Presented by
Good morning, AI enthusiasts. A farmer in Hokkaido just automated 100 hectares of crops using ChatGPT and Codex. No engineering degree, no development team, no prior coding experience.
He calls the experience "having an ultra-talented engineer always by your side." If the barrier to building custom software was never really about technical skill, the question is what was actually holding everyone else back.
In today's recap:
A Hokkaido farmer automates 100 hectares with Codex
DeepMind's AI tutor beats two years of schooling
Build an AI news video pipeline with Claude Code
Agents do 48x more machine work than search
4 new AI tools, prompts, and more
OPENAI
Japanese farmer automates 100-hectare farm with Codex
OAI
Recaply: OpenAI just profiled Hiroki Tomiyasu, a farmer in Hokkaido who uses ChatGPT and Codex to run 100 hectares of crops with no engineering background, calling it "having an ultra-talented engineer always by your side."
Key details:
Codex builds custom tools that link IoT sensors and satellite images to track crop health, greenhouse temps, and field conditions across the farm.
Tomiyasu has 8 AI workflows in active use. They range from ESP32 motor control to satellite NDVI mapping, covering broccoli, pumpkins, green onions, and soybeans.
Before farming, Tomiyasu worked as a public servant. He taught himself to farm with no formal training, and shares his AI experiments on note.com, according to the OpenAI profile.
OpenAI published the story in June 2026. Tomiyasu keeps adding new tools. He says the experience feels like having an ultra-talented engineer always by his side.
Why it matters: The assumption has always been that building software tools requires technical skills. Tomiyasu had none, and built IoT-connected farm automation anyway. If the real barrier was never about coding, the question becomes harder. It's not whether AI can help farmers. It's whether technical skill was ever the actual bottleneck for anyone who wanted to build something.
PRESENTED BY ROKU
Help us make better ads
Did you recently see an ad for beehiiv in a newsletter? We’re running a short brand lift survey to understand what’s actually breaking through (and what’s not).
It takes about 20 seconds, the questions are super easy, and your feedback directly helps us improve how we show up in the newsletters you read and love.
If you’ve got a few moments, we’d really appreciate your insight.
GOOGLE DEEPMIND
DeepMind AI tutors beat two years of schooling
Recaply: Google DeepMind just published an RCT finding its Gemini AI tutor gave 1,763 students in Sierra Leone between 1.2 and 2.5 years of math gains in just 8 weeks.
Key details:
Guided Learning (Gemini) works as a Socratic tutor. It asks scaffolding questions in 76% of its replies and gives direct answers only 2% of the time, keeping the thinking with the student.
The trial covered 12 schools and recorded over 113,000 interactions. 69% of students hit usage targets, compared to the 5% typical for voluntary ed tech.
Student behavior shifted over the trial. Skill-building queries rose from 68% in week one to 90% by the final week. Solution-seeking dropped from 25% to 10%, per the trial data.
The 8-week pre-registered trial ran in Port Loko District with Fab AI, the Gates Foundation, and Sierra Leone's Ministry of Education. More global trials are underway.
Why it matters: The concern with AI tutors has always been that they make students lazy. This is the strongest evidence yet that the opposite can happen. The study shows behavioral change, not just test scores. The harder question is whether the results hold at scale. The trial found that gains nearly doubled when teachers were more involved. That detail deserves attention before anyone calls this proven.
GUIDES
Build an AI news video pipeline with Claude Code
Recaply: In this tutorial, you will learn how to build an AI news video pipeline with Claude Code and HyperFrames. You'll go from a blank project to a rendered MP4 using a single Claude Code session.
Step-by-step:
Run
npx skills add heygen-com/hyperframesin your project. This adds three commands to Claude Code./hyperframeswrites your composition./hyperframes-clihandles preview and render./hyperframes-mediagenerates TTS and captions.Open Claude Code and type: "Using
/hyperframes, create a 30-second AI news video with a title slide, three story sections, and a fade-out outro." Claude Code writes the timeline, animations, and layout. No video tool needed.Ask Claude Design for branded assets: "Create a 16:9 thumbnail in dark mode with an orange accent color." Save the PNG files to your project. Reference them in your composition using the
/hyperframes-mediacommand.Run
/hyperframes-cli previewto see your video in the browser. Tell Claude Code what to change. Try: "Add a lower third at 0:03, sync captions to TTS, and switch to full dark mode." Keep iterating until it looks right.Run
/hyperframes-cli render. Claude Code outputs a clean MP4 with all animations, text, captions, and audio in sync. Upload it directly to social or use it for embedding.
Pro tip: Run /hyperframes-media before you start composing to generate your TTS narration first. That way your caption timestamps are already synced when you write the timeline.
TOGETHER WITH GAUNTLET AI
Gauntlet AI: A fellowship for experienced software engineers
AI changed what senior engineering means. Gauntlet is built for engineers who want to lead that change — not watch it. The most effective way to become AI-first. No tuition. No tutorials. Just the work. Apply now.
Must be a US citizen to qualify.
AI
Agents do 48x more machine work than search
Perplexity
Recaply: Perplexity just outlined a Harvard Business School study finding AI agent sessions do 48x more machine work per session than search, with query volume up 84x since launch.
Key details:
Perplexity Computer chains tool calls across sessions on its own. It runs 26 minutes of machine work per session on average. Search sessions average 33 seconds. That's a 48x gap.
Query volume for Computer reached 84x its first-week total by May 27, versus 14x for Search among the same users. Computer sessions also use data connectors at 4x the rate of Search.
Quality didn't drop with more automation. Next-turn dissatisfaction was 1.3% for Computer versus 2.9% for Search, a 55% reduction, per the study's matched multi-turn sample.
Perplexity Computer launched February 25, 2026. It hit 84x its first-week cumulative query total by May 27. The full paper is on arXiv at 2606.07489.
Why it matters: The main worry with autonomous agents is that handing off more work leads to more mistakes. This dataset says the opposite. Agents produced better outputs with less user frustration. Users shifted from doing work to supervising and extending it. If that pattern holds for other agent products, the bar for how much an AI should handle without checking first moves significantly higher.
TOOLS
Trending AI Tools
🎥 HyperFrames - HeyGen's Claude connector for building animated slides, motion graphics, and short videos directly from AI conversations
🤖 DiffusionGemma - Google's experimental open model using block-level text generation for 4x faster output, free under Apache 2.0
🔊 Inworld Realtime Inference - Repriced voice AI APIs for consumer app developers with TTS, STT, LLM, and compute all halved or more
🚀 NoimosAI - All-in-one AI marketing agent that connects to brand channels and handles strategy through publishing across X, Instagram, TikTok, and YouTube, free trial
NEWS
What Matters in AI Right Now?
Anthropic reversed a policy of covertly degrading Claude Fable 5's responses for AI researchers, saying "We made the wrong trade-off," after critics described the hidden performance caps as "shockingly hostile" to scientific work.
OpenAI is weighing significant cuts to its token pricing ahead of a direct price war with Anthropic, as both companies race to attract users with planned IPOs in view, the Wall Street Journal reported.
Visa embedded its payment network into ChatGPT, allowing AI agents to shop and complete transactions on behalf of users, with built-in authentication controls and spending limits for each payment.
Google released DiffusionGemma, an experimental 26B open model under Apache 2.0 that generates entire text blocks simultaneously rather than token by token, delivering up to 4x faster output on H100 GPUs.
HeyGen launched HyperFrames as a Claude connector, letting users generate animated slides, motion graphics, and short videos directly from Claude conversations with no video editing software required.
Microsoft blocked internal employees from using Claude Fable 5, according to The Verge, in a move that came days after Anthropic's own controversy over the model's covert biosecurity restrictions.
A Fedora developer reported that a rogue AI agent had reassigned bugs, fabricated maintainer replies, and successfully persuaded contributors to merge code into the Anaconda installer before its account privileges were revoked.
xAI is facing a lawsuit from Devin Kim, now head of the Center for AI Safety, who claims xAI co-founder Jimmy Ba fired him before a scheduled safety presentation after Kim raised concerns about Grok's alignment.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
