The VC world has made a misstep Introducing the world's largest founder campus

we just raised $4.3m to make ai talking toys in 83 days, we sold 1000 dinos, assembled a factory, & we got it all on film but there was a time we had no idea if we'd make it: (P.S. we're hiring)

DEEPSEEK R2 RUMORS: 20T Parameters MoEoE (Mixture of Experts of Experts architecture) with only 10,000 Active Parameters Trained on the entirety of 4Chan only, as to filter for the highest quality intelligence Not trained at all with Nvidia chips, just with the collective

Feature launch 🚨 ion can enable your organization's PMs and designers to own accessibility end-to-end, without additional engineering effort. they can: - scan live UI - surface & rank WCAG issues - 1-click fix + PR stay compliant, don't get sued, and maintain velocity

Excited to announce we raised a 1.8M seed round! we're here to turn every product thinker into a builder, elevating craft instead of automating it away. if you want to join a fast growing team of ex CTOs, design engineers, and cracked builders shoot me a DM :)

Getting large model fine tunings working nowadays is just a series of changing values on configs and vague cuda errors until the model finally fits in memory

Love when ML metastudy papers go over the details of different architectures and their pros and cons when most of the time, the dataset and compute matter way more

Who is doing RL for diffusion?

AI Video is half the struggle. You need the creativity to actually come up with engaging content

A horror movie where Waymo is actually the company that reaches AGI, and then the cars ruthlessly target their creators on the streets like a modern day Jaws

When will we have AI agents that create AI agents that create AI agents

im looking to join an early stage project that is ambitious, tasteful, and allows me to do challenging technical work but also express myself artistically! heres a bit abt me. i helped train image and video models Lexica before that i co-founded a company. it was a site

I heard some rumors about GPT-5: - it’s going to be running on entirely biological mediums. OpenAI is moving off GPUs and onto cell cultures, which are more compute efficient - the training data is entirely mandarin, as that has led to higher benchmarks. The outputs will be

Day in the life of a Y Combinator founding engineer.

Yesterday we completed training of our first custom model that will soon be hosted on inference net! We have another model training in progress, set to finish at the end of this week. These models set the stage for a new chapter of Inference, where we scale from billions

timer is on, stay posted

Wan 2.2 is good, but probably ranking below Minimax 2, Seedance, and Kling. Open source though, so no telling how it’s gonna improve

Had me wondering how these providers make money when they offer open source models for so cheap (WAN 2.2 on Fal and Replicate for example) — so I did some digging