Because the scale of change AI is poised to deliver is so great, our desire to know exactly what’s coming and when is equally urgent.
It’s also inevitably speculative.
Various industry leaders and pundits have recently asserted that artificial general intelligence — a hypothetical stage at which a computer can perform any cognitive task at a human level — will arrive within three years. Instead of speculating on AGI, I frame AI’s progress with a different question.
Will AI be significantly more useful, more powerful, and more integrated into every aspect of life three years from now?
The answer is: yes, of course.
For proof, consider recent history. In the spring of 2022, OpenAI released key updates to GPT-3 that dramatically boosted the model’s coherence compared with previous versions. Still, the lack of an accessible interface meant that most people had a better chance of naming all nine Supreme Court justices than explaining what GPT-3 was or how they might use it.
Six months later, OpenAI released ChatGPT and changed everything. With a simple chat interface, OpenAI made advanced language models usable by virtually anyone. This catalyzed a new era of wide deployment, rapid innovation, and accelerating feedback loops. It also fueled a flood of investment and interest across sectors, making “AI assistant” functionality a default expectation in software, from word processors to customer service portals.
The question was no longer whether large language models could be useful but how quickly they might transform the economy. Serious conversations about AGI became less hypothetical. Following the release of GPT-4 in March 2023, predictions that AGI might arrive within two or three years began to move from the fringes to the mainstream.
That hasn’t come to pass yet, but consider all that has. In 2022, AI models could generate images or process text, but they couldn’t do both in a seamless unified workflow. Today, multimodal systems like GPT-4 and Gemini integrate text, vision, and audio capabilities in ways that make interactions feel more natural and even trivial — which, as the TV remote control taught us, is exactly how massive transformation starts.
In 2022, models forgot everything from one session to the next. Today, memory features allow continuity across conversations and tasks, and AI increasingly adapts to you — following your instructions, fine-tuning on the documents and other media you give it, and performing workflows that once required juggling multiple tools or doing everything yourself.
GPT-3’s context window in 2022 was just 2,048 tokens, or approximately 1,500 words, meaning it could “remember” only a few pages of prior text within a single interaction. That was enough for answering basic questions or maintaining short conversations, but it often lost the thread in longer interactions or failed to connect ideas across sections. In effect, it had the memory of a goldfish.
Many of today’s best models have, metaphorically, the memory of an elephant — or even a small herd of them. Google’s Gemini 2.0 Flash has a 1 million-token context window. Llama 4, Meta’s newest model, has 10 million. With this exponential increase in capacity, these models can easily process and analyze multiple books and technical manuals in a single prompt. They can track long, multistep conversations without losing the thread and even work through detailed legal contracts or computer source code while preserving coherence and relevance over hundreds of pages.
If you’re a writer, you can feed a full book-length draft into a model, then ask it to identify inconsistencies in argument structure, tone, or factual claims. If you’re a software engineer, you can drop an entire codebase into a prompt and debug a persistent issue in a single pass. What’s changed isn’t just how well a model can “think.” It’s how much it can think about at once.
All of these attributes are essential for making AI feel truly adaptive, personalized, and context-aware.
Put memory and multimodality together, and you get more than just an incremental upgrade. Today’s AIs are already fundamentally different from their recent predecessors in how they process inputs, handle context, and track time. Even without exponential leaps in underlying pattern-matching capabilities, the shift is underway from autocomplete to coauthor, from database to chatbot to trusted confidante and creative partner.
Imagine what happens when multimodal models can ingest and even generate video during interactions. Current models simulate empathy and tone based on prompts. But with better conversational memory and the ability to detect affect via vocal inflection, facial expressions, word choice, and syntax, future AIs will be able to effectively respond to your varying emotional states — if you desire that.
Thanks to such advances, interacting with AI in 2028 will feel as qualitatively different from today as today does from 2022. Even if we don’t reach AGI in the sci-fi sense, we’ll be living in a world that feels increasingly like science fiction. More of us will utilize machines to effectively manage our side hustles. We’ll rely on AI health advisers that schedule appointments based on subtle signs — before we even notice any symptoms. We’ll watch DIY blockbusters we cowrote with our digital doubles.
In 2025, you may think you can easily live without turning your selfies into Ghibli portraits or having an AI summarize your meeting notes. In 2028, going an entire day — or even a few hours — without engaging with an AI will be as inconceivable as going without your phone or internet is now.
Reid Hoffman is the cofounder of LinkedIn, Inflection AI, and Manas AI and a partner at Greylock. He’s the cohost of the podcast “Possible” and the coauthor of “Superagency: What Could Possibly Go Right With Our AI Future.”