Hacker News Reader

Comment by pclowes | Hacker News Reader

Having listened to him give a couple of talks I am always struck by how much he talks and writes like Claude code.

I don’t think he farms out all of his writing and talking to LLMs. I don’t think Claude code was trained to emulate him or anything.

I think his “voice” has been filed LLM smooth by years of agent based interactions. He claims Anthropic engineers use an average of 500+ agents a day. They are human interfaces to token generators more than human to human communication. They are picking up the tendencies of their most frequent communication partner.

Unfortunately, I see Claude being wrong often enough that I view it as a faulty narrator. Often helpful, sometimes totally full of it.

And now I subconsciously apply this filter to anything that sounds like Claude.

Makes me nervous that my voice may be becoming that of a faulty narrators.

View on HN

>"They are human interfaces to token generators more than human to human communication. They are picking up the tendencies of their most frequent communication partner."

I see this happening at work, too. All communication and collaboration between engineers has broken down. Juniors don't have questions anymore. Seniors don't discuss architecture anymore. Everyone is just off in their own world of feature work with the only interaction coming at merge time. Really depressing what has happened to our trade.

To err is human.

Higher level thinking has fewer constraints on being wrong than lower level subconscious thinking. Low level stuff tends to have some repetitive and evolutionary basis. Higher level thought tends to be more one off, less evidence based, and more exploratory until we get to the point of being an expert around particular concepts.

A huge amount of human thinking is being a parrot and repeating statements without a deeper understanding. For example when I say 1+1=2 I'm not thinking about it. If I say 638483+949948= calculation is necessary.

So in this, your voice has always been one of a somewhat faulty narrator, LLMs just may be further degrading the quality of the narration.

Left, right, higher, lower, male, female. People love to explain the brain by dividing it in half, or into poles.

Nothing more human than to label, simplify and categorize, even in the absence of higher quality information.

Lets say you have a colleague named Bob. Bob is a pretty good engineer, but he is as confident when he doesn’t know as when he knows.

You can never tell his degree of confidence in his answer. For low-level things he is pretty good but for critical high-level things it is more ambiguous.

You work around this by checking Bob’s work. Having formal verification steps to check what he says.

Now let’s say you meet Bob’s twin brother Bill. Who talks and sounds just like Bob. He makes fairly grand high-level statements that you can neither test nor verify.

How much do you intuitively distrust him?

Bob also knows how to quickly check things with a little bash scripting and tests. Can quickly switch branches to confirm regressions and consult documentation if something requires more information. His twin can also quickly scan for all the common security vulnerabilities in any code, and will do it without complaining as many times as you want him to. I’ve come to trust Bob with code more than any human.

> I think his “voice” has been filed LLM smooth by years of agent based interactions.

I've wondered this for a long time now: how does one differentiate between your voice getting "filed smooth to match LLMs" or your brain getting "filed smooth to match LLMs"?

After all, if we accept (as the token providers want us to) that evidence of language use as we see in LLMs is evidence of intelligence, then surely the reverse is true - evidence of machine-only language use is evidence of machine-only intelligence.

IOW, it's the LLM talking through you, not you talking through it.

> I see Claude being wrong often enough that I view it as a faulty narrator.

I'm reminded of <insert-someone-in-your-life-that-generates-friction-here>

"Often wrong, never in doubt"

The problem here is that the latest Claude models have gone off the deep end in their communication, and there seems to be neither awareness nor willingness to fix this. "Having your writing and talking turn into Claude is the risk nothing names" or what vapid strange meaningless thing would Opus say.

> He claims Anthropic engineers use an average of 500+ agents a day.

What are they doing all day with that, I wonder?

Does that mean 500 separate agents or 500 agent sessions I wonder.

the former

What does that mean though? Does it mean there are more than 500 different "agent" systems within Anthropic and employees interact with most of those in a daily basis?

That isn't humanly possible, you will blow out your neurotransmitters and not be able to think. Only being semi serious, but we can't make that many decisions a day.

Right, but you can imagine eg a hierarchy, in which case you might only talk with -say- 10 reports directly.

That isn't typically how abstraction is tallied. The ceo doesn't say they use 15k employees, or the programmer uses 30B transistors. Or the director uses 120 workers. You enumerate what you actually interact with.

Well, trying to make RSI for one so humans don't have to do that work.

LLMs are still a very new technology in relation to human timeliness. There is a whole lot of exploration to be done.

I mean, in relation to human timelines, so are computers in general. It's not close why you think that this is unique to LLMs.

There might be some confusion on your part on just how much AI compute has been installed in the past few years in relation to the total amount of CPU compute and the speed at which it was installed.

If the CPUs life was an hour, the GPU AI accelerator would have been around 5 minutes and is around 5x larger in total compute.

It should not be underestimated in any way the scale of what is occurring.

No, I think the confusion on your part is around what it sounds like when you say "in relation to human timelines". Humans have been here for around a million years. We've been living in cities for several thousand. It sounds like what you're actually comparing to is AI computing to pre-AI computing, which is reasonable, but I don't really see how that is accurately described by "in relation to human timelines".

Reply on news.ycombinator.com