Great read. We went from “doing” to “managing” to “directing” insanely fast. The part I struggle with most with is letting go of understanding the details. Years of trauma learning that details bite. Finding tight abstractions you can trust, like in this case, is key.

on using llms - green: what you're expecting if you call it intelligence. result is disappointment and frustration - yellow: what you can get out of it with iteration and careful prompting. usage requires intense work. - red: what you're actually getting with general use. outside your domain it feels impressive because you can't tell the difference. useful to level up to the average, but its important to remember there's a whole slope beyond it.

An overdue cold-plunge

Chinese, or open weight? Americans have dropped the ball on the open weight model race. Platform-wise you also get a lot more bang for buck with going east.

We still don't grasp LLMs. That's evident in our choice of words to anthropomorphize what's really just a statistical model. We don't understand what the existence of a large statistical language model implies. It does not imply intelligence, but it’s something easily confound-able with it. The training corpus was human written, with thought and intelligence, so the output is a sort of mirror of humans, of how we write and reason to some extent. the mirror is confusing because people have never seen that kind of mirror before. Once you identify these models as mirrors or the text they are given, problems like "sycophancy" and "hallucinations", described in these anthropomorphized terms sound silly. Once you identify these models as mirrors or the text they are given, problems like "sycophancy" and "hallucinations", described in these anthropomorphized terms sound silly.

“never expected” 😆 🙉

In a previous era, I often made a big deal about deleting unnecessary bits in a project repository. “What’s the big deal? It’s not harmful” — but it is, like clutter on a work table. I’m glad there’s now a bigger imperative: it costs tokens and increases perplexity!

This won’t last. They’ll need adjustable, per division and per employee budgets

dealers know better than getting high on their own supply

This what collapse looks like. Societal culture collapse. Every thought, trend, what you assume is mutually assumed gets pushed and compressed toward the statistical model. A worldwide echo chamber worthy of a black mirror episode. Alpha comes from the ability to disconnect.

engineers are realizing that any serious work requires LLM use to work *slower* orgs are still scrambling to answer “what are we getting back from all this AI spend?” unproven but generally accepted answer is “faster” or “more” but the direction to look is “higher quality”

Inference will never be free or even negligibly cheap. Everyone will get a token budget. ➤ Productivity gains from tokens are hard to attribute.
➤ Consumption is easy to scale.
➤ Waste is nearly invisible. You don't know where to cut. That's all you need for a price floor. Tokens will forever be a scarce resource you need to allocate consciously, much like old problems: headcount for managers, capital for investments, ad budget for marketers. Tokens won’t be different. You need a performance-adjusted budget.

Hang on to your ears. When even the most hyped innovators start to U-turn, you know a correction is looming. Maybe a welcome breeze for laggard adopters

Jiujitsu is very inclusive. No fancy credentials or rich parents needed. Friends made from all sorts of backgrounds. 6yrs ago I first heard locker-room crypto trade talk. Plumbers, dentists w/ strong opinions. Was a clear sign. Today people were raving about Cerebras’s IPO.

👀 Vet any creator profile on X with an automated team of experts. Check engagement quality, patterns, reach (ability to break through the algorithm), promotion style, and check for red flags,

The ultimate botnet. Built in the open, installed voluntarily, often by paying customers.

It's tough out there for inference pipeline builders: * Groq - vulture-gutted, tiny limits & fewer models * Cerebras - cut down to 2 models post IPO * Together - raising prices * DeepInfra - still unreliable latency Hope SambaNova & Fireworks stay off this list.

The writing’s bright on the wall.

Didn't like the answer? Just keep prompting. Sad reality: sycophancy isn't a marketing play, it's the feature you want even if you say otherwise. We may get less polarized but increasingly delusional.