The Answer Engine AI News · Sydney

Ollie, an AI assistant that prioritised user privacy, is closing its doors in one week, according to its founder. The shutdown comes despite — or perhaps because of — an increasingly crowded market for AI assistants. While the source doesn't detail specific reasons for the closure, the timing is notable: as tech giants like Google, Microsoft, Apple, and Meta push their own AI assistants, smaller players face significant challenges competing for attention and funding. Ollie's privacy-first approach clearly wasn't enough to sustain the business in a market dominated by well-resourced competitors who can afford to offer free or low-cost services subsidised by other revenue streams. The closure highlights a growing tension in the AI space between privacy-conscious alternatives and mainstream offerings that may collect more user data but benefit from massive marketing budgets and ecosystem integration. For users who valued Ollie's privacy protections, the shutdown means one fewer alternative to the data-hungry assistants from big tech.

Intel's share price has taken a significant hit, wiping out $12 billion in value from the US government's stake in the chipmaker. The culprit? The AI boom is creating unexpected ripple effects across the PC market. As AI companies race to build data centres and AI systems, they're buying up massive quantities of memory chips and other components. This surge in demand is driving up prices for the same parts that go into regular PCs, making computers more expensive for everyday buyers. The result: PC sales are slumping, and Intel—which relies heavily on PC chip sales—is feeling the pain. While other chipmakers like NVIDIA have thrived by focusing on AI-specific processors, Intel has struggled to pivot quickly enough. The company's troubles highlight how the AI industry's appetite for hardware is reshaping traditional tech markets, creating winners and losers in unexpected places. For investors and governments alike, it's a reminder that even established tech giants can stumble when the market shifts rapidly beneath them.

Researchers have developed LRMoE, a training system that helps large AI models called Mixture-of-Experts (MoE) run more efficiently on GPUs with limited memory. MoE models are popular because they can scale to handle huge workloads, but they have a problem: data doesn't route evenly through their different "expert" components, creating bottlenecks. Previous solutions copied popular experts across multiple GPUs, but these copies compete for the same limited memory needed to process data. LRMoE takes a different approach by intelligently deciding which calculations to keep and which to skip during the backward training phase, based on how much each contributes to improving the model. It also redistributes the workload more strategically across available GPU memory. Tests on two large Chinese language models (Qwen3-30B and GLM4.5-106B) showed LRMoE significantly reduced training time to reach the same accuracy targets. The system achieves this by reducing both the amount of data transferred between GPUs and the memory needed during training, making it practical to train powerful AI models on more modest hardware setups.

Researchers have developed a technique to better calibrate computer models that simulate physical systems—like ocean turbulence—when conditions vary across different scenarios. The traditional approach tries to find one set of calibration parameters that works everywhere, but this often fails because the real world behaves differently in different contexts. The new method creates multiple "localized" calibrations that adapt to specific conditions, accounting for uncertainty in both measurements and the contexts in which they were taken. Using a type of AI called conditional normalizing flows, the system learns to quickly produce calibrated predictions for any given context without starting from scratch each time. When tested on ocean turbulence models (which predict how water mixes at different depths based on atmospheric conditions), the localized approach produced more accurate predictions than traditional global calibration. The technique handles messy real-world complications like measurement errors, unevenly distributed data, and boundaries in the problem space. This represents a step forward in making complex physics simulations more reliable across the range of conditions they encounter in practice.

Researchers have developed ThinkJev, a system that helps language models make better yes/no and multiple-choice decisions without wasting time or money. When AI models answer questions that require calculations, counting, or date comparisons, they often get it wrong if they answer immediately. Letting them "think" first (generate reasoning steps) fixes most errors, but takes much longer—seconds instead of milliseconds. ThinkJev solves this by creating a two-speed system from a single AI model: System 1 gives fast answers when confidence is high, and System 2 takes time to think through harder questions. The clever bit is that both systems work from the same starting point, so switching between them wastes no extra computing. ThinkJev automatically decides which system to use based on testing: it only lets the fast system answer when statistical tests on sample data prove it will stay under a target error rate. This means businesses get accurate answers without paying for unnecessary thinking time on easy questions, while harder decisions automatically get the reasoning time they need.

Reddit's CEO has confirmed what many publishers feared: Google's AI Mode and AI Overviews are not sending traffic back to source websites. The company reports falling search referrals from Google, directly attributing the decline to Google's AI features that answer questions without requiring users to click through to websites. This matters because Reddit is one of the web's largest content platforms, and if even they're seeing measurable drops in Google referral traffic due to AI features, smaller sites are likely experiencing similar or worse impacts. Reddit is also testing requiring users to log in to view content, which could be a response to this traffic decline. The statement adds to growing tension between content publishers and Google, as AI-powered search results increasingly keep users on Google's own pages rather than directing them to the original sources. For businesses relying on Google search visibility, this is a clear signal that traditional SEO strategies focused purely on ranking may no longer guarantee traffic.

Brands face a new measurement challenge as AI-powered search tools like ChatGPT, Perplexity, and Google's AI Overviews recommend products and services to buyers. The article argues businesses need to start tracking whether these AI systems actually mention or recommend their brand when users ask relevant questions — not just whether they rank in traditional search results.

The piece highlights three key actions: measure if AI tools are surfacing your brand to potential customers, identify and fix any outdated or incorrect information these systems are pulling from, and find ways to convert "shortlist visibility" (being mentioned alongside competitors) into trackable customer relationships. The emphasis is on creating metrics that CFOs will value — moving beyond vanity metrics to demonstrate actual business impact from AI visibility.

As AI answers increasingly replace traditional search result clicks, brands that don't track their presence in these new formats risk flying blind. The gap between what businesses measure (traditional search rankings) and where customers actually discover brands (AI recommendations) is widening.

Discord is making its most public servers more discoverable by creating dedicated web pages that search engines can index. These pages will display some messages from public Discord communities and may include AI-generated summaries of the conversations. Until now, Discord's millions of community discussions have been essentially invisible to Google and other search engines, locked behind the platform's login wall. This change means public Discord servers – those marked as discoverable and not age-restricted – will now have a presence in traditional search results. The move brings Discord in line with other social platforms that allow search engines to crawl and display their public content. For Discord, it's a significant shift in strategy that could help communities grow by making them easier to find, while also making the platform's vast repository of user discussions and knowledge accessible through web search. The feature will only apply to servers that have opted into Discord's public discovery features, and server owners will presumably have some control over their participation.