@Aaron Parker asked how my animations come out the way they do. I said I brain dumped and went YOLO. That was half true, so here's the proper answer. First, Jake's kit. I run it as it came and haven't touched the engine. The thing it does that matters most is timing: it turns the voice into word timestamps, so animations land on the exact word being said. That's why it feels designed and not just pasted on. Otherwise, I just started a session in my project folder that contained it and told it to use that workflow. I did not start a session at the proper folder ICM level. Second, the drawings aren't the kit's stock characters. Claude wrote custom drawing code on top of it. My character, a race car driver mechanic type, is drawn once in a shared file and reused in every scene, so he never drifts. He came from me rejecting two attempts and saying what I wanted. Though it felt very brief and quick based on some general desires. Third, and I think this is the real answer, the way it got built. Every video is split into short scenes. One AI builds a scene, a different AI looks at still frames of it and sends it back. Most first builds got sent back. In one pass, all eight scenes did. When a reviewer noticed the crew drawn three different ways, we locked the drawings into one library and every scene had to use it. I also wrote the look down first: my colors, one highlight at a time, big text, calm motion. This was the automated part of my mess. My part was almost all "what", not "how" and honing it in. Stuff like "the hand looks off around 1:55", "too much green", "the ending's abrupt", "wrong theme, give me options". About 16 rounds of that across five videos. That's the YOLO part I undersold: I didn't write any technique, but I did a lot of looking and saying no. This is also the part that I will be able to create my own process from, and it only took a couple tries. The ROI on YOLO mode is well worth it. Two things that saved me the most time. Picking from options instead of describing a look (four themes, a few character styles), and getting the idea out of my head first. My first versions were wrong because Claude guessed what I meant - here is where I learned just from one video how to better review a transcript before we generate the audio, and to generate the audio before we generate the preview slide deck of the animations, and then tie it all together, and then preview the whole thing. And that led to less steps because I noticed getting to the end result and modifying different elements of it was. More tedious and time-consuming and less productive.

Hi all! Video two is done. The first one was the human side of how I work. This one is the Claude side: how I run it so the right model does the right job, without me having to tell it. It's 3 minutes 45 seconds, and I explain it as a mechanic's shop: - Haiku is the parts runner - Sonnet is the mechanic - Opus is the foreman - The Fable models are the specialist I call in on purpose You don't pay the foreman to fetch a filter, and you don't ask the runner to decide if the engine is worth saving. Then it covers how my rules sit in layers (one for everywhere, one for each project), why that keeps the cost of a request down (the cache), how helpers and workflows use it, and my own numbers from the last month. About 96% of my input tokens were cache reads, which at public list prices works out to roughly 90% cheaper than paying full price for every token. That's list-price math, not my bill, and the video says so. I'm not saying it's perfect or right. It's just what we built. I'm still learning and I still get things wrong, so if you spot something I got wrong, please tell me. Video: https://youtu.be/6RonYbHcbbw First video, if you missed it: https://www.skool.com/cliefnotes/the-most-productive-procrastination-period-ive-ever-had Heads up: the narration is an AI clone of my own voice (ElevenLabs), every drawing is an illustration, and I wrote the words and directed Claude to build it. Which model do you reach for first, and what makes you switch up or down?

my 9B model (Ornith 1.5, 8-bit, one Tenstorrent Blackhole card) uses speculative decoding: a small head built into the model (MTP) guesses the next few tokens, and the big model checks them all in one pass. every token is still the big model's own choice, so the output never changes. the speed depends only on how often the guesses are right. the official head's first guess was right 84% of the time. but the 2nd and 3rd guesses dropped to 50-69%. - the reason: the head was trained to guess from the big model's own hidden state. when drafting, it chains on its own guesses instead, and it never practiced that. - the fix (the FastMTP idea): let the big model write ~100k tokens of its own answers, greedy and sampled. save its hidden state at each position. then train the head to chain on its own outputs, the way it's actually used. - cheap: a LoRA on my M4 Mac, ~9 minutes per run. the data was ~80 minutes of card time. the result, through the real server, single user: - greedy: 40.0 → 41.7 tok/s (+4.2%) - sampled: 31.3 → 32.9 tok/s (+4.9%) - short answers: 45.4 → 48.8 tok/s - all 50 test outputs byte-identical to before. the head only drafts; it can't change what the model says. my weak spot: offline it predicted +7.6%, the card gives about half that. the card runs an 8-bit copy of the head, and the offline test set was small. the gain was still climbing with more data, so it isn't done. small model, small trick, free speed. anyone else tuning their draft heads?

I'm currently using AI for a pet project writing a private weekly newsletter for my fantasy football league, and right now I'm at the [copy-paste] stage. The thing that really caught my attention from this lesson is the entire concept of separate files for tasks, because that level of breaking things down step by step is something that I've had to do most of my adult life. I often find myself in roles that require creating SOP's or training materials and I feel like talking to an LLM is like talking to the 'new guy' or a child that read the internet! I'm excited to learn how to automate this process and see how structured approaches to AI can help me save time in the process. It currently takes me anywhere from 3 - 5 hrs to put it all together because even with my current prompts/skills, I need to heavily edit/audit the output. Figured this would be a great place to start as it's a simple test based project that can be broken into steps, and has a measurable output. I promise my next project will be far more exciting whatever it is! lol.

Hi all, got a question about security. In particular, existing content and stopping it from being stolen. Some of my clients live in the "paid course/paid monthly report" teaching/info space. Lately, nearly all the measures we've had in place for stopping duplication of product and reselling at ridiculously low prices has become ineffective. Cease and desist is ignored (not that it was ever really listened too) and of course its expensive and time consuming. Water-marking has become a joke with AI being able to remove them. Placing product behind any type of existing protection also seems to be easily subverted. We've even found out about a $10,000 product being bought collectively by 10 people, copied by hand/AI and re-packaged across the globe at a fraction of the cost (making them a fortune by the way). It's a racket I tells ya. So I thought I'd come to you, the brains trust, to see if anyone has some up-to-date sneaky-ass (maybe even AI driven) methods to stop these thieving $%&#*%s. Thank you in advance.