|
AI Builders Digest
Friday, July 10, 2026
|
|
There's a quiet argument running through today's payload about what AI is supposed to replace and what it isn't. Replit's CEO thinks we're asking the wrong comparative question entirely. Google is asking users what Gemini breaks. And Microsoft Research shipped a tool that tries to make AI output look less like AI output. The thread connecting all three: the gap between what these systems promise and what they actually deliver is still very much open for business.
|
|
---
|
|
01
|
Google's Gemini lead is asking what his own product can't do
|
|
|
Josh Woodward, who leads Google's Gemini app, posted a public question this week asking users what Gemini should have fixed "a long time ago." The post pulled 1,586 replies, which suggests the frustration backlog is real.
|
Why it matters: When a product leader posts this publicly rather than mining support tickets internally, they're either genuinely crowdsourcing priorities or running a perception exercise. Either way, the replies are a free map of where Gemini users feel stuck. If you've hit a wall with Gemini on something specific, this is the moment your complaint might actually land on someone's roadmap.
|
|
Source →
|
|
---
|
|
02
|
Microsoft built a charting language so AI stops mangling your data visualizations
|
|
|
Microsoft Research released Flint, an open-source visualization language designed specifically for AI agents. The problem it solves is real: when an AI tries to generate a chart, it typically either produces something ugly with default settings or writes fragile, verbose code that breaks when the data changes. Flint sits in the middle, letting agents write compact specs that compile to Vega-Lite, ECharts, or Chart.js with proper color choices, scaling, and labels handled automatically. There's also a model context protocol server so agents can generate and render charts directly inside coding or chat environments.
|
Why it matters: If you're building any kind of AI-powered reporting or dashboard tool, this removes one of the more tedious failure modes. An agent that can reliably produce a readable chart from a data description without you hand-correcting the axis labels is a meaningful step up from what most teams are duct-taping together today.
|
|
Source →
|
|
---
|
|
03
|
Replit CEO: stop judging agents by the standard of hand-written code
|
|
|
Amjad Masad made a pointed comparison: nobody evaluates a compiler by asking whether it writes assembly as elegantly as a human expert would. So why do we benchmark autonomous agents against hand-written code? His implication is that the comparison itself is the wrong frame, and that agents should be judged on a different axis entirely.
|
Why it matters: This is a convenient argument for anyone shipping agentic products that produce messy output, but it's not entirely wrong. The real question is what the right benchmark actually is, and right now nobody has a good answer. If your team is evaluating whether to deploy an AI coding agent based on code quality reviews, you may be measuring the thing that matters least.
|
|
Source →
|
|
---
|
|
|
|
**Peter Steinberger demos agents asking for help mid-task** — Steinberger shared an example of agents using something called "nameplate" to surface clarifying questions to users when they need more input, rather than guessing and proceeding. The post is light on detail but the interaction pattern it shows, agents that pause and ask rather than barrel forward, is one of the more important open problems in agentic UX right now.
|
|
Source →
|
|
**Shawn Wang on slides that don't look like they were made by an AI** — Latent Space founder Shawn Wang noted that Theo's recent keynote used hand-drawn Excalidraw-style slides, a deliberate contrast to the "claudeslopped thin color border professional" look that's become the default aesthetic for AI-assisted decks. The observation is minor but the underlying point isn't: audiences are already developing a visual instinct for AI-generated polish, and it doesn't read as thoughtful.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|