@bnjmn_marie and it's apache 2.0 with a 1M context window. for anyone putting a model inside a product, the licence matters more than the next point on a benchmark
@ArtificialAnlys 3.1B unique weights doing 5.4B active of work through three loop passes is the interesting part. curious whether tokens per second on a macbook tracks the 3.1B or the 5.4B
@TheAhmadOsman the number i'd watch is 5.4B active out of 26B. it should run closer to a small model's speed than the 26B suggests, which is the whole laptop pitch
@teortaxesTex 594B to 26B in one release, and unsloth had 16GB-RAM GGUFs up the same evening. the gap between "open weights" and "on my laptop" keeps shrinking
noticed something rewatching explainers that work: they say the answer in the first line, then spend the rest proving it.
most scripts save the answer for the end and lose people by second 5.
do you lead with the answer or build up to it?
@kimmonismus "first step" is doing a lot of work there. the real question is whether harness use draws from the same limits as the app or from a smaller pool
@unusual_whales the flat part is the interesting bit. 36% in year one, 48% by year two, then stuck near 50% for five quarters. looks less like a takeover and more like a ceiling
built kinereel around one box this week. type a topic and it makes an explainer you can edit, or drop in a podcast episode and it cuts clips of the best moments. here's the 30s film https://t.co/0dzzC0noGm
every video starts in one box.
type a topic, get an explainer you can edit in a studio. drop in a podcast episode, get clips of its best moments, with who's speaking and your caption style.
explainer or clips, which would you try first?
(clips: Linux Foundation, CC BY 4.0)
every video starts in one box.
type a topic, get an explainer you can edit in a studio. drop in a podcast episode, get clips of its best moments, with who's speaking and your caption style.
explainer or clips, which would you try first?
(clips: Linux Foundation, CC BY 4.0)
@thegreatola feels less gatekept than unglamorous: one narrow topic, a fixed schedule, and making video 40 better than video 1. which part felt hidden to you?
@Da7_Tech comparing against what people chose is the right anchor. do you also count how many turns it took? two models can land on the same answer with very different amounts of hand holding
@MatthewBerman for browser use i'd love a recovery score next to it: how often it notices a click landed on the wrong element and fixes it. single pass success hides that
@pdicarlotrader the tell for a trap is how fast it gets reclaimed. spy topped at 739.90 this morning and closed 715.87, a few cents above the 715.65 low of the day
noticed something making explainers this week: the good ones answer one question.
the ones that try to answer three end up answering none, and you can feel it by scene 4.
so now the first line i write isn't the hook, it's the question.
how do you pick yours?