Tech news picked on 2026-10-24

2026-10-24

1. Using a mask in PyTorch attention inference spiked GPU memory; the author reduced it by changing the fallback path ↗

   • The model Xiaomi-Lens-0, which creates masks marking image regions, switches to the MultiheadAttention path when flash-attn is not installed. When a mask was passed, that path created a matrix sized as 16 heads × the square of the patch count.
   • According to the author's measurements, for a 3-megapixel image the extra VRAM peak in the mask step was 18.02GB on the original path and 2.28GB on the SDPA path. The test environment was RTX 5090 32GB, Windows 11, with flash-attn not installed.
   • The PyTorch documentation says an optimized path is used when need_weights=False. However, when the author read the CUDA source, the path was not entered when a mask was present. In CPU profiling, the same layer in training mode used SDPA, so the issue reproduced only in eval inference.
   • The author reported that across a comparison of 6 samples, tokens and boxes matched, and mask IoU ranged from 0.9976 to 0.9998. This is a limited comparison, not a full quality evaluation.

   zenn.dev19d agoSummary and discussion

2. When Work Split Stalls, a One-Page Responsibility Map Turned Cross-Team Talks into Revision Work ↗

   • In an external service integration feature, one team owned requirements definition and the neighboring team owned implementation. The author explains that no one was positioned to judge whether the requirements were feasible and how much effort they would take, so work kept going back and forth between the two teams.
   • The author says that abstract requests to change how work is divided between teams always ended with the claim that it was practically difficult. In response, they created a one-page responsibility boundary map showing the owner of each stage and brought it to a meeting.
   • In the diagram, engineers gather materials for preliminary research and pass them to a product manager (PdM, the role that sets a feature's direction and priorities). The PdM decides the requirements definition, and engineers take charge of design. The author states that after this diagram was used, the meeting shifted to concrete revision discussion. The source does not state that the effect was measured numerically.
   • Two objections came up in the meeting. One was that the receiving team had no spare staff. The other was that the knowledge needed for research and requirements remained with the former team. The author says they handled these separately from the discussion about responsibility division.

   zenn.dev19d agoSummary and discussion

3. Apple Opens AI Assistant Beta in iWork Apps, Available Only to Creator Studio Subscribers ↗

   • An AI assistant beta is now open in Pages, Keynote, and Numbers to help with document drafts, slides, and table work. According to the source, only Apple Creator Studio subscribers can use it, while the three apps themselves are free.
   • The source says that in Pages, users can draft text, adjust tone, generate images and shapes, and search for recent information before inserting it into a document. Keynote helps with slide creation, presenter notes, and animations, and Numbers helps with chart suggestions, formula writing, and table generation.
   • The subscription costs 12.99 dollars per month or 129 dollars per year, and new subscribers get a 1-month free trial. The subscription also includes Final Cut Pro, Logic Pro, and Pixelmator Pro.
   • Right after the update, the three apps briefly disappeared from the iOS and Mac App Stores. The source says it is still unclear whether the removal was intentional or a bug.

   cultofmac.com20d agoSummary and discussion

4. Mahanine raises $34 million for AI CAD that automatically turns scan data into civil engineering drawings ↗

   • Mahanine is civil-engineering-specific software. Survey teams scan roads, pipes, and construction sites with equipment and upload 3-dimensional data. The software automatically moves elements such as curbs, signs, utility poles, lane markings, and power lines onto drawings. Users review the results, make corrections, and then finalize them.
   • The company says this method is up to 10 times faster than existing manual work, its cost is about 1 out of 5, and it produces centimeter-level accuracy. This is the company's own claim and not an independently verified figure.
   • The product is currently in a private beta, and about 100 customers have used it on more than 2,500 projects. Customers include civil engineering firms such as WSP, HNTB, and NV5, as well as transportation agencies such as Caltrans and the Minnesota Department of Transportation.
   • This Series A round of $34 million was co-led by First Round Capital and Bessemer Venture Partners, and total funding has reached $49 million. The company was founded in 2021 by Carnegie Mellon University researchers and shifted from an early hardware business to software.

   wowtale.net20d agoSummary and discussion