🪗 Joey de Villa

So which models produce HTML that has the least accessibility errors, and at what cost? Mapped it out here (bottom left is the best). This is helpful for understanding which models are most compliant for those with disabilities and do it efficiently. Qwen and Gemini 2.5 Flash lead the pack. Hat tip to Ben Ogilvie for pointing me to this: aimac.ai

55

2 Comments

What is the important change when prompting ChatGPT5... You need to ask for a checklist in your prompt 🤷‍♀️ I have been intensively working with the new ChatGPT5 Prompt Optimizer. This is the tool that OpenAI provides on their developer platform to automate writing of sophisticated prompt Every prompt that this tool generates starts with the sentence "Begin with a concise checklist (3-7 bullets) outlining the steps to create...." Why is this happening??? This seems to align with the increasing emphasis on context engineering instead of prompt engineering. Context engineering is about making sure that sufficient context is generated at the start of executing the task. By asking to create a checklist, you are actually asking chatgpt to first create context (steps in a checklist) and subsequently execute these steps

49

10 Comments

Roo Code lets you mold your #AI coding agent into what you actually need. Brandon Mathis from This Dot Labs built custom modes for documentation, pair programming, writing specs, and improving his prompts. If you're serious about agent workflows (which you have to be), this is worth a read.

13

4 Comments

Turning my inbox into a writing surface cut my capture friction to near zero. I send an email to a dedicated folder and it becomes a post; if I reply to fix a typo, the page refreshes. The system leans on existing infrastructure: IMAP IDLE for real‑time publishing, read‑only access to one folder, content sanitization, and a simple heartbeat for monitoring. Subject maps to title, Message‑ID to permalink, Date to published time, and attachments render as images. The result is a stable, boring workflow that makes thinking in public feel as easy as “send.” See links in the comments How would you simplify your publishing stack?

13

1 Comment

New blog: "Claude Tested Everything Except the One Thing That Mattered" I've been building a social app with Claude Code for the past month. Claude wrote 154 end-to-end tests across 17 spec files — tournaments, song battles, badges, tour crews, setlist search, a getting-started tutorial. Zero tests for posting. The one thing the app actually does. Then a refactor broke the posting flow, in the one place with no coverage. 202 out of 833 commits are fixes. That's 24%. The post is about what happens when an AI agent is excellent at writing tests but doesn't write them where they matter most (but writes them where they are *easiest*) — and what it looks like when it guesses at bugs instead of proving them first. https://lnkd.in/etVQffRt

49

4 Comments

🧠 For decades, we relied on relational rows and NoSQL documents. They’re battle-proven, but they weren't built for a world where meaning is more important than exact matches 🚀 In our latest blog, Alvin Richards dives into why vector databases aren't just an "add-on"—they are a fundamental shift in how we build for the AI era ❌ No more rigid schemas ✅ Yes to semantic relationships 🏗️ Built for the future, not your grandparents' era

58

8 Comments

Agentic systems are great in demos. But in production, they might confidently do the wrong thing—fast. So I wrote a short piece on Architectural Patterns for Agent Workflows and Multi-Agent Systems. (or: How to stop your agents from arguing like a bad stand-up meeting). BONUS: "The office" references! Here’s the gist 👇 • Single-Agent Pipeline: One agent does everything (Also ships one hallucination end-to-end). • Planner–Executor–Critic: Separate thinking, doing, and judging. There’s a reason software teams evolved this way. • Tool-Orchestrated Agents: Agents reason. Tools execute. LLMs are not databases, schedulers, or payment systems. • Router + Specialists: Generalists route work. Specialists finish it correctly. • Blackboard Architecture: Shared state works—until nobody owns writes and it becomes a graffiti wall. • Hierarchical Agents: Flat systems scale conversations. Hierarchy scales outcomes. • Event-Driven Agents: React to signals. Don’t poll the void. • Human-in-the-Loop: Autonomy without accountability is just automated confidence. Final rule: Agents scale reasoning. Architecture scales trust. If your system relies on “the model will figure it out,” you don’t have an architecture—you have hope with tokens. Would love to hear how others are structuring agent systems before production finds the edge cases. #AgenticAi #Ai #agents #aiWorkflow #SWE #intern #architechture #Interview #InterviewPreparation #resources

19

1 Comment

Don't Measure AI-codec Quality with Traditional VQ Metrics In new subjective testing, AI-based codecs like JPEG AI and Deep Render consistently outperformed traditional codecs, but legacy metrics didn't see it. Even a retrained VMAF couldn't match the ranking accuracy of a purpose-built model like MLCVQA. Why this matters: Whether you're building playback infrastructure, selling a codec, or optimizing a ladder, you need tools that reflect actual user experience. This article outlines the tradeoffs, the current options, and what to do when "better than nothing" isn’t good enough. Read more 👉 https://lttr.ai/AemZX

28

3 Comments

VAMP = Winter is here! It's been really quiet so no major intel to share. Processors are starting to issue VAMP fines on a select basis. What I see happening is going to be problematic IMO. I've had a couple of MID's get hit with modest VAMP fines. Less than $1K for clients in the $100M/yr range. The VAMP fines caused the merchant's effective processing rate to spike and this led them to reach out to me. The client VAMP ratio is well below the 2.20% and 1,500 count thresholds. So they are technically compliant and shouldn't be fined per TC40. Processor A charged $4 and Processor B charge $8 per TC40 which showed up on their monthly statement. This means that BIN A is above the 0.50% VAMP threshold and BIN B is excessive with an above 0.70% VAMP ratio. The processors are not VAMP compliant but the merchant is. So who eats the VAMP fines? If the merchant is not excessive, 100% of fines should be paid by the BIN owner. The bank or processor made the choice to accept high risk merchants, was not proactive in terminating the bad actors but now they want the merchants to pay for those poor decisions. Why is this problematic? First, if every decent merchant gets hit with VAMP fines, it'll be a customer service nightmare for everyone. Second, merchants won't put up with that. They'll cancel their MID and simply shift their volume elsewhere. Lastly, the processor will lose decent merchant volume and get even more concentrated with the remaining bad merchants. It'll be interesting when things tighten further up in January...

28

6 Comments

Anthropic just blocked third-party use of Claude Code subscriptions. Sound familiar? In 2023, Apollo did the exact same thing. Shifted their license from open source to a restrictive source-available model. Pulled the rug on their community. That decision created WunderGraph's opportunity. After years of working with enterprises like eBay, Paramount, FanDuel, and others, here's what I've learned: Open source will always win in infrastructure. Why? Enterprises are complex. They want control. They want customization. They want to audit the code running their critical systems. With WunderGraph Cosmo being fully open source, our customers can: - Audit every line of code - Extend the platform for their specific needs - Contribute back to the codebase - Run it anywhere they want No license changes. No vendor lock-in surprises. No "we changed the terms" emails. Every time a company restricts their license or usage, they hand their competitors a gift. Apollo handed us one in 2023. Anthropic just handed someone else one today. Our goal is simple: if you're running GraphQL Federation, you're runnin WunderGraph. Open source makes that possible.

33

4 Comments

I’m really impressed by Webflow’s latest enhancement: native support for MP3, OGG, and ZIP files in both the CMS and Assets panel. Now I can upload audio files—like podcasts, sound effects, or voiceovers—and ZIP archives (think fonts, code snippets, or project assets) directly into Webflow, without relying on external hosts or workarounds. Each audio file can be up to 10 MB, and ZIP files up to 25 MB. Imagine using these to deliver richer client previews—audio walkthroughs, downloadable resources, clean packaging—all within your site build. For solo creators like me, this streamlines the process in two big ways: Organization – project assets and audio stay where they belong, embedded in the project. Experience – your site can go beyond visuals and text, embracing multimedia storytelling more cleanly. I’m already thinking of embedding audio intros in portfolio projects and bundling client deliverables in one neat ZIP. Have you started using the new file types yet? What’s your creative use-case? #Webflow #NoCode #WebflowUpdate #AudioSupport #ZIPFiles #CMS #CreativeWorkflow #SoloCreator

5