jbusted

Creator of surreal and dreamlike artwork with using neural networks, using a modified version of #BigSleep and #VQGAN clip by @advadnoun and @RiversHaveWings.

[jbusted @jbusted1]

Hello! In this account I will be posting Artwork generated using a modified version of BigSleep, a method innovated by @advadnoun.

Access to the full sized works, as well as high quality upscale will be available soon!

[jbusted @jbusted1]

@BlancheMinerva Subataging ML development with this model is disgusting specifically

[jbusted @jbusted1]

@MikeE_3_14 @MathMath901 2^(x) - 2^(x-1) = 2×2^(x-1) - 2^(x-1) = 2^(x-1)

[jbusted @jbusted1]

@taha_moji @itsreallyvivek I disagree, how many points here are from Vivek's actual experience and how many are filler from the llm being "insightful"?

there is a lot of noise here. You can't tell what to keep and what to ignore.

[jbusted @jbusted1]

@Creative_Math_ The I'm pretty sure cosine similarity goes to -1 so the two directions are orthogonal. I guess something is strange with the geometry? Perhaps the momentum one is picking up on a subtle direction in a weirdly curved loss surface?

[jbusted @jbusted1]

@liadagmon כן אני כבר מדלג על תוכן אוטומטית כשאני קורא , it's not xits y

[jbusted @jbusted1]

@VictorTaelin You should also write your perspective on the entire thing when you are not feeling frustrated. Some people are getting mixed messages:(

[jbusted @jbusted1]

@EricLevitz I have not noticed it until the last few months, it's possible it's due training on synthetic data of other models. Hopefully not a start of a representation collapse

[jbusted @jbusted1]

@VictorTaelin reinforcement learning, they probably spend an enormous amount of resources on playgrounds for llms and alot of synthetic data.

[jbusted @jbusted1]

@FixingEducation This seems ai generated, the tone of the voices feel inconsistent and sort of disconnected from what being said. Also feels a little AI scripted.

[jbusted @jbusted1]

@tammarzohar אוקיי אבל מה הייתה השאלה?

[jbusted @jbusted1]

@VictorTaelin comprehension debt?

[jbusted @jbusted1]

@VictorTaelin Why no qwen 3.6? I keep hearing it's better?

[jbusted @jbusted1]

@SoftEngineer oh! that's pretty good!

[jbusted @jbusted1]

@theo It searches google -> get sponsored link for "the cheapest flight" ->r ecommends it. Everything is working as expected, nothing here is very suprising

[jbusted @jbusted1]

@Birchlabs yeah seems so

[jbusted @jbusted1]

@VictorTaelin having the ai do the proving sounds good in concept, but would that require many token usage? If it does, people might prefer cheaper options.

[jbusted @jbusted1]

@VictorTaelin Models also take a really long time to train, retraining is not a very good answer

[jbusted @jbusted1]

@VictorTaelin But it's too big to run even on a DGX Spark :(