The Claude Code model highlights the power of simple search (grep) and selective reads (only reading in excerpts). The only time I vectorize is when I explicitly want to similarity-based searching, but that's actually pretty rare.

> Vocabulary is just the surface.

Yes, but with LLMs, sometimes simply mentioning the right words is enough to prime the model in the direction you want to take it. If you start a prompt talking about leading and type pairings, it will take greater care with typography. You don't need to be an expert typographer to take advantage of this phenomenon.

How will an LLM "take greater care with typography" if it can't see the page it is creating? How will it "improve" leading if you need a human to see that there's too much distance between lines or too little?

With playwright, it can see what it is creating. Unsurprisingly, it works much better if you hook it up to a browser.

There's also the Claude extension for Chrome which integrates with Claude Code.

Because humans have already annotated diagrams and examples of what ‘too much’ and ‘too little’ look like, and these have been incorporated into the model. It tries to reproduce the content that is associated with humans indicating that they are taking greater care, and that content has the ‘not too much / not too little’ judgement already baked into it.

I'm assuming this is in the context of an agent that can see what it's doing. And I wouldn't assume humans have a monopoly on judging leading.

Honestly it sounds like they went above and beyond. Does this solve the trifecta, or is the network still exposed via connectors?

Looks like the Ubuntu VM sandbox locks down access to an allow-list of domains by default - it can pip install packages but it couldn't access a URL on my blog.

That's a good starting point for lethal trifecta protection but it's pretty hard to have an allowlist that doesn't have any surprise exfiltration vectors - I learned today that an unauthenticated GET to docs.google.com can leak data to a Google Form! https://simonwillison.net/2026/Jan/12/superhuman-ai-exfiltra...

But they're clearly thinking hard about this, which is great.

> Does this solve the trifecta, or is the network still exposed via connectors?

Having sandboxes and VMs still doesn't mean the agent can still escape out of all levels and still exfiltrate data.

It just means the attackers need more vulnerabilities and exploits to chain together for a VM + sandbox and permissions bypass.

So nothing that a typical Pwn2Own competition can't break.

Honestly lack of true multithreading (without the Web Worker hack) is the biggest downside for me. Every major project I work on needs the concept of a main thread for UI and a separate thread for processing.

Looks interesting, but operator overloading is an anti-pattern. Leading with that is like leading with "full support for null pointers."

Now just get Claude Code to build a hUGETracker exporter, and you could actually bring one of these songs into GB Studio!

Just a hot take, but if you ask someone to complete a rote task that AI can do, you should not be surprised when they use AI to do it.

The author does not mention whether the generated project plan actually looked good or plausible. If it is, where is the harm? Just that the manager had their feelings hurt?

Interesting how similar this is to the tradeoff of using AI coding agents

Only institutional shareholders matter, and they're on the same page as the CEOs.

Having seen a lot of AI-generated ads, I'm so skeptical that AI is actually improving marketing metrics. Every time I see one of these abominations on YouTube I think "is this working for you?"

With that said, I'm long-term bullish on AI. A lot of companies will over invest, just as they did during the dot-com bubble. But some of those investments will actually pay off, because this technology is not going anywhere.

Regarding your first paragraph, I've even talked with people who go out of their way to actively _avoid_ said product after encountering AI-generated advertising. So that'll probably continue to have an effect for as long as average people with good eyes can still distinguish "AI"/generative media from "real"/traditional footage.

I have observed this as well, and we've already seen some pushback when major brands use AI in their creative. I wonder if we're entering an era where AI will actually taint a brand.

> Having seen a lot of AI-generated ads, I'm so skeptical that AI is actually improving marketing metrics.

I'm sure that firing tonnes of marketing people for prompt engineers will be a good return on investment... Perhaps it reveals something deeper - which is that a 30 second Youtube ad doesn't generate that much revenue at all.

> With that said, I'm long-term bullish on AI

In the long run, yes. But a lot of people will end up losing their shirts over this.

> But a lot of people will end up losing their shirts over this.

Oh absolutely. In particular, it's hard for me to see a path forward for OpenAI. If they implode, it will be a tsunami.

> Having seen a lot of AI-generated ads, I'm so skeptical that AI is actually improving marketing metrics. Every time I see one of these abominations on YouTube I think "is this working for you?"

they likely track conversions, so someone is clicking and buying.

Measuring the effectiveness of ad campaigns, particularly in the short term, is notoriously difficult. They likely mostly don't _know_ if it's working at this point (though, yeah, I'd kind of assume it isn't.)

Really? It shouldn't be difficult to measure the performance of a YouTube campaign. There are very clear metrics related to watch time, CTR and conversion rate (if applicable).