Google is replacing Gemini’s single Keep Activity setting with separate controls for chat history and AI training. Users can choose whether chats are saved and for how long, independently of whether chats and attachments are used to train Google’s generative AI models. The training control covers content such as text, photos, files, audio, location information and connected-app data. The change is rolling out over the coming weeks; one report says it is due to take effect on November 17. Users with Keep Activity off will have both new controls off by default, while those who want to retain history but opt out of training must adjust the training setting. With History off, Google says it will still retain future chats for about two months to respond, process feedback and support safety—longer than the previous 72-hour period. Google is also planning to share more information about data use. ChatGPT and Microsoft Copilot offer options to disable training while keeping conversation history, while Claude’s training option is off by default.
AI is appearing in 911 dispatch centers across the United States, where it is being used in dispatchers’ work and described as part of frontline crime fighting. The reports do not identify specific systems or explain how they are used.
TSMC, which manufactures chips designed by Nvidia and Broadcom, is positioned to benefit from demand for AI chips. The articles cite a 73% share of the AI chip market, and one also reports a 73% share of the semiconductor foundry market and a near-monopoly position in advanced AI chips. TSMC forecasts fourth-quarter revenue of $52.3 billion to $54.3 billion, up 55% to 61% year over year, and plans capital spending of up to $64 billion this year to expand capacity. Its shares have risen 31% this year but fell more than 7% in the month through November 13. The company also makes chips for smartphones, TVs and car infotainment systems.
Researchers’ Proactive Co-Creator (PaCC) aims to give people more influence over creative AI before it generates media. It exposes its assumptions and uncertainty through questions, belief graphs and media attributes that users can revise. In a small exploratory study, participants reported a stronger sense of agency and control than with familiar prompt-based tools; some said the visible assumptions helped them notice overlooked details. Views on ownership and authorship varied.
Optiv is expanding its threat exposure management program as software vulnerabilities mount. The company cited 1,415 Common Vulnerabilities and Exposures in October and warned that attackers can use AI models to combine smaller flaws into larger threats. New services include managed exposure validation and cloud exposure management, with ongoing testing for attack paths in cloud and on-premises systems. Product Director Ben Spencer said the services help customers prioritize risks and address them when straightforward patching is not possible.
Researchers have proposed subprocess constrained Markov decision processes (S-CMDPs), which require expected discounted costs to stay within limits after every reachable history—not just from an initial state, as in traditional constrained MDPs. They find stationary policies can be suboptimal and that computing even an approximately optimal stationary policy is NP-hard. Their polynomial-time method uses history-dependent policies and can deliver expected reward within ε of the best feasible value while exceeding each cost limit by no more than δ.
Investor Steve Eisman said limits on data-center construction could pose a larger risk to AI companies than Nvidia-related concerns. Nvidia was due to report earnings after the market close on Tuesday amid scrutiny of its China business and a Commerce Department move to close a loophole allowing remote use of data centers in China. Eisman is watching for local or state restrictions, including a possible moratorium. He said a 10% to 30% slowdown could help hyperscalers become cash-flow positive again, while a complete halt could push the country into recession overnight. He also identified debt-funded AI neocloud companies building data centers as potential losers.
Netflix has built code-based controls for AI Discover, a steerable TV tab for exploring series, films, games and live programming. A controller gathers context and selects a plan—tested in advance or created by an AI planner—and checks it before execution. Each validated plan must produce exactly one usable page; because multi-step generation can take seconds, it runs in the background. MemberLLM prepares summaries of users’ tastes, while Memory Hub stores those summaries and direct preferences, such as requests not to recommend a title. Netflix applies safety rules through code but did not disclose speed, fallback-rate or user-impact figures.
Former Ford employee Kurt Kromm is suing Ford and Aramark after Ford fired him over an alleged $1.95 cookie theft flagged by an AI-powered surveillance system. His complaint says his debit card was charged at 3:38:33 a.m. on May 9, but Ford acted on the alert without checking Aramark’s payment records. Ford later reversed the firing, restored his seniority and offered back pay after Aramark confirmed the charge. Kromm declined to return and seeks more than $75,000 in compensatory damages plus unspecified punitive damages. The complaint says other workers also faced theft accusations, with at least three reinstated after investigations.
European private-equity firm Inflexion is exploring AI to identify potential acquisitions and improve client work. Employees have held hackathons and are considering an AI-powered “deal scout” to search data for targets. The firm is working with an AI startup and encouraging staff to learn how to use the technology.
Key Takeaways
- Google is separating Gemini chat-history controls from AI-training permissions; chats may still be retained for about two months when History is off.
- Gemini’s rollout is expected over the coming weeks, with one report specifying November 17; existing Keep Activity settings determine the initial state of both controls.
- ChatGPT and Microsoft Copilot also let users disable training while retaining conversation history; Claude’s training option is off by default.
- AI is being used in 911 dispatch centers, but the reports provide no details on specific systems or applications.
- TSMC forecasts fourth-quarter revenue of $52.3 billion to $54.3 billion, up 55% to 61% year over year, and plans up to $64 billion in capital spending this year.
- PaCC lets users inspect and revise a creative AI system’s assumptions before generation; a small study found participants reported greater agency and control.
- Optiv’s expanded security program adds managed exposure validation and cloud exposure management; the company cited 1,415 CVEs in October and warned AI could help attackers combine flaws.
- A new S-CMDP method uses history-dependent policies to meet per-history cost limits, with reward within ε of the best feasible value and cost-limit exceedance no greater than δ.
- Steve Eisman says data-center construction restrictions could threaten AI neocloud firms and says a 10%–30% slowdown could benefit hyperscalers.
- Netflix’s AI Discover uses code-checked plans, MemberLLM taste summaries and Memory Hub preferences; Inflexion is exploring an AI deal scout, while a former Ford worker is suing over a firing triggered by an AI surveillance alert.
Gemini adds separate controls for chat history and AI training
Google is rolling out two privacy settings for the Gemini app over the coming weeks. The History setting lets users choose whether chats are saved and how long they remain available, while a separate control decides whether chats and attachments can train generative AI models. When History is off, Google says future chats will still be kept for around two months to respond to users, process feedback, and support safety. Google is notifying users by email.
Google separates Gemini chat history from AI training controls
Google is replacing Gemini’s single Keep Activity setting with separate controls for chat history and AI training. History determines whether chats are saved and for how long, while the other setting covers use of chats, photos, files, audio, location information, and connected app data to train generative AI models. When History is off, Google will keep chats for around two months instead of 72 hours. The update is rolling out over the coming weeks, with changes set to be implemented on November 17, and Google is notifying users by email. If Keep Activity is off now, both new settings will be off.
Gemini lets users save chat history without training Google’s AI
Google is rolling out new controls that let Gemini users choose how long chats are kept and whether chat content is used to train generative AI. Turning off History will not delete future chats immediately, as Google will still save them for about two months to respond to users, process feedback, and support safety and security efforts. The controls are rolling out globally over the next few weeks. Users who already have Keep Activity turned off will have both new settings off by default.
Google lets Gemini users opt out of AI training without deleting chats
Google plans to replace Gemini’s Keep Activity setting with two controls in the coming weeks. One will let users manage chat history and choose how long it is saved, while the other will decide whether chats, including text and attachments, are used to train AI models. If Keep Activity is on when the change arrives, both new settings will be on; if it is off, both will be off. Users who want to keep their chat history but stop model training will need to change the training setting in Gemini’s Settings. The article notes that ChatGPT and Microsoft Copilot also offer ways to turn off training while keeping conversation history, and says Claude’s training option is off by default.
Google adds Gemini controls to separate chat history from AI training
Google is rolling out a Gemini app setting that lets users decide whether their data is used to train the AI, while keeping chat history separate from training. The change is expected to reach all users in the coming weeks, though Google has not given a specific date. Users can find the controls in the app’s settings. The article also says Google plans to share more information about how it collects and uses data and to give users more control over it.
AI tools are moving into 911 dispatch centers
Artificial intelligence is being used in many 911 dispatch centers across the country. The report looks at how dispatchers use AI in their work. The article provides no further details about the tools or how they are used.
AI joins crime fighting in 911 dispatch centers
Artificial intelligence is being used on the frontlines of crime fighting in many 911 dispatch centers across the country. The report focuses on how 911 dispatchers are using the technology. It does not provide further details about specific systems or their use.
TSMC stands out as an AI chip investment
Taiwan Semiconductor Manufacturing, or TSMC, makes chips designed by companies such as Nvidia and Broadcom. The article says TSMC has a 73% share of the AI chip market and benefits as demand for chips grows. It expects fourth-quarter revenue of $52.3 billion to $54.3 billion, up 55% to 61% from the same quarter last year. TSMC also makes chips for products including smartphones, TVs, and car infotainment systems.
Why TSMC may be a stronger AI chip pick than Nvidia
Nvidia and Broadcom design chips, while Taiwan Semiconductor Manufacturing, or TSMC, manufactures them. The article reports that TSMC holds 73% of the semiconductor foundry market and has a near monopoly on advanced AI chips. It expects fourth-quarter revenue of $52.3 billion to $54.3 billion, up 55% to 61% from the year before. TSMC's stock rose 31% this year but fell more than 7% in the month through Nov. 13, while the company plans capital spending of up to $64 billion this year to expand capacity.
A system helps creators shape AI before it generates
Researchers developed the Proactive Co-Creator, or PaCC, to help people guide creative AI before it produces media. The system shows its assumptions and uncertainty through clarifying questions, belief graphs, and media attributes, so users can revise how it understands their ideas. In a small exploratory study, participants reported a more positive sense of agency and control than with familiar prompt-based AI tools. Some said the system’s visible assumptions helped them consider details they had missed, while views on ownership and authorship varied.
Optiv expands security services to address AI driven vulnerabilities
Optiv announced an expansion of its threat exposure management program on Thursday to help customers manage a sharp rise in software vulnerabilities. The company cited 1,415 Common Vulnerabilities and Exposures in October and said attackers can use AI models to combine small flaws into major threats. The expanded program adds managed exposure validation and cloud exposure management, with ongoing testing to find possible attack paths in cloud and on premises systems. Product Director Ben Spencer said the services help teams choose important risks and apply fixes when simple patching is not possible.
New MDP framework requires safe costs after every decision history
Researchers introduce subprocess constrained Markov decision processes, or S CMDPs, which require expected discounted costs to stay within a limit after every reachable history. This is stricter than traditional constrained Markov decision processes, which limit expected total cost from an initial state. The researchers show that stationary policies can be suboptimal and that finding even an approximately optimal stationary policy is NP hard. Their polynomial time method uses history dependent policies and can produce a policy whose expected reward is within ε of the best feasible value, while exceeding each cost limit by no more than δ.
Steve Eisman sees data center limits as a bigger AI risk than Nvidia news
Investor Steve Eisman said AI neocloud companies, which borrow money to build data centers, could be among the biggest losers as concerns around Nvidia grow. Nvidia was due to report earnings Tuesday after the market closed, amid scrutiny of its China business and a Commerce Department move to close a loophole involving China’s remote use of data centers. Eisman said he was more focused on whether local and state governments would slow data center construction, including through a possible moratorium. He said a slowdown of 10% to 30% could benefit hyperscalers because they would become cash flow positive again, while a complete halt could put the country into recession overnight.
Netflix uses code based controls to guide its AI Discover feature
Netflix built a controller and executor for AI Discover, a steerable TV tab that helps users explore series, films, games and live programming. The controller gathers context and selects a plan, which may be tested in advance or created by an AI planner, and the plan must pass checks before it runs. Netflix says each validated plan must produce exactly one usable page, while multi step generation can take seconds, so it runs in the background. MemberLLM creates taste summaries ahead of time, and Memory Hub stores those summaries and direct signals such as requests not to recommend something. The system applies safety rules through code, but Netflix did not share figures on speed, fallback rates or user impact.
Former Ford worker sues over firing after a $1.95 cookie purchase
Kurt Kromm, a former Ford worker, sued Ford and Aramark after the automaker fired him over an alleged theft of $1.95 cookies from a factory kiosk. His lawsuit says an AI powered surveillance system flagged the purchase, even though his debit card was charged at 3:38:33 a.m. on May 9. Kromm says Ford relied on the alert without checking Aramark’s payment records, then later reversed his firing, restored his seniority and offered back pay after Aramark confirmed the charge. He refused to return to work and seeks more than $75,000 in compensatory damages, plus unspecified punitive damages. The complaint also says other workers faced theft accusations and that at least three were reinstated after investigations cleared them.
Inflexion turns to AI to find new companies
European private equity firm Inflexion is exploring artificial intelligence to help identify and acquire companies. Employees have held hackathons and are considering an AI powered “deal scout” that could search data for possible targets. The firm is working with an AI startup and encouraging employees to learn how to use the technology. Inflexion is also exploring ways AI could improve its work with clients as part of a wider effort to modernize and expand.
Sources
- New Gemini setting lets you keep chat history, but turn off AI training
- Gemini separates history from AI training but keeps chats longer
- You can now keep your Gemini history without feeding Google’s training pipeline
- Google is making it easier to opt out of Gemini AI model training - here's how
- Google makes it easier to use Gemini without training AI on your data
- How 911 dispatchers are using AI
- How 911 dispatchers are using AI
- Not Nvidia. Not Broadcom. This Is the AI Chip Stock I’d Buy Before the Year Ends.
- Not Nvidia. Not Broadcom. This Is the AI Chip Stock I’d Buy Before the Year Ends.
- Shaping AI's Beliefs Before It Generates: Human Sense of Agency in Creative AI
- Optiv expands threat exposure services amid AI-driven security flaws
- Subprocess-Constrained Markov Decision Processes
- 'Big Short' investor Steve Eisman to discuss Nvidia, Anthropic and AI bull on CNBC
- Netflix puts code, not prompts, in charge
- Ford worker fired over $1.95 cookie sues automaker, Aramark — claiming AI falsely flagged him as thief
- Inside Inflexion’s AI push
Comments
Please log in to post a comment.