I wrote my takeaways from LeadDev #LDX3 in this post. * Gains from AI are real but uneven across companies * Existing bottlenecks in the SDLC become blockers in AI-enabled orgs * Human relationships are still very important to build large successful products

#LDX3 The shadow culture: Why engineering principles fail under load by Manogna Machiraju We all want to build things fast with the right level of quality. The reality: the code is archaic, pipelines are broken, security is an afterthought. Engineering is always catching up.

#LDX3 What makes an effective EM in the AI era? Panel: Vernon Richards, Priscilla Nagashima, Alicia Collymore, Scott Carey

#LDX3 How will we deal with the new drudgery of AI-generated code? Panel: Lawrence Jones, Maude Lemaire, Randy Shoup, Birgitta Böckeler

#LDX3 AI in the trenches: Real-world wins without breaking things by Yiğit Darçın Deja Vu vs. Vuja De - the feeling that you have never seen it before. MIT study: the "GenAI Divide" - 95% of AI pilots fail to generate ROI, while a successful 5% utilize a "Vuja de" approach to rethink workflows.

#LDX3 AI killed the coding interviews. Here’s what Meta built instead by Danit Nativ Navon What broke? * Candidates used AI tools during interviews * Evaluations and ladders didn’t evolve * Leaders had no idea how to lead in this new context

#LDX3 Technology advances; history repeats itself by Zoe Cunningham A talk emphasizing the need to keep focusing on humans when everything is changing 🥰 The rate of change is increasing. The current message around the tech industry is impacting managers drastically: less managers, more IC work.

#LDX3 How to grow your engineers into great leaders? by Charles Duncan Jr. Your best engineer may not be your best managerial candidate The IC and manager jobs are different, require different skillset

#LDX3 Engineering at scale: Why developer experience is your competitive advantage by Dr Nicole Forsgren Teams "move" at 10x speed but the system moves at 1x The speed/friction mismatch * 15 minute builds * Manual approvals * Unclear deployment criteria * No documentation of decisions

#LDX3 What will AI-accelerated engineering teams actually look like? Panel with Luke Marsden, Hannah Foxwell, Yenny Cheung, Liz Fong-Jones What practices that predate AI help us move faster now? Documentation, comments, tests, small PRs Spec driven dev in all parts of SDLC: from PRD to tests

#LDX3 Building resilient engineering teams when failure is the default by King-Immanuel Edoh In Nigeria, infrastructure is highly unreliable, starting with the power grid. This forces good engineering practices to be the default Fantastic and lively talk by King-Immanuel.

#LDX3 What does a CTO even do? by Dee Kitchen 1. Do we have the right people? 2. Do we have the right technology? 3. Do we have the right product vision? 4. Do we have the right balance between innovation and liability 5. Do we have the right enablement and sales motion?

#LDX3 Up and down the management track: Equalising your leadership style across the pressure of scale by Karen Lee Lessons from Karen who had an went up and down the management track, starting as engineering, growing to EM, MoM and Director, back down to MoM, EM and back up, twice.

#LDX3 Moving Accessibility from debt to done at giffgaff by Abi Harrison-Nye Accessibility represents an 8T$ opportunity globally, that’s the buying power of people with any disability. Making the product right, accessible, from the start, does not cost more or add time to a project.

#LDX3 The mechanics of scaling: Why delivery slows as you grow and what do to about it by Maryia Tarpachova As the company grows, features keep being added to create more growth, and pace starts to slow down A deep dive revealed specific teams were impact more than others, mostly core teams

#LDX3 Quantifying tech debt to modernize critical systems by Ejber Ozkan Tech debt negatively impacts productivity Image from Tushar Sharma in Four Strategies for Managing Technical Debt Tech debt is net asset to build features for the business and becomes a liability when it reaches a threshold

#LDX3 30 to 70 PRs a day: How we managed to not wreck our systems by @lizthegrey.com Honeycomb followed the example set by Intercom to 2x R&D productivity rather than try to 10x (linked article) Honeycomb mandated AI usage from Aug 2025 and only saw marginal changes until Opus 4.6 in Feb 2026

I’m coming from Tokyo to attend LeadDev LDX3 in London. (1/2) I’m particularly interested in how organizations are transitioning into AI-native organization, as we are in the middle of that process at Mercari. Let’s meet to discuss these topics or if you want to learn about Japan’s tech scene.

I tried Claude Design to design a new feature in my app. It speeds up UI prototyping over vibe coding prototypes. The prototypes are interactive and it’s possible to run them side-by-side. The hand-off to Claude Code is janky but it’s quite new so it’ll improve. Check the post for details.

Oh, what happened to Google’s AI overview?

Has Google killed the cheap Gemini models? From Gemini 1.5 Flash to Gemini 3.1 Flash Lite, the cost increased between 5x and 20x depending on the language used, English or Japanese.

1/2 In the current times of AI, I really like this advice from Charity Majors from the article below: The best advice I can give anyone is: know your nature, and lean against it.

I released my first mobile app on iOS and Android. There is such a contrast between how fast building with AI is, and how slow the mobile app release process is. I wrote about the whole experience in this post.

Google is hot right now, back to back days announcing big things 😳 Sunday: Universal Commerce Protocol Monday: Apple picked Gemini for Siri

A very easy to understand review of the paper Too Big to Think: Capacity, Memorization, and Generalization in Pre-Trained Transformers,". Large models simply memorize the training data. Smaller models learn the rule because they don't have enough "brain space" to memorize every individual answer.

Ben Thompson argues that while AI is obviously in a bubble, it will create long term benefits in the form of investments in chip fabs in the U.S. and power plants. This is similar to how we got dark fiber and train tracks as infrastructure out of the internet and railway bubbles.

Quick intro to Claude Skills Simplicity is the point. The most significant limitation of MCP is token usage Almost everything an MCP can do, can also be handled by a CLI tool instead. LLMs know how to call cli-tool --help, which means you don’t have to spend many tokens describing how to use them

What happens to a driving Waymo when AWS us-east-1 goes down? 😅 Is everything processed locally? It’s an Alphabet company, so if something is remote, I’d expect it to be on GCP or Google internal platforms, but with external dependencies, isn’t every product 6 degrees of separation from us-east-1?

Okay, for the folks who asked: here's the majority AI view, writing up the reasonable, thoughtful view on AI that the vast majority of people in tech hold, that gets overshadowed by the bluster and hype of the tycoons trying to shill their nonsense. anildash.com/2025/10/17/t... Please share!

This is the first article I read about spec-driven development and this is a good introduction. I thought spec meant PRD, but it’s more of a detailed technical design. As a profession, we were not good at documentation before AI, so this will not suit many orgs. And it seems overkill in most cases.

I wrote an article on how we translate user-generated content at scale at Mercari. It summarizes two years of work with various models, LLMs but not only, and how cost evolved, reducing it 100x over the period. I also go over non-AI features and user experience.

Our engineering team took on the massive challenge of rebuilding our marketplace from scratch to make it scale globally 🌏, and we just launched it. 🚀 Check out the first article in our blog series to learn how we did it!

The best description of Product vs Platform engineering 😂 Like the tail of an enthusiastic puppy: The product work might go a thousand directions in a minute. They're at the end of the puppy's tail. Infrastructure is the part of the tail anchored to its butt, not changing direction nearly as fast

Great article explaining what atproto is about in non-technical terms. Excerpts: We are at a similar juncture with social apps as we have been with open source thirty five years ago […] I like to call it “open social” What open source did for code, open social does for data.

A good thread on retaining top engineers, how it sometimes means encouraging them to take on roles outside your org I’d rather keep top talent in the company by sharing opportunities outside my org, rather than keeping them locked and they leave later out of boredom or lack of growth opportunities

Maybe social media needs circuit breakers like when the stock exchange crashes. If the mod team can’t remove hateful content fast enough, block the entire system: no posting, maybe even no viewing. Stop the snowball effect and let the mod team catch up. Put a SLO on the number of unack’ed reports.

I've been reading this on the Month of AI bugs from @wuzzi23.bsky.social and I am losing my fucking mind. The series is excellent, and Johann does great work, but SWEET FUCK people, you cannot be serious here. This is absolute clownshoes territory embracethered.com/blog/posts/2...

Nice real world application of AI agents and lessons learned, full of interesting advices. The analogy to Map Reduce is clever. Keep it simple with two layers and no state in the sub-agents. Caching is smart. It attempts to make non-deterministic LLMs into agents as deterministic as possible.

I love how one can share a bug on Bluesky and the main maintainer not only answers but also promises to fix it.

Great new feature in VS Code, checkpoints in chat sessions with Copilot. We finally can chat with the tool without having to commit between every request. If the next prompts mess up, just revert to the previous stable state.

Great summary by @simonwillison.net of @wuzzi23.bsky.social ‘s findings on AI tools vulnerabilities. In short, all AI tools are vulnerable if one attaches external files and links to their prompts, leading to secrets leaks and remote code execution. Johann publishes daily until the end of the month.

LinkedIn is broken. All your data is trapped behind a walled garden, job applications never seem to go anywhere, and the whole site increasingly seems to be an ad to try to upsell you LinkedIn Premium. We're here to help. For Pros on Proto to connect Pro to Pro!

I love how @bsky.app gets built in public, to the point that core contributors use the app to post bugs and tag colleagues 😄

The first job of a leader is to make the business succeed. This is a common point of these three articles from 2011, 2024 and 2025 - a16z: Peacetime CEO / Wartime CEO - Charity Majors: Pragamatism, neutrality and leadership - Lara Hogan: Balancing direction and empowerment

I laughed so hard I thought I was going to cough up an organ. Every line is gold.

On the impact of AI generated code. From the article: The code that we rely on for bank transactions, package deliveries, medical results, satellite launches, airline flight paths, self-driving cars, mortgage payments, and nuclear power plants. This is durable code, and it’s going to stay that way.

💯 It’s always a trade-off. The Career Tug-of-War: To do meaningful work with people you respect and admire. To lead with integrity. Part of you wants to climb.

Great reading that provides a clear structure. Amongst them, I believe Avoid “us vs. them” language is pretty important to highlight. The negativity can really make things worse.

It’s awesome, we do that at Mercari and there are ~2000 employees. It’s low noise channels, great to get context from areas of the company I don’t work with regularly, or catch news that I missed in the 100s of important but noisy channels. And it’s great to get to know your colleagues personally.

I love documentation. I’ve constantly contributed a lot wherever I was. It’s been a struggle trying to convince many of its usefulness until recently. Humans could handle ambiguity. AI is not as tolerant. Looks lime AI has convinced many that it is useful after all 😄