Mikael Svenson

If you read only one thing about AI/LLMs this week, make it this: How do AI software engineering agents work? by Gergely Orosz and Elin Nilsson https://lnkd.in/d3STueiE This is a premium post from the Pragmatic Engineer, but even if you don't subscribe, the free preview (the intro plus a nice overview of SWE-Agent) is worth a skim. Lot of great stuff in here, but a couple things I'd highlight: • One of the harder problems in agent SW dev is finding the right code to provide as context! You can get lucky with RAG, but we've found that more targeted code searches and snippets work best. Agents can get confused by too much code -- much like humans. • Linters and test frameworks are critical in agentic code workflows; the easier it is for agents to identify errors and check work, the quicker to get to usable output. • Finding the balance between prompt engineering and self-healing flows is another challenge. We've found that precision prompting can produce really good zero-shot performance (meaning that you get what you want the first time you try), but allowing agents to detect and resolve issues autonomously is part of the LLM magic. The sweet spot is somewhere in between! Have a great weekend!

26

https://lnkd.in/dxqQa5kn Just read a fantastic article by Christian Heilmann 👀 that I couldn't help but share. In this article, Christian passionately defends the role of frontend developers and highlights the critical importance of specialized expertise in web development. Christian perfectly captures the essence of what it means to be a frontend developer today. We're not just writing HTML, CSS, and JavaScript. We're crafting the very interfaces that define user experiences across all devices and browsers. His experiences with major companies like Yahoo, Bing, and Microsoft, as well as his work on browsers like Firefox and Edge, underscore the significance of having dedicated frontend professionals. One standout point was the common misconception that frontend work is somehow less "real" than backend or full-stack development. Christian dismantles this myth by highlighting the deep expertise required in performance optimization, cross-browser compatibility, and accessibility. I highly recommend reading this article, especially in an era where the push for full-stack generalists can sometimes overshadow the need for dedicated frontend expertise. It’s a powerful reminder that having specialized skills is crucial and that investing in experts can save time, money, and headaches down the line. #TechInsights #FrontendDev #LGUW #LGUW14dayschallenge

7

4 Comments

Mistral now offers the capability to finetune AI models to meet the unique needs of their customers. To make finetuning easy and accessible, there are three options: 1. On-Premise Finetuning: Utilize open-source SDK available on GitHub (https://lnkd.in/gam7d9wb), to finetune models within own infrastructure. 2. Mistral's La Plarforme: Leverage this platform to avoid any infrastructure hassles and focus on what matters most - achieving the goals. 3. Custom Training Service: For advanced model training on proprietary data, their service provides a tailored solution to meet specific needs #MistralAI #GenerativeAI #LLMs #Finetuning

6

1 Comment

I've been doing Codewars katas in Rust to try an cement my (quite limited) knowledge and really thought I was on the right track with the top solution! The problem was to return the sum of all the numbers between a and b. - So I added a mutable sum variable on line 5. - Returned 1 value if they were the same on line 6. - Determined which value was highest/lowest on lines 9-10. - Created an inclusive range between the two on line 12... Then forgot the basics of coding on line 13! I'd been so focused on using the syntax of my new language that I forgot the basics! I failed the very first test! That had passed the values 0 and 1 - should be simple right? I think what I'd been thinking of was mapping over an array in JavaScript, or something..., because I know there are methods like next() in Rust. I thought I was being clever - how do you add one value to the next. 🤔 I ran the code, and was totally puzzling over how 0 and 1 could somehow equal 4. Then I remembered I wasn't using an array, but a simple for loop. One that, for those params, would return 1 for the first iteration, then 3 the last. 🤦‍♀️ And, of course, I didn't figure that out right away. Nope, me, being me, had to question my Rust specific code. So don't forget the basics, like little old me!

16

Have you tried the FEDn API Client to interact with your Studio project? Learn about it from Niklas Holmström in this mini-tutorial: https://lnkd.in/d6fEUM5Q #federaredlearning

1

Will AI replace software engineers? Absolutely not. Do you know why? Software engineers are most of the time not struggling writing the code itself. What takes time is communication, stakeholder management, keeping the Fowler-drinking architects and the Cagan-coolaid-drinking product people at bay so work can be done. (I really believe in AI-assisted development btw. AI can be used to automate repetitive tasks, pattern recognition and much more. SW engineers will still be needed for creative problem-solving, high-level system designs, understanding nuanced human needs etc.) :) #ai #softwareengineering

122

9 Comments

GitHubNext is just the best place to work. Back in May 2023, I explored ideas around natural language coding with Johan Rosenkilde. Our ideas ended up complementary and best explored in two projects, but because everyone at Next is lovely, we made two projects, no problem. One was SpecLang, one was "EEA". Over the next months EEA merged with new ideas and become Copilot Workspace. And, with the most amazing team I've worked with, we did lots of exploration and research and user testing through to technical preview today. https://lnkd.in/gtBV9uPt I've captured a little of this with the team's blessing here https://lnkd.in/g-R839TP We're open and transparent about where we're at, and you can see that here https://lnkd.in/g-YMGakh My prog. languages work with C# generics, F#, C#/F# async programming and the simplicity of types, functions and data still have my heart, as do my ongoing responsibilities as F# language design approver. But right now I'm getting to be an applied researcher again with a wonderful team, and loving it. p.s. if you'd like a look at a use of Copilot Workspace, try this one from Jonathan Carter on our team https://lnkd.in/gvQ_VYEX

60

5 Comments

TDD haters are everywhere. For more than a decade, I have dealt with them, and I still do. The curious thing is that I could classify TDD haters into two categories: Group 1 - Practiced TDD for several months and several projects and they still don’t see the value. Group 2 - Tried TDD for two weeks and claimed it is garbage, unnecessary, and anti-agile. Here is another curiosity, based on my own empirical observations: Group 2 represents 99% of all the TDD haters I have to deal with.

8

2 Comments

Last night at Microsoft, the Stockholm AI community hosted a great talk by Sid Mohan and Arthur Lambert from Artefact on building ethical layers on top of LLMs.   I think ethical AI is one of the biggest challenges in AI today. Despite the great number of available guidelines and frameworks for building AI ethically and responsibly, the reality of building ethical AI is very complex. Ensuring that generative AI, the type of AI that powers products like ChatGPT and Copilot, produces output that is as non-discriminatory as possible is tough. It's not just a technical challenge but also a deeply human one, as different cultures have different values and, therefore, varying expectations of what an AI should produce. In some countries, we are already starting to see people of different political affiliations use different AI models for this reason. It was very interesting to learn more aspects of attempting to gather datasets and fine-tuning AI models to detect microaggressions in text. Thanks again, Sid and Arthur! #artificialintelligence #machinelearning #stockholmai

51

3 Comments

GPT-4o was released yesterday. It is claimed to be both faster and cheaper than its predecessor. Is it worth switching to the new model? At DEMA, we have compared GPT-4-Turbo and GPT-4o on our test suite for generating specific API queries from user questions in the domain of data analysis. From our initial testing, we have observed: ~3x Faster Inference: Significantly faster response times across a wide variety of prompts. A slight trade-off in precision with prompts containing more elaborate instructions. Is it worth the switch? It probably depends on the use case: For simpler tasks, GPT-4o performance seems to (at least) match its predecessor while being both faster and cheaper. For more complex tasks, some differences start showing, and careful evaluation and testing need to be carried out before adopting the new model. We'll be investigating ways to maintain and improve accuracy while enjoying the benefits of much faster inference. Exciting times ahead, as this enhancement promises significant improvements in user experience.

35

4 Comments

First there was the API. They said all you need is prompting. No coding. Prompt engineers were born. Then they said use a library. Then we needed to have guardrails. Knowledge graphs? DAGs? Oh. Libraries are not enough, we need whole frameworks. Prompt engineers are dead. Regular programming is back on the cards. AI is just a tool. It will magically not make you a successful startup. Unfortunately you still need to struggle and work hard to make a successful startup. And btw, count on being lucky too.

272

9 Comments

The most common complaint I hear about serverless is that it "is difficult to test". And this is where I most strongly disagree. In fact, I will try to prove that serverless actually is much easier to test and debug than the more common solution of using a container or a local server. Why do I say this? Simple: a serverless application is built out of isolated functions. Each function is a component, that receives an input and an output, and the services it uses are injected as arguments.   If anyone remembers the basics of the Spring Framework (which is an Inversion of Control, a testing framework at its core), the concept is the same, just at the infrastructure level. With this structure it is extremely easy to write tests: all you need to do is to pass arguments, and, if you are using external services, provide mocks for them. Serverless also makes it easy to write integration tests because the deployment is fast and can be integrated into the test code itself. However, in my experience. people who are used to building with containers are upset by this experience and complain that they "can't debug it". They expect to be able to start the server (which they cannot do with functions) and tweak with the application by interacting with it, often manually playing with the UI. The truth is that they are not used to writing tests at all . They don't do that, period. So they start the whole thing, interact with it often at the UI level and attach a debugger to find out what is wrong.  Of course this is a bad practice because you do not get a repeatable test suite, and what is worse, you have to test the whole thing instead of unit tests. Unit testing an entire server is much more harder than testing functions. So the truth is: is not that serverless is hard to test, it is that you usually do not write tests for your code at all.

22

3 Comments

This is such an impactful talk. The main gotcha for me was the study on the acceptance rate of women's PRs to open source projects. Women's PRs have a higher acceptance rate, so no discrimination there, right? Except, when they are recognizable as women on their public profile, the acceptance rate plummets. Which means... women are secretly more competent at coding, but the bias against them assumes the opposite. Thank you Andrew Harmel Law for sharing this and many more important thoughts. I would love to see more studies on this topic.

16

1 Comment

𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐖𝐨𝐫𝐤𝐟𝐥𝐨𝐰: 𝐏𝐢𝐨𝐧𝐞𝐞𝐫𝐢𝐧𝐠 𝐭𝐡𝐞 𝐅𝐮𝐭𝐮𝐫𝐞 Recently I happened to see this nice video ( 𝘩𝘵𝘵𝘱𝘴://𝘸𝘸𝘸.𝘺𝘰𝘶𝘵𝘶𝘣𝘦.𝘤𝘰𝘮/𝘸𝘢𝘵𝘤𝘩?𝘷=𝘴𝘢𝘭78𝘈𝘊𝘵𝘎𝘛𝘤 ) by Andrew Ng in which he talks about Agentic workflows. In Language Model (LLM) development, we provide a zero-shot prompt and receive an answer. We may review it and add more prompts to get it to a desired shape. This is exactly what the agentic workflow does. Basically, 𝘢 𝒃𝒖𝒏𝒄𝒉 𝒐𝒇 𝑨𝑰 𝒂𝒈𝒆𝒏𝒕𝒔 𝒘𝒐𝒓𝒌 𝒕𝒐𝒈𝒆𝒕𝒉𝒆𝒓 𝒊𝒏 𝒂𝒏 𝒊𝒕𝒆𝒓𝒂𝒕𝒊𝒗𝒆 𝒘𝒂𝒚 𝒕𝒐 𝒑𝒓𝒐𝒅𝒖𝒄𝒆 𝒂 𝒅𝒆𝒔𝒊𝒓𝒆𝒅 𝒐𝒖𝒕𝒑𝒖𝒕. Moreover, he even claims the output produced by agentic workflow is much better than the output produced single LLM which might sound obvious but what surprised me was, as per the “𝑯𝒖𝒎𝒂𝒏𝑬𝒗𝒂𝒍” coding benchmark, agents-based approach beating plain LLM by quite a margin (see the picture below). The plain 𝐆𝐏𝐓 3.5 was at 48% but 𝑳𝑫𝑩+𝑹𝒆𝒇𝒍𝒆𝒄𝒕𝒊𝒐𝒏 (𝘈 𝘓𝘢𝘳𝘨𝘦 𝘓𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘔𝘰𝘥𝘦𝘭 𝘋𝘦𝘣𝘶𝘨𝘨𝘦𝘳 𝘷𝘪𝘢 𝘝𝘦𝘳𝘪𝘧𝘺𝘪𝘯𝘨 𝘙𝘶𝘯𝘵𝘪𝘮𝘦 𝘌𝘹𝘦𝘤𝘶𝘵𝘪𝘰𝘯 𝘚𝘵𝘦𝘱-𝘣𝘺-𝘴𝘵𝘦𝘱) fared at impressive 95.1% with the same GPT 3.5. The plain 𝐆𝐏𝐓 4 was at 67% but 𝑨𝒈𝒆𝒏𝒕𝑪𝒐𝒅𝒆𝒓 (𝘔𝘶𝘭𝘵𝘪-𝘈𝘨𝘦𝘯𝘵-𝘣𝘢𝘴𝘦𝘥 𝘊𝘰𝘥𝘦 𝘎𝘦𝘯𝘦𝘳𝘢𝘵𝘪𝘰𝘯 𝘸𝘪𝘵𝘩 𝘐𝘵𝘦𝘳𝘢𝘵𝘪𝘷𝘦 𝘛𝘦𝘴𝘵𝘪𝘯𝘨 𝘢𝘯𝘥 𝘖𝘱𝘵𝘪𝘮𝘪𝘴𝘢𝘵𝘪𝘰𝘯) fared at whooping 96.3% with the same GPT 4.  You can get more information about the evaluation and the models at 𝘩𝘵𝘵𝘱𝘴://𝘱𝘢𝘱𝘦𝘳𝘴𝘸𝘪𝘵𝘩𝘤𝘰𝘥𝘦.𝘤𝘰𝘮/𝘴𝘰𝘵𝘢/𝘤𝘰𝘥𝘦-𝘨𝘦𝘯𝘦𝘳𝘢𝘵𝘪𝘰𝘯-𝘰𝘯-𝘩𝘶𝘮𝘢𝘯𝘦𝘷𝘢𝘭. He also mentioned the following are the 4 design patterns that are emerging in the agentic workflow. 1.    Reflection 2.    Tool use 3.    Planning 4.    Multi-agent collaboration If you are interested in learning more, instead of me explaining each of these patterns, I would encourage watching the video :) #ai #agent #agenticworkflow #gpt

22

Thank you Sigurd Meldgaard for the boost 🙌🏻 3 weeks ago Ive raised a suggestion request for pub.dev on LinkedIn and on GitHub too, luckily some of the Flutter Dev and Google community members liked my idea and within a week they implemented the idea on https://pub.dev website ✨ ⭐ What's new on pub.dev: Now you can copy the latest version of any of your package from https://pub.dev website *directly from the listing* 😋 You dont have to dive into the detail page 'just to copy' the latest version ⚡ Also thanks for the approval István Soós 🌟 and thanks Lukas Klingsbo for the Github issue reminder 🦺 🔗 Heres the reference of my previous post: https://lnkd.in/em9vHCaW 🔗 Github Issue reference: https://lnkd.in/eUcj5grc #flutter #google #dart #flutterdev #flutterdeveloper #opensource #pubdev #linkedin #github #flutterdeveloper #fluttercommunity #support #suggestion #enhancement #developer #hybriddeveloper #userexperience

59

4 Comments

There's been a lot of talk in the API industry lately about design-first versus code-first. Today Tom Akehurst from WireMock suggests a third option: prototype-first. "An API prototype, unlike a design document or specification, can be used immediately by its consumers. The client code is written, and app features are built early on," says Tom. Upfront design is usually a good idea for API development, but it's not always used in practice. So, could more dynamic prototyping allow the best of both worlds? What do you think? Check out his post on the Nordic APIs blog today for some thoughts about how to mock prototypes and the potential benefits of this approach. https://lnkd.in/eGqGeV7D

11

1 Comment