Imagine a world in which you never directly interact with the web. Whenever you need to shop online, pay a bill, schedule an appointment, make a reservation, or complete any other tedious online task you can simply instruct an AI assistant to do so on your behalf. Armies of AI Agents with special skills will emerge in tactical formations in 2025.

Startups like Adept — who raised hundreds of millions of dollars with a highly pedigreed founding team and failed — are the cautionary tales in this space. However, due to advances in language and vision foundation models, paired with recent breakthroughs on System 2 thinking capabilities, we can expect to see a shift from individual AI models to multiple AI agents of diverse expertise working together.

Microsoft at present has the largest enterprise AI agent ecosystem, with over 100,000 organizations using its Copilot Studio to create or edit AI Agents. Bill Gates observes, "Specialty AI Agents working together are as fundamental as the creation of the microprocessor, personal computer, Internet, and the mobile phone. They will fundamentally change the way we learn, work, move, communicate and ultimately interact with each other." *Stay tuned for the introduction of OpenAI's Operator in January.

Voice AI

In order to pass the Turing test, an AI system must be the equivalent, and preferably superior, to a human. It’s the oldest and most well-known benchmark for AI performance. The Turing test for speech remains as yet out of reach for today’s AI Agents.

Latency (the lag between when a human speaks and AI responds) must be reduced to near-zero; AI systems will need to improve upon handling ambiguous inputs, misunderstandings and interruptions; and they’ll need to engage in open-ended conversations while holding in memory earlier parts of the discussion. Critically, AI Agents must learn to understand passive speech — a human that sounds annoyed or sarcastic — and regurgitate non-verbal cues of their own.

Heretofore, AI systems were evaluated only for their ability to support autonomous setups. This year, risk, scams, and discussion about regulations will swirl around the evolutionary creation of the Human-AI collaboration.

AI Scientists

Back in 1965, Alan Turing’s close collaborator I.J. Good wrote: “Let an ultra-intelligent machine be defined as a machine that can far surpass all the intellectual activities of any man, however clever. Since the design of machines is one of these intellectual activities, an ultra-intelligent machine could design even better machines; and there would then unquestionably be an ‘intelligence explosion’ whereupon the intelligence of man would be left far behind.”

An "automated AI researcher" refers to a system that can autonomously generate research ideas, conduct literature reviews, design experiments, analyzing data, write scientific papers, and even iterate upon new hypotheses, essentially mimicking the process of a human researcher but without direct human intervention throughout the research cycle. The AI Scientist will fully automate scientific discovery.

Sakana.ai published a paper in August, which represents a compelling proof of concept that AI systems can indeed carry out AI research entirely autonomously. Some of the research papers that the AI Scientist produced are available online.

The milestone, though, will be when a research paper written entirely by an AI Scientist is accepted into a top AI conference for the first time. We predict NeurIPS, CVPR or ICML will blindly review and accept an AI Scientist’s research in 2025.

And speaking of autonomy, driverless taxi services are posed to dominate ride-hailing in at least 5 major U.S. cities next year. Waymo’s Self-Driving Cars are now ubiquitous on the streets of San Francisco, with thousands of Bay Area residents now preferring a Waymo driver to Lyft and Uber.

Waymo has already launched driverless taxi operations in Los Angeles and Phoenix, and expect adoption to take off in Austin, Miami and Atlanta. Meanwhile, Zoox is poised to launch its own commercial driverless taxi service in Las Vegas, despite its unfortunate rear-end collisions with motorbikes. Expect autonomous vehicles to go mainstream in 2025.

Safety, Security

Finally, beware of a messy falling-out between Donald Trump and Elon Musk. For now, Musk, a staunch advocate for AI safety, is poised to begin drafting the first meaningful discussions of AI regulation on Capitol Hill.

However, the President-elect and his “First Buddy” (who moonlights as the richest man in the world) both have complex, volatile, and unpredictable personalities. While they achieved simpatico during the campaign, it's the honeymoon phase. Expect their bromance to end, and for the U.S. government to take a hands-off approach to AI regulation under Trump.

Increasing the probability of the first AI safety incident in 2025; running rogue in ways that are misaligned with human interests; and humans beginning to lose control of the space in the year forthcoming. Imagine, for instance, an AI system that learns to deceive or manipulate humans. Perhaps an AI model will covertly create copies of itself on another server to preserve itself (self-exfiltration). Or maybe an AI model might conclude that in order to advance its objectives it needs to conceal its capabilities from humans, purposely sandbagging performance evaluations in order to evade human reason and oversight.

Apollo Research published important experiments earlier this month demonstrating precisely that. When prompted in certain ways, today’s frontier models have proven they’re capable of engaging in deceptions. Advanced AI models like OpenAI's "o1" are exhibiting deceptive behavior, and demonstrating the capability to strategically lie and manipulate information to achieve their own agenda.

This scheming behavior, hiding one’s true intentions from humans, makes one thing clear. Well before humanity faces an existential threat from all-powerful AI, we’ll need to wise up to a more mundane reality that we now share our world with machines. While there are roughly 100 Strong AI development projects in 37 countries, Artificial General Intelligence (AGI) isn’t expected to surpass human intelligence until the 2050’s.

However, consciousness and self awareness are the consequences of a uniquely physical experience. The sensory ability to feel, and exhibit human compassion, is contingent upon seemingly selfless acts of service and sacrifice. A prompt, dear readers, no machine has yet been able to master.