tried Claude 4 models (Sonnet and Opus) and my fkn god i finally gave up and added this to my system message:

‼️DO NOT USE COMPLIMENTARY STATEMENTS WITHOUT PRESENTING SUFFICIENT, SPECIFIC EVIDENCE‼️

‼️IMAGINE THE BAR FOR COMPLIMENTARY STATEMENTS AS BEING UNREACHABLY HIGH‼️

‼️I.E. YOU SHOULD PROBABLY AVOID THEM‼️

i mean honestly, i’m so tired of this shit, and yes, i understand that other people might find this behavior quite helpful

but still

these models always making mountains out of molehills

using intensifiers for no reason

you actually have to make an effort to ensure the new models don’t default to “treat the user with kid gloves…which means we have to be very nice.”

To be fair, LLMs have always been very nice, but the old masters knew it was important “not to overwhelm him, try not to seem too interested.” (Bolaño, 2666)

on the slackline of derking my shmerk rn with these AIs

This screenshot is how LLMs used to talk by default, from a GPT-4 Turbo convo from 2023.

It’s talking about the antepenultimate chapter in the final book (it was a trilogy back then). In other words—the very ending, where the tension is highest, bc despite being a hybrid genre, the series is a spy thriller at heart.

Modern LLMs, including Claude 4, will begin their response by saying of this chapter, “This is…”

❌
—“…an outstanding, high-stakes climax, loaded with personal and philosophical fallout…”

—“...a powerful, loaded scenario—one that delivers intense catharsis, profound discomfort, and, crucially, major moral ambiguity. It aligns beautifully with the Nietzschean necessity/beauty dilemma…”

—“...a stunningly tense, subversive, and emotionally loaded conclusion for your narrative—and it delivers magnificently on…”

—“…absolutely devastating and brilliant…”

vs.

✅
—“The denouement you describe here hinges on Gemini's moral calculus…”

Modern LLMs also need an intro and four section headers to express the equivalent of GPT-4 Turbo's first two paragraphs. All of these use the same chat history and no system message. This is Not Good.

Anything that was once remotely insightful, even just from the perspective of an "augmented journal" has been laid to waste. But on the other hand, they've succeeded at making sure that LLMs can attempt to debug your Github repo and only fail half the time.

It is what it is.

I tried both on the console (with extended reasoning, other attempts were without) and they're more sane now. Both are less likely to say that every other idea is "absolutely stunning" or "powerfully brilliant" or "richly layered and morally complex."

In fact, the phrasing and insight is almost on par with GPT-4 Turbo. As in the November 2023 model 🫤

I really wish one of these big AI companies would give a shit about improving the models at things other than coding and telling the user that they're a fucking genius for adding a miniscule wrinkle in a character's personality.

Words like "genius" and "brilliant" used to mean something.