Artificial intelligence

AI models make decisions for reasons that no one fully understands. As a result, AI models are often thought of as a black box—data goes in and bodies of text, generated images or videos, and more come out. 

But if researchers can’t understand why models do what they do, it’s difficult to fix them when they generate bad or useless information. That’s where Neel Nanda, 26, focuses his work: “I see my job as: Do research such that by the time we make human-level AI, it is safe and good for the world.” Nanda leads a team at Google DeepMind working on a subfield of AI safety called mechanistic interpretability, often shortened to “mech interp,” which involves using mathematical techniques to better understand what an AI model is doing internally.

A popular approachis to divide an AI model into layers of computation, and use tools called sparse autoencoders to pull out traits and concepts the model is implicitly learning within each layer. Last year, Nanda and other Google DeepMind researchers published Gemma Scope, a collection of over 400 sparse autoencoders. Each was trained on Google’s Gemma 2 models to represent a distinct concept that Gemma interprets in pieces of text. This publicly available collection, which can be demoed online, allows researchers to get a kind of X-ray view into the behavior of the Gemma models, uncovering associations the models made completely on their own.

Nanda got into AI because of his growing concern about how quickly artificial general intelligence, or AGI, could arrive—which he believes could pose major risks without a proper understanding of how to make it safe. He believes getting more people involved in the field is critical to ensure people understand AGI before they build it. To this end, Nanda writes explainers on mech interp, makes YouTube walkthroughs, and works as a mentor in the independent ML Alignment & Theory Scholars program. 

Nanda suspects this outreach has helped popularize mech interp as a field. “I’ve seen professors complaining on [X] that too many of their PhD applicants want to do mechanistic interpretability,” he says. “I like to think I helped.”

Generative AI models make mistakes—they can confidently get facts wrong or combine bits of truth into nonsense. These so-called “hallucinations” occur even if AI models train on vast amounts of true information. 

For Akari Asai, 30, this is a big problem, especially when the facts matter, such as in scientific research or software development. The solution, she says, is to stop focusing on making bigger and bigger models that simply spit out answers in response to prompts. “We need to do a transformative switch from just scaling monolithic language models to developing augmented language models,” she says, meaning models that can interact with other entities and analyze their own outputs and behavior. 

Asai works on retrieval augmented generation (RAG), a technique for language models that makes them consult stored reference materials, called a datastore, before generating a response. Checking a datastore can help the model detect when it’s about to generate a falsehood. Then it can correct its response using the retrieved information.

Self-RAG, a behavioral framework that Asai and co-authors introduced in 2023, takes this approach a step further by having the model work with different parts of the datastore in parallel to decide which are most relevant. Self-RAG doesn’t completely prevent hallucinations, but it tries to limit them while also making sure the machine doesn’t sound like it’s reading from an encyclopedia. From their team’s testing, Self-RAG trained on Meta’s Llama can answer short-form questions 10-25% more accurately than Llama with plain RAG, depending on the type of questions, and the improvement over Llama without any RAG is even more stark. 

Asai, who just completed her PhD at the University of Washington and will begin a professorship at Carnegie Mellon University in 2026, is also building custom datastores, which could yield better fact-checking results than general databases like Wikipedia. So far she and her colleagues have built datastores for scientific literature, with 45 million papers, and coding, with 25 million documents. She wants to explore how this approach could work with sensitive biomedical data, too. 

This post has been updated to correct Asai's affiliation.

In December 2022, as he was finishing his PhD in artificial intelligence at UC Berkeley, Tim Brooks sensed the time was right to make his mark on AI video generation. With the launch of ChatGPT a month before, generative AI was having its moment. Through simple prompts, users could engage chatbots in fluid conversation—and some such systems could create high-resolution, realistic images, too. AI-generated video, however, still didn’t really work. Early models had made some progress in simulating specific scenery or giving still images a bit of motion. But a high-quality, generalized model remained elusive. 

When Brooks, 31, joined OpenAI soon after, the race was on: Together with Bill Peebles, a former Berkeley colleague, he began to engineer a model that could generate high-definition clips up to a minute long. Their strategy involved a novel way of breaking images and videos into smaller bits of information, which allowed them to train their model on a broader range of visual data. They also leveraged a transformer architecture similar to what underpins most chatbots, which enabled their model to get progressively better as it scaled. The end result was Sora, a groundbreaking photorealistic AI video generator released to the public in December 2024. 

Like similar products now available from Google, Meta, and others, Sora caused alarm as well as fascination. Some critics worry these models will lead to job losses across advertising, film, and other creative industries; and most Sora users are prohibited from making videos that depict real people, due to fears over deepfakes and the spread of disinformation. There are also worries that it will lead to the production of more “AI slop,” low-quality content that’s proliferating across the internet.

Brooks, however, believes these tools will open up new possibilities for digital creators. He also sees them as an important step toward the goal he’s pursuing in his new role as a research scientist at Google DeepMind: building a more far-reaching “world model” that will help AIs better understand our physical surroundings and more closely approximate the functioning of the human brain. 

ChatGPT is fluent in text, audio, and images, capable of taking prompts in one format and generating results in another. Much of this fluency is attributable to Mark Chen, 34, who is now chief research officer at OpenAI.

After joining the company in 2018, Chen led a team that pioneered the techniques many leading AI models now employ to ingest and generate visual data. In particular, he figured out how to adapt the transformer architecture, which researchers had successfully used to generate natural language, to handle images. The pixels that make up an image, it turned out, could be encoded as a series of tokens, similar to words in a sentence.  

“Once you have this representation that treats images like a strange language, then you can use it in the transformer,” says Chen. The team incorporated its method first in ImageGPT, released in 2020, which was followed by the DALL-E series. Now, they’ve deployed it in GPT-5, OpenAI’s flagship model.

Besides his work on images, Chen also spearheaded Codex, OpenAI’s model that generates computer code from prompts. Even though code is written in text, a model that produces it is held to a different standard than other language models—because the code it produces must perform the desired function when executed rather than sound correct. 

Now, Chen leads OpenAI’s effort to make a model capable of more complex reasoning than earlier iterations were. The company’s strategy is to have the model slow down and break a prompt into steps, known as chain of thought, which OpenAI first demonstrated in the release of its o1 model in 2024. Chen aims to soon build models to underpin agents that work autonomously for long periods of time to generate more-nuanced outputs, such as a research plan to carry out a science experiment. 

In his new position, Chen also works on product safety. A safe AI model, he says, is one that does what the user wants, without “going rogue,” such as by sending emails to people without the user’s consent. He will also have to contend with criticism of the company’s models for exhibiting cultural and political bias, and ongoing lawsuits about intellectual property infringement in its training data. 

English is spoken by less than 20% of the world’s population, but some experts estimate that it accounts for over 90% of the training data used to build large language models. The result is AI models that perform worse in the roughly 7,000 non-English languages spoken around the world, reinforce the cultural norms and values espoused in the English-language data, and create hard-to-detect harm. 

As a senior research scientist at Google Research, Sunipa Dev, 32, is trying to change that with more inclusive, multilingual, and multicultural datasets to train and evaluate AI.

Starting in 2023, Dev and her colleagues published a multilingual and multi-regional dataset of stereotypes, a pair of papers called SeeGULL, which made up the largest dataset of its kind at the time. Using a combined methodology of synthetic and community-contributed data, they include examples from 178 English-speaking countries, as well as 20 non-English languages in 23 regions. 

To ensure that generative AI’s outputs are relevant to local users, her team worked with individual data annotators around the world, including in the Middle East. In some underrepresented regions, including across India, Latin America, and sub-Saharan Africa, they partnered with local nonprofit organizations, UX designers, and others to include additional insights. 

Google is already using SeeGULL datasets to evaluate how well its LLMs are able to avoid reproducing harmful stereotypes. It’s also publicly available for broader AI safety evaluations. Since SeeGULL is open-source, Dev and her peers hope it will ensure that the concerns of non-Western communities are included in AI safety testing.

Dev is looking to expand the reach of her mission by helping foster a community of like-minded AI practitioners. The ultimate hope, she says, is that, in the next five years, 90% of the speakers of the world’s major languages will be able to access coherent, relevant, safe, and ultimately beneficial AI; and that one day that number will get closer to reaching everyone. “Artificial intelligence has to be globally intelligent,” Dev says, “and not just intelligent in some contexts.” 

In March this year, Yichao “Peak” Ji appeared in a launch video that quickly went viral. Speaking in fluent English, the 32-year-old introduced Manus—an AI agent developed by Chinese startup Butterfly Effect. It’s built on top of various models, including Anthropic’s Claude.

“This isn’t just another chatbot or workflow,” Ji says in the video. “It’s a truly autonomous agent that bridges the gap between conception and execution.”

As AI agents become Silicon Valley’s latest obsession, what sets Manus apart is its promise of true autonomy. Most agents require constant supervision, but Manus is designed to operate independently—navigating tasks, adapting mid-process, and even responding to new instructions without needing a restart. Users can close their laptops and come back to completed work.

Within a week of its debut, Manus attracted a waiting list of around 2 million users. The hype quickly translated into funding: A $75 million round brought Butterfly Effect’s valuation to around $500 million. The launch energized China’s startup scene and drew attention to a wave of AI applications emerging from the country.

Ji’s approach is rooted in years of building tools that combine technical depth with real-world usability. A longtime open-source contributor and product obsessive, he has been shipping software since high school—most notably Mammoth, an iPhone browser that briefly became the most downloaded of its kind in China. In his early 20s, he secured backing from Sequoia Capital and ZhenFund to launch Peak Labs, where he developed Magi, a knowledge search engine powered by a custom language model. It’s inspired by frontier AI research and constructs knowledge graphs, a “mind map” showing interconnected knowledge.

But Ji stands out not just for his products but also for what he represents: a new generation of Chinese technologists who are global in outlook, steeped in open-source culture, and fluent in both technical architecture and product vision.

Now based in Singapore, Ji is part of a new wave of founders building world-class AI from outside the traditional US tech hubs. Manus itself is a product of global fluidity—developed by a Chinese team, powered by US models, and designed for a worldwide user base.

Large language models are brilliant with words, partly thanks to the huge amount of text data they’ve been fed. But the frontier of AI is building systems that can interpret the world around them beyond text—seeing, hearing, and responding to complex situations. It’s an ambitious leap, and one of the field’s biggest challenges. One of the top researchers dedicated to pushing this frontier further is Northwestern University’s Manling Li, 33.

Li’s work focuses on a key challenge in AI: translating language into real-world action. While traditional AI systems specialize in a single type of input—like text—Li makes systems that integrate perception, reasoning, and action. She created a framework that can allow AI to piece together what is happening from multimedia information, such as images, audio, and video, as well as text. The ability to “perceive” across various data formats is essential for building AI that can make more well-rounded judgements in the real world.

Rather than just identifying what happens in AI systems’ surroundings, her work helps the  systems “understand” why things occur and how they connect. Instead of just tagging objects in a video or picking out keywords from a sentence, her systems can follow what’s happening, figure out how different actions are related, and explain why something occurred. This transparency is increasingly crucial as AI systems make more consequential decisions in our daily lives.

Her work is already being used beyond the lab. Government agencies, including DARPA, have adopted her systems, and through open-sourced tools, she has helped make advanced AI techniques more widely available. She’s also created new benchmarks for evaluating AI performance in real-world settings, such as navigating physical environments and answering complex questions about what’s going on in a video.

As AI becomes embedded in everything from smart assistants to autonomous vehicles, Li’s work ensures these systems are not only powerful but also trustworthy and transparent.

In the early 2020s, when the large language models (LLMs) that power ChatGPT and other chatbots were under development in Silicon Valley, Maithra Raghu was thinking a step ahead. As an AI research scientist at Google, Raghu, 34, saw the appeal of using these models to create platforms that could engage in human-like conversation. But she was most excited about the possibilities this tech held to automate more research-intensive tasks, such as the data-gathering and analysis that underpins the world of finance.

But ChatGPT and other general-purpose LLMs weren’t great at handling specialized, real-time information. So Raghu, encouraged by friends working in finance, set out to build her own: Her startup, Samaya AI, launched in 2022. 

The company’s first AI-powered tool functions like a personal research assistant: It scours the web and users’ internal data to deliver research and analysis, which it can output in a variety of formats, including reports and presentations. It’s now in use at several financial institutions, including Morgan Stanley. 

Unlike most general-purpose chatbots, which rely on one large LLM, Samaya uses several specialized smaller models built in-house. They’re trained in a way that allows them to evolve together, improving their ability to retrieve high-quality information, extract insights from it, and place it all in context. This approach, Raghu says, enables tools that can more accurately sift through large volumes of data, thereby minimizing the risk of hallucination, or the proffering of false information.

For now, Samaya’s tech is being used primarily by research analysts: It can pinpoint a single figure buried in reams of documents, analyze more sources than they ever could, or keep tabs on real-time information. Yet early tests suggest the models might also be used to make automated performance forecasts—of companies or the wider economy.