AI digest: Small models, big breakthroughs
Compact vision encoders rival giants, AI writes GPU kernels, and OpenAI sketches out the post-work economy.
This week’s standout theme is efficiency. Everyone’s racing to pack more capability into smaller packages.
Meta’s EUPE proves tiny vision encoders can punch above their weight
Meta released EUPE, a family of vision encoders under 100M parameters that matches specialist models across image understanding and VLM tasks. The clever bit is how they’ve cracked the efficiency problem without sacrificing capabilities. This matters because it makes powerful vision AI practical for edge devices and mobile apps where every parameter counts. MarkTechPost has the details.
AutoKernel lets AI write its own GPU optimisations
RightNow AI dropped AutoKernel, an open-source framework where LLM agents autonomously optimise GPU kernels for PyTorch models. Writing fast GPU code is notoriously difficult, so automating it could be huge for democratising performance optimisation. The autonomous agent approach is particularly interesting because it can iterate and improve without human expertise in CUDA programming. More from MarkTechPost.
OpenAI sketches out the post-work future
OpenAI published a policy paper outlining their vision for a superintelligence economy: public wealth funds, robot taxes, and four-day work weeks. It’s a mix of redistribution and capitalism that feels more realistic than the usual tech utopia nonsense. Whether policymakers will listen is another question, but at least someone’s thinking beyond “AI will create new jobs” platitudes. TechCrunch covered the proposals.
Google quietly ships offline dictation app
Google released an offline-first dictation app using Gemma models, taking aim at apps like Wispr Flow. The offline-first approach is smart given privacy concerns around voice data. It’s another example of edge AI getting genuinely useful, though launching quietly on iOS suggests Google isn’t entirely confident about market fit yet. Details on TechCrunch.