Microsoft Copilot Adoption Outpaces Security at MSPs

Levent Alpöge says Claude helped develop a proof of the Riemann Hypothesis, including work involving automorphic forms on higher-rank groups and a theta-lift construction. The proof was formalized in Lean, and Lean FRO and the Mathlib Initiative report that its 17-million-line code compiles, proves Mathlib’s stated version of the hypothesis, and passes four verification tests. That confirms the formal proof matches the theorem encoded in the system; it does not validate Alpöge’s separate Generalized Riemann Hypothesis claim or offer a readable mathematical explanation. The result has not completed the publication and community-review process required before the Clay Mathematics Institute would consider its $1 million prize.

AI adoption is also running ahead of security at managed service providers. A survey of more than 400 MSPs found 82% sell Microsoft 365 Copilot at scale or run customer pilots, but only 46% offer data protection services. Among reviewed Copilot tenants, 56% had neither data-loss-prevention policies nor sensitivity labels, and 86% of MSPs lacked full visibility into customers’ AI use. Only 28% consistently apply security practices across customer environments, despite 80% naming automation and standardisation as a priority. Just 31% can apply a policy change across all managed tenants within minutes.

Meta AI is funding five LSU-system projects, including Eyes-On, an $80,000 infant airway monitor. It uses green, orange and red indicators to show whether an artificial airway tube appears in place, whether the system is uncertain, or whether it is out and an alarm should sound. LSU pediatric otolaryngology head Michael Dunham stressed that people must assess the data and make decisions based on the AI output. Elsewhere, two Fond du Lac-area businesses use AI for practical tasks: Wonder brainstorms cafe events and specials, while mobile acai-bowl business The Berry Bowl uses it to manage inventory and limit fruit waste. The Better Business Bureau advises businesses to adopt AI gradually and where it suits their needs.

Researchers testing smartphone earthquake detection trained five deep-learning models on 35.4 million two-second data windows from about 186,000 recordings. With no more than 100 false alarms per device daily, the best model detected 76% of seismometer events and 61% of smartphone events. Performance fell to chance on phones sampling at 25 hertz, too slowly to capture the relevant seismic frequencies. The models can run on Android, with inference taking 3.2–6.4 milliseconds per window.

Two projects focus on computing efficiency. Tufa Labs’ tufazip reportedly reduced one gigabyte of Wikipedia data to 105.39MB without information loss, using a 4.33MB starter model that learns while processing text. The reported process requires two Nvidia B200 GPUs and about a week to compress or decompress. Separately, a builder raised generation speed for the 2.8-trillion-parameter Kimi K3 Q2_K model from 2.20 to 3.11 tokens per second by moving 45 of its 94 shards—414GB—into a Linux RAM drive. The 865GB model runs on two RTX 3090s, and cooling the remaining Optane memory is still a concern.

AI-generated political ads and AI training policies are prompting disputes over transparency and consent. California campaign ads by former Los Angeles mayoral candidate Spencer Pratt and gubernatorial candidate Steve Hilton used AI-generated images. Experts warn that targeted content can make it harder for voters to judge what is real; some Hilton ads disclosed their AI origin, but much content on TikTok and Instagram was not flagged. Researchers cited in the article say people have not shown a reliable ability to distinguish realistic AI content from human-made material. In Canada, some Etsy sellers are removing work or closing shops over terms taking effect Dec. 16 that allow Etsy and partners to use seller content to develop and improve AI. Etsy executive James Ossman said sellers own their work, denied using it to train models intended to copy or replace their creativity, and acknowledged the change needed a better explanation.

At Northeastern University, researchers Babak Heydari and Hongan Zhu are studying whether AI agents can better reproduce human choices by reflecting how people actually make decisions, including their flaws. They had agents from seven AI models play the beer game, a supply-chain exercise involving a retailer, wholesaler, distributor and factory. The aim is to make simulations useful for testing business and policy choices—such as evacuation instructions, transit fares or pharmacy locations—before implementation. Their work builds on agent-based simulations, including Thomas Schelling’s 1971 demonstration of how small individual choices can produce urban segregation.

Key Takeaways

  • Claude assisted Levent Alpöge’s Lean-formalized Riemann Hypothesis proof; Lean and Mathlib communities verified that the code proves the theorem as encoded, but the result is not yet an officially accepted, prize-eligible result.
  • A survey of more than 400 MSPs found 82% sell Microsoft 365 Copilot at scale or run pilots, while 56% of reviewed Copilot tenants lacked both data-loss-prevention policies and sensitivity labels.
  • Meta AI is funding five LSU projects; its $80,000 Eyes-On project monitors infant airway tubes with status indicators and an alarm, while clinicians remain responsible for interpreting its output.
  • Fond du Lac businesses Wonder and The Berry Bowl use AI for event brainstorming and inventory management, respectively; the Better Business Bureau recommends gradual, needs-based adoption.
  • Smartphone earthquake models detected up to 76% of seismometer events under a limit of 100 false alarms per device per day, but performance fell to chance at a 25-hertz sampling rate.
  • Tufa Labs says tufazip compresses 1GB of Wikipedia data to 105.39MB without information loss; the reported process uses two Nvidia B200 GPUs and takes about a week.
  • Moving 414GB of the 865GB Kimi K3 Q2_K model into a Linux RAM drive increased generation speed from 2.20 to 3.11 tokens per second on a two-RTX-3090 setup.
  • Experts warn AI-generated campaign ads can make it harder to assess what is real; disclosure varies across ads and social platforms, and research cited in the article finds people cannot reliably spot realistic AI content.
  • Some Canadian Etsy sellers are removing work or closing shops over terms allowing seller content to be used to develop and improve AI; Etsy says it does not use work to train models intended to copy sellers.
  • Northeastern researchers are testing agents from seven AI models in a supply-chain game to see whether simulations that reflect human decision-making can help evaluate business and policy choices.

AI’s Riemann Hypothesis Breakthrough Came Decades Before Tao’s Estimate

On October 15, mathematician Terence Tao told Brady Haran in a Numberphile interview that AI might not solve the Riemann Hypothesis until around 2050. On November 20, Levent Alpöge announced a proof developed with help from Claude, saying the AI helped with parts involving automorphic forms on higher-rank groups and a theta-lift construction. The proof was reportedly formalized in the Lean proof assistant. It is not yet an officially accepted result, since the Clay Mathematics Institute requires a proof to be published and reviewed by the mathematical community for at least two years before considering its $1 million prize.

Lean and Mathlib Communities Verify Alpöge’s Riemann Hypothesis Proof

The Lean FRO and Mathlib Initiative say Levent Alpöge’s 17-million-line Lean proof correctly proves Mathlib’s definition of the Riemann Hypothesis. The communities checked that the code compiles, proves the stated theorem, and passes Lean’s official kernel and several independent proof checkers. They said the proof passed all four verification tests. The statement does not certify Alpöge’s separate claim about the Generalized Riemann Hypothesis, and it does not provide a human-readable explanation of why the hypothesis is true. Alpöge has said that version would exceed 250,000 pages and that he does not fully understand parts of it.

Survey Finds AI Adoption Outpacing Data Security at MSPs

A global survey of more than 400 managed service providers found that 82% are selling Microsoft 365 Copilot at scale or running customer pilots. However, only 46% offer data protection services, and 56% of Copilot tenants reviewed had neither a data loss prevention policy nor data sensitivity labels. The report also found that 86% of MSPs lack full visibility into customers’ AI use, while only 28% consistently apply security practices across customer environments. Although 80% said automation and standardisation are a top priority for the next 12 months, only 31% can apply one policy change across all managed tenants within minutes. inforcer Chief Strategy Officer Christian Nagele said MSPs need to maintain and enforce strong, standardised security for Microsoft 365.

LSU and Meta AI Develop Infant Airway Monitor

Meta AI is funding five projects with the LSU system to address healthcare challenges, including Eyes-On, a project for infants who need artificial airways. The project received $80,000 and is designed to monitor whether a baby’s airway tube is in place. Green means the tube is in place, orange means the system cannot tell, and red triggers an alarm if the tube is out. LSU pediatric otolaryngology head Michael Dunham said people must check the data and make decisions based on the AI’s output. The project aims to help keep babies healthy and eventually breathing on their own.

Fond du Lac Small Businesses Use AI for Ideas and Inventory

Two small businesses in the Fond du Lac area are using AI to help with day-to-day work. Wonder, a play cafe that opened in downtown Fond du Lac in 2023, uses AI to brainstorm event ideas and specials, but makes its promotional graphics without AI. The Berry Bowl, a mobile business recently launched near Ledgeview Lanes, uses AI to manage inventory and avoid wasting fresh fruit for its acai bowls. Owner Max Ramker said Facebook and Instagram have helped the business, with its Facebook posts nearing 40,000 views. The Better Business Bureau recommends that businesses take a gradual approach and use AI only when it fits their needs.

Smartphone AI can spot earthquakes but struggles with low sampling rates

Researchers tested five deep learning models that use smartphone accelerometers to detect earthquakes. Their dataset contained 35.4 million two second windows from about 186,000 recordings, including seismometer, smartphone and human activity data. A model with Bahdanau attention detected 76 percent of seismometer events and 61 percent of smartphone events under a limit of 100 false alarms per device each day. Detection dropped to chance levels on devices sampling at 25 hertz, which cannot fully capture the key seismic frequency range. The models were small enough to run on Android devices, with inference times of 3.2 to 6.4 milliseconds per window.

AI tool compresses Wikipedia to 105.39MB without losing data

Tufa Labs’ tufazip compressed one gigabyte of Wikipedia data into a 105.39MB file without losing information, according to a research paper. The package includes a 4.33MB starter model, and the system learns as it processes text rather than using hand built models. Tests found that a 4K word tokenizer compressed better than 8K and 16K versions, while a 188 million parameter model slightly outperformed a 267 million parameter model. Increasing the text window from 15 million to 35 million tokens reduced the file size by only 1.21 percent while using 40 percent more computing power. The article says the process needs two Nvidia B200 GPUs and about a week to compress or decompress the data.

RAM drive boosts giant Kimi AI model speed by 41 percent

A homelabber improved the speed of the 2.8 trillion parameter Kimi K3 Q2_K model by moving 45 of its 94 file shards into a Linux tmpfs RAM drive. With two RTX 3090 graphics cards, generation speed rose from 2.20 to 3.11 tokens per second, while prompt evaluation rose from 13.3 to 16.8 tokens per second. The model file is 865GB, and 414GB now sits in the RAM drive instead of Optane memory. The model has 896 experts, with 16 selected for each token and two shared experts. The builder still uses fans to cool the Optane memory, and cooling remains a concern.

AI election ads raise concerns about voter targeting and trust

California political campaigns used AI generated images in election ads, including ads by former Los Angeles mayoral candidate Spencer Pratt and gubernatorial candidate Steve Hilton. Experts told Times of San Diego that AI can create political content aimed at people using personal data and may make it hard for voters to tell what is real. One expert said people should be told when online content is AI generated and supported watermarking images with a clear label. Some of Hilton’s ads included a line saying they were AI generated, while much AI content on platforms such as TikTok and Instagram was not flagged. The article says early research has not found strong evidence that people can reliably distinguish realistic AI content from human made content.

Etsy sellers remove work over AI training policy

Some Etsy sellers in Canada are removing products and artwork after the company added AI training language to its terms. The policy, rolling out Dec. 16, says Etsy and its partners can use seller content to develop, train, test and improve AI and machine learning technologies. Vancouver seller Tsuki Matsuba is pulling her custom stationery products, saying the policy felt like a kick in the teeth. Ottawa artist Tina Nawrocki said she was closing her shop, while Halifax artist The Calendork removed its artwork but kept its shop open. Etsy executive James Ossman said sellers’ work belongs to them and the company does not use it to train AI models meant to copy their work or replace their creativity. He also said Etsy should have done more to explain the policy change.

AI Agents Aim to Recreate Human Choices in Supply Chains

Researchers Babak Heydari and graduate student Hongan Zhu at Northeastern University studied whether AI models can better reproduce human decisions by reflecting how people actually make choices, including their flaws. Realistic simulations could help businesses and policymakers test changes, such as evacuation instructions, transit fares and pharmacy locations, before putting them into practice. The research builds on agent-based simulations, an idea demonstrated by Nobel laureate Thomas Schelling in a 1971 experiment on how small choices can lead to urban segregation. To test AI agents, Heydari and Zhu had agents from seven AI models play the beer game, a supply chain exercise involving a retailer, wholesaler, distributor and factory.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

Claude Riemann Hypothesis Lean theorem proving Mathlib AI security Microsoft 365 Copilot Managed service providers Healthcare AI Earthquake detection Deep learning Data compression AI inference optimization AI-generated political ads AI transparency AI training data consent Etsy AI agents Agent-based simulation Human decision-making AI adoption

Comments

Loading...