"AI roundup" is a curated collection of links I've gathered in the past related all aspects of AI, including LLMs, generative AI, etc.. I've added a brief summary (mostly for myself) next to the links so I have some memory what was the reason I collected the link in the first place. Consider these roundups as public bookmarks.
Summary:
https://huyenchip.com/2025/01/16/ai-engineering-pitfalls.html
Matt Pocock has a tutorial for MCP to get started.
https://www.aihero.dev/model-context-protocol-tutorial
Connect Vercel's AI SDK to model running locally, or anywhere.
https://www.aihero.dev/use-local-models-with-vercel-ai-sdk
When using AI consider its four "superpowers": content creation, summarization, basic data analysis, and perspective taking.
As you approach the use of AI in your product, don’t race to adopt a new technology for its own sake. Instead, you should aim to craft solutions that help users achieve their goals effectively and efficiently. The most innovative products of the future will not just be powered by AI, they will be designed with a deep understanding of human needs, with AI as one part of the equation. By combining clear intention, thoughtful design, and the right application of AI’s strengths, your team can create experiences that are not only functional but also transformative.
https://www.nngroup.com/articles/ai-superpowers/
When designing an AI feature, its scope (how broad or narrow its capabilities are) influences its usability. Our research shows that narrower AI features are (typically) easier for new users to understand and adopt.
https://www.nngroup.com/articles/scope-ai-features/
UX ranks among top fields adopting AI, mostly in writing, design, and coding tasks — though complex or human-centric UX activities remain largely AI-free.
https://www.nngroup.com/articles/ux-ai-adoption/
System-generated suggestions for AI prompts must be contextually relevant, personalized, and specific to the task and the user’s level of experience.
https://www.nngroup.com/articles/prompt-suggestions/
This one is a bit more philosophical.
While AI excels at both splitting ideas and bringing them back together, the final convergence—that last mile of refinement—remains distinctly human.
https://www.unknownarts.co/p/creativity-flows-like-light
Checklist for AI tool evaluation for design:
https://uxmag.com/articles/is-your-team-ready-for-ai-enhanced-design
Typescript agent framework from the team behind Gatsby.
In context of building AI agents focus on: context management, accumulated "knowledge" files and reliability. Do periodic e2e evals and cut every feature that is not essential.
https://jamesgrugett.com/p/what-i-learned-building-an-ai-coding
Agents are overhyped and overused. Most cases needs simpler patterns. Agents excel in human-in-the-loop scenarios. Build with observability and explicit control.
https://decodingml.substack.com/p/stop-building-ai-agents
Prompt-driven interfaces can overwhelm users. Designers should ease input with templates and suggestions. Since AI outputs can be opaque or incorrect, it's vital to show reasoning, confidence levels, and sources to build trust. Transparency – like memory controls and feedback channels – helps users feel in control. AI tools must fit into workflows, while addressing bias and keeping responses fast with smart loading strategies.
https://www.uxness.in/2025/06/challenges-in-thinking-uiai.html
This one is a take for "AI native employee" – I am honestly a bit lost with the core of the idea.
No headcount asks. No project briefs. No handoffs. Just action.
Plain action, no strategy? It feels the rationale is centered around powerful AI tooling and implied efficiency benefits. It's like having access to Ferrari and thinking that would be good for setting up logistics operation to transport cargo.
Don't get me wrong, I am all about finding the essence and cutting of the fat out of any process, but there is a reason these activities exists. Removing them does not resolve needs behind them.
The comment section has interesting take:
Today I saw this at a YouTube video of young team working by slack with cursor and MCPs… for changes and updates of their site without even the engineer being there (good practice or not, they are moving and improving).
I think that doesn't sound like a good practice and not all "movement" is good movement.
https://www.elenaverna.com/p/the-rise-of-the-ai-native-employee
Achieving product-market fit and practicing user-centered design become difficult when AI is presented as a product’s main value.
We should think about how to deliver product value only after firmly establishing what that value is going to be. This is because the experience we want to create determines which technologies and form factors are most suitable.
https://www.nngroup.com/articles/powered-by-ai-is-not-a-value-proposition
I'm not sure I'm fully on board with the sentiment, but the general idea is intriguing.
Don’t start by copying an existing product. Start by asking what should exist if you had no constraints. Then use AI to make that possible. The biggest winners are not adding AI to software. They are replacing software entirely.
https://www.productmarketfit.tech/p/why-most-startups-are-building-ai
For all the time we spend talking about GenAI generating code, we spend very little time talking about GenAI generating and executing tests. I believe they are so much better at the latter than the former.
https://carloarg02.medium.com/why-im-betting-on-llms-for-ui-testing-ac44e30e14c1
VC incentive is to make every investor out there to believe that this is the biggest opportunity in history. You are either in or out. And you do not want to be the guy who missed AI.
https://aimode.substack.com/p/the-ai-replaces-services-myth
OpenAI has released an agent.
ChatGPT now thinks and acts, proactively choosing from a toolbox of agentic skills to complete tasks for you using its own computer.
https://openai.com/index/introducing-chatgpt-agent
Language modeling is fundamentally compression. Better models compress data more effectively. This one goes academic.
https://blog.jxmo.io/p/there-is-only-one-model
More prototypes are now part of Shopify's product-building process — specifically, increasing the ratio of prototyped attempts to build attempts. This enables one of Shopify’s principles, the “green path of product-building,” where the only way to figure out a problem is by trying many things. … “There are an infinite number of bad solutions and probably 10,000 good ones. Your job is to find the best solution among the 10,000. What you just showed me is the first one that worked vs. the best one. Why did you stop there?”
https://www.firstround.com/ai/shopify
Interesting read about author's short journey at OpenAI.
https://calv.info/openai-reflections
Designers think through the creative act of doing.They shape ideas by moving materials, trying options, and reacting to feedback in real-time. The act of creation is not just the moment you press “render” – it’s the whole messy, intuitive, back-and-forth path to getting there.
Practical AI must be built inside that act. Inside the creative flow, and enhance the flow with practical heavy lifting. It must live inside the creative loop — not stand off to the side with a clipboard.
That’s why the right metaphor isn’t, ‘’the future is a robot assistant.” It's an exoskeleton.Extra arms. Extra speed. Better reach. But still your hands. Still your craft. Still your vision is in control.
Decision fatigue kicks in with the overload of choices. How many of these things we need? When are we expected to try these things out AND do actual work AND to effectively incorporate AI into our workflows? Maybe the best strategy is to sit down during this phase and be in the early-majority, once the dust settles.
Kiro helps you do your best work by bringing structure to AI coding with spec-driven development.