Expert profile
Founder & AI Product Leader
Yash Thakker is a Generative AI expert with over 12 years of experience in product leadership and technical strategy. As the founder of explainx.ai, he has taught over 300,000 learners and built AI platforms serving millions of users globally. He specializes in Agentic AI, Multimodal RAG, and the intersection of LLMs with consumer hardware.
Researchers ran emergent misalignment and subliminal learning on Talkie, a 13B model trained only on pre-1931 text. What worked and what was fragile.
Tether QVAC released TranscribePsy, a 254M-parameter Apache 2.0 speech recognition model for 25 European languages. Numbers, how to run it, and caveats.
Ten Gemini 4 Argon use cases from Sundar Pichai's thread and Google's posts: CAD animation, Mars data decks, Blender pagodas, bug triage and more.
VRSD fixes on-policy self-distillation collapse with a teacher-only system prompt: Pass@1 up 3.9 points on average, 11 of 12 settings. Code released.
The RISE paper reaches up to 13% human-verified jailbreak success on image models; newer versions reportedly worse. Paper facts, caveats, what builders do.
WTF (Wasserstein-Tilted Flow Maps) fine-tunes image flow maps for reward without simulation. Claimed 50-100x vs paper 280x, tables, code, checkpoints.
YouTube Music now labels songs made or altered with generative AI. How the Fully, Partly and No Gen AI tags work, who decides, and how rivals compare.
406 scientists including Hinton, Bengio and Choi urge airgapping cyber-capable AI models and an international body to track containment failures. What it says.
Surveys from Sage, Workday and Foxit suggest checking and redoing AI output eats much of the time AI saves. The numbers, their limits, and how to cut rework.
AMD unveiled Helion at OCP 2026: an open rack with up to 128 EPYC Venice CPUs and 32,768 cores for agent sandboxes. Specs, timing, and the Verano 2027 teaser.
American Express wants a signed Proof of Intent Token before an AI agent can pay. How the Intent Intelligence Framework works, and what it leaves open.
Anthropic cut distiller data harvesting 90-99% with preserved thinking, a new classifier and bans. Here is what developers must change.
