In a DNS security NRD (newly registered domains) is a very efficient security protection, but how annoying it's in active communities like Software Development swe or following new startup(s)...
Knowing an efficiency, I cannot decide to disable this feature, but having most of new announcements blocked is really painful...
For now only browser plugin VPN helps to temporary access the website and turn it off without allowing it in the system globally.
🌗 許多通過 SWE-bench 測試的 PR 不會被合併至主分支
➤ 基準測試分數與真實開發環境的鴻溝：AI 代理的實用性迷思
✤ https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main/
該研究指出，即便 AI 代理產出的程式碼通過了 SWE-bench 的自動化測試，仍有約半數無法通過真實專案維護者的審核。研究人員邀請了來自 scikit-learn、Sphinx 與 pytest 的活躍維護者，針對 AI 生成的補丁進行盲測，結果發現維護者的合併意願與基準測試分數存在 24 個百分點的落差。此結果顯示，單純依賴基準測試分數可能會高估 AI 代理在實際開發場景中的實用性。作者強調，AI 缺乏人類開發者在審核反饋後進行迭代修正的過程，因此基準測試應被視為評估 AI 能力的參考指標之一，而非唯一的決策依據。
+ 這項研究非常及時。我們經常看到 AI 在基準測試上刷高分，但在實際專案的 PR 審
#人工智慧評測 #軟體工程 #AI 代理 #SWE-bench
SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via CI
https://arxiv.org/abs/2603.03823
#HackerNews #SWE-CI #AgentCapabilities #Codebases #ContinuousIntegration #Research
New post on our blog!
Let's look back in time a bit, let's see how far Software Engineering has come, we have a timeline of WHEN and WHY Software Engineering came to be.
https://blog.codeminer42.com/a-short-story-on-how-modern-software-engineering-came-to-be/
#Codeminer42 #SoftwareEngineering #Software #SWE #Programming
🥌 🇩🇪 🇸🇪
Wieso macht #GER die Guarden nicht weg gegen #SWE? #Curling
https://www.olympics.com/de/milano-cortina-2026/results/cur/m/team4-------------/prel/000803--/team-shotbyshot
One could argue that some languages address very specific needs like SQL or Erlang or even Rust for its memory safety because they represent a point of view on how to build software. But AI doesn't erase those differences, again it just makes them implementation details: if you are good enough in describing a problem, an AI agent can then decide "this part needs a relational model, this one an actor model, here Rust is better..." and can stitch together multiple paradigms under the hood while we stay one layer up talking about behavior, contracts, risks, constraints...
I hear a voice from the bottom of the room "HEY! AI is not deterministic!!"
Fair point. But I can win this. Networks are unreliable, disks fail, clocks drift, processes crash, users mess up. We don't fix reality, we build protocols, retries, idempotency, tests, monitoring and abstractions to achieve a predictable system.
AI is no different! You will not just ship the raw generated code into production and hope for the best. You have specs, constraints, evaluation suites, safety rails, reviews. The non-deterministic part will be boxed, tested, observed and, when needed, overridden by humans.
AI will not cancel our work as engineers. It will cancel the illusion that our value lives in the choice of tools, instead of in the quality of the systems we imagine and the responsibility we take for them.
If you feel like UML diagrams aren't very concise or helpful, here is an infographic from a 2014 study about how diagramming works in practice, and what is helpful and gets archived vs not. I found it very educational! #softwaredevelopment #swe #programming #uml
https://web.archive.org/web/20221206170620/https://empirical-software.engineering/assets/pdf/sse14-sketches-poster.pdf
END OF 1ST: #SWE 🇸🇪1 - #SVK 🇸🇰 1
We are tied again after It’s ruled NO GOAL!
What a save from Markstrom!
Slovakian fans are not gonna like that one as the puck was as close to “fully across” as perceivably possible.
Friday's here and so is another Dev Weekly entry!
From worrying Heroku product updates and resumable CSV imports in Rails, to groundbreaking AI models and Toyota’s new game engine, dive in for a quick roundup of what’s new!
https://www.youtube.com/watch?v=b9EbCb5A408
Today's find on the impact of LLMcoding to maintainability of the result.
Assumption 80% of a systens cost arises from.maintenance, thus maintainability is still relevant in the prssence of LLMcoding.
TL;DR: A fool with a tool is still a fool. And LLMcoding is just that: a tool
Given the confirmation bias I'm curious to see reproduction and follow up studies and papers.
The video mentions that the results were published as a peer reviewed paper. Unfortunately I couldn't immediately find said paper. If any one finds it, please post a link/DOI below.
#swe #research #softwareengineering #LLMs #aiassistedcoding #claude #ai
Every tech company has an A team and a B team.
The A team gets to work on all the flashy features, the things that management and the C suite usually brag about. They're technically very capable, not dogmatic, and fast. Management sees them as the "rockstar engineers". Their promo docs write themselves. They have an easy time surfacing business metrics to augment their argument for promotions.
Then there's the B Team, equally if not even more competent. They are deeply accustomed to the implementation details of the abstractions they use, they can be counted on to fix P1 incidents.
Both of these teams suffer from positive feedback loops. The A Team will continue to get work that gets them accolades, and those services will end up being maintained by the B Team.
The B Team will increasingly be given the A team's services to maintain after the A team moves on to new flashy projects. The issue here is that maintenance is an invisible part of software to management, and often treated as a cost centre. For that reason, the positive feedback loop for Team B is that they'll be relegated to be the unsung heroes of maintenance. Its much harder to argue for a promotion when "nothing happened this quarter" is an incredibly successful metric for maintaining a service.
AI is an exponential function for both of these positive feedback loops: the A team can pump out more code and flashy features faster, the B team has exponentially more software to maintain.
#Endstand #SWEGER 4:1 Der Sieg geht absolut in Ordnung 👍 (37:16 Torschüsse) #Eishockey 🏒 #OlympicGames #Italien 🇮🇹 #SWE 🇸🇪 4:1 #GER 🇩🇪 🏆🏒🏆🏒🏆🏒🏆🏒🏆🏒🏆🏒🏆
