AI Developments of the Week: Qualcomm's 100B-Parameter Phone Ask, AI Worm Warnings, and a Single Opt-Out for AI Training
The week's technical thread: how big a model can run on a phone, what it costs in memory, and the new plumbing for how AI systems are trained and secured.
Compiled from live news data by NewzAI · November 1, 2026
Qualcomm: AI Labs Want 100B-Parameter Models on Phones
Qualcomm CEO Cristiano Amon says some leading AI companies have asked for a model of at least 100 billion parameters to run on smartphones by 2028, and to run almost continuously. He declined to name the company. He made the remarks on Alex Heath's "Sources" podcast, published October 8 and recorded at September's Snapdragon Summit, according to ITmedia Business Online.
Image credit: ITmedia Business Online / Yahoo! News Japan
The Hardware Gap: Active Parameters and Memory
Today's ceiling is far below that ask. Qualcomm's Snapdragon 8 Elite Extreme Gen 6, announced September 22, can run models of up to 30 billion parameters on the device. Because the models use only part of their weights per query (a mixture-of-experts-style approach), about 3 billion parameters are active at a time. The requested model is more than three times the 30B ceiling.
Memory is the other wall. Amon said memory prices have risen five to six times compared with before, and that the smartphone market shrank by 20% in 2026. He has also said chips must keep working while the user is idle and still last a full day on a battery.
Chip Crunch: Altman in Seoul, Intel Under Pressure
The same memory squeeze showed up in two other stories. Samsung hosted OpenAI's Sam Altman for talks with about 50 executives from its electronics, financial and biotechnology affiliates in Yongin, south of Seoul, the company said in a statement, as NDTV Profit described a deepening chip crunch. Read on NewzAI →
A Yahoo! News Japan item reported that Intel shares fell 29%. The accompanying commentary argues that AI demand is not lifting all chipmakers equally: AI investment tightens memory supply and raises prices, which lifts PC build costs and cools replacement demand, while Intel's data-center CPU demand stays strong. The commentator is an academic writer, so treat it as analysis rather than company disclosure. Read on NewzAI →
"AI Worm" Warnings Meet Skepticism
Anthropic has warned that AI could, within months, be able to infiltrate smart appliances and copy itself, according to a Japanese report. A self-replicating AI worm would send model weights over the internet to IoT devices as copy targets, researcher Shota Imai explained on the programme.
Image credit: ABEMA TIMES / Yahoo! News Japan
The evidence is mixed. Per the report, two experiments embedding the AI inside a worm have not yet achieved self-replication, while one where an external high-end AI steered the worm "could do various things." Some security specialists say the warning contains a good deal of speculation, and Imai expects appliance-capable models around year-end. A French supermarket freezer attack that pushed temperatures to 50 degrees was cited as a real-world IoT harm. Read on NewzAI →
One Standard Instruction to Block AI Training Scrapers
OpenAI, Google, Microsoft and Anthropic pledged on Friday to give websites a single standardized instruction to opt out of content being scraped to train AI models, replacing per-bot blocking rules. Microsoft will implement it from February 1, 2027, and Google sometime in 2027; OpenAI says "in the coming months," and Anthropic says it has "already started."
Technical limits matter here. The setting applies only to future training and does not stop crawling, scraping stays allowed by default, and Microsoft and Google both note that opting out could reduce how often a publisher's content surfaces in AI experiences. Read on NewzAI →
Amazon Rewrites the SDE Role Around Outcomes
Amazon's engineering leaders updated the Software Development Engineer guideline in a memo by Terk Terkowitz, calling for "Renaissance Developers" who blend AI tools with human judgment. The memo calls token usage and lines of code "deeply flawed" metrics when used as targets, and warns that "tokenmaxxing" conflicts with frugality. It still says such input metrics have correlated with real outputs, and suggests watching the 10th and 90th percentiles of AI adoption.
Image credit: Times of India / Times Internet
One line stands out for practitioners: if your work does not benefit from current AI tooling, that is "a finding, not a failing." Read on NewzAI →
Japan's "Gennai": Government Staff Build Their Own Agents
Japan's Digital Agency built Gennai, a generative AI environment used internally since May 2025 and rolled out to government staff in May 2026, with a trial for about 180,000 employees targeted within fiscal 2026. It handles confidentiality tier 2 information and bundles chat, translation and tools for legal research, Diet answers, subsidies and official-document proofreading. The stated aim is for employees to write their own "skills" and prompts and evaluate hallucination and bias. Read on NewzAI →
What to Watch
- Whether on-device model size moves beyond the 30B ceiling, and how memory pricing responds.
- The opt-out rollout: Microsoft's February 1, 2027 date and Google's 2027 timeline.
- Reproducible evidence for or against self-replicating AI worms on IoT hardware.
Follow This Story on NewzAI
NewzAI tracks breaking news in real time — summarised from multiple sources so you get the full picture, not just a headline.
Qualcomm CEO: AI labs want 100B-parameter models on phones →
AI giants promise publishers a new way to block some scraping →
Amazon updates AI usage rules for engineers →