[handoflixue]

"We wrap LLMs in Agent harnesses because AI isn’t enough."

This feels equivalent to saying you're not a real coder if you need a compiler to hold your hand. What, a human coder isn't enough?

Of course we're going to improve tools by building more tools on top of them. That's how we went from punch cards to assembly to high-level languages!

[oblio]

Anthropic isn't selling that, that's the point. Anthropic is selling "coding is solved", worth $1tn.

If they keep hiring developers and LLMs need huge procedural harnesses, then the hypocrisy is open for all to see. And worse for them, it probably means their business model is worth maybe 5-10x what Jetbrains is worth, not $1tn.

[BetterThanSober]

I agree that "coding" isn't solved but LLMs needing a harness is not necessarily a huge hit on AI

As for valuation, it's all a grift, always has been for anything hip in tech. SpaceX should've been in the $500b range and they have an actual rocket and working satellite constellations

[oblio]

LLMs needing harnesses is a hit on AI because it means LLMs can't be extended a lot (or indefinitely as AI boosters hope). We're back to (enhanced) expert systems from the 80s and nobody is excited about those.

And regarding SpaceX, SpaceX even without xAI didn't make a ton of money and it's business initiatives are super long horizon things. Moon bases won't be profitable for at least 10 years and Mars bases probably for 20+.

Also, even though they are behind, competition is moving. China caught their first rocket yesterday (and we know they can move fast), BlueOrigin is behind but they have the money to keep going. The EU and India will keep investing and developing as rockets are a core concern for both, at some point they will have reusable rockets.

These huge valuations for SpaceX are predicated on winner takes all mechanics, which are mostly software based. Many markets don't work like that since the cost to get another unit of revenue increases linearly with that unit. Not like software where extra copies and customers are basically free.

[handoflixue]

Coders needing compilers is a hit on coding because it means coding can't be extended a lot?

I'm just not following the logic at all.

We build vastly more ambitious software today than we did in the 80s

[oblio]

That's an intentionally obtuse comment.

Compilers are widely available as open source software not coming from companies that need to justify a $1tn valuation ASAP for their development of LLMs.

Also the LLM = compiler comparison is generally just ridiculous. Their complexity, both internal and in usage for coding are multiple orders of magnitude apart.

[handoflixue]

Oh yeah, nothing complex about a compiler, that only takes 4 years of college to understand. That's soooo much simpler than typing "make me a Tetris clone".

Literally no one has 4 years of experience with LLMs, because they haven't existed for four years. Today's experts are tomorrow's college graduates.

[ozozozd]

Umm, yeah.

Obviously the neckbeard who wrote C in vim and compiled when he is done is a better coder than the one who kept compiling and fixed issues when told by the compiler.

[handoflixue]

The neckbeards that wrote C did so precisely because they saw how C was a better way to get things done, and they were probably the first people to start using compilers, because, again, they're the ones that built them.

The neckbeards that wrote C tend to view LLMs as the obvious next step, and a lot of them actively use them for hobby projects even if they're not quite "maintain the Linux Kernel" tier yet.