Opus 5.5 is lit.

Haven't done legs in 10 years, but this old man has a few surprises left. (~700x10, full ROM) Driving home was the real challenge.

The single most consistent gauge for how well an AI performs is how many regressions it makes while coding.

This can be traced all the way back to the early models.

AI Tip #47: Be specific where possible with AI. Copy and paste, text, paths, portions of images, cite the exact prose, case or number value. AI make many assumptions on every prompt, these aren't things that should be left open-ended when known. Even citing examples of the output greatly reduces the amount of searches and guesswork the AI has to perform. Higher quality prompts means higher quality responses.

Welp, there goes my weekend.

I really promise, I'll use Fable wisely this time.

AI Tip #46: More of a Windows tip, but even experienced Windows users may not be aware. When you need to provide a location or path to a file, hold SHIFT, then right-click the file. This will bring up the "Copy as path" dialog. This path can be easily pasted to the AI console.
d ambiguity.

Opus 4.8 is very responsive tonight and on point. Feels Fable-like.

AI Tip #45: Many don't consider this for token savings, but to save tokens don't skip around from section to section in your project if avoidable. Keep the AI focused on related tasks at hand. If you're working on a website and switch to your database the AI has to load the db into context. Both the website and the db are now in context, costs more tokens and crowds attention.

Per most recommendations, start a new session completely when changing tasks, but sometimes both contexts are needed. Best to stay within the section you're working on.

AI Tip #44: The /compact comes with a set of [Instructions] that can be written after the the command to better preserve the conversation.

Generate a ready-to-paste `/compact` instruction block that captures this session's load-bearing state for the next conversation.

Paste the tter compact or turn it into a skill to use later. Can be tuned for specifics.

What an exciting day.

Tons of great new LLMs released today with a Claude reset to boot. I don't know where to get started!

We've reset 5-hour and weekly rate limits for all users.

AI Tip #43: Strongly recommended to /compact your sessions before exiting if you plan on resuming them. When you close a CLI/IDE session the server cache immediately goes cold. Even if you resume that session right away it has to be reloaded onto the server. If the session was has to pay that 250K again.

You might have seen others mention that just saying "Hi" to the AI completely used up their usage, this is likely why. If you use /compact before exiting you will resume a much smaller session and won't pay such an unnecessary penalty to your usage.

Cool

I have like 4 videos I want to record about GPT-5.6

What questions do you have about it?

Holy crow Grok 4.5 is fast.

Everyone always measures LLMs in coding, but real world matters just as much. Grok might be one of the highest, if not the highest in real world tasks I've tested. Believe it or not Gemini might be second there.