Research Explained

An AI captioner can only look at a few frames of a video — PEEK is a tiny, fast model that picks the most useful ones, learning what matters from a “cheating” expert it then throws away. Best when the budget is one or two frames, for ~5% added time instead of rivals' 65–212%.

AI now writes most of the new code, drowning human reviewers. RADAR is a layered funnel that auto-reviews and ships the low-risk changes — 535K+ reviewed, with a revert rate 1/13 and incident rate 1/50 that of normal changes.

One model that maps text, images, video, and audio — and any mix of them — into a single shared vector space, beating specialized models at their own games and proving native audio embedding beats transcribe-then-embed.

How to make a diffusion model animate portraits at 15–20 FPS live — by distilling 20 denoising steps into 4, and replacing batch generation with a sliding window that streams frames continuously.

How to make AI agents get better at their jobs the more tasks they see — by distilling reusable strategies from their own successes and failures, with no extra training.

GPT-style models store ~3.6 bits per parameter. This paper measures exactly how much models memorize, explains double descent, and predicts when privacy attacks fail.

How to fine-tune a 175-billion-parameter model by training only 0.01% of the weights — with no inference latency penalty.

How HeyGen generates talking avatar videos that preserve identity, talking style, and micro-expressions from a short reference video.

How to shrink a language model's memory by up to 50x in seconds, plus Ramp Labs' Latent Briefing for efficient multi-agent context sharing.

How to give AI agents persistent memory across conversations — 91% faster and 90% cheaper than processing full context.

The 2017 paper that threw out step-by-step recurrence and rebuilt sequence models out of pure attention — letting every word look at every other at once. It set translation records, trained far cheaper, and became the architecture behind nearly every modern AI model.