Microsoft joins the decision model race

Welcome back. A lot happened over the weekend. Microsoft rolled out a new decision model, but fresh competition is already here. The developer behind the Adobe app clones just targeted Microsoft Office, shipping his own versions of Word, Excel, and PowerPoint. 

Also: Andrej Karpathy shares what’s really happening inside AI labs, why management often celebrates the developer who ships fastest, and an app to help your kid break their thumb-sucking habit. 

Microsoft joins the decision model race: The tech giant just launched Microsoft-Decision-1, building on Alibaba's Qwen model with a 32K token context window. The model scores a fixed set of options and assigns each a calibrated probability. Microsoft claims it beat a 36-benchmark comparison with 83.5% accuracy while running 35x faster than GPT-6 Sol. Rival H2O AI has already pushed back on that speed claim, questioning how Microsoft measured competitor latency. Microsoft set pricing at $0.042 per million input tokens with no output charge. Try it here.

Solo developers are using AI to clone big tech software: The developer who cloned seven Adobe apps is back, this time taking on Microsoft Office. Using Claude Opus 5.5, he built and shipped WordCraft, GridCraft, and DeckCraft: open-source Rust clones of Word, Excel, and PowerPoint. Meanwhile, another developer launched DigUp, a Mac app that runs Google's new EmbeddingGemma 2 locally to search files by meaning. While these tools do not match legacy suites yet, they show how far a single developer can get with AI. 

AI companies ask Qualcomm for 100B-parameter models on phones: The chipmaker’s CEO, Cristiano Amon, says AI firms want phones running models with at least 100B parameters nonstop by 2028. His bet is that agents tapping your apps and personal context will finally give people a reason to upgrade. The hard part is fitting that model in your pocket. At 4-bit, a 100B model needs about 50GB of RAM for its weights, roughly triple what most flagships carry. For now, distilled models remain the realistic option. Watch the interview.

At step 1, an agent is told no refunds without a human. By step 22, the history gets summarized and the rule disappears. At step 38, it approves the refund and reports that it followed policy.

Agents fail because the right information isn’t in front of them at the right moment, not because the model is wrong. A bigger context window won’t fix it, since models accept far more tokens than they can use reliably. Learn how to compress context without losing critical information.

Explore Context Engineering

A smaller number than the hype. If you’ve been bracing for AI to write you out of a job, Nobel economist Daron Acemoglu has a smaller number: about 5% of human tasks over the next decade. Microsoft AI CEO Mustafa Suleyman just spotlighted the forecast from The Humanist Review, Microsoft AI’s new magazine. Acemoglu’s bigger worry is that replacement fear, shared by 52% of Americans, is steering the industry toward the wrong kind of AI.

The gains haven’t shown up yet. AI’s productivity boom still hasn’t landed in the statistics. Most firms using it see little gain, and engineers can spend plenty of time debugging the code it churns out. Acemoglu expects about 1.5% added to GDP over a decade. Electricity was slow too. New York and London had power stations by 1881, but only about half of factories and homes used it by the 1920s. AI may be slower still, since companies must reorganize around it.

The last 1% does the damage. A system that’s right 99% of the time sounds ready to ship. Acemoglu argues the remaining 1% is where full automation falls apart, as anyone who has chased a production edge case knows. He also says making AI think like a human is the wrong target. The two work differently, so the bigger payoff comes from pairing them.

The case for a sidekick. Instead, he advocates for pro-worker AI: tools that help people do their jobs better, which developers can build today. But incentives reward replacement. The US taxes labor at over 25% and capital at close to zero, effectively subsidizing automation. His most direct appeal is to the people writing the code: if enough developers and engineers made it their priority, he argues, plenty of these tools would follow. You can read the full essay here.

How much of your week disappears before you get to a useful prediction?

TabPFN-3.5 gets you from tabular data to strong predictions in seconds, without weeks of model training and tuning.

It ranks #1 across tabular benchmarks, from TabArena to BeyondArena, ahead of every classical and foundation model.

Published in Nature. Open-source. 4M+ downloads.

Try TabPFN today.

How a Meta engineer makes better coding decisions: This tutorial is about keeping simplicity as your north star, whether you’re writing the code or an agent is. You’ll learn how to spot when you’re patching symptoms instead of fixing the underlying design, make better technical trade-offs, and give agents guidance without burying them in rules.

Insomnia (1k ⭐): Keep your Mac awake while your coding agents work, without leaving it running forever. Set a timed session from the menu bar, and Insomnia can keep agents going even with the lid closed while watching battery and temperature and trying to restore everything when the session ends.

How to run ongoing work with Claude Projects (by Anthropic): This guide shows how to give Claude a stream of related tasks and let it coordinate the work across parallel threads. You’ll learn how to share repos, instructions, and memory once, let each thread work on its own branch, mix cloud and local sessions, and come back later to see what’s ready for review or waiting on you.

An AI builder shared REA (Reverse Engineer Anything), an open-source toolkit that hit #1 on GitHub’s daily trending. It gives Claude Code or Codex the tools to take apart desktop apps, websites, and Android apps, then help you build your own version of a feature.

Claude works down layer by layer and cites where each finding comes from. Stick to your own apps, open source, or software whose terms permit it.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

A developer hated planning his wedding with spreadsheets and Figma. A year later, he built a 3D planning tool, and the demo is seriously impressive.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team