Teaching Tower Thai: our plan for a TowerInstruct LoRA
We are building a small LoRA on Unbabel TowerInstruct-7B so English→Thai machine translation can improve on our own GPU—without burning frontier tokens on every paragraph.
When a local coding agent hallucinates success: Qwen-Coder on our GPU desk
We handed a real multi-file React Native feature to a local Qwen2.5-Coder 14B coding agent on our Gaia GPU. It reported success. Git stayed clean. Here is the failure mode — and how we route AI engineering work now.
Local GPU bulk MT: Tower-Plus on a 3090 cuts frontier token cost for 3DN and PolitiCap
Hybrid Grok desk: frontier judgment, gaia-tower for EN→ZH/NL. 59 articles translated locally; ~$12 saved this batch.
Hybrid GPU desks: Grok Build makes local Qwen a first-class worker under Grok API direction
After open weights lost to Grok 4.5 on our coding desk, a quad-3090 + EPYC box looked like a desinvestment. Hybrid is a first-class Grok Build feature: keep Grok on the API for judgment, offload bulk turns to local inference — with cost calculus.
Open weights are not open minds: the neural net cosplay
Open-weight models are sold as more open than API models. A neural net is billions of opaque floats either way. Open mostly means run it yourself — usually on NVIDIA, sometimes the Ascend bet.
DeepSeek returns on Huawei iron — we told you the bus was the story
DeepSeek’s new model generation, tuned for Huawei Ascend, is the software half of the NVLink-vs-open-bus fork. How likely is Chinese hardware independence? High for a sovereign inference lane; contested at the deepest manufacturing layers.
Why long sessions make AI coding cheap: prompt cache hit-rate
If you run coding agents day in and day out, you eventually notice something odd. Some weeks the bill looks almost boring. Then a few sharp spikes appear — usually on days when someone reopened an old chat on another … Continued
