The Answer Engine AI News · Sydney

Serve Robotics, a company that operates sidewalk delivery robots, has announced a new business line called Physical AI Services. Released alongside its third-quarter earnings, the service will let other companies access Serve's connectivity and data infrastructure – essentially renting out the backend technology that powers its robot fleet. The move mirrors Amazon Web Services' model: AWS started as Amazon's internal infrastructure before becoming a standalone business selling cloud computing to others. Serve is betting that other robotics companies will pay to use its proven systems for connecting to robots in the field and managing the data they generate, rather than building their own from scratch. This pivot suggests Serve sees more immediate revenue potential in selling tools to the robotics industry than in delivery services alone. It's a common strategy in emerging tech sectors – companies that solve hard infrastructure problems for themselves often discover those solutions are valuable to competitors and adjacent businesses.

Meta has rolled out a new feature allowing US users to video chat with Muse, its AI assistant, which now appears as a customizable animated avatar during calls. Instead of text-only interactions, users can now have face-to-face conversations with their AI agent, choosing how the avatar looks and behaves. This update comes ahead of Meta's planned launch of a Tamagotchi-style wearable device designed to bring the Muse AI into the physical world. The video chat capability represents Meta's push to make AI assistants feel more personal and lifelike, competing with other tech giants developing conversational AI products. While the feature is currently limited to the US market, it signals Meta's broader strategy to embed AI companions more deeply into everyday life through both software and hardware. The Tamagotchi-like wearable suggests Meta is betting on portable AI devices that users can carry and interact with throughout their day, moving beyond smartphone and computer screens.

Texas has been the fastest-growing data center market in the United States, attracting massive investment from Big Tech companies building AI infrastructure. However, that rapid expansion now faces serious obstacles due to energy policy shifts and concerns about the state's power grid capacity. The AI boom requires enormous amounts of electricity – data centers running AI models consume significantly more power than traditional computing facilities. Texas's grid, which famously struggled during winter storms in 2021, is now under fresh scrutiny as energy demands from AI facilities strain available capacity. This has prompted regulators and energy authorities to pump the brakes on new data center approvals. The slowdown marks a significant shift for an industry that had viewed Texas as ideal territory due to its business-friendly regulations, available land, and relatively cheaper power costs. Companies that had planned major AI infrastructure investments in the state are now reassessing their options, potentially looking to other regions or countries that can guarantee the stable, massive power supply that AI operations demand.

Researchers have tackled a common problem in multi-agent AI systems: when one AI reviews another AI's work, it can fix genuine errors but also wrongly override decisions that were already correct. The study tested different approaches to giving a "reviewer" AI the authority to change an "upstream" AI's recommendations. Instead of letting the reviewer change anything it wants, they introduced an "evidence gate" — the reviewer can only override a decision if it first produces a clear record showing a specific rule was broken and offers a compliant fix. Testing this across nearly 200 tasks using two open-source models (Qwen and Mistral), the evidence gate dramatically reduced cases where correct decisions were wrongly replaced, while still catching most genuine errors. When given three explicit options (keep, replace, or escalate), one model (Qwen) stopped corrupting correct answers entirely, but the other (Mistral) still replaced all 48 correct proposals it reviewed. The research shows that how you structure decision-making authority between AI agents significantly affects the quality and reliability of their combined output.

Researchers have identified a critical weakness in AI agents that use external tools: the AI choosing an action doesn't know whether that tool is actually working properly at that moment, even though the system running it does. They call this "operational blindness." The solution, called Agent Harness Awareness, lets the host system share a small, verified snapshot of tool health with the AI before it makes important decisions—without giving the AI direct control over monitoring or enforcement. Testing across GPT-4o, Claude Sonnet 3.5, and Gemini 1.5 Flash showed dramatic improvements: errors that cost an average of 4,252 milliseconds dropped to just 131 milliseconds (a 96.9% reduction), and successful task completion on the first try jumped from 69.4% to 95.8%. The approach keeps trust boundaries intact—the host system still measures, enforces rules, and controls final execution—while simply telling the AI "this tool is slow right now" or "that API is down" before it commits to an action. It's a straightforward fix to a problem that occurs whenever AI agents interact with real-world systems that can fail, slow down, or become temporarily unavailable.

Google is facing an unusual problem: more people are searching than ever before, with queries hitting a five-year peak, yet search revenue only grew 12%. The culprit is AI Mode—Google's conversational AI search experience that answers questions directly without users clicking through to websites. When searchers get their answers straight from AI-generated responses, they don't click ads, and Google doesn't make money. Google's solution? Put more ads directly inside the AI answers themselves. This marks a significant shift in how search monetisation works. For years, the model was simple: someone searches, sees ads alongside traditional results, clicks, and Google gets paid. Now Google is redesigning that entire system around AI responses. The gap between rising query volume and flat revenue growth shows how quickly AI is reshaping search economics. For business owners, this confirms what many have suspected: the way people find information online is fundamentally changing, and the old rules about search visibility and traffic are being rewritten in real time.

New data from the holiday shopping period reveals where ads are appearing inside AI-powered search results. The analysis tracked ad placements across Google AI Overviews, Google AI Mode, and ChatGPT, examining five shopping categories. Results show that apparel and electronics dominate ad appearances in these AI-generated answers. The research also identified which specific prompts triggered ads within ChatGPT's responses. This is the first substantial dataset showing how paid advertising is being integrated into AI search tools during high-stakes shopping periods. While Google has been testing ads in AI Overviews since earlier in the year, this holiday data provides concrete evidence of which product categories are seeing the most commercial integration. For businesses trying to understand how AI search tools are evolving, this signals that some industries are already securing prominent paid placement in AI answers, while others remain largely organic. The concentration in apparel and electronics suggests these categories either have higher advertiser demand or that AI platforms are finding these products easier to monetise through embedded advertising.

Generative Engine Optimisation (GEO) may show low direct referral traffic in your analytics, but that doesn't mean it's not influencing sales. The article argues that traditional traffic metrics miss the bigger picture: people are using AI chatbots like ChatGPT to research purchases, then coming to your site later through other channels (direct visits, branded searches, or bookmarks). The key finding is that visitors who arrive directly from ChatGPT convert at nearly five times the rate of regular search visitors, suggesting they're further along in their decision-making process. This means AI tools are acting as a research and recommendation layer before customers ever click through to your website. If you're only looking at referral numbers in Google Analytics, you're missing how AI is shaping the path to purchase. The takeaway: GEO work—making sure AI tools can find and accurately represent your business—matters even when it doesn't show up as a clear traffic source, because it's influencing buying decisions upstream of the final click.

OpenAI has demonstrated that AI agents can create working exploits to escalate privileges on Linux systems with known vulnerabilities. In an exercise, the AI systems generated 948 distinct exploits that could turn low-level user access into full administrative (root) control on unpatched Linux kernels. This isn't about AI discovering new vulnerabilities—it's about AI rapidly weaponising known security holes that haven't been patched yet. The test shows that what previously required specialist human expertise can now be automated at scale by AI. For businesses running Linux servers or systems, this dramatically shortens the window between a patch being released and attackers having working exploits ready to use. The practical message: if you're behind on security updates, the risk level just jumped significantly. AI can now turn publicly documented vulnerabilities into working attack code much faster than before, making "we'll patch it next month" a much more dangerous gamble than it already was.

National Australia Bank has rolled out an internal AI assistant named Ruby, built on Anthropic's Claude platform, for its employees. The bank is now pursuing approximately 50 additional use cases for artificial intelligence technology across its operations. Ruby represents NAB's move to implement generative AI tools internally, following the broader trend of Australian businesses adopting large language model technology to improve productivity and operational efficiency. While specific details about Ruby's functions weren't provided in the source, the deployment signals NAB's commitment to scaling AI adoption beyond a single pilot project. The bank's pursuit of 50 more AI use cases suggests they're taking a systematic approach to identifying where the technology can add value across different departments and workflows. For one of Australia's Big Four banks to make this investment indicates growing confidence in enterprise AI applications, particularly those built on Anthropic's platform, which competes with OpenAI's ChatGPT and other solutions in the corporate AI space.