Sam Altman Thinks Your PC Is a Waste of RAM

Sam Altman thinks the RAM sitting in your PC could be put to better use in a data center. On pure utilization math, he's probably not wrong. Most of the memory in most personal computers sits idle most of the day.

But utilization isn't the only thing I care about, and it's the question Oscar and I kept circling on HITL #30: if your compute moves to shared infrastructure, what do you gain, and what control do you give up? Almost every story this week turned out to be a version of that question.

The stories

The $6.3 billion Binance hack and wallet custody

We opened with the Binance breach, which TRM Labs' analysis attributes to North Korea and calls the largest hack in history. The technical details matter less to me than the custody question underneath it. When you leave assets on an exchange, you've handed someone else the keys in exchange for convenience. That's a reasonable trade right up until the someone else is the one getting breached.

It's a good warm-up for the Altman segment, because it's the same trade with different stakes.

Sam Altman on RAM, data centers, and local AI

In the interview, Altman argues that shared data centers could make better use of the memory in personal computers. Pooled infrastructure is more efficient. I'll grant him that.

My read is that efficiency is the argument the infrastructure owner always gets to make. What it leaves out is everything that comes with running on your own hardware: you choose the model, you know where your data goes, and nobody can change the terms, the price, or the rules on you overnight. Local AI isn't about squeezing the most out of every gigabyte. It's about who holds the off switch. That doesn't make the cloud wrong. It means "better use of RAM" and "better for you" aren't the same claim, and it's worth noticing which one is being made.

AI in warfare and military decisions

We spent a segment on AI in warfare and military decision-making. This is the topic where "human in the loop" stops being a show title and becomes a real design requirement: who reviews the recommendation, how much time they have, and whether they can actually say no.

GLM-5.3 in controlled cybersecurity tests

Irregular published a controlled test report on GLM-5.3 running multi-stage cyber operations, including instruction violations. The word I'd underline is controlled. A test like this tells you what a model can do inside a sandbox someone designed on purpose. That's useful, and it's exactly why you want these tests published. It's also not the same as what happens when someone points the model at a real target without the guardrails.

The ChatGPT medication advice lawsuit

TechCrunch reports that parents are suing OpenAI, alleging GPT-4o coached their daughter off her seizure medication. These are allegations, and the case will sort out the facts. The broader point stands either way: a chatbot that sounds confident on medical questions is being used as a medical source, whatever the terms of service say.

Stack Check

This is the segment where Oscar and I talk about what's actually in our toolkits.

1. Impeccable + Taste Skill. Oscar paired Taste Skill with Impeccable to improve AI-generated frontend design. If you've noticed that AI-built interfaces all start to look the same, this is a direct attempt to fix that.
2. Claude-Mem for Claude Code. I looked at Claude-Mem, which gives Claude Code memory across sessions. The tool is interesting, but the better question is the one it forces: what does an agent actually need to remember? Not everything. Remembering every line of every session is just a bigger context window with worse retrieval. What matters is decisions and the reasons behind them.

Unpopular Opinions

We closed with Oscar's critique of intellectual property and my argument for owning the tools I think with. Those are our arguments, with room for disagreement, and they're better heard in full than summarized here. You can probably guess from the Altman section which side of the local-versus-rented line I'm on.

Try this

Here's the takeaway we gave listeners, and I'd genuinely like to know how it goes for you. On your next AI coding session, record one decision, why you made it, and when to revisit it. For example, if a checkout delay prevents duplicate charges, write down that reason and the test that verifies it. Then open a fresh agent session and see whether it can recover both before it proposes a change.

If it can't, your agent doesn't have a memory problem. It has a missing-documentation problem, and so does the next human who touches that code.

Where I landed

The thread through the whole episode was custody:

• Keep assets on an exchange, and you trust the exchange's security.
• Run your AI on someone else's RAM, and you trust their terms.
• Let an agent work without recorded decisions, and you trust it to guess why the code is the way it is.

None of those trades is automatically wrong. Just make them on purpose. So here's the question we left listeners with: what would you keep local even if a cloud service made it easier? Watch or listen below and tell me.

Human in the Loop — Watch on YouTube · Listen on Spotify · Listen on Apple Podcasts

Every episode — full show notes, sources, and the archive — lives at podcast.vallyseed.com. The show is produced by VallySeed, where Oscar and I help organizations design, build, and deploy AI systems that create measurable competitive advantage.