MONTHLY · DIGEST
September 2026 digest
2026-09-01 – 2026-09-30 (Beijing time) is in progress. This report updates with the period's records; its data cutoff is shown below.
Data through · Beijing time
Record sources and gaps
This period is ongoing; its evidence and analysis may change.
The month, synthesized
Review 6 recorded days this month. Browse recorded items by category or open the day-by-day record.
Item descriptions use report materials through 2026-09-06. Collection and selection counts describe our coverage; they do not measure industry importance or real-world outcomes.
This month is still ongoing; focus records continue to accumulate with daily updates.
The month's rhythm
One cell per day; yellow depth shows focus picks. Solid cells have records (including zero picks), dashed cells have no saved record, and faded cells are future dates. Recorded dates open the archive.
Category review: where focus went this month
Browse recorded topics by category, alongside each item’s sources, conditions and follow-up records.
Related topics: Zhipu GLM, NVIDIA / AWS GPU supply, Anthropic alignment security.
Related topics: Qwen, DeepSeek, @huggingface/kernels.
Related topics: AfterQuery, OpenAI GPT.
Related topics: Gradio workflows, ChatGPT Work.
Related topics: Google Gemini.
Worth revisiting this month
- 01Gemini 3.5 Transcribe enters public previewOfficial source · 4 days in focus
For interview, subtitle, and meeting-tool developers, timestamps help readers return to the recording and check who said what. Test representative recordings against existing tools to assess the benefit for Chinese transcription.
September 5 · Open the day → - 02Qwen3.8-Flash-Next weights offer a Qwen4 architecture previewOfficial source · 2 days in focus
Developers working on deployment, quantization, or inference compatibility can inspect the architecture and measure task performance. A replacement decision also needs hardware, output-quality, and licensing checks.
September 5 · Open the day → - 03Perplexity open-sources Lily, a Metal inference server for exactly one checkpointOfficial source · 2 days in focus
An official open-source capability teams can exercise the same day: one checkpoint plus fail-closed load validation is a pointed engineering bet for local inference on Apple Silicon, and the pinned weight revision under Apache-2.0 makes the benchmark reproducible by anyone. The 1.35x figure is aggregator-reported, but the repo ships reproduction steps to check it.
September 5 · Open the day → - 04funes helps coding agents recall past decisionsOfficial source · 2 days in focus
If you switch sessions, coding tools or machines, try retrieving three known decisions and their reasons. Check whether the results lead to the correct original records before relying on the memory.
September 5 · Open the day → - 05Gemini adds agentic video understanding at up to 66% less costOfficial source · 2 days in focus
The Gemini mainline Flash series enters the radar with a capability-level event: three in-service models gain dynamic video analysis at once, pulling down the unit-cost curve for video understanding — an immediate cost-structure change for teams doing batch video, monitoring, or retrieval.
September 3 · Open the day →
Our selection record
These counts describe this month’s coverage and selections. Use them to find material; they do not measure market opportunity.
- More focus objects in Platforms & policies and Open-source projects
Of this month's 22 focus objects, 7 + 6 of them sit in these lanes. This describes our selection mix and offers a starting point for peer comparisons; an opportunity still needs evidence of demand, costs, and adoption.
- 7 objects selected on multiple days
“Gemini 3.5 Transcribe enters public preview” appears on 4 focus days, the most this month. Repeated selection records continued editorial attention; inspect the evidence added between those days to judge whether its value holds.
- 18 objects first selected for focus
First selection is a starting point in our record, which may come after launch. Follow the linked thread and check later versions, comparisons, and pricing changes before deciding how much attention it deserves.
Selected on multiple days
These objects appear on multiple focus days. Compare the evidence added between them to assess whether continued attention is warranted:
Tools & products
Open-source projects
Open-source projects
Platforms & policies
Practice & methods
Open-source projects
Open-source projects
Theme distribution
- Platforms & policies: 7
- Open-source projects: 6
- Practice & methods: 4
- Founder experience: 4
- Tools & products: 1
Our tracking coverage
This preserves coverage status at the report cutoff. Existing details and sources remain available after active tracking ends; coverage status is not a judgement of project quality.
Of the 22 focus objects this month, 1 were outside live tracking at the report cutoff; the watch pool and history library then retained all 35 threads.
Evidence boundaries
All judgements come from public observation snapshots; attention is not quality, and continued tracking is not continued correctness. The opportunity-and-risk read is for your own judgement — not advice, and not a decision made for you.
The summary, classifications, counts, and card wording use the same report record. Later additions are listed separately. Missing historical content is not replaced with current wording.