This week in tech and AI
Open weights close the gap, agent formats settle, and Khmer speech-to-text gets usable. Six things worth ten minutes.
Updated 4 Sep 2026
What to know
- Three open-weight models now fit on a single accelerator.
- The reasoning-benchmark gap with frontier models has narrowed enough to re-run your own evaluations.
- Serving cost, not capability, is now the deciding factor for most workloads.
Latest stories
Inference prices drop again for batch workloads
Overnight batch is now cheap enough to change how you schedule enrichment jobs.
Flexitech Digest8 Sep3 min read
Container queries are now safe to use without a fallback
The last holdout shipped support. Component-level responsive rules no longer need a media-query shim.
Flexitech Digest7 Sep5 min read
Khmer speech-to-text gets usable on commodity hardware
Word error rates on conversational Khmer are down again. We ran the two open models against our own call recordings.
Flexitech Digest6 Sep6 min read
Agent frameworks converge on one tool-call format
The competing schemas are settling. If you build integrations, this is the week to standardise.
Flexitech Digest5 Sep3 min read
Ten minutes, then test yourself
A short quiz on the week’s stories.
This week in AI: three questions
Three questions on the week’s stories. Under two minutes.
By category
Every story we publish lands in one of these.
Models, agents and what it costs to run them.
DataDatabases, pipelines and the shape of the numbers.
FrontendBrowsers, frameworks and interface engineering.
Khmer techCambodian products, Khmer-language tooling and local research.
MobileNative and cross-platform, and when each is worth it.
SoftwareLanguages, tooling and the craft of shipping.
Ten minutes of tech and AI, once a week.
One email a week, Monday 8am ICT. Unsubscribe any time.