Small language models are quietly winning the enterprise
Companies that spent 2025 chasing frontier models are now shipping 7B-parameter systems that run on a single GPU and cost a tenth as much.
13 stories
Companies that spent 2025 chasing frontier models are now shipping 7B-parameter systems that run on a single GPU and cost a tenth as much.
Automation was supposed to end the labeling industry. Instead the work got harder, more specialised and better paid.
Dozens of competing orchestration libraries have collapsed into three serious options. Developers finally get to stop rewriting glue code.
Reinforcement learning from human feedback made assistants polite. It also made them hedge, flatter and over-explain.
Phone makers are shipping speech models that run offline, handle crosstalk and know when to stop writing.
Accelerators are finally available. The grid connections to run them are not.
Bootcamps and universities are rewriting curricula around tools that write most of the boilerplate.
A new evaluation suite rotates its questions weekly and publishes the answers afterwards. Scores dropped overnight.
A lab has trained manipulation policies from home footage instead of lab demonstrations, and the robots generalise surprisingly well.
Theorem provers, planners and rule engines are being bolted onto language models, and the combination is beating either alone.
No spam. One email a week. Unsubscribe anytime.
One email a week: the breaches that matter, the patches to apply, and the tools we actually tested. No filler, unsubscribe anytime.