The AI Friday Brief
Good Monday, NOLA. Today has a satisfyingly builder-shaped mix: a delightfully weird frog-SVG benchmark, a practical look at AI carrying bugs through a COBOL-to-Java migration, and a startup called June trying to make organizational AI adoption less painful. There’s also a useful reminder that the most compelling AI stories are often about judgment: where to trust an agent, how to evaluate its output, and what work gets easier once the first draft is cheap.
Things People Are Trying
A tiny benchmark for the visual taste of AI models
Ask a model to generate an SVG frog with a Habsburg jaw, then compare the results. It is intentionally silly, but it makes a real point: compact, memorable tests can reveal whether a tool follows a creative brief better than a leaderboard can. The Hacker News discussion has plenty of variations to steal for your own model bake-offs.
Hacker News
Claude Code tries rewriting a production app
A firsthand account of using Claude Code to rewrite the Claude app in Swift. The interesting takeaway is not “press button, receive app”; it is how much leverage a coding agent can provide when a team gives it a bounded, reviewable job. HN’s discussion is a useful companion for the inevitable tradeoffs.
Hacker News
AI migrated COBOL to Java—and carried bugs along for the ride
This study is a useful reality check for anyone eyeing AI-assisted modernization: translation can preserve existing behavior, including the behavior nobody meant to keep. The practical move is to treat generated migrations as a fast starting point, then validate them against representative business cases before declaring victory. Discussion on HN.
Hacker News
The productivity gap is usually a workflow problem
A thoughtful argument that AI’s value does not automatically appear because a team bought licenses. The gains show up when people redesign the handoffs, review habits, and repeatable tasks around the tool—good fuel for a team retro or an AI Friday experiment this week. HN discussion.
Hacker News
Tools, Models & Product Moves
June wants to make AI adoption easier to ship
June emerged from stealth with a $20 million pre-seed round and a pitch aimed squarely at the awkward middle between “we tried a chatbot” and “this is how the company works now.” Worth watching if you have seen promising internal AI pilots stall at the rollout stage.
TechCrunch
Perplexity offers a remote connection for AI tools
The Neuron flags Perplexity’s remote MCP server—a standard connection for AI tools. In plain English: it aims to let an agent use Perplexity’s research capabilities without every developer building a custom integration from scratch.
The Neuron
Kimi K3’s weights are now open
Moonshot has released Kimi K3’s open weights, giving builders another model they can evaluate and adapt outside a single hosted product. Open releases matter most when they expand your choices for experiments, privacy-sensitive prototypes, or model comparisons—not when they become another tab you never open.
The Neuron
Gemini opens video generation more broadly
The Neuron reports that Gemini’s video capabilities are opening to everyone. For builders, broader access means it is becoming reasonable to prototype video explainers, social assets, and rough storyboards inside the same AI workspace where the script starts.
TAAFT - There's An AI For That
Wispr Flow turns talking into ready-to-send text
Wispr Flow’s pitch is straightforward and attractive: speak naturally, then get text that is close to ready for an email, message, or draft. Voice input is one of those small workflow upgrades that can compound quickly if typing is the bottleneck between you and a useful first draft.
TAAFT - There's An AI For That
Interesting Work Beyond the Chat Window
AI helps researchers study wild primates at scale
Researchers are using AI to make observations of wild primates more tractable, opening a new window into cognition outside the lab. It is a lovely reminder that the best AI applications are not always another assistant—they can make previously impractical research possible. HN discussion.
Emory University
A conversation about an agent designed to improve itself
Behind the Craft talks with Hermes co-founder Karan Malhotra about building an agent that improves its own performance. Worth a listen for builders thinking beyond one-off prompts toward systems that learn from feedback and iteration.
Behind the Craft
Can royalties give artists a workable AI bargain?
The Verge examines whether paying artists can create a more workable relationship between creative professionals and generative tools. For anyone building with media models, the product question is concrete: can clear attribution and compensation become a feature people actually trust?
The Verge
Also
- condense-json 1.0 — Simon Willison releases a tidy utility for making large JSON easier to inspect.
- DeepSeek V4-Flash follow-up — A follow-up to Saturday’s release coverage.
- Gemini Robotics 2 — A look at better robot manipulation and coordination.
- OpenAI API price drop — Worth checking against the economics of any API-heavy prototype.
- The Ohio State Fair’s AI poster winner — A small but vivid sign of how fast AI-assisted work is reaching mainstream creative contests.
- HN discussion: the frog benchmark — More benchmark ideas and model outputs.
- HN discussion: COBOL-to-Java migration — A useful debate about verification and legacy behavior.
- A $100 team-credit offer — If it fits your stack, free credits are a good excuse for a small experiment.
- The new home for real AI workflows — The Rundown highlights a community-submitted workflow.
- A Fender conversation about AI and music — A prompt for thinking about what creative tools should augment, rather than replace.
Pass it on