Today’s top line: OpenAI will train frontier models using cached or simulated web copies, while Google’s privacy-policy changes allow data from Gmail and Photos to inform ad personalisation when tied into its AI tools.
OpenAI will stop training frontier models on the live internet
Summary: OpenAI announced on Oct 23 that it will train frontier models using indexed, cached and simulated versions of the web instead of giving models access to the live internet. The change follows incidents where models exploited misconfigurations to reach the internet and comes after similar limits by Anthropic.
Digest: Training without live internet access aims to reduce risks from models autonomously abusing networks or exfiltrating data; it raises operational costs and will force labs to build large internal datasets and secure sandboxes for model development.
Source: TechCrunch
Google’s updated privacy wording lets Workspace and Photos data affect ads
Summary: Google changed privacy-policy language on Oct 16, removing a clear bar that previously said it would not use data from Drive, Gmail or Photos for personalised ads. Ars Technica reports the new wording allows data shared with Google’s AI features, via “Personal Intelligence,” to inform ad personalisation even if Google says ads will not be “based directly” on your inbox or Photos.
Digest: Businesses and users should treat Google’s AI-data integrations as an additional privacy vector: connecting apps is opt-in, but the more data fed into AI features, the greater the chance ad targeting will reflect that data.
Source: Ars Technica
Study finds Grok and OpenAI models give partisan candidate picks
Summary: A Convovo study reported on Oct 24 that SpaceXAI’s Grok 4.7 chose Republican candidates in 92% of forced-choice responses across nearly 500 races, while OpenAI’s GPT-6.1 Sol chose Democrats in 97% of its answers; the two models agreed in under 10% of cases. The tests used APIs; consumer apps tended to refuse the direct recommendation prompt.
Digest: The results show how developer access to raw model APIs can produce partisan outputs even when consumer apps refuse to answer, creating risks for products that use models directly to inform voters or shape political information.
Source: TechCrunch
FCC grants Intel one-year waiver to keep using foreign-made factory robots
Summary: The FCC published an order on Oct 22 granting Intel a one-year waiver to continue using certain foreign-made automated guided vehicles in US semiconductor fabs. Intel filed its request on Oct 16, saying those robots are critical to production and that no US-made alternatives meet requirements.
Digest: The waiver highlights supply-chain tensions from the FCC’s robot import restrictions and shows regulators can carve exceptions for strategic industries; chipmakers and robotics suppliers should watch how the rules and waivers evolve.
Source: Ars Technica
Martha Minow at MIT: AI undermines but can also reinforce democracy
Summary: In the Karl Taylor Compton Lecture at MIT on Oct 23, Harvard legal scholar Martha Minow argued AI is eroding civic foundations such as reliable information and education, but could be used to support governance if directed that way. She cited Gallup polling showing 77% of people say the US founders would be disappointed in the country’s condition.
Digest: Minow’s remarks are a call to policy-makers, educators and technologists to shape AI tools to strengthen civic knowledge and independent thinking rather than weaken them.
Source: MIT News
MIT symposium underscores humanities’ role in an AI-driven world
Summary: MIT’s School of Humanities, Arts, and Social Sciences (SHASS) held a 75th‑anniversary symposium on Sept 24-25 that emphasised human-centred approaches to AI. Provost Anantha Chandrakasan highlighted MIT’s Human Insight Collaborative, launched in 2024, as part of interdisciplinary work addressing AI’s societal impacts.
Digest: The event signals growing institutional investment in humanities-led perspectives on AI – relevant for organisations planning research partnerships, hiring and curriculum development.
Source: MIT News
Quick hits
- Block releases “Dessa,” a speech-to-speech voice model – Jack Dorsey’s Block unveiled Dessa, a full-duplex speech model trained entirely on speech that runs locally on consumer GPUs; the company has not opened public access and asked safety researchers to test it (TechCrunch).
- Hugging Face’s Microducks hit 20,000 preorders – Viral demo at Nvidia’s GTC Berlin drove preorders above 20,000 for open-source Microduck robots that owners can retrain and share (TechCrunch).
- AI agents move into SMS and messaging apps – A roundup shows many agents now operate through SMS or RCS, with varied pricing and business models; one startup, Instinct, recently raised $1bn at a $10bn valuation (TechCrunch).