An easy way to think about AI tools is to separate them into specialized “media” systems and general-purpose chatbots.
In media tools, you have image generators, such as Midjourney and Google’s Imagen; AI text-to-video models are led by OpenAI’s Sora and Google’s Veo 3. Suno and Udio are market leaders in AI-generated music, and both companies are currently in discussions with major music labels—including Universal Music Group, Warner Music Group, and Sony—regarding licensing. ElevenLabs, which released that new music generation tool last week, is most notable for its voice AI, which allows users to clone their voice or generate a new one for narration and voice acting.
“General” AI tools are those capable of performing a wide range of functions—notably, research, writing, and coding—and the most familiar versions of these are chatbots. At their core, these are large language models (LLMs): computer programs trained on vast amounts of text, enabling them to recognize patterns in language and generate responses that sound like a human wrote them. You give them a “prompt”—your question or instruction—and the AI produces an output of words that it thinks will answer the prompt.
However, the latest chatbots are now multimodal, meaning they’re capable of understanding and producing information presented in a large variety of forms. You can share a photo with Anthropic’s Claude, have a call with OpenAI’s ChatGPT, or livestream video to Google’s Gemini. They can also search the web, generate images, and write code. At this point, developers can create entire functional apps without writing a single line of code themselves.
Even the classic text function has undergone fundamental changes, with the development of reasoning models—which “think” through their answers first, in lengthy trains of thought—and research modes, which allow the AI to spend time reading hundreds or thousands of sources and synthesizing that information into a lengthy report.
Earlier chatbots weren’t just far less capable but also far less reliable, often making up—or “hallucinating”—information. The ability to “hallucinate” is necessary for LLMs to function—it seems like a byproduct of linguistic creativity—but there are new measures to reduce its frequency, such as “temperature” controls on hallucination and error-checking protocols. With regular use of a model, users can also begin to identify the types of prompts that will provoke the model to hallucinate, and how.
The leading general AIs are OpenAI’s ChatGPT, Anthropic’s Claude, Meta’s Llama, and Google’s Gemini. All have their quirks and advantages, and determining which is the most competent for coding, answering general questions, and solving complex, reasoned responses largely depends on which was most recently updated. However, there are many more chatbots beyond the Big Four. Inflection makes the conversational Pi.AI, France boasts Mistral, and Mr. Musk has his infamous and frequently tagged Grok 4, to say nothing of Chinese alternatives such as Baidu’s Ernie, Alibaba’s Qwen, ByteDance’s Doubao, and DeepSeek R1. These Chinese models are not yet as powerful or capable as the leading Western models, but they’re not far off.
Perhaps the most interesting point is not simply that the tools are changing, but so are their uses. The first publicly accessible version of ChatGPT, released in late 2022, was largely a tool to cheat on university essays, but in just a few short years, its use cases have exploded. Whereas AI companies used to warn that their tools couldn’t provide medical advice, OpenAI’s GPT-5 rollout prominently noted that it was “the best model yet for health-related questions.” AI therapy is growing in popularity, as is chatting to the model as a digital friend. Following last week’s GPT-5 rollout, many emotionally-attached users complained to OpenAI about being barred from their more personable, informal model, 4o. “Losing this direct access would mean an irreversible emotional loss for me,” one user wrote. “It’s mentally devastating.” The sentiment was widely shared, and in response, OpenAI allowed paying users to access older models.
Where does this all go? The goal of every AI company is to build artificial general intelligence (AGI): an ill-defined term that refers to an AI’s ability to perform any task a human could, better than a human would, and therefore replace or dramatically augment human labor. But how far we are from this milestone remains unclear. If we are to believe the widely circulated assessments by former OpenAI developers Daniel Kokotajlo and Leopold Aschenbrenner, AGI will arrive in 2027 as AI tools start to code themselves, leading to exponential, unstoppable self-improvement.
Kokotajlo’s report is a speculative forecast of how that breakthrough could affect politics and Western societies, either producing a spacefaring, robot-powered, science-fiction future or an extinction-level apocalypse. His assessment predicted that, by Q3 2025, companies would start publicly releasing AI agents—AIs that autonomously plan, decide, and act toward goals, using their own tools and scrolling through browser windows, just as a person does—and right on schedule, OpenAI, Anthropic, and Manus have all released agentic models. They’re clunky and unreliable, but as the last years of AI development have taught us, this won’t be true for long.
Skeptics of such reports, and the “doomer” concern about AI acceleration, argue this talk is hype in an AI bubble —that you can’t justify multi-billion-dollar pre-money valuations unless you claim to be building the most important tool ever—and that the current development trajectory doesn’t support the concerned view. On Saturday, David Sacks—the White House’s AI and crypto czar—released a lengthy X post arguing that the models have been improving at a reliable and consistent pace, becoming cheaper and faster, but without a feared intelligence explosion. The Trump administration broadly seems to share its czar’s view, unveiling an “AI Action Plan” in late July, which sought to “remove red tape and onerous regulation” to hasten development and guarantee America’s AI dominance.
But as a reminder, ChatGPT came out less than three years ago. The future is approaching quickly.
