pain compared to a ready-to-go API.

Are you guys actually moving your daily workflows and tech stacks to local alternatives, or is cloud still your go-to for convenience? Where do you stand on this?

Building and launching products, testing them in real markets and building a business are like mini-MBAs. It teaches you a lot of things about human behavior, finance, marketing, sales, entrepreneurship, management and more.

We become wiser; and wish someone had given us the right advice at the right time.

I've been thinking about AI product economics, and I think "cost per API call" is only part of the problem.

An AI feature can look cheap from the infrastructure side and still become expensive to operate.

I spend a lot more time on PH at the moment to see what indepedent makers are spending their time on. I've noticed some patterns and also want to share a little bit about my journey at South Park Commons. Most startup stories begin at zero when there s already a team, an idea, maybe even a prototype. But at South Park Commons (SPC), the philosophy is different: people gather in the -1 to 0 stage. That liminal space where you don t yet know what you re building or even if you should build at all. It s a place for exploration, experimentation, and being brutally honest about what s working and what s not.

A hallmark of SPC is how often industry leaders drop by to share what they ve learned in the wild. Recently, I was in a small chat with Tyler Payne former Google and LinkedIn AI lead, startup builder, who has spent the last decade helping teams actually ship real-world ML systems. We're always talking about what's being launched at SPC.

Last week I attended Tech Week for the first time in my city and had a great time!

I was surprised by how much value came from conversations that weren't planned.
It feels that in a world accelerated and misconstrued by technology, the best way to establish trust is often to meet face-to-face.

For those who regularly attend Tech Weeks, startup events, or hackathons:

More big tech companies bet on AI instead of employees.

• @Duolingo is starting to prioritise AI over employees and contractors.

• @Shopify is planning on going "AI first."

• @Fiverr CEO Micha Kaufman just sent his employees the most brutally honest email: "AI is coming for your jobs. Heck, it's coming for my job too."

• Firecrawl hires AI agents and has a Slack channel solely for AI agents.

We re working on smart home tech that aims to be invisible, useful, and quietly perfect in the background.

It s got me thinking