Watch the video, transcript and sources
- Android 17 Pixel libc bug debugged via disassembly
- AOSP Q2/Q4 schedule lags shipped QPR1 binaries
+ Gemini 4 Argon reaches paid API and AI UltraAndroid is open source, yet the public source code is months behind the binary on a Pixel phone. A developer had to disassemble a shipped system library just to report an emoji bug. Bionic is Android's core C library, the layer every native app relies on for memory, files, and text. Google publishes Android through the Android Open Source Project (AOSP), while also shipping developer AI tools like Gemini 4 Argon.
The bug surfaced in Visual Pinball on Android. After September's update on Pixel 7 and Pixel 8a phones (CP3A.260905.009), pinball tables failed to load whenever a VBScript comment contained an emoji. The culprit sits in two wide-character search functions in Bionic: wcscspn() and wcspbrk(). A program hands them a string and a list of delimiters, such as a newline, and asks how many characters appear before the first match. On September's Android 17 build, when the delimiter list has two or more characters below 0x100, the scan stops at the first character at or above U+0100. That breaks 1,111,808 Unicode code points, including emoji, Cyrillic, and Chinese characters.
When open-source code breaks, you normally read the repository. Developer francisdb checked June's public android17-release branch and found a clean llvm-libc loop that worked fine. The phone binary ran different code. So the developer pulled libc.so off the Pixel and disassembled the ARM64 instructions. To speed up multi-character searches, the binary builds a 256-byte lookup table on the stack. The loop compares each character against 0x100 (cmp w8, #0x100) and keeps scanning only while the character is lower. Any wider character exits the loop and reports a match, which works for keeping characters (wcsspn) and fails when rejecting them (wcscspn).
Google's Bionic maintainer replied within three hours and confirmed the bug. He was about to tell the reporter to read the source, until he saw the public branch lacked the April changes. Thirty-five minutes later, he posted a follow-up: Bionic did have a wide-character unit test in bionic/tests/wchar_test.cpp at line 1475 (wcscspn(L"he\ufffelo", L"lo")), and it asserted the buggy return value 2u instead of 3u. AOSP remains open source, and under Google's 2026 policy Google publishes Android source twice a year, in Q2 and Q4. Bionic's public android17-release branch still sits on a snapshot from March 28, 2026 while Pixel phones run September's quarterly release.
While outside developers wait on source drops, inside Google the code moves faster with AI. Google claims Gemini 4 Argon helped the Chrome team untangle a 6-million-line legacy folder inside Chromium's 40-million-line codebase, turning months into weeks. Google Cloud engineers also ran Argon for 11 hours overnight on their open-source Python Agent Development Kit (google/adk-python), spanning 205,000 lines. Argon proposed six optimizations across 32 files, cutting in-memory session overhead by about 40%. Human maintainers shipped two of those six proposals and rejected four in code review: one cache risked returning stale session data after garbage collection, and deleting top-level imports broke 25 internal Google tests until engineers rewrote them as lazy loaders.
Which brings us to the model itself. When Google announced Gemini 4 Argon on September 30, I stamped it NEEDS REVIEW because regular developers could not use it. On October 13, Google marked the gemini-4-argon endpoint generally available. DeepMind's landing page says Argon is available today for Google AI Pro and Ultra subscribers, while the AI Studio plans and API pricing docs state in a warning box that Argon is unavailable on the Free tier and on Google AI Pro. In AI Studio and Antigravity, you need AI Ultra or a paid API key. The introductory API price of $2 per million input tokens and $10 for output expires after November: on December 1, the price doubles to $4 in and $20 out. On the Artificial Analysis leaderboard, Argon (High) scores 53, tied with GPT-6 Astra and Claude Fable 5.1, behind Claude Opus 5.5 at 58, and averages roughly 44 seconds of thinking before the first answer token arrives. Google also showed Argon controlling a Boston Dynamics Spot robot to drop cans into a foot-pedal bin after practicing 3-leg balance inside a MuJoCo physics simulation, guided by a human-written skill and camera-verifier prompts.
Verdict: NEEDS REVIEW — Argon API is live ($2/$10 till Nov 30); shipped Pixel code still outruns public AOSP.
Sources
https://github.com/android/ndk/issues/2260
https://news.ycombinator.com/item?id=50083130
https://source.android.com/docs/whatsnew/site-updates?year=2026#aosp-changes
https://android.googlesource.com/platform/bionic/+log/refs/heads/android17-release
https://github.com/vpinball/libwinevbs/pull/41
https://ai.google.dev/gemini-api/docs/changelog
https://ai.google.dev/gemini-api/docs/pricing#gemini-4-argon
https://ai.google.dev/gemini-api/docs/google-ai-plans
https://deepmind.google/models/gemini/
https://deepmind.google/models/model-cards/gemini-4-argon/
https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon/
https://developers.googleblog.com/gemini-4-argon-for-robotics/
https://artificialanalysis.ai/models/gemini-4-argon
https://www.youtube.com/watch?v=Q_-jycDLV7I
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
YouTube · Newsletter · thedailydiff.dev · forward this to the intern who deployed on Friday.