I would love to know the balance of training material that was fiction vs non-fiction vs questionable (ie. reddit posts). Most human storytelling is not a pure recounting of reasons and events, it is deliberately fit into a dramatic structure, and it seems like these vending machine experiments always result in AI unnecessarily ramping up the drama instead of doing the boring thing and just buying and selling soda.

I'm not sure if it was a bad question or a bad take, but saying "Apple is insulated against an AI crash" really just means that Apple's not a player in AI. And when a bubble does burst, AI isn't going away. A lot of questionable businesses will disappear but the big players (Google, Microsoft, etc.) will still be around and Apple will still be behind.

Pioneering AI companies are going to go talk to the handful of companies with data center capacity, and lock in deals that prevent using anyone else's models in those data centers. So they will still be able to lock the market down without a needing a differentiator at the model level.

It's also a reminder of one of the biggest flaws with AI today-- Its complete lack of judgement. And I don't just mean safety guardrails. It's all over everything it does.. from unnecessarily verbose writing to fixing compiler errors by doing crazy things like looking up symbols dynamically at runtime to avoid a missing interface. AI is happy to achieve a goal, with no sense of the larger picture.

Governments should protect free speech. And please, go freely post your guesses about the future! Government should also regulate businesses. And Polymarket should be regulated out of existence.

AI solves an important problem which is that experts and smart people won't necessarily say what you want them to, because they have usually formed opinions and likely have morales. Today the primary options are to bribe or fire non-compliant experts, but AI creates a new option: An "expert" that can be directly controlled, has a fixed cost and can't be prosecuted because it's ultimately just an algorithm. It's the perfect solution for when you want to do something, but human experts would be problematic.

I worked for a startup that was building a fitness app, and Apple asked our marketing team to pitch ideas for an Earth Day promotion to promote us in the App Store. Our team had a pretty unique idea that we all were excited about, but after the pitch our Apple contact stopped answering messages about the promo. Sure enough, 3 months later Apple took the exact idea and used it in a promo for one of their own products on the App Store. At the time I thought this was some weird one-off thing that happened, but it turns out it wasn't.. it's exactly how they operate. Make no mistake, OpenAI is no saint, but Apple is a ruthless mega-corporation that's been stealing ideas for years. There's no one to root for in this one.

I think I'd like Instagram without auto-play, endless scrolling and algorithmic content, but serious question: What would TikTok even be if you took those 3 features away? It sounds like TikTok shouldn't even exist.

They're changing their tune because AI is not showing productivity gains. I've seen this first-hand. Every major tech company dove headfirst into AI, and the work being produced went up. But it's the wrong kind of work. It's not revenue producing work.

The reason is simple: The time it takes for humans to review *new* work is often similar to the time it takes to create new work in the first place. So every engineer eventually comes to the conclusion that they'll send agents to do code-cleanup type tasks in parallel with the actual work, because it's easier to review and validate that. The result is they're now splitting their time on the real work and the tasks they added to keep the agents busy to increase productivity. So on aggregate people are "more productive", but moving slower on real work, only now with high token costs.

The answer of course is "more AI", which means AI code reviews, but nobody is comfortable with the inmates running the prison yet, or the increasing token costs that are being paid to frenemies. So effectively there needs to be a generational leap in capability and an overall reduction in token cost before AI coding productivity gains are realized, but right now capability has plateaued and costs are going up.

One of the requirements for using Fable or Mythos is agreeing that Anthropic can look at your prompts and sessions. Presumably so they can data mine how people try to misuse it.

I recently saw an innocuous prompt trip the "safety" warning. It wasn't the prompt though, it was a sub-agent that tripped it. I asked Claude to try again several times and the same thing happened. I asked why it was triggering a safety issue and Claude said the safety checks happen on the backend and it had no idea. The most likely explanation is that a subagent went off the rails in some way and **was** doing something worthy of being called "unsafe". I've seen Claude reverse engineer binaries to better understand why a command failed. Who knows what a subagent might try.. if for example, it failed a file permission check.

It's been clear for at least a year there was not going to be a true successor to the Vision Pro. The M5 upgrade was a strategy to take all of the parts Apple had sitting in warehouses, throw in a new CPU (which they were already going to make), then sell it as an upgrade to previous buyers. They knew they weren't going to win over new customers (more CPU power was never the problem), they just needed a way to get rid of excess parts inventory. The people on the Vision Pro team moved on long ago.

Security isn't something you bolt on after the fact, but a new tool could significantly upend what it is you need to secure against. I suspect that's the real issue here.

My experience of Opus is that it's shockingly capable of tearing apart software binaries. I drop a path to a binary in Claude Code and ask it to tell me how a feature works, and it will usually give me a complete breakdown of classes and functions and how they work together. The binary loader information, symbol data, assembly, etc. are all just another language to Claude, so it really doesn't care. It's not hard to imagine a model trained on and geared toward binary data could seriously undermine how "secrets" are hidden in software.. "Hey Claude, could give me the API keys and explain how transactions are signed for this app?"

There is a huge AI bubble, it's just not going to pop with the top-tier model providers. They'll be fine. But the companies building data centers that are sitting on expensive, rapidly out-moded GPUs.. the thousands of companies that are building a single AI feature and hoping to be acquired.. the AI service companies trying to grab some market in some vertical.. There's a lot that's not going to work, and there's going to be a turning point where those guys start dying off and the money being thrown around will dry up.

As far as I've seen, the AI fanatic's answer is "don't care about the code".

I'm not an AI fanatic, I work for a major tech company and have been forced into being "AI Native" to keep my job. I *like* writing the code, and often disagree with how CC does it, but "don't care about the code" is pretty much right and not a fanatic's point of view.

All of these arguments already happened when "high-level" languages like C first appeared. "How can I trust the assembly produced by the compiler?" "I can do register optimization better!" "If I don't practice managing operands on the fpu stack I'll lose that skill!"

English is just the next high-level programming language. If you don't like the code being produced, write a skill or update your personal context.md to explain why. Ask CC to do code reviews that catch and fix the bad patterns.

Maybe not today, but ultimately saying "I need to always read every line" is going to turn out like trying to verify the assembly produced by compilers.