2025 In Review: Technology - When AI Hit Escape Velocity

 THE SIGNAL FOR the age of Artificial Intelligence (AI) was the launch of ChatGPT in 2022. It might be tempt­ing to call 2025 the beginning of its mature phase but that does not consider just how much run­way AI has left. It is still in its infancy taking its first steps. The year did however begin with something unexpected and it was not so much the technology as who wielded it. Large Lan­guage Models, enormous expensive datasets on which AI is trained, were a preserve of the US tech behemoths. On January 20, a Chinese startup, DeepSeek, announced that it had cre­ated a new model at a fraction of the cost while outperforming existing ones. It even momentarily overtook ChatGPT in free downloads in Apple’s App Store for the US. The companies in the American ecosystem had first-mover advantage and bottomless capital but now, it was evident, others could catch up. The story will unfold in future with many twists but for the rest of 2025 the old guard—OpenAI, Google, Meta, X, An­thropic—went on to shape the contours of the revolution.

A key development was in these models increasingly chang­ing how they processed their responses. When generative AI first arrived on the scene, they relied on prediction. The data they were trained on told them what the next word to a sequence must be. And so, you had human-like answers, but often the AI would falter doing simple multiplication. That began to change. In February, for instance, Anthropic launched Claude 3.7, OpenAI their GPT-4.5 and X had Grok 3, all of which were now doing more ‘thinking’ to come up with answers to queries of users. They were using logic and reasoning step by step. By August, when GPT-5 was launched, this phenomenon had become more fine-tuned, and depending on the time the user allowed, it could think more and give better an­swers. An interesting offshoot of this development was that the models became more honest. If it didn’t know an answer, it wouldn’t hallucinate and lie as much as earlier. OpenAI wrote on its website, “Alongside improved factuality, GPT‑5 (with thinking) more honestly communicates its actions and capa­bilities to the user—especially for tasks which are impossible, underspecified, or missing key tools.”

A giant leap that happened in AI in the year had to do with image generation. It was tied to another phenomenon. AI could earlier come up with text responses and also create images but you had to go to different programmes for them. Now it had become multimodal. The same tool, say ChatGPT or Gemini, could create text, images and even videos. That meant image generation came into the hands of users without going through an additional layer of effort. Its potency became apparent in the viral trend of Studio Ghibli. This was when, on March 25, OpenAI launched an update to its ChatGPT-4o model with significantly improved image generation abili­ties. During the demo, selfies were turned into Ghibli-style anime art images. The internet went on to explode with simi­lar images in the days to come. It has got more than 7 billion views on TikTok alone. Sam Altman, the CEO of OpenAI, posted on X that they couldn’t handle the load. He wrote, “It’s super fun seeing people love images in ChatGPT, but our GPUs are melting. We are going to temporarily introduce some rate limits while we work on making it more efficient. Hopefully, it won’t be long!” But the image war was eventually won by Google in August when its own image generation tool Nano Banana, which could be used in their AI platform Gemini, proved to be more advanced than anyone else’s. So confident was Google that it didn’t even have a high-profile launch and allowed the popularity to swell from the ground up as people started being amazed by its features. Nano Banana could deliver results faster, it was extraordinary at editing and modifying images according to the demands of users. The model could scour online and bring that knowl­edge while executing commands, and it could maintain character and style across multiple changes that a user prompted. An updated Nano Banana Pro in November solved the very prickly problem of generated images not being able to get spell­ings accurately. Combined with high-resolution outputs, it meant someone could create professional-level infographics, posters, and ad­vertisements using nothing but the AI from their phone or laptop.

AI also saw great progress in video generation with OpenAI’s Sora and Gemini’s Veo3, though it is still taking baby steps. Elon Musk has been furiously investing money and resources into Grok to become an AI leader. It also came out with an Imagine tool. It had fewer safety filters in keeping with Musk’s adherence to free speech, and that led to some rough weather because of a Spicy mode that could be used to make sexually explicit videos. There were media reports of it being used to make such videos of real people without consent. Tech pub­lication The Verge did an article on how Grok had been used to make deepfakes of Taylor Swift. However, X has kept going with it. A post by Musk that soon AI will be able to spot deepfakes and trace them to source still did not address what was already happening using the tool.

Another area that showed progress was in audio conversations with AI assistants now available on personal devices. ChatGPT and Gemini became good enough to have regular conversations with in near real time. A change in all these voice assistants was that they became better at deciphering tone and intent of speak­ers, and instead of converting interactions to text, processing them and then responding with audio, it was doing it directly via audio, like human beings. A large number of people were even using these AI voice assistants for personal therapy uses. AI also made a leap from being just a responder to actually doing tasks. Agentic AI, as this development is known, has made itself appar­ent in different mediums, from individuals to businesses. If you have to, for instance, go on a trip, just a simple command can now make most GenAI bots look up flight options and compare fares.

While many big developments were happening at OpenAI, Gemini and X, an area in which Meta showed how it is planning for the future was in wearables. AI is already present in smart­watches and fitness trackers, but Meta brought an iteration of its smart glasses that had AI integrated into it. A small, unobtrusive screen in the glass itself served as the display interface. The glasses came with a band that had to be worn around the wrist, enabling the use of finger gestures for navigation. The AI could see what the wearer of the glasses was seeing and could answer questions on it. For instance, if you were looking at a flower or a dish, it could identify it. In a new locality, it could provide visual directions to an address. Meta also gave a glimpse of how AI could change lives radically. In November, in a blog on its website, it wrote that their AI glasses were helping people with disabilities “navigate their daily lives by enabling hands-free calling and texting, translating speech, describing their surroundings, connecting them with volunteers who can provide assistance, and more. People with dis­abilities also use our AI glasses to create photo and video content, track their fitness training, and listen to music”.

From software coding in which Anthropic’s Claude AI has made a reputation for doing in minutes what human beings would take days and weeks, to Google’s NotebookLM which has changed the nature of research by automatically taking in inputs from a variety of sources and creating explainer podcasts and mindmaps, the number of ways in which the world is being reshaped increased exponentially in 2025. And behind all of it were the silicon chips of Nvidia, which became a more valuable company than all those multinationals that were using its GPUs. Its market valuation even crossed $5 trillion at one point. It holds a near monopoly at present but the idea was shaken somewhat when Google’s TPU chips began to match its efficiency at a lower cost and companies like Meta announced that it was buying size­able quantities. Increasingly, there are also concerns about an AI bubble considering the steep market valuations of the companies leading in this arena. It remains to be seen what surprises next year has in store, but if 2025 showed anything it was that AI is now firmly interwoven with humanity.