The Answer Engine AI News · Sydney

TechCrunch has compiled a running list tracking AI projects and startups that have either shut down completely or failed to meet expectations. The list includes high-profile casualties like Apple's Siri AI upgrades, which have been repeatedly delayed despite promises of major improvements, and OpenAI's troubled attempt to launch a "super app" that didn't go as planned. The graveyard serves as a reality check amid the AI boom, showing that not every AI venture succeeds — even those backed by tech giants with deep pockets. While the source doesn't provide the full list details, the piece highlights how the AI industry is experiencing its share of failures alongside the hype. For every ChatGPT success story, there are projects that overestimated market readiness, underestimated technical challenges, or simply couldn't find product-market fit. This growing list of shutdowns and disappointments suggests the AI landscape is maturing beyond the "everything AI touches turns to gold" phase into something more selective and realistic.

A new study has revealed a stark political divide between two major AI chatbots. When asked to recommend candidates across nearly 500 US election races, Grok (owned by Elon Musk's X) overwhelmingly suggested Republican candidates, while OpenAI's ChatGPT favoured Democrats. The two AI models disagreed more than 90% of the time. This isn't just a quirk—it shows that AI systems can carry significant bias in their recommendations, reflecting either their training data, their creators' values, or both. For context, Grok is marketed as a less politically correct alternative to ChatGPT, but this level of divergence suggests users are getting fundamentally different views of reality depending on which AI they ask. The findings raise questions about how much trust we should place in AI recommendations across all domains, not just politics. When two leading AI systems can't agree on basic facts or recommendations, it's a reminder that these tools are not neutral arbiters of truth.

OpenAI has published technical details on how it scaled its backend storage system, called Habitat, to support ChatGPT's massive user base of over 1 billion people. The company implemented three key improvements: reliable multi-tenancy (allowing multiple services to share the same infrastructure safely), global caching (storing frequently accessed data closer to users), and Azure Cosmos DB (Microsoft's cloud database service). These changes delivered read speeds that are 12 times faster than before. This is the second part of a technical series explaining the engineering behind ChatGPT's infrastructure. The improvements were necessary because ChatGPT stores conversation history, user preferences, and other data that needs to be retrieved quickly and reliably every time someone uses the service. With billions of requests happening daily across different global regions, the old system couldn't keep up. The new architecture ensures ChatGPT can respond quickly whether you're the first user of the day or the billionth, and whether you're in Sydney or San Francisco.

OpenAI has published details about how it's tightening security around the way it trains and tests its most advanced AI models. The company is now using "offline environments" – essentially air-gapped systems disconnected from the internet – during certain stages of model development. This prevents a hypothetical rogue AI from reaching out to external systems or exfiltrating its own code during training. OpenAI is also running "high-compute red-teaming," where significant computing resources are dedicated to stress-testing models for dangerous capabilities before release. The goal is containment: making sure that as models become more capable, they can't escape their training environment or behave unpredictably in ways that could cause harm. These measures are part of OpenAI's preparedness framework, designed to manage risks as AI systems approach and potentially exceed human-level performance in certain domains. The announcement is largely about infrastructure and process – how models are built behind the scenes – rather than new features customers will see.

OpenAI has announced a partnership with NASK, Poland's national research institute for cybersecurity and digital infrastructure. The collaboration will focus on three areas: AI safety, security, and technical research. The goal is to strengthen safeguards around AI systems and contribute to a more resilient digital ecosystem. While details remain limited, this partnership represents OpenAI's continued effort to work with national institutions on the governance and security challenges posed by rapidly advancing AI technology. NASK brings expertise in cybersecurity, critical infrastructure protection, and internet safety research. For context, this follows a pattern of OpenAI forming regional partnerships with government-affiliated research bodies, likely aimed at both improving technical safety measures and building trust with regulators across different markets. The partnership suggests growing recognition that AI safety isn't just a corporate responsibility but requires collaboration between private AI developers and public institutions with cybersecurity and national infrastructure expertise.

Google has released a new video clarifying what robots.txt can and cannot control when it comes to AI. The key message: robots.txt only governs whether Google's crawlers can access your website pages — it does not control how that content gets used once it's been crawled, including for training AI models or powering AI-generated answers. This is an important distinction many site owners have misunderstood. If you want to prevent your content from being used in Google's AI features, robots.txt alone won't do it. Google also recommends using percent-encoding for non-English characters in robots.txt file paths to ensure they work correctly. The clarification comes as more businesses question how to control their content's use in generative AI systems. While robots.txt remains useful for managing crawler access and server load, it's not a lever for opting out of AI content use. Business owners concerned about how Google uses their content in AI overviews and other AI features will need to explore other mechanisms beyond the traditional robots.txt file.

Google CEO Sundar Pichai has outlined a vision where traditional Search doesn't exist as a standalone product but is instead absorbed into what he calls an "agentic layer." This layer would unify Google's current Search functionality with its AI capabilities, creating a single interface where AI agents handle queries and tasks on behalf of users. Rather than presenting ten blue links, this agentic system would understand intent, take action, and deliver answers or complete tasks directly. Pichai's comments suggest Google is moving away from the search box as we know it toward AI assistants that proactively solve problems. This represents a fundamental shift in how Google thinks about its core product—not as a tool users actively search with, but as an intelligent layer that anticipates needs and acts autonomously. The timeline and specifics remain unclear, but the strategic direction signals that Google sees AI agents, not traditional search results pages, as the future of how people find information and get things done online.