AI digest: Reality bites back
Massive funding rounds for coding agents, Google's spelling problems, and new techniques for faster AI inference.
The hype train keeps rolling with billion-dollar valuations, but practical problems are piling up faster than solutions.
Cognition raises $1B as coding agent fever peaks
Cognition just doubled its valuation to $26 billion in under nine months, pulling in another billion for their Devin coding agent. They’re claiming $492 million in annualised revenue, which sounds impressive until you remember most coding agents still struggle with basic tasks. The money suggests investors believe the hype, but we’re not seeing the productivity gains that justify these numbers.
Google’s AI still can’t spell “Google”
Google’s AI models have an embarrassing spelling problem that makes them struggle with basic text rendering. This isn’t just awkward, it’s a fundamental issue with how these models understand language versus visual representation. For a company that’s betting everything on AI search, having models that can’t reliably spell their own name is properly cringe.
EAGLE 3.1 fixes speculative decoding drift
EAGLE 3.1 tackles attention drift in speculative decoding, a real production issue that’s been causing inference instability. Speculative decoding should make models faster, but attention drift was making outputs unreliable. This joint effort from the EAGLE team, vLLM, and TorchSpec actually addresses a technical problem that matters for deployment.
Nvidia’s Polar trains agents without breaking them
Nvidia released Polar, a framework for training language agents with reinforcement learning whilst keeping their existing harnesses intact. It sits between the agent and inference server, capturing token interactions for training without disrupting the agent’s behaviour. The SWE-Bench improvements are solid, though we’d like to see how it performs beyond coding tasks.