I've always sensed that Duolingo doesn't effectively teach anything and it's main value is providing the "feel good" of keeping a streak or "learning" while failing to retain any knowledge.
It's a very smart business model for an app masquerading to help "language learners."
The loudest cheers didn't come from social media, they came from school children watching India's future lift off.
Nothing is more inspiring than seeing school children cheer for a rocket launch. The excitement, the applause, and the pride on their faces after the successful launch say it all.
Good to see the next generation setting its priorities right.
Now, how does Claude act if the model isn't aware of the testing harness?
It would be really cool to dig into how Claude reached the conclusion that it was being evaluated.
This is both scary but also promising - the AI has developed a compass (moral? scared of it's teachers?)
New Anthropic research: Natural Language Autoencoders.
Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read.
Here, we train Claude to translate its activations into human-readable text.
Our newest model, π0.7, has some interesting emergent capabilities: it can control a new robot to fold shirts for which we had no shirt folding data, figure out how to use an appliance with language-based coaching, and perform a wide range of dexterous tasks all in one model!