LG makes you liable for anyone its TV's AI mic records
Two hardware stories and one biology model this hour: a TV that listens when it looks off, a phone chip that runs a 30B model locally, and a four-month-old startup that says it beat AlphaFold 3 where it hurts.
LG's smart TVs were caught capturing voice audio while the screen was off — and its terms put the legal risk on you. Hardware teardown channel Gamers Nexus published a 135-minute investigation that Notebookcheck summarized: during bench tests the TV recorded clean microphone audio while the panel appeared to be in standby, and when the Ethernet cable was pulled the set kept saving voice input locally and uploaded the stored files once the network came back. LG has not commented. The finding lands on top of a separate contract problem: section 6(d) of LG's updated terms, headed "Voice Recognition and Privacy Compliance," makes the owner solely responsible for getting consent from anyone the product's AI voice features might capture — including notifying household members and guests that their voices may be recorded.
The clause would be merely odd if the exits were clean, and they are not. The AI voice services ship enabled by default, so owners are opted in before they are asked to manage anyone else's consent. According to Notebookcheck, accepting the terms also lets LG monitor how the product is used and share that data with third parties — the consent obligations flow toward the customer while the data flows toward the company. And declining is penalized: older sets can dodge the new terms by refusing webOS updates, but a TV that stops updating also stops receiving security patches. For an internet-connected device sitting on a home network, "consent or accept a less secure device" is not really a choice. What to watch: whether a regulator with an actual wiretapping statute reads section 6(d) the way it reads on paper.
Huawei's Kirin 9050 Pro is the first phone chip Huawei claims can run a 30B-parameter MoE model on device. Announced alongside the Mate XT 2 trifold, the chip pairs a redesigned nine-core CPU — 24% higher single-core peak, 52% better multi-threaded throughput, 42% faster overall than the previous generation — with a Da Vinci NPU that Huawei says is the industry's first to run a total-30B, 2B-active on-device multimodal MoE. That model is a native Pangu build, and the demo use case is local photo-library organization that works with no network. The GPU, a self-designed Maleoon part with hardware ray tracing, is rated 142% faster at rendering. The Mate XT 2 starts at 19,999 yuan; the Pura X View, on the older Kirin 9030S, starts at 5,999. HarmonyOS 7 opened public beta at the same event, adding an AI entry point on the home dock and agent skills for its XiaoYi assistant.
The number that matters isn't the render percentage, it's the 30B/2B claim. Putting a mixture-of-experts model of that size behind 2B of active parameters on a phone is the difference between "on-device AI" that summarizes a notification and on-device AI that does real multi-step work without a round trip — and Huawei is doing it on a chip it designs and fabricates domestically. Treat the benchmark figures as vendor claims until someone runs them, but if a 30B MoE genuinely runs locally, the constraint on phone agents stops being bandwidth and starts being battery.
A four-month-old Shenzhen startup says its all-atom biology model beats AlphaFold 3 on peptides and ligands. Atomelody (深旋科技), founded in Nanshan and working on AI-driven large-molecule drug discovery, released Melo-1 preview, which it says refreshes core metrics on public benchmarks for protein-peptide and protein-small-molecule structure prediction and exceeds AlphaFold 3 on most evaluated categories — peptides, small molecules, antibody-antigen, and RNA — with DNA the exception. The pitch is physical validity as much as accuracy: the company says Melo-1 bakes in chemistry constraints so its high-scoring structures don't come out with clashing atoms and impossible bond angles, which is what makes some leaderboard-pretty models useless in a wet lab.
Why a four-month-old company is doing this is the interesting part. AlphaFold 3's weights are still released only to non-commercial users, so a Chinese drug company that wants an AF3-class all-atom model for real commercial research cannot license one — it has to build one. Atomelody's core team claims prior wet-lab-validated wins on more than ten GPCR targets, one of the hardest target families in pharma. Melo-1 is a structure-prediction model, not a designer; the company says the full generate-then-validate loop is still to come.
What to watch: whether an independent group reproduces the peptide and ligand numbers once the full version opens, and whether LG answers the standby-microphone finding at all.
Would you accept a smart TV that only gets security patches if you consent to voice capture? Tell us in the comments.
Sources: Notebookcheck · Gamers Nexus investigation · International Cyber Digest · Hacker News discussion · Zhidx (智东西) · IT之家 ITHome · Sina Tech · BlockBeats · AI Era (新智元) via Sina · Sina Finance · Atomelody Melo-1 report
