READ THE NEWSROOM Choose your reading view.

OBVS / A wider view. A closer read.

Intelligence.

The monthly edition2026-11

Ideas worth a closer look.BRAND / TECHNOLOGY / BUSINESSPublished reporting
01 / commerce

Images 2.6 adds targeted edits to the API
One product. A changing setting. A fixed label.An original illustrative bottle sits within a selectable background. The scene changes from warm paper to botanical shapes; its label remains the review target. This is a conceptual explanation, not an output from Images 2.6. FIELD NOTESBotanicalSTUDY NO. 01250 ml REVIEW THE DETAILSName. Type. Volume. AN ORIGINAL OBVS ILLUSTRATION
THE BRAND EDIT01—03
Select the setting. Keep the product outside the requested change. Conceptual illustration, not an Images 2.6 output.

OpenAI’s new image release lets developers mark where an edit should happen. For product teams, the review starts with the details that must stay fixed.

By OBVS Intelligence Team

Read the story
02 / ai search

Writer’s sparse-attention study cuts a wait during generation
WORK IN PARALLEL
Layer n
Attention
Layer n+1
SelectAttention
Illustrative sequence, not a measured timing chart. The research moves the next selection alongside current work.

Writer researchers let one AI operation use an earlier token selection while another prepares the next. Their reported speed gains depend on model and workload.

By OBVS Intelligence Team

Explore the visual story

A CLOSER LOOK / AI RESEARCH

The wait
between words.

A long answer is built one token at a time. Writer’s researchers ask a smaller question: must every layer wait for a fresh selection?

Continue to the other stories ↓

01 / THE DEPENDENCY

First choose.
Then attend.

Earlier tokens → a selected subset

IndexerSelect
AttentionwaitUse selection

In a conventional layer, attention depends on that layer’s selection. The waiting is the point.

02 / THE HANDOFF

Use one selection.
Prepare another.

Preceding indexer layer’s
selection
AttentionUse preceding selection
IndexerPrepare later selection
To a later layer

Both rows can run together, within the same decode step. The first retained indexer layer still runs serially.

03 / THE MEASURED RESULT

A gain, with
conditions attached.

DeepSeek-V3.2up to14.5%

Fused kernel · 120K context

GLM-5up to9.8%

Deferred streams · 195K context

Decode throughput: IndexCache + MLI versus IndexCache, both at ½ indexer retention. Single request, eight H200 GPUs.

Original OBVS explanation. Timing and token shapes are schematic, not measured durations or actual selections.
01

A small pause.
Repeated many times.

An indexer chooses which earlier tokens a layer should attend to. Conventionally, attention waits for that choice. This happens as the model generates its answer.

The opportunity is in the dependency, not in making the answer shorter.

Read the method in the paper ↗
02

Let the work
overlap.

Neighboring layers often select similar tokens. The Multi-layer Indexer lets attention use the preceding indexer layer’s selection while the current indexer prepares a later layer’s choice.

Similar does not mean identical. This reuses a selection from the same decode step, not from the previous generated token.

Follow the researchers’ design ↗
03

Now ask what
the test proves.

The reported gains are specific to the model, implementation and workload. More concurrent requests reduce the benefit and can produce regressions. Quality varies; equivalence is not established.

For a brand assistant, time the whole sourced answer and check the cited passage. Faster token generation alone cannot prove a better customer experience.

Inspect the tests and limitations ↗
BEYOND THE HEADLINEMore perspectives from this edition.
03 / commerce

Zalando’s partner share grows alongside its advertising business
33.6%PARTNER SHARE
OF CONSUMER GMV
Q3 2026 · up 2.9 percentage points year over year.
Source: Zalando’s November 3 release.

Partners accounted for 33.6% of Zalando’s Q3 consumer merchandise volume. Retail media grew too, leaving brands to check how much growth reaches their own margins.

By OBVS Intelligence Team

Read the story
04 / commerce

Stretch Gen 2 targets the hours lost to repairs
ABOUT A DAYAbout an hour.
Swerve drive replacement, according to Boston Dynamics. Not a full-shift productivity result.

Boston Dynamics says one drive replacement now takes about an hour rather than a full day. Fulfillment buyers still need to see what happens over a receiving shift.

By OBVS Intelligence Team

Read the story
05 / policy

Two data-center votes, two meanings of “Yes”
01 Land use02 Local process03 Power
Separate decisions. An agreement in one lane does not settle the others.

Unofficial returns in Boulder City and Wiscasset favored restrictions through opposite ballot wording. A separate OpenAI power pledge concerns a different stage of development.

By OBVS Intelligence Team

Read the story
Reporting 2–4 NovCompiled 6 Nov 2026

November,
in brief.

Early November brings more precise image edits, faster token generation and more connected commerce. The shared task is to preserve product facts, define permissions and measure what the whole workflow delivers.

Read the November roundup
Up to14.5%

Higher decode throughput

DeepSeek-V3.2, single request, fused kernel, half the indexers retained: IndexCache + MLI versus IndexCache.

Authors’ tests. Gains narrowed with concurrency; quality varied. MLI added no first-token improvement.

Writer AI Research
33.6%

Partner share of B2C GMV

Zalando’s reported Q3 2026 partner share of business-to-consumer gross merchandise volume.

A share of merchandise volume, not company revenue or a seller’s profit.

Zalando Q3 results
≈1 hour

One swerve-drive replacement

Boston Dynamics says this Stretch Gen 2 repair takes about an hour, compared with a full day on the earlier design.

Manufacturer’s repair example, not measured fleet uptime or a faster-delivery promise.

Boston Dynamics

Patterns across the reporting OBVS Intelligence Team

Keep the facts intact

Images 2.6 adds targeted edits; Perplexity describes connected campaign work. Both make the review boundary important: preserve approved product details and require permission before publishing, sending or spending.

Image edits Campaign agents

Measure the whole job

Writer targets a wait during token generation; Stretch targets repair downtime. Each gain concerns one stage. Check answer quality and first-token delay for the assistant, and order-ready time for the warehouse.

Sparse attention Repair time

Name the milestone

Zalando’s agent checkout is planned for 2027. Data-center votes concern land use, while Grid Protectors concerns electricity demand. Commercial access, local permission and operating results each need their own evidence.

Commerce plans Data-center votes
Scroll to look closer All reports ↓

OBVS INTELLIGENCE / READING EDITION

A little more perspective.
On the agentic web.

Browse the same OBVS Intelligence stories in a reading grid. Follow the reporting, practical implications and original sources.

All stories.

Reporting and briefings, month by month.

Search and topics

November 2026

  • Zalando’s partner share grows alongside its advertising business

    Partners accounted for 33.6% of Zalando’s Q3 consumer merchandise volume. Retail media grew too, leaving brands to check how much growth reaches their own margins.

    Coverage –
    Published

  • Writer’s sparse-attention study cuts a wait during generation

    Writer researchers let one AI operation use an earlier token selection while another prepares the next. Their reported speed gains depend on model and workload.

    Coverage –
    Published

  • Stretch Gen 2 targets the hours lost to repairs

    Boston Dynamics says one drive replacement now takes about an hour rather than a full day. Fulfillment buyers still need to see what happens over a receiving shift.

    Coverage –
    Published

  • Images 2.6 adds targeted edits to the API

    OpenAI’s new image release lets developers mark where an edit should happen. For product teams, the review starts with the details that must stay fixed.

    Coverage –
    Published

  • Two data-center votes, two meanings of “Yes”

    Unofficial returns in Boulder City and Wiscasset favored restrictions through opposite ballot wording. A separate OpenAI power pledge concerns a different stage of development.

    Coverage –
    Published
  • Weekly briefing

    Agents move into marketing—and need clearer limits

    New campaign workflows, planned agent checkout and web-grounded decisions make evidence and approval boundaries a practical business priority.

    Coverage –
    Published

Return to the front page ↑
OBVS / Intelligence