I am very curious about how the "threat" of local inference and open-weights-models will change the trajectory of the model labs. The air will become pretty thin
I noticed that Claude's reasoning summaries show a phenomenon when I'm using the model in German (on claude.ai) - it mixes English grammar with German vocabulary!
"Evaluating Spülenposition gegen Wasseranschlussabstand" and "Analysierend die Platzierungskonflikte und Rohrleitungszwänge klären" are examples of generated summaries.
I wish someone could explain this - because I see no way that sentences like these would show up in training data. But maybe the RL for thinking has some quirks?
great point, we haven't really spent too much thoughts about that.. changing the logos to companies we know well makes more sense, we'll do that to save us trouble
that gave the best results so far - tweaking the system prompt to make sure the agent respects the project's ui system, is aware of dark mode and responsiveness, etc.
Haha, u're right, I recently had a chat w a friend about how high our threshold for excitement has become. A few years ago, a new iPhone release would have meant the world
It's so funny how many people tell us that - it's one of the obvious ideas someone just needed to start with. What kept you from building the first version?
Thanks! There are also some other tools out there that give 'vision' to your coding agent via MCP, but we figured that prompting the agent directly by clicking on elements is the most intuitive way to interact with it.