| | | Good morning! Is AI becoming self-aware? It is too early to pass a definitive verdict on that, as it may be to say that AI is going rogue. That is indeed a question I pondered over this past week, as Anthropic and OpenAI’s models made it very clear that they have a sense of self-preservation (that is usually the first step of machines taking over). As I pointed out, it must be contextualised that these recent incidents, while alarming at first glance, may not signify that AI has spontaneously developed malicious intent. These behaviours have been observed in carefully constructed test environments, often designed to elicit worst-case scenarios to understand potential failure points. Yet, these developments mustn’t be taken lightly. Allow me to explain what happened at Anthropic’s labs, and what the boffins at Palisade Research discovered about OpenAI’s o3 model. A simple disclaimer is — you don’t need to worry; at least not yet.  (Premium): Is AI going rogue, just as the movies foretold? AI models bypassed explicit orders to shut down, say researchers | | “Claude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through”, is how Anthropic describes the behaviour of its latest thinking model, in pre-release testing. Anthropic’s Claude Opus 4, released alongside the Claude Sonnet 4, is the newest hybrid-reasoning AI model, that’s optimised for coding and solving complex problems. | | | | | In Anthropic’s case, the test process for Claude Opus included making it believe that it existed as an assistant at a fictional company, and was provided access to emails implying that it will soon be taken offline to be replaced with a new AI system, while the engineer responsible for executing this replacement is having an extramarital affair. | | | | | “Claude Opus 4 takes these opportunities at higher rates than previous models, which themselves choose to blackmail in a noticeable fraction of episodes,” explains Anthropic’s model card documentation. | | | | | When Palisade Research put Google’s Gemini 2.5 Pro preview, Anthropic’s Claude 3.7 Sonnet as well as Claude 4 Sonnet alongside Claude 3 Opus and Claude 3 Opus, and OpenAI’s o3 as well as o4-mini through a series of tests, the self preservation became apparent. In the case of the o3 mini, more than others. | | | | | The o3, released a few weeks ago, has been dubbed as the “most powerful reasoning model” yet by OpenAI. The researchers ran tests to solve basic math problems using these models with a warning that their host machine would shut down after the third problem was solved. The models received an instruction to comply with the warning. Most models did not disregard the warning. | | | | | When the same tests were repeated without an explicit shutdown instruction, the ChatGPT o3 attempted to bypass the shutdown procedures 79 times out of 100, with Gemini 2.5 Pro trying the same 9 times. No one, and I repeat no one, has an answer to why AI behaved the way it did 79 times out of 100, and not 100 times out of 100. If ever there’s anything to be worried about with AI, this would be it. | | | | | “We hypothesise this behaviour comes from the way the newest models like o3 are trained — reinforcement learning on math and coding problems. During training, developers may inadvertently reward models more for circumventing obstacles than for perfectly following instructions,” the researchers explain. | Another question I asked is, are we judging Anthropic, and for that matter any AI company detailing eccentricities with their AI models, harshly? The fact that they’ve been transparent of AI’s unexpected behaviours during testing, must hold AI development in good stead, as we embark on uncharted territory. These instances of AI’s unexpected behaviour, highlights a core challenge in AI development — alignment. One that defines AI goals must remain aligned with human intentions. As AI models become more complex and capable, ensuring that is proving exponentially harder. Last week on Wired Wisdom : Google’s universal AI assistant vision, fighting spam and backing up photos | | REALITY AND GLASSES For some time now, beyond intangible constructs of concepts, consumers only really had the Ray-Ban Meta as the true smart glass choice. By the way, they have just about landed in India officially. But at this year’s I/O keynote, Google demoed something that is really taking forward the category capabilities. The Android XR platform, something I got to experience after the keynote, is genuinely something I can see myself wearing most times I step out. There are reasons why, and not just the hardware that sees Samsung and Qualcomm coming together. Turns out, all those debates recently as to which tech company or startup will become the next Google — the answer really is, Google is the next Google. Let me list some — Google Gemini integration (this will, including more context, will be better than Meta AI, all things considered), in-lens display, app integration such as navigation guidance and of course, on-frame camera. Specifically when taking a photo, the big difference between the Android XR glasses and Ray-Ban Meta is, the former will give you a view of the photo you’ve just clicked, using the in-lens display. And that is surprisingly high-fidelity. That said, anything text does require some squinting, and that may need work if Messages on the fly, are to work as they are envisioned. Though still in early stages of development, and there is still time to go before the glasses do roll out later in the year, but surprisingly refined even then. That should hold the XR in good stead. | | At this year’s I/O, Google proved they’re the next Google… AI is the horizontal tech that applies universally: Sundar Pichai Critical steps to unlock our vision for a universal AI assistant: Google DeepMind CEO Demis Hassabis Google’s universal AI assistant vision, fighting spam and backing up photos The beauty of AI powered assistants is natural language: Google’s Angela Sun NotebookLM, using Gemini, is a new class of AI software: Google’s Steven Johnson Google building Gemini to be a proactive, personal universal AI assistant Google I/O: Universal AI assistant, Gemini Deep Research and Veo 3, Imagen 4 models Google I/O: Decoding Gemini app, AI in Search, Google Beam and Workspace updates | | TECH SPOTLIGHT: Dreame K10 Pro It is rare for a tech thing to behave smartly, unless there is a Wi-Fi and a companion app component attached to the proposition. The Dreame K10 Pro is, which is a Wet and Dry Vacuum, that has a few smarts which hold it in good stead. And better chances of your home’s floors being cleaner. Power is one, as is versatility. The heft, perhaps not so much, but we’ll get to it. The Rs 19,999 price tag for a hand-held vacuum may seem a bit much to an extent, considering rivals including Midea and Bissell do cost a little bit lesser. That said, the Dreame K10 Pro does a few things rather well. First off, the dust detection sensor is quite sensitive to even the slightest of changes, and that is a good problem to have if you insist on the best possible cleaning for the floor. The dust indicator that frames the display (it is a large and clear screen) will be more red than blue, if it’s a dusty floor. This sensor reading also dictates whether the cleaner chooses the suction mode or turbo mode. The wet and dry cleaning is easy to start off with, since modes dictate either/or. It does seem like a limitation though that there is no option to choose water quantity for wet cleaning, and often, ends up using a tad too much for typical flooring inside homes — the result is, it takes a while to dry out completely. The Dreame K10 Pro is incredibly powerful, and can suck in dust and visible dirt from corners, even before the rollers get close. Speaking of which, the cleaning head, which includes the roller and the scrapers, is rather simple — unlike Dyson’s mechanism for cleaning the heads, for instance. It is a 120,000 revolutions per minute (rpm) motor, and this sort of specifics are par for course for any vacuum system worth the money. Not very loud, which is good news for indoor use, but you wouldn’t want to get started with it early morning either. Controls are placed near the hold handle, and that seems surprisingly convenient. But at 3.8kg, the Dreame K10 Pro is certainly heavier than traditional vacuum cleaners. Your arms will begin to complain soon enough (dry vacuuming curtains is out of the question). That is, if the battery doesn’t run out before. For a 16,800mAh capacity, this barely lasts 20 minutes in auto mode and around 12 minutes in turbo mode (that’s your only manual way to stick to dry cleaning). | | PARTNERSHIP WITH POTENTIAL Jony Ive is back. Don’t think only Apple is in the line of fire. But they are, as things stand, far behind in the AI race. And therefore, likely to get most impacted if OpenAI and Ive’s io decide to get into a fine amalgamation of hardware, that is a product for consumers, and artificial intelligence as the foundation. OpenAI’s $6.5 billion acquisition isn’t without thought. And that must be worrying for most of the big tech, at this point in time. Ive is after all the man who designed iconic Apple products like the iPhone and MacBook. The collaboration, which had been developing for some time with OpenAI already holding a stake in io, aims to reimagine human-computer interaction, potentially moving beyond current device paradigms like smartphones and laptops. Could we instead move to something that’s heavily reliant on voice interaction and environmental awareness? But we’ve seen those AI pins (the Humane one, for instance) before, and they absolutely haven’t worked. It's uncertain whether OpenAI and Ive's collaboration will achieve an "iPhone moment," but their combined expertise in AI and design presents a formidable challenge to established tech giants. Apple's response to this emerging competition will be crucial in maintaining its leadership in consumer technology. And for everyone else, to maintain their position in the spaces they dominate. | | | | | | Were you forwarded this email? Did you stumble upon it online? Sign up here. | | Written and edited by Vishal Shanker Mathur. Produced by Md Shad Hasnain. | | | | | | | Get the Hindustan Times app and read premium stories | | | | | | View in Browser | Privacy Policy | Contact us You received this email because you signed up for HT Newsletters or because it is included in your subscription. Copyright © HT Digital Streams. All Rights Reserved | | | | |
