Seth Kravitz

Coming to a strip mall near you.

Trying to read an article in 2026.

A lot of these devices have a safety features coded into them, related to overheating or exceeding motor limits. I LOVE the jailbreaking devices movement happening now, but not ALL of the software on these devices is garbage that needs to be replaced.

Not sure how more people haven't figured this out. Over the past year I have migrated over $800/mo in services, apps, sites for a single $5/mo Workers plan and a few bucks worth of extra usage.

What stage of capitalism is this?

Push to main or nah?

Ummmm... sure...

Took me almost two weeks to go through my $5 @typesafeai credits. The value is insane.

I have switched back to Opus 4.8, as Opus 5 is almost unusable. I cannot be the only person that feels that 5 was a rush job to race something out well before it was ready.

Yea, that's great. But can it understand "Toilet Paper" math? 🤔

The billing team @AnthropicAI needs fix their policy around card declines and service access. You are launching a $1.5 billion "Services" division, but your curent billing policy is so anti-customer. I use virtual cards in @mercury and assign every SaaS its own card, which have spending limits on them. The card assigned for Anthropic exceeded its daily spending limit, so the transaction was declined. That is the first time I have had that happen with them, so it's the first time I got to experience their "card declined" process. You INSTANTLY lose access to ALL Anthropic services the second your card is declined. That means web & desktop app, mobile app, Claude Code CLI, and API access. Across my various companies and consulting I probably have 40-ish SaaS products I use. I cannot think of another company I use that has a policy of instantaneous denial of all services and access the second a card is declined. EVERYONE has a grace period of at least several hours to several days. You need to do better.

People worry that AI coding agents generate poor quality code. Yes, yes the existing software we have is human made perfection.

Deepseek v4 Pro inside of Pi is almost too good to be true. Getting Opus 4.7 level results with what might as well be uncapped unlimited usage.

To the people wondering how the Copilot rollout has been so terrible. This is the company you're talking about...

I vote “Unwrapping a Starburst” for the next Summer Olympics sport.

The Forbes 30 Under 30 prison pipeline continues to be undefeated.

Anyone else think about this paper at least once a week at this point? To me, the simple premise of what they did here is both ground breaking and seems so obvious at the same time. arxiv.org/html/2511.09030v1

Imagine being able to have unlimited use of a free LLM, running on your own computer, as powerful as Opus 4.7 or GPT-5.4. That will be reality this year.