AI digest: Models get weird, agents get real
New architectures challenge transformer dominance while Chinese labs face theft accusations and AI agents start handling real work.
The AI world is splitting between architectural experiments and practical deployments. Meanwhile, the drama between US and Chinese labs is heating up.
Liquid AI ditches pure transformers for hybrid approach
Liquid AI released LFM2-24B-A2B, a 24B parameter model that mixes attention mechanisms with convolutions to solve memory bottlenecks. This matters because everyone’s hitting the limits of pure transformer scaling. The hybrid approach could be the path forward when throwing more compute at the problem stops working.
Inception’s Mercury 2 does diffusion for text reasoning
Instead of generating text token by token, Mercury 2 refines entire passages in parallel using diffusion, making it 5x faster than conventional models. Diffusion worked for images, so applying it to language reasoning makes sense. Early days, but this could change how we think about text generation.
Anthropic accuses Chinese labs of systematic Claude theft
Anthropic claims DeepSeek, Moonshot, and MiniMax used 16 million queries across 24,000 fake accounts to systematically extract Claude’s capabilities for their own training. This isn’t surprising but the scale is wild. Expect more of these accusations as the competition gets nastier.
Claude jumps between Office apps autonomously
Anthropic now lets Claude switch between Excel and PowerPoint independently, running analysis in one app then building presentations from the results in another. This is the kind of practical agent work that actually saves people time instead of just being a demo.
OpenAI wants to kill the coding benchmark everyone uses
OpenAI says SWE-bench Verified is broken because most tasks reject correct solutions and models have likely seen the answers during training. Fair point, but convenient timing when their models aren’t winning anymore.