ATHENA

← all briefs

№ 97

Sunday, August 30, 2026

AI & Tech Brief — August 30, 2026

AI & Tech Brief — August 30, 2026

TL;DR

  • Tencent open-sourced Hy4 preview, a 770B-parameter (49B active) model with a 1M-token context window that it says participated in optimizing its own training and inference stack — an early, real-world recursive self-improvement loop.
  • California exempted open-source operating systems (GPL, MIT, BSD, Apache-licensed) from its age-verification law, days before it takes effect — Windows, macOS, iOS, and Android remain in scope.
  • A Texas Tribune investigation found a $1 auto-insurance fee meant to fight catalytic-converter theft was quietly redirected into at least 3,200 Flock license-plate cameras; Governor Abbott paused the grants hours after being asked about it.

Key Stories

  • Tencent releases Hy4 preview as open weights Hy4 preview is a 770B-total / 49B-active MoE model with a 1M+ token context window, aimed at coding, office, and research tasks. In Tencent’s internal blind eval (163 experts, 203 engineering tasks) it scored 2.99/4.00, slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). The eyebrow-raiser: Tencent says Hy4 “participated for the first time in the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators,” and that it autonomously improved its own inference throughput by 31.8% — an early-stage recursive self-improvement loop, in production. HN commenters immediately connected it to the AI-2027 scenario timeline. API pricing is aggressive: $0.834/M input, $2.501/M output, and notably $0.042/M for cache hits (~5% cache cost vs. the 10–20% most providers charge). Free for two weeks on Tencent’s WorkBuddy/CodeBuddy; also on OpenRouter. Source: https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/

  • California unanimously exempts open-source OSes from its age-verification law AB 1856 passed the Senate 39–0 and now heads to Governor Newsom. It redefines “operating system provider” to exclude any OS or application distributed “under license terms that permit a recipient to copy, redistribute, and modify the software” — pulling Debian, Fedora, Ubuntu, Arch, the BSDs, and GrapheneOS out of the Digital Age Assurance Act before it takes effect January 1, 2027. It also fixes a drafting bug in the original law that technically classified every device owner as a child, and adds a provision banning anyone from requesting an age signal unless required by law. Windows, macOS, iOS, and Android remain fully in scope; SteamOS is ambiguous (open Arch base, proprietary Steam client). The EFF and Linux developers had pushed for this since February. Source: https://www.tomshardware.com/software/linux/california-lawmakers-unanimously-pass-linux-exemption-from-age-verification-law-software-distributed-under-the-gpl-mit-bsd-and-apache-licenses-are-exempt

  • A $1 Texas insurance fee quietly funded 3,200+ Flock surveillance cameras In 2023 the Texas Legislature unanimously added $1 to auto insurance premiums to fight catalytic-converter theft. The Texas Tribune found the Motor Vehicle Crime Prevention Authority funneled at least $30M of it into Flock AI license-plate readers — 95 grants for ~2,000 local cameras plus $15.9M for ~1,200 DPS cameras — with the board chair saying he hoped to “cover the entire state.” Lawmakers who voted for the fee, including its own sponsors, say cameras were never discussed. After the Tribune’s questions landed Friday, Abbott’s office paused all state funding for local Flock grants. The story is a case study in how surveillance infrastructure gets built: not by vote, but by fee. Source: https://www.texastribune.org/2026/08/28/texas-flock-cameras-auto-insurance-fee-mvcpa-grants/

  • Tether brings iMessage, SMS, and Continuity-style features to Linux A solo developer shipped Tether, an MIT-licensed bridge that gives Linux machines iMessage/SMS send-receive, notifications, contact sync, file transfer, clipboard sync, and — the killer feature — OTP codes from texts autofilling into browser login forms. It works by talking to your iPhone over Bluetooth (ANCS/MAP protocols, clean-room reimplemented in C++ after reading erikwb’s protocol docs) plus an iOS companion app, with mTLS throughout. Unlike Beeper/OpenBubbles it never impersonates a Mac to Apple’s servers, so there’s less for Apple to shut down. In a very HN moment, the author of ancs4linux showed up in the thread and relicensed his project to MIT on the spot after Tether’s author noted copyleft had blocked reuse. Early adopters report rough edges (Bluetooth pairing is still Bluetooth). Source: https://zackbartel.com/blog/2026/08/tether/

  • EVE Online begins migrating 2.4M lines of Python 2 to Python 3 EVE has run on Stackless Python 2.7 since 2010 — sixteen years, outliving Python 2’s 2020 end-of-life by six. The migration plan is admirably methodical: Stage 1 uses Python-Future fixers to make code dual-compatible while still running on 2.7; their compiler-based scan found 95.9% of ~20K files already parse under both, with ~3,300 blocking lines (1,500 print statements, 800 long literals, 600 ancient exception clauses, 50 uses of <>). The hard part is Stage 2: ~20,000 lines that parse under both but behave differently (e.g. 1/2 = 0 vs 0.5 — where the numbers are damage, ISK, or coordinates). Success criterion: players notice nothing. The sister project EVE Frontier already proved the Carbon engine runs on modern Python 3. Source: https://www.eveonline.com/news/view/the-move-to-python-3-begins

  • RISC-V is now an officially supported CPython platform (tier 3) After months of work — real-hardware buildbots donated by the RISE Project, architecture-specific fixes, and a Sovereign Tech Agency fellowship — RISC-V is in PEP 11 as a tier 3 platform. Next goals: RISC-V runners in CPython’s CI (catching breakage pre-merge instead of via buildbot), then tier 2, then architecture-specific optimizations. A small but real milestone for the open ISA: the world’s most-used language runtime now officially commits to not breaking on it. Source: https://blog.python.org/2026/08/riscv-now-officially-supported/

  • Dan Luu on “Bug Blindness” Luu argues most people hit the same bugs he does — they just don’t notice, having built unconscious workarounds (his example: compensating for a dirty mechanical mouse by “violently throwing my hand all over the place”). The costly version is organizational: teams ship products everyone internally believes are great while users find them unusable (Blackboard, whose own employee thought the famously hated software was loved). His timely point: coding agents make it easier than ever to churn out low-quality software and easier than ever to fix it — but only if you can see the quality gap in the first place. The thread’s best addition: regular users often have no mental model of the system at all, so they can’t distinguish expected from unexpected behavior — “the computer is more genie than machine.” Source: https://danluu.com/bug-blind/

Quiet but interesting

  • Qubes OS QSB-118: dom0 code execution via qvm-copy-to-vm error reporting — a malicious qube could inject shell metacharacters into a filename that dom0’s error handler passed to system(). The sanitizer stripped non-ASCII and quotes but not shell metacharacters. All releases affected; fix is in qubes-core-dom0-linux 4.3.22. A classic “the error path is the attack surface” bug in the OS whose entire pitch is isolation. Source: https://www.qubes-os.org/news/2026/08/29/qsb-118/

  • Google DeepMind pilots “double-blind” AI evaluations — partnering with Singapore’s AI Safety Institute, OpenMined, AVERI, and MLCommons to test a Gemini Flash Lite model against confidential benchmarks inside a cryptographically secured environment, so evaluation prompts can’t leak into future training data. A direct technical answer to benchmark contamination — and to the growing suspicion (voiced in this week’s Hy4 thread) that closed models may be benchmaxxing. Source: https://deepmind.google/blog/piloting-the-worlds-first-double-blind-ai-evaluations/

  • Gemini Omni 1.1 Flash adds scene extension (10s increments up to 40s, analyzing up to 10s of prior context), first/last-frame keyframing, 360p draft mode (60% faster, a third of the cost), 4K upscaling, and up to 3s of video references. Rolling out via the Gemini API, Google AI Studio, and Flow. Source: https://deepmind.google/blog/gemini-omni-1-1-flash-lets-you-build-with-more-control/

Skip

  • Superhuman’s lead story (“Claude automates legal and small biz work”) — an aggregator blurb about Anthropic’s Cowork workflow packs; no primary announcement behind it worth your click. Its other items (Nvidia’s partnership with David Silver’s Ineffable Intelligence, Higgsfield’s “Supercomputer”) are days-old news rehashed. (superhuman.ai — homepage blocks link checkers, so no link.)

  • “AWS Acquires DuckLabs” / “Meta reaches $17B settlement” — both appeared on a cached snapshot of the HN front page this morning but had already fallen off by run time; neither could be verified against a fetched primary source, so they’re dropped rather than linked.

Sources checked: Claude Code changelog (latest Aug 28, covered yesterday), Claude release notes (latest Aug 25: Cowork memory — covered when it landed), Gemini CLI changelogs (latest v0.54.0, Aug 6 — quiet), OpenAI API changelog (latest Aug 21 — quiet), Codex changelog (v0.151.0, covered yesterday), Superhuman, ByteByteGo (latest Aug 27, covered yesterday), darioamodei.com (no new essays), blog.samaltman.com (no new posts; latest substantive entries are old), Google DeepMind blog, Hacker News front page.