The cofounder and CTO of Perplexity, @denisyarats just said internally at Perplexity theyโre moving away from MCPs and instead using APIs and CLIs ๐
sol vibecoded an MCP for my oven and now Claude Code can Claude Cook. it can access sensors, a temp probe (if inserted), and cooking. it has air frying, steaming & combi modes
produced via a light touch of decompiling the official app
(pic from MCP camera tool)
As someone who trains and fine-tunes LLMs, this loss landscape video is the closest thing to a "why Llama and GPT-5 don't get stuck during training" explainer I've ever seen released for free.
Everyone thinks training an LLM is rolling a ball downhill. It's not. A wormhole opens under your model and drops it into a valley that wasn't there a second ago - the effect that lets Claude and GPT-5 train at all instead of freezing on step one.
Bookmark & watch this weekend. Same gradient descent Meta and OpenAI use today, from an 1847 paper to trillion-parameter frontier models.
std::remove doesnโt remove
std::unique doesnโt make it unique
std::async isnโt reliably async
std::vector<bool> isnโt a vector of bool
std::move doesnโt move
std::forward doesnโt forward
std::endl isnโt just a newline
std::map::operator[] might insert
C++ ๐