okay so confirmed the new Ideogram model is a SenseNova style model (not using a tokenizer like Chameleon)
is it just me that sees like a super good GPT Image 2.5 image and is like hmmm i wonder what patch size this is
Just noticed that since i switched from claude code to codex ive been using codex with fast mode. it now makes sense how i got through my weekly limit in 1 day
Tesla is one of the most innovative companies on earth but sucks at showing this to people. They are currently launching the Cybercab without a launch event while Dyson has a huge launch for the same shit vacuum cleaner every year.
GLM5.3 Flash is actually insane running on my mac and using my own ComputeUse harness it configures me a mac. Just a few days ago i used this to automate a work compliance training that claude and codex both refused to do for me.
just read the blog post of @dwarkesh_sp on the open ai HF hack and i couldnt stop to imagine what could have happend if there were humanoid robots that the agent swarm could have hacked into and maybe triggered the fire alarm to distract researches from an ai overtake.
Wow
Met a German founder this week and asked him if all the stories one reads about the challenges of startups in Germany are exaggerated. "No, they're understated." Proceeded to describe spending a full day having a 90-page investment contract read to him (mandatory under German law; § 13 BeurkG) by a notary that then charged €30,000. That was for his first company. His second company, needless to say, was not incorporated in Germany.
LAION is so back.
[1/7] 🚨 We're releasing LAION-BVD: a 10-million-hour open video dataset for multimodal pre-training.
- 1.3B video URLs from CommonCrawl
- 80M downloaded videos
- 10M video hours
- 55M captioned clips
- 300M frame-caption pairs
🌐: projects.laion.ai/bvd/
🧵👇
Claude Code and me the second after wandb decreases the free storage to 5gb and deleting runs takes 10 hours
over the last 2 months i ablated every single tiny design choice on the model i am working on. today everything seems to finally have clicked into place
Thought my ImageNet model was brokens until i finally increased the CFG. CFG=[1.0, 1.5, 2.0, 2.5, 3.0, 4.0]
why arent LLM people using things like REPA? i think SRA could be usefull. aligning an early layer with a late layer
what do we think about these gpt image 2 artifacts. Do you guys think the decoder is trained as a GAN/LDM/ PixelDM




















