OpenAI 等 AI 巨头实习生日薪曝光

据 Odaily 报道，全球 AI 巨头实习生薪酬呈现显著分层。顶尖项目中，OpenAI Residency 项目月薪达 1.8 万美元（日薪约 5625 元），面向核心研究团队，门槛极高，被视为顶尖研究员预招聘；Anthropic 的 AI 安全研究员周薪 3850 美元（日薪约 5198 元），另配每月 1.5 万美元算力经费，首期学员超八成有论文发表。美国常规技术实习岗中，Meta 日薪约 3780 元，Google 约 3400 元，英伟达博士岗上限可超 5000 元。相比之下，中国顶尖专项计划如字节跳动 Top Seed 日薪 2000 元，小米顶尖岗日薪 1000-1100 元，而普通岗如 DeepSeek 日薪 500-1000 元，阿里 350-550 元，智谱 200-300 元，部分公司以高比例员工持股作为长期补偿。网传 DeepSeek 实习生日薪 5500 元未经官方确认，属顶尖个例，中美薪酬差距仍达 6-8 倍。 #OpenAI #AI #实习 #薪酬 #科技 #人才竞争 #Anthropic #DeepSeek #字节跳动 #大模型

Paper on Architecture for AI-Assisted Software Development

Focus to learn more arXiv-issued DOI via DataCite Submission history From: Hartwig Grabowski [ view email ] [v1] Thu, 25 Jun 2026 13:51:22 UTC (21 KB) Full-text links: Access Paper: View a PDF of the paper titled The Spec Growth Engine: Spec-Anchored, Code-Coupled, Drift-Enforced Architecture for AI-Assisted Software Development, by Hartwig Grabowski View PDF HTML (experimental) TeX Source view license Current browse context: < prev | next > new | recent | 2026-06 Change to browse by: cs References & Citations NASA ADS Google Scholar Semantic Scholar export BibTeX citation Loading... BibTeX formatted citation × loading... Data provided by: Bookmark Bibliographic Tools Bibliographic and Citation Tools Bibliographic Explorer Toggle Bibliographic Explorer ( What is the Explorer? ) Connected Papers Toggle Connected

Papers ( What is Connected Papers? ) Litmaps Toggle Litmaps ( What is Litmaps? ) Toggle scite Smart Citations ( What are Smart Citations? ) Code, Data, Media Code, Data and Media Associated with this Article alphaXiv Toggle alphaXiv ( What is alphaXiv? ) Links to Code Toggle CatalyzeX Code Finder for Papers ( What is CatalyzeX? ) DagsHub Toggle DagsHub ( What is DagsHub? ) GotitPub Toggle ( What is GotitPub? ) Huggingface Toggle Hugging Face ( What is Huggingface? ) ScienceCast Toggle ScienceCast ( What is ScienceCast? ) Demos Demos Replicate Toggle Replicate ( What is Replicate? ) Spaces Toggle Hugging Face Spaces ( What is Spaces? ) Spaces Toggle ( What is ? ) Related Papers Recommenders and Search Tools Link to Influence Flower Influence Flower ( What are Influence Flowers? ) Core recommender toggle CORE Recommender ( What is CORE? ) Author Venue Institution Topic About arXivLabs arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

The AI Situation in Software Development

Random thoughts about prompting, context windows, compression, and working with AI. You want it to do something you have in mind, and you know how/what to do, or sometimes you don’t. Now there are 3 options: you tell it everything down to the details, every specific thing. Or you just tell it to do something at a high level and expect the thing to understand. Or you can go the middle way. I feel this is the go-to way, explaining the important parts that you think might be difficult for it. You can feed it examples. It’s a faster way to do things, but it depends on the example being close to what you want. All of these are time-consuming. Some you spend time before giving to AI, some after. A common pattern is easy for LLMs to implement, considering they must have seen it before in their training set, for example implementing user auth. A n

ew problem you’re imagining or telling it is of course hard for it and needs hand-holding. Then there’s the context window problem. You can’t just give a 3000-word, 4-page detailed dense spec and expect it to follow everything, and the larger the codebase, the less it can pack everything in, nor are the vast documents you can feed it worthwhile. Not only for writing detailed specs ~ you also want it to summarize patterns, draw conclusions from a large dataset, be it something like analyzing vast amounts of numerical data, for example a historical dataset for a stock. So the cost is on you: you still need to spend the time to write a detailed guide for your project, its goals and its issues, and more importantly the blueprint of the thing you want. Then there’s domain-specific expertise of AI models. You need to pick and choose the right one. As you work on bigger problems and as you integrate AI into your applications, a need for compression arises, packing as much useful information as possible, if not all, into your AI agent to solve a particular problem or to draw a conclusion, make a decision or whatever. I think there will be companies in this space that’ll do this effectively, or the model builders will just solve this once and for all. There must be feedback loops in terms of tests, tooling (purpose-built or otherwise), and refining its approach as the codebase grows large. And ways for improving the signal-to-noise ratio in your codebase. I feel that great explainers or natural teachers find it easy to engage with AI and produce better outputs. Bottom line: you still need to spend time. The implementation time is gone. Now the time you spend has shifted to designing the system upfront, changing assumptions, and refining your dev setup. But implementation is not really gone. I feel I am still implementing in words instead of code.

ders but instead something else with a voracious appetite for new information: AI. The idea that the secondhand sales boom is being driven by the explosive growth of AI can be traced back to a court ruling in the US. In 2025, a judge ruled using books purchased in this way to train AI software was not a violation of US copyright law. The decision was the result of a lawsuit brought against AI firm Anthropic by three authors. In his ruling, Judge William Alsup said Anthropic's use of the authors' books was "exceedingly transformative" and therefore allowed under US law. When court documents were unsealed last month it also emerged books were being destroyed in the process of training Anthropic's AI chatbot, Claude. "Claude is trained on a mix of publicly available web data, commercially acquired datasets, and data we generate ourselves," a spokesperson said. They insisted sourcing books for training was a widely used approach across the AI industry. "None of our data acquisition programs buy and destroy rare or antiquarian books," they added. Nonetheless, the idea that books are being pulped is causing unease. David Tobin, runs Walden Books in north London, has also had unusual sales. "In some ways it's very nice to sell some of these titles which haven't been sold for many years, but it would be sad if they are ultimately destroyed," he says. The court documents relating to the Anthropic case revealed the project of ingesting old books was referred to in internal company communications as "Project Panama." The documents indicated the company's aim was to "destructively scan all the books in the world". Destructive scanning is the process of shipping books to locations where they can be digitised at an industrial scale. It includes removing a book's spine so all the pages can be scanned rapidly - and the remains recycled. "A lot of mystery surrounds Project Panama," Manley, from Barter Books, says. "The name is new to me, but the reality of the project is not, and has been the subject of much discussion on the bookseller forums." He does not know that his books are being bought for it or similar projects by other AI firms. But he says it's also difficult to account for the sales, which appear random with "no rhyme or reason." They have varied from obscure Latin texts to cowboy novels. Experts say the diverse subject matter also points to AI, as unusual and rare texts could provide fresh material to improve the training of large language models, the tech which underpins generative AI tools like chatbots. Professor Emily Hudson, intellectual property specialist at Oxford University, says copyright laws in the UK are different to those in the US. "The starting point in the UK is that all these acts of copying - creating the training library and doing the training – require the permission of the copyright owner," she says. For the booksellers, it poses a dilemma. They are uncomfortable with the idea of books being destroyed - even if they admit not every title needs to be saved. "A recent academic text published in only 100 copies, 75 of which are already in libraries, may be very rare on the market - but it is perhaps not such a great loss if one copy is destroyed," says Derek Walker, owner of Edinburgh bookshop McNaughtan's. "But we have, and have sold, books which are for example the only known surviving example of an edition from the 18th century. "It would be a much more significant problem if one like that were to be bought for

destruction, having survived this long." And Manley says recycling books is a good solution for many titles which the public no longer want on their shelves - especially when it comes with a bump in trade. "Some may have ethical concerns about where the books end up and if they're destroyed," he says. "But the world no longer needs five million copies of The Da Vinci Code. "I've had books advertised for 20 years on the web which haven't sold until now". What is AI and how does it work? Sign up for our Tech Decoded newsletter to follow the world's top tech stories and trends. Outside the UK? Sign up here .

Model your business once – for humans and AI alike

nexusx allmonday/nexusx Guide Guide Overview — One Model, Two Graphs Quick Start Data Graph Data Graph GraphQL Mode Auto-Generated Queries GraphQL Pagination Response DTOs Response DTOs Core API Mode Core API Advanced Relationships beyond the ORM Relationships beyond the ORM ER Diagram & Non-ORM Relationships Custom Relationships Virtual Entities Operation Graph Operation Graph UseCase Service UseCase + FastAPI AI Delivery AI Delivery Compose MCP for AI MCP Service (entity-first) MCP & Context Efficiency Beyond One Database Beyond One Database Federation ComposedErManager Tooling Tooling Voyager Visualization ER Diagram Visualization Troubleshooting Articles Articles Design Highlights Clean Architecture Comparison API Reference API Reference GraphQLHandler Core API Cross-Layer Data Flow Relationships & ER Diagram MCP

API UseCase API Reference Reference Changelog Migration Guide DTO-first Execution Migration UseCase GraphQL 3.0 Migration Table of contents Run It in 60 Seconds What You'll Get Who Is This For Learning Path Guide (Tutorial Path) Advanced Guides API Reference Back to top

Yadda 3.0.0: BDD in the Age of AI Agents

I’ve just published Yadda 3.0.0 to npm. That means that instead of writing something like: Given a university, The University of Bouvet Island And The University of Bouvet Island offers a degree course in Computer Science with entry requirements of ABB And an A-Level graduate, Steve And Steve has a D in Physics And Steve has a D in Maths When Steve applies to study Computer Science at The University of Bouvet Island Then The University of Bouvet Island rejects the application you can write: The University of Bouvet Island offers a degree course in Computer Science The entry requirements for which are ABB Steve is an A-Level graduate With a D in Physics And a D in Maths When Steve applies to study Computer Science at The University of Bouvet Island They reject his application Both are executable specifications. I find the second considerabl

y easier to read. Most of Yadda 3.0 is a modernisation exercise. All useful, but not especially interesting to write about. There are two things about the release that I think are much more significant. I modernised Yadda using Claude Code with Opus 4.8. The Yadda 3.0 epic , which was itself written by Claude, broke the work into a series of deliberately separated phases: remove obsolete functionality, update the toolchain, perform mechanical formatting separately from behavioural changes, modernise the source, explore API changes, update examples and CI, then finish the metadata, documentation and TypeScript definitions. We planned each phase before implementing it, and then I largely let Claude get on with the work. It made remarkably few mistakes and, more impressively, identified some fairly subtle edge cases that would have been easy to miss during what initially looked like a mechanical modernisation. I made very few interventions. One important factor was that Yadda already had a comprehensive test suite. I also deliberately avoided asking Claude to modify production code and the corresponding tests in the same step. If an agent changes both simultaneously, a green test suite becomes weaker evidence because it is free to change the definition of “correct” at the same time as the implementation. Keeping those changes separate gave Claude a much firmer external constraint. From starting the work to having the package published was roughly a day of elapsed time, and I was doing other things in parallel. At the beginning of this year I wrote about an experiment asking why experiences of vibe coding were so polarised . My conclusion then was that the results depended enormously on how the agent was used. A tightly constrained and supervised Claude could produce extremely good results very quickly. Left to its own devices, it tended towards architectural drift, unnecessary code and operational debt. That was only seven months ago, and the capability has moved on enormously. Even so, saying that Claude can now write this code with very little intervention barely scratches the surface of what is changing. To appreciate where this is going, it helps to stop thinking about a single developer having a conversation with a single coding agent and instead consider several agents working in parallel. There are already several ways to do this. You can simply run multiple Claude Code sessions. Git worktrees let each agent work against an isolated working copy. Tools such as cmux make running a collection of Claude sessions more manageable, while Claude Code Agent View provides another way of seeing what multiple sessions are doing and which ones need attention. All of these let you build significantly faster than working serially, but I fairly quickly hit another limit: my own ability to manage the parallel work. I can comfortably keep three tasks moving at once, and sometimes four or five. Beyond that, I start losing the context of what each agent is doing, which decisions have been made, which task is waiting for me and what I need to review next. At that point, the model is not overloaded and the machine is not overloaded. The bottleneck is the human coordinating the work. I’ve become convinced that good orchestration is the next important layer. I’m not alone in reaching that conclusion. My colleague Marco describes almost exactly this progression in My AI Engineering Journey , moving from AI as autocomplete, through supervised and trusted