OpenAI has yet to fulfill its ambitions of creating superintelligence or artificial general intelligence, nor has it successfully built an autonomous "AI researcher." However, it has accomplished a small victory: getting ChatGPT to refrain from misusing the em dash.

In a recent update shared on X, CEO Sam Altman celebrated this milestone, stating, "If you tell ChatGPT not to use em dashes in your custom instructions, it finally does what it’s supposed to do!" He referred to this achievement as a "small-but-happy win." The company also demonstrated ChatGPT's new capability on Threads, where the chatbot attempted to craft a formal apology for “ruining the em dash,” though ironically, it still included the punctuation in its apology.

This situation highlights a significant nuance in OpenAI's approach. Rather than training ChatGPT to use the em dash appropriately or minimizing its usage by default, the company allows users to deactivate the em dash through the chatbot's personalization settings.

This development comes on the heels of the introduction of GPT-5.1, OpenAI’s most recent model. A key feature emphasized during the launch was the model's enhanced ability to adhere to user instructions and offer greater personalization options. Thus, the em dash adjustment serves as just one illustration of how users can tailor the model's behavior rather than a comprehensive fix to its overall output.

However, because this em dash adjustment must be configured individually, it underscores the opaque nature of many large language models (LLMs). Some users responding to Altman's post on X reported that their versions of ChatGPT continued to produce em dashes despite the explicit instruction. OpenAI’s portrayal of personalization as a solution suggests that achieving uniform resolutions across the board remains a complex challenge.

It appears that the company has discovered a method to prioritize custom instructions when generating responses, allowing for a personalized version of ChatGPT that no longer uses em dashes. Yet, it is clear that OpenAI still grapples with understanding both the root causes of these issues and why they persist. This ongoing focus on personalization and a decrease in discussions about AGI indicates the company is navigating complexities on multiple fronts.

ChatGPT Reaches an Enhanced Level of Intelligence: Avoiding the Em Dash

Prompt: A bright sunlit study with a sleek laptop on a wooden desk; on its screen is ChatGPT’s interface, and from the keyboard golden streams of glowing text and neural-network filaments rise like a blossoming tree of knowledge; the word “OpenAI” appears on the laptop chassis with no logo; warm natural light, cozy atmosphere

ChatGPT Achieves a New Level of Intelligence: Not Using the Em Dash - Gizmodo

- OpenAI's ChatGPT now allows users to disable em dashes through custom instructions.  
- CEO Sam Altman described this update as a "small-but-happy win" for users.  
- Despite improvements, ChatGPT's em dash issues persist, indicating ongoing challenges in AI development.  

OpenAI has not achieved its goal of developing a superintelligence or artificial general intelligence, nor has it cracked its planned construction of an autonomous “AI researcher.” But it has figured out how to get ChatGPT to stop misusing the em dash. So, that’s something.

In a post on X, CEO Sam Altman announced, “If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do!” He called the development a “small-but-happy win.” The company confirmed the capability to cut off the chatbot’s reliance on the punctuation mark in a post on Threads, where it made ChatGPT write a formal apology for “ruining the em dash.” Notably, the chatbot was not able to write the apology without using an em dash.

To that point, it seems like there’s a noteworthy distinction to be made here. OpenAI has not figured out how to get ChatGPT to use the em dash in a more appropriate manner or to deploy it more sparingly by default. Instead, it has simply given users the ability to tell ChatGPT not to use it, a change that can be made within the chatbot’s personalization settings.

That ability follows the release of GPT-5.1, the latest model from OpenAI. One of the primary points of improvement that the company hammered in its rollout of the new model was the fact that GPT-5.1 is apparently better at following instructions and offers more personalization features. So the em dash clampdown appears to just be one example of how users can make use of the model’s more compliant sensibilities rather than some fix to the underlying model’s overall output.

The fact that the em dash fix is something that has to happen on a user-by-user basis probably speaks to just how much of a black box most LLMs are. In fact, there are users in Altman’s replies on X showing that, despite the instruction, their instance of ChatGPT continues to spit out em dashes. OpenAI’s presentation of personalization as a solve would seem to suggest that finding a solution at scale is still really, really hard.

The company has seemingly figured out a way to weight custom instructions more heavily within its calculations when producing a response to a prompt, which can produce a result like a person’s version of ChatGPT no longer using em dashes. But it still seems like the company can’t figure out why the problem happened in the first place or persists. No wonder the company is leaning heavily into personalization and talking less about AGI lately.

ChatGPT Achieves a New Level of Intelligence: Not Using the Em Dash

In recent years, the prevalence of em dashes has emerged as a potential indicator of AI-generated content. Users have noted that these punctuation marks appear frequently in outputs from ChatGPT and other AI chatbots, leading some to believe that they can recognize AI writing simply based on its excessive use—though it can also be overused by human writers.

On Thursday evening, OpenAI's CEO Sam Altman announced via X that ChatGPT could now adhere to custom instructions to minimize the use of em dashes. “Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do!” he shared.

This update, which followed the release of OpenAI’s latest GPT-5.1 AI model just two days prior, elicited mixed reactions from users who have often struggled with getting the chatbot to follow specific formatting guidelines. This development, while labeled a "small win," raises a significant question: If OpenAI, the leading company in AI technology, has faced challenges in controlling something as straightforward as punctuation after multiple years of effort, does this suggest that what many refer to as artificial general intelligence (AGI) is still a long way off?

One user on X commented on Altman’s announcement, stating, “The fact that it’s been 3 years since ChatGPT first launched, and you’ve only just now managed to make it obey this simple requirement, says a lot about how little control you have over it, and your understanding of its inner workings. Not a good sign for the future.”

While Altman often discusses concepts like AGI, superintelligence, and his imaginative notion of "magic intelligence in the sky"—terms he uses when seeking investments for OpenAI—it is evident that reliable artificial intelligence remains elusive at this point in time.

Forget AGI—Sam Altman rejoices as ChatGPT finally adheres to em dash formatting guidelines

A bright, sunlit modern office scene: Sam Altman seen from behind, arms raised in triumph as he faces a giant wall-mounted digital display showing a ChatGPT conversation with perfectly placed em dashes (“Welcome — how may I assist you — today?”). Confetti flutters in the warm natural light streaming through large windows. The far wall features “OpenAI” text with no logo.

Forget AGI—Sam Altman celebrates ChatGPT finally following em dash formatting rules - Ars Technica

- Em dashes are increasingly seen as indicators of AI-generated text in writing.  
- OpenAI's CEO announced ChatGPT can now follow custom instructions to avoid em dashes.  
- User reactions question the reliability of AI amid struggles with basic punctuation control.  

Em dashes have become what many believe to be a telltale sign of AI-generated text over the past few years. The punctuation mark appears frequently in outputs from ChatGPT and other AI chatbots, sometimes to the point where readers believe they can identify AI writing by its overuse alone—although people can overuse it, too.

On Thursday evening, OpenAI CEO Sam Altman posted on X that ChatGPT has started following custom instructions to avoid using em dashes. “Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do!” he wrote.

The post, which came two days after the release of OpenAI’s new GPT-5.1 AI model, received mixed reactions from users who have struggled for years with getting the chatbot to follow specific formatting preferences. And this “small win” raises a very big question: If the world’s most valuable AI company has struggled with controlling something as simple as punctuation use after years of trying, perhaps what people call artificial general intelligence (AGI) is farther off than some in the industry claim.

Credit: X A screenshot of Sam Altman’s post about em dashes on X.

“The fact that it’s been 3 years since ChatGPT first launched, and you’ve only just now managed to make it obey this simple requirement, says a lot about how little control you have over it, and your understanding of its inner workings,” wrote one X user in a reply. “Not a good sign for the future.”

While Altman likes to publicly talk about AGI (a hypothetical technology equivalent to humans in general learning ability), superintelligence (a nebulous concept for AI that is far beyond human intelligence), and “magic intelligence in the sky” (his term for AI cloud computing?) while raising funds for OpenAI, it’s clear that we still don’t have reliable artificial intelligence here today on Earth.

Forget AGI—Sam Altman celebrates ChatGPT finally following em dash formatting rules

In a recent discussion, Microsoft CEO Satya Nadella provided a balanced perspective on the emerging issue of artificial general intelligence (AGI) at a time when global conversations are heating up around its implications. For those unfamiliar, AGI refers to a theoretical AI that can learn, understand, and operate with a level of intelligence comparable to that of humans.

Nadella expressed enthusiasm for AGI's transformative potential during his dialogue with Dwarkesh Patel and Dylan Patel, the founder of SemiAnalysis. However, he repeatedly emphasized familiar concepts associated with technological advancements—tools, productivity enhancement, integration into the real economy, and the necessity for diversified models instead of reliance on a single, overarching framework.

When prompted about AGI's significance in relation to the industrial revolution, Nadella did not deny its potential impact but maintained a cautious optimism, stating, “I start with the excitement that perhaps this is the biggest thing since the Industrial Revolution. However, it’s essential to remember that we are still in early innings.” He believes that while impressive advancements have been made and certain patterns are showing promise, the journey is just beginning, suggesting we might be in the initial stages of this technology's evolution.

Nadella also shared a memorable definition of AI from computer scientist Raj Reddy, which he found particularly relatable. Reddy described AI as either a "guardian angel" or a “cognitive amplifier,” a metaphor Nadella appreciates as a straightforward way to conceptualize what AI represents.

Addressing the potential of achieving AGI, Nadella shifted the focus from whether AGI is attainable to understanding its practical applications for humanity. "Ultimately, what is its human utility? It’s meant to serve as a cognitive amplifier and a guardian angel. Viewing it as a tool is crucial for understanding its purpose," he explained.

While the interviewers sought a unified vision of AGI development—envisioning a single, all-encompassing model that learns and performs across numerous sectors—Nadella offered a more nuanced view. When asked about the implications of a dominant model monopolizing value, he acknowledged that it could be true under specific circumstances, saying, “If there's only one model serving everywhere, then, yes: that's game over. The playing field would be dramatically limited.”

Yet, he is confident that the reality is more complex. “Despite the prominence of any single model, the landscape is becoming increasingly diverse. Take coding as an example; there are various models thriving at the same time. The narrative of one dominant model being widely deployed is becoming less likely,” he remarked.

Nadella believes that limitations imposed by specific domains, geographic factors, and varied use cases will hinder the emergence of a singular 'god model.' He expressed skepticism about whether such dominance could occur across all sectors simultaneously.

When it comes to the distribution of value in an AGI-driven future, Nadella indicated that both model companies and those creating supporting infrastructures, tools, and workflows would be crucial players. He noted that model creators might face a ‘winner’s curse,’ warning that even with significant innovation, they risk being just one step away from commoditization. Those who possess contextual data could potentially leverage this to advance their advantages.

Ultimately, Nadella suggests that the winners in this evolving landscape will be powerful open-source models, as well as various advanced models that effectively merge with data, tools, and workflows. Furthermore, he envisions a future not limited to human interaction or chat interfaces alone but one where autonomous agents seamlessly integrate across corporate infrastructures.

Satya Nadella enthusiastic about AGI while prioritizing human utility and challenging the 'one-model-for-all' concept.

A sunlit stage in a modern glass-walled auditorium. Satya Nadella stands center-stage in business attire, gesturing toward a softly glowing orb labeled “AGI.” Around him, diverse professionals collaborate over tablets and notebooks, embodying human utility. Behind him, a large Microsoft logo is mounted on a minimalist white wall. To one side, a looming dark monolith representing a “one-model-for-all” AI fades into shadow as Nadella points away from it. Warm natural light streams through floor-to-ceiling windows, casting optimistic highlights.

Satya Nadella excited for AGI, but puts human utility first and pushes back on the ‘one-model-for-all’ idea - The Indian Express

- Satya Nadella believes we are in the early stages of AGI development and utility.  
- He emphasizes that multiple models will likely compete rather than a single dominant model.  
- Future success will depend on integrating models with data, tools, and workflows, not just AI.  

At a time when the world is debating the arrival of artificial general intelligence (AGI), Microsoft CEO Satya Nadella offered a surprisingly grounded view of the technology. For the uninitiated, AGI is a conceptual AI that has the ability to understand, learn, and apply intelligence much like humans do.

While he is clearly excited about its potential, in his interview with Dwarkesh Patel and SemiAnalysis founder Dylan Patel, Nadella kept going back to the usual themes – tools, productivity, diffusion through the real economy, and the growing need for multiple models and ecosystems rather than one model ruling everything.

Is AGI the biggest thing after the industrial revolution?

Though Nadella does not entirely dismiss that AI could be as important as the industrial revolution, he somehow leans more towards the premise, adding a qualifier that we are still early. “I start with the excitement that I also feel for the idea that maybe after the Industrial Revolution this is the biggest thing. I start with that premise.”

Story continues below this ad

“At the same time, I’m a little grounded in the fact that this is still early innings. We’ve built some very useful things, we’re seeing some great properties, and these scaling laws seem to be working. I’m optimistic that they’ll continue to work,” Nadella said.

Nadella asserts that AI has the potential to be transformative, but he acknowledges that we are still in the early stages, perhaps even in the first or second innings.

During the conversation, Nadella also shared that his favourite definition of AI came from computer scientist Raj Reddy, which he claimed was very human-centric. “He had this metaphor for AI: it should either be a guardian angel or a cognitive amplifier. I love that. It’s a simple way to think about what this is.”

Will we get AGI?

For the Microsoft chief, the key question is not ‘will we get AGI’ but ‘what is its human utility?’. “Ultimately, what is its human utility? It is going to be a cognitive amplifier and a guardian angel. If I view it that way, I view it as a tool,” he said.

Story continues below this ad

He also acknowledged that people may even go ‘mystical’ about AI, meaning they may say that it is a tool that does things that only humans did before so far. However, he also added that this has been the case with many technologies in the past.

During the conversation, the interviewers pushed for a common AGI vision, such as a single model that gets deployed everywhere, continuously learning, and doing almost every job in the economy.

When asked if that happens, will it let the leading model company capture almost all the value? Nadella responded, saying, “If there’s literally one dominant model, deployed everywhere, ingesting all the data and continuously learning, then yes: that’s game, set, match. You’d basically stop shop.”

However, he also believes that is not how things are playing out actually. “For all the dominance of any one model, that is not the case. Take coding; there are multiple models. In fact, every day it’s less the case. There is not one model that is getting deployed broadly.”

Story continues below this ad

Nadella feels that constraints by domain, geography, and use case will prevent one single ‘god model’ from owning everything. “Is it going to happen in all domains? I don’t think so. Is it going to happen in all geos? I don’t think so. Is it going to happen in all segments? I don’t think so. It’ll happen in all categories at the same time? I don’t think so.”

The CEO asserts that “design space” is too big for a single winner.

Where does AGI’s value go?

When asked who gets the margins in an AGI world, the model companies or the ‘scaffolding companies’ that build tools, platforms and workflows around these models, Nadella said that both would matter and models alone are vulnerable to commoditisation.

He also went on to suggest that model companies may even face a ‘winner’s curse’. “You may have done all the hard work, done unbelievable innovation, except it’s one copy away from being commoditised. Then the person who has the data for grounding and context engineering… can then go take that checkpoint and train it.”

Story continues below this ad

His answer to who wins in all these is that there will be powerful open-source models, multiple frontier models, and the real advantage would be combining models with data, tools, and workflows.

Meanwhile, the Microsoft CEO also thinks that the future would not be simply about humans and chat UI. He hopes for a world with autonomous agents working across a company’s infrastructure.

Satya Nadella excited for AGI, but puts human utility first and pushes back on the ‘one-model-for-all’ idea

Google DeepMind unveiled SIMA 2, a cutting-edge AI agent designed to function like a virtual companion within digital realms, on Thursday. This new iteration marks a significant leap from merely executing basic commands to an AI capable of planning, self-explanatory interactions, and learning through play. The advancement puts the company closer to achieving Artificial General Intelligence (AGI), suggesting transformative possibilities for the future of robotics and AI implementations.

Initially launched in March 2024, the first version of SIMA (Scalable Instructable Multiworld Agent) mastered numerous fundamental skills by observing the screen and using keyboard and mouse inputs. The newly developed SIMA 2 takes autonomy to a new level, enabling it to make decisions independently.

Reinforced by Google’s Gemini model, SIMA 2 can decipher high-level objectives, articulate its planned actions, and engage in cooperative gameplay with a reasoning capacity that previous versions lacked. According to DeepMind, SIMA 2 has demonstrated improved adaptability across various virtual environments, successfully completing more challenging tasks which include reasoning puzzles, visual sketches, and emoji-based interactions.

DeepMind highlighted that SIMA 2’s ability to solve problems is closer to human performance, achieving a 65% task completion rate compared to just 31% for its predecessor. The AI agent has also shown proficiency in navigating entirely new 3D spaces crafted by Genie 3—another DeepMind initiative launched last year that generates interactive worlds from simple text or image prompts. SIMA 2 effectively familiarized itself with these environments, understood its objectives, and initiated significant actions shortly after they were created.

“SIMA 2 is now far better at carrying out detailed instructions, even in worlds it's never seen before,” stated Google's announcement. The system can leverage learned concepts, such as ‘mining’ in one game, and apply them to new tasks like ‘harvesting’ in another context, showcasing an impressive capacity for associative learning.

After observing human demonstrations, SIMA 2 transitioned into autonomous gameplay, utilizing trial and error alongside feedback from Gemini to enhance its experience. This process includes a feedback loop, where SIMA 2 generates and executes tasks, refining its data for future iterations.

Despite the significant progress, the research uncovered areas that warrant further development, such as difficulties with extensively long, multi-step tasks, limitations imposed by a restricted memory framework, and challenges in visual interpretation typical of 3D AI technologies.

Nonetheless, DeepMind believes that SIMA 2 embodies a promising foundation for future applications in robotics and navigational systems, representing a crucial milestone toward achieving real-world AGI. “Our SIMA 2 research offers a strong path towards applications in robotics and another step towards AGI in the real world,” the company concluded.

Google DeepMind's Latest AI Agent Learns, Adapts, and Plays Games in a Human-like Manner

A bright, sunlit study room with a rich wooden gaming table at its center. A sleek silver robotic arm reaches out to move a knight on a mid-game chessboard. Nearby lie a Go board with neatly arranged black and white stones, a pair of dice, and a handheld game controller. A small tablet screen on the table glows with an adaptive neural-network graph in progress. The DeepMind logo is prominently etched on the base of the robotic arm. Warm morning sunlight streams through a window, casting soft shadows and highlighting the scene’s natural tones, as a lone adult observer stands to the side, arms folded in awe of the AI’s human-like gameplay.

Google DeepMind’s New AI Agent Learns, Adapts and Plays Games Like a Human - Decrypt

- Google DeepMind introduced SIMA 2, an advanced AI agent for virtual 3D worlds.  
- SIMA 2 learns through self-directed play and adapts to new gaming environments quickly.  
- The AI agent's performance approaches human level, completing 65% of complex tasks successfully.

In brief The system used Google’s Gemini model to reason about goals, explain its plans, and act across unfamiliar games.

SIMA 2 learned new skills through self-directed play and adapted to worlds created moments earlier by Genie 3.

DeepMind planned a limited research preview for developers and academics.

Decrypt’s Art, Fashion, and Entertainment Hub. Discover SCENE

Google DeepMind introduced SIMA 2 on Thursday—a new AI agent that the company claims behaves like a “companion” inside virtual worlds. With the launch of SIMA 2, DeepMind aims to advance beyond simple on-screen actions and move toward AI that can plan, explain itself, and learn through experience.

“This is a significant step in the direction of Artificial General Intelligence (AGI), with important implications for the future of robotics and AI-embodiment in general,” the company said on its website.

The first version of SIMA (Scalable Instructable Multiworld Agent), released in March 2024, learned hundreds of basic skills by watching the screen and using virtual keyboard and mouse controls. The new version of SIMA, Google said, takes things a step further by letting the AI think for itself.

SIMA 2 is our most capable AI agent for virtual 3D worlds. 👾🌐 Powered by Gemini, it goes beyond following basic instructions to think, understand, and take actions in interactive environments – meaning you can talk to it through text, voice, or even images. Here’s how 🧵 pic.twitter.com/DuVWGJXW7W — Google DeepMind (@GoogleDeepMind) November 13, 2025

“SIMA 2 is our most capable AI agent for virtual 3D worlds,” Google DeepMind wrote on X. “Powered by Gemini, it goes beyond following basic instructions to think, understand, and take actions in interactive environments–meaning you can talk to it through text, voice, or even images.”

By using the Gemini AI model, Google said SIMA can interpret high-level goals, talk through the steps it intends to take, and collaborate inside games with a level of reasoning the original system could not reach.

DeepMind reported stronger generalization across virtual environments, and that SIMA 2 completed longer, more complex tasks, which included logic prompts, sketches drawn on the screen, and emojis.

“As a result of this ability, SIMA 2’s performance is significantly closer to that of a human player on a wide range of tasks,” Google wrote, noting that SIMA 2 had a 65% task completion rate, compared to 31% by SIMA 1.

The system also interpreted instructions and acted inside entirely new 3D worlds generated by Genie 3, another DeepMind project released last year that creates interactive environments from a single image or text prompt. SIMA 2 oriented itself, understood goals, and took meaningful actions in worlds it had never encountered until moments before testing.

“SIMA 2 is now far better at carrying out detailed instructions, even in worlds it's never seen before,” Google wrote. “It can transfer learned concepts like ‘mining' in one game and apply it to 'harvesting' in another—connecting the dots between similar tasks.”

After learning from human demonstrations, researchers said the agent switched into self-directed play, using trial and error and Gemini-generated feedback to create new experience data, including a training loop where SIMA 2 generated tasks, attempted them, and then fed its own trajectory data back into the next version of the model.

While Google hailed SIMA 2 as a step forward for artificial intelligence, the research also identified gaps that still need to be addressed, including struggling with very long, multi-step tasks, working within a limited memory window, and facing visual-interpretation challenges common to 3D AI systems.

Even so, DeepMind said the platform served as a testbed for skills that could eventually migrate into robotics and navigation.

“Our SIMA 2 research offers a strong path towards applications in robotics and another step towards AGI in the real world,” it said.

Google DeepMind’s New AI Agent Learns, Adapts and Plays Games Like a Human

The senior AI reporter at The Verge has been covering artificial intelligence for over five years, with contributions also featured in media outlets like CNBC, MIT Technology Review, and Wired UK.

Articles authored by this reporter will be included in your daily email summaries and will appear on your homepage.

Google DeepMind has unveiled its latest AI agent designed to engage with various video games, including titles such as No Man’s Sky, Valheim, and Goat Simulator 3, positioning it as a potential "interactive gaming companion."

This new tool, dubbed SIMA 2, enhances its predecessor, the SIMA (Scalable Instructable Multiworld Agent), which was launched in March 2024. For the first time, it integrates Google’s Gemini AI, enabling the agent to not only follow instructions but also comprehend higher-level user objectives, perform intricate reasoning, and adeptly conduct goal-driven actions in games, even those it has never encountered, according to a recent blog post from DeepMind. Currently, it is being rolled out to a select group of researchers and developers as part of a limited research preview.

Despite its advanced gaming capabilities, the DeepMind team clarified during a briefing that the primary objective isn't to develop a consumer-facing gaming assistant. Jane Wang, a senior staff research scientist at DeepMind, referred to this as an excellent training ground for possibly adapting these skills for real-world applications in the future.

As always, this development is part of the escalating competition in the quest for Artificial General Intelligence (AGI) involving tech giants like Google, Meta, OpenAI, and Anthropic. "This represents a crucial advancement toward AGI, with meaningful ramifications for the future of robotics and AI embodiment in general," stated DeepMind’s blog.

Echoing this sentiment, Joe Marino, a research scientist at DeepMind, emphasized that SIMA 2’s capability to operate within unexplored virtual environments is a "fundamental" milestone on the path to AGI and potentially laying the groundwork for the creation of a versatile robotic entity in the future.

See Google DeepMind's latest AI agent mastering video games.

A bright, sunlit studio where a sleek metallic humanoid AI sits at a gaming desk, leaning forward as dual monitors display colorful pixelated video games interwoven with floating streams of algorithmic code. Warm morning light pours through a nearby window, casting dynamic shadows across the scene. The Google DeepMind logo is emblazoned prominently on the desk’s front panel, telling the story of a cutting-edge agent learning to play through immersive gameplay.

Watch Google DeepMind’s new AI agent learn to play video games - The Verge

- Google DeepMind's SIMA 2 learns to play games as an interactive gaming companion.  
- The AI agent develops complex reasoning and understanding of user goals in gaming.  
- SIMA 2 signals progress toward Artificial General Intelligence and real-world applications.  

is The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.

Posts from this author will be added to your daily email digest and your homepage feed.

Google DeepMind’s new AI agent learned how to play a bunch of video games — including No Man’s Sky, Valheim, and Goat Simulator 3 — to become a viable “interactive gaming companion.”

The new agent tool, SIMA 2, builds on its earlier iteration, SIMA (Scalable Instructable Multiworld Agent), which DeepMind released in March 2024. It also incorporates Google’s Gemini AI for the first time, meaning the agent can go beyond simply following instructions to “understand a user’s high-level goal, perform complex reasoning in pursuit, and skillfully execute goal-oriented actions within games,” even ones it hasn’t seen before, according to a DeepMind blog post. It’s currently being released to some academics and developers as a limited research preview.

Despite SIMA 2’s gaming prowess, creating a consumer-facing gaming helper isn’t the broader goal here, members of the DeepMind team told The Verge during a Wednesday briefing. Jane Wang, a senior staff research scientist at DeepMind, called it “a really great training ground” for potentially transferring the skills to real-world environments one day.

And, as usual, it all comes back to the ever-intensifying AGI race between Google, Meta, OpenAI, Anthropic, and others. “This is a significant step in the direction of Artificial General Intelligence (AGI), with important implications for the future of robotics and AI-embodiment in general,” DeepMind’s blog post states.

Joe Marino, a research scientist at DeepMind, doubled down on that, saying that SIMA 2’s ability to take actions in a virtual world and handle environments it has never seen before is a “fundamental” step toward AGI — and potentially toward building a general-purpose robot down the line.

Watch Google DeepMind’s new AI agent learn to play video games

We’ve made several attempts to reach out to you regarding an issue with processing your payment. To ensure the continuity of your subscription, please update your payment information either through the My Account section or by clicking on the option to update your payment details.

The question is not if AI will replace judges, but when it will happen.

A dim, wood-paneled courtroom under low-key lighting; the judge’s bench sits empty except for a sleek black AI core pulsing with red circuitry where the judge would be; a single gavel floats before it, casting a long shadow; two silhouetted lawyers stand on either side, staring up in tense anticipation as the scene asks whether AI will soon preside over justice.

The question should be when not if AI will replace judges - The Times

- Multiple attempts made to contact for subscription payment issue.  
- Update payment details via My Account to continue service.  
- Action required to prevent subscription cancellation due to payment failure.

Act now to keep your subscription

We've tried to contact you several times as we haven't been able to take payment. You must update your payment details via My Account or by clicking update payment details to keep your subscription.

The question should be when not if AI will replace judges

A notable transition appears to be underway among those advocating for OpenAI, as the phrase "artificial general intelligence" and the concept of an all-encompassing AI have seemingly lost some of their appeal. Instead, the emphasis is shifting toward tailored AI experiences and dedicated researchers.

This transformation is prominently illustrated in the recent dialogue between OpenAI’s CFO, Sarah Friar, and the Wall Street Journal. Following her previous remarks about the necessity for government support regarding their hefty financial responsibilities, she was probed about the future of AI applications. Her response underscored the potential of personalized chatbots functioning like digital assistants.

Friar highlighted memory as a fundamental priority for future chat applications, stating it could significantly enhance user utility. She provided familiar examples, illustrating how ChatGPT could draw from past travel preferences to curate a more personalized vacation experience on platforms like Expedia.

Additionally, she made an intriguing analogy to OpenAI's vision by invoking the startup trope of being the "Uber for X," where X can represent virtually any service. She related this to the idea of everyone having access to services like personal doctors, tutors, stylists, and travel agents, akin to the way Uber democratized access to private transportation.

Invoking the concept of healthcare is a curious choice, especially considering OpenAI's recent decision to limit its provision of medical advice to users in a bid to manage liability risks. The comparison to Uber is also relatively fresh and reflects a strategy akin to what Uber employed historically—operating at a loss to eliminate the competition before ultimately seeking profits as the dominant player.

While prioritizing memory in her discussion makes sense from a marketing perspective, it raises substantial technical challenges. The issue of memory in AI remains complex and resource-intensive, with uncertain benefits in terms of chatbot efficacy. Furthermore, personalized chatbots can lead to users forming unhealthy attachments, which might increase engagement but could have detrimental effects.

This focus on personalized AI seems to mark OpenAI's evolving strategy, as the company appears to be moving away from the pursuit of an all-powerful AGI. Last month, during a conversation about organizational restructuring and near-term objectives, CEO Sam Altman mentioned the development of an "AI researcher" capable of making independent scientific discoveries, with AGI largely sidelined in that discussion.

Friar also echoed this sentiment, expressing a view that excitement around AI’s capabilities is still lacking. She suggested a potential future scenario where cancer patients could benefit from novel drugs—hypothetical breakthroughs facilitated by AI that could accelerate the drug development and FDA approval processes. While this vision is undoubtedly promising, it falls short of creating an omnipotent AI. Nevertheless, it signals a movement toward more grounded expectations for the technology.

What's More Impressive Than Creating a Literal Deity? Developing the New Uber

A bright, sunlit design studio with floor-to-ceiling windows overlooking a modern city skyline. Centered on a large wooden table are two unrolled blueprints side by side—one depicting a towering classical god statue, the other a sleek autonomous ride-share pod emblazoned with a bold Uber logo. A small group of adult designers in casual business attire lean in, animatedly pointing at the Uber blueprint. Warm natural light, cinematic angle, high-detail.

What’s Cooler Than Building a Literal God? Building the New Uber - Gizmodo

- OpenAI shifts focus from artificial general intelligence to personalized AI experiences and applications.  
- CFO Sarah Friar emphasizes memory features and digital butler-like capabilities for future chatbots.  
- The company aims to develop an AI researcher while moving away from AGI ambitions.  

An interesting shift seems to be happening in the language of OpenAI’s biggest hype people. The term “artificial general intelligence” and the idea of the do-it-all AI seem to have fallen out of favor a bit. In its place: personalized AI experiences and researchers.

Perhaps the clearest example of this change can be seen in OpenAI CFO Sarah Friar’s latest conversation with the Wall Street Journal. Fresh off her “when I said we expect a government backstop for our massive financial commitments, I didn’t mean it” media tour, she was asked by the publication what the future holds for AI applications. In her response, she leaned heavily into the idea of personalized chatbots serving as a sort of digital butler.

“When we think about the underlying first priorities of where weâre going with chat, number one is definitely memory, because we think this adds an incredible amount of utility to the consumer,” she said. She busted out a few examples that have become a bit of the standard schtick at this point: if you’re looking for a vacation on Expedia, for instance, ChatGPT will remember how you have liked to travel in the past and customize your experience in that direction.

She also drew a very interesting comparison for OpenAI’s model by bringing back the old startup standby of being the Uber for X (when solving for X here, it basically equals “everything”).

“I always think about when Uber went public, they went public with this tagline about being everyoneâs private driverâsomething only really affluent people had experienced, the black car that showed up for you, was now available to everyone,” she said. “When I think about that same thing for ChatGPT, I think about a world of where everyone can now have a concierge doctor, everyone can have a personal tutor, everyone can have a stylist, you can have a travel agent, you can have all of these things. Now we need to get better, I think, at manifesting it.”

Invoking doctors is an…interesting choice, given OpenAI recently announced that it would back off giving medical advice to users, likely as a way to avoid some liability concerns. But the Uber comparison is also something relatively new. It definitely seems like OpenAI’s financial strategy is the same as Uber’s was for most of its existence: operate in the red for as long as possible to kill your competition and then monetize when you’re the only game in town. So it’s fitting they’d lean into the likeness.

Mentioning memory as a top priority makes sense as part of a sales pitch, but it does present some real technical problems. AI’s memory problem has been one that has proven incredibly difficult to solve and resource-intensive, and it’s not clear that the outcomes from chatbots with longer memories are actually any better. There’s also the problem that personalized chatbots tend to be more likely to result in people forming unhealthy relationships with their chatbots, which probably drives up usage numbers but likely isn’t great for anyone.

But this personalized AI experience seems to be OpenAI’s latest calling card, as, in recent months, it has laid off the whole mission of developing an all-powerful “AGI.” Last month, when CEO Sam Altman talked about the company’s restructuring and near-term goals, he referenced developing an “AI researcher” capable of making autonomous discoveries. AGI was, at most, in the periphery of the conversation.

Friar leaned into that, too. When she busted out her new favorite refrain, that there isn’t enough exuberance about AI capabilities yet, she gave an example of what the future could hold: “A cancer patient can get maybe drugs that were never thought of before. A model was able to come up with a new molecular compound and a distribution mechanism that gets it to the cancer, and gets you through FDA approval multiples times faster than we could even imagine.” That certainly sounds promising, but it also isn’t quite building god. Neither is building a “personalized” AI that is better at selling you things. But at least it feels like we’re getting closer to a realistic expectation for these machines.

What’s Cooler Than Building a Literal God? Building the New Uber

Superintelligence refers to an advanced AI system that would outperform human intelligence in nearly every facet, including coding, video creation, surgical procedures, and driving vehicles, all at once—an ability absent in today's AI technologies. Current AI tools are prone to errors and require extensive data training to execute tasks effectively. In contrast, Artificial Superintelligence (ASI) would have the capability to solve intricate problems with superior reasoning and contextual awareness. Philosopher Nick Bostrom, who brought the term into public discourse, describes superintelligence as "any intellect that significantly surpasses the cognitive performance of humans across nearly all areas of interest."

Today's AI systems depend on human intervention for enhancements; engineers routinely refine code and provide expansive datasets to elevate the accuracy of their outputs. ASI, by contrast, would be able to autonomously enhance itself—potentially rewriting its algorithms, inventing new capabilities, and restructuring its control mechanisms without needing guidance.

The feasibility of superintelligence remains uncertain. However, should machines achieve higher intelligence than humans and possess self-improvement capabilities, there is a growing concern among experts that they may operate beyond human oversight if precautions are not taken. Many predict that superintelligence could emerge within the next ten years, leading to an escalation in investments, with billions being funneled into ventures aiming to realize this technology. Notably, OpenAI co-founder Ilya Sutskever departed from the organization in 2024 to establish a startup dedicated to the safe development of ASI, managing to raise billions in funding prior to even launching a product.

In a noteworthy collaboration in 2023, Sutskever, along with OpenAI CEO Sam Altman and President Greg Brockman, advocated for regulatory frameworks for superintelligent AI, cautioning that it could present an "existential risk." They suggested it is "conceivable" that within a decade, AI could surpass expert-level proficiency in numerous fields.

Understanding AI Superintelligence.

A sleek research chamber bathed in warm morning sunlight: an intricately detailed, brain-shaped network of golden circuitry hovers above a minimalist table; an adult scientist in neutral attire stands back in awe, shafts of natural light illuminating each glowing synapse, hints of faint equations etched on glass walls.

What Is AI Superintelligence? - bgr.com

- Superintelligence could surpass human intelligence, solving complex problems across multiple domains simultaneously.  
- Unlike current AI, superintelligence could self-improve without human intervention or extensive data input.  
- Experts warn that superintelligence may pose existential risks, necessitating urgent regulation and oversight.  

Superintelligence is a hypothetical AI system that surpasses human intelligence across every domain. From writing code and generating videos to performing surgery and driving cars, it would do everything simultaneously, something the current AI systems lack. The AI tools we use hallucinate, and require training with massive data to perform tasks in a specific way. ASI would find solutions to complex problems with better reasoning and the knowledge of context. Philosopher Nick Bostrom, who popularized the term, defines superintelligence as "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest."

Current AI systems need humans to improve them: Engineers work on their code, feed the models vast amounts of data to improve their predictions and responses. ASI, on the other hand, could improve itself. It could hypothetically rewrite its own algorithms, design capabilities, and control systems without instructions.

It's unclear if superintelligence is possible or not. However, if machines would become smarter than humans and improve themselves, some say that AI could exceed human control if they aren't careful. Some experts predict superintelligence could arrive within a decade, and the race to develop it has intensified already, as investors pour in billions of dollars into companies looking to build it. OpenAI co-founder Ilya Sutskever left the company in 2024, and founded a startup focused on building ASI safely. Sutskever has already raised investment running into billions of dollars without even launching any product so far.

Interestingly, in 2023, Sutskever joined OpenAI CEO Sam Altman and President Greg Brockman in calling for called for regulation of super-intelligent AI, warning it could pose an "existential risk," adding that it's "conceivable" that AI would exceed expert skill level in most domains within 10 years.

What Is AI Superintelligence?

Leading figures in artificial intelligence have high ambitions for the technology's future. Dario Amodei, CEO of Anthropic, projected that by 2026 or 2027, AI systems will surpass human capabilities across a broad range of tasks, as he stated at this year's World Economic Forum in Davos, Switzerland.

Elon Musk, in a bold assertion made late last year, suggested that by the end of 2023, AI will outperform any individual human, and by 2030, an AI might even outstrip the cumulative intelligence of all human beings. Sam Altman, CEO of OpenAI, shared with Bloomberg earlier this year that he anticipates the emergence of artificial general intelligence (AGI) before the conclusion of Donald Trump’s presidential term.

AGI, sometimes referred to as superintelligence, is envisioned as an advanced AI capable of excelling in all cognitive domains, serving as a guiding vision for the tech sector. For instance, Meta has launched an entire division and is investing billions in the quest for superintelligence, with CEO Mark Zuckerberg claiming that its realization is on the horizon.

However, a recent study presents a more tempered viewpoint on the prevailing optimism among AI leaders. The "Longitudinal Expert AI Panel," a report from the Forecasting Research Institute, reveals a much more gradual timeline for the arrival of superintelligent AI than commonly projected. This study, directed by experts from various fields including economist Ezra Karger, involved consultations with computer scientists, economists, industry professionals, and AI researchers who are in high demand among policymakers and business leaders regarding AI’s development trajectory and impact.

The findings suggest that the technology leaders' fast-paced predictions have only a 23% likelihood of materializing as estimated. The report details a scenario where AI would author Pulitzer Prize-winning literature, drastically reduce research durations, outperform human software developers, and independently discover new cancer therapies — aligning closely with the aspirations promoted in Silicon Valley.

One expert characterized the situation, stating, “Major systems undergo radical changes over longer periods than just four to five years. Moreover, in various fields, even swift advancements in AI might not result in significant immediate improvements due to unforeseen challenges.” Another highlighted potential bottlenecks, predicting that AI's expansive influence would likely be hindered by obstacles in areas yet to be dominated by the technology, emphasizing that there may be thousands — if not millions — of such barriers that will only become apparent as AI advances.

Concerns concerning AI's progress have also surfaced, particularly following an Oxford study that cast doubt on the reliability of common benchmarking tools used to evaluate AI performance.

Not all leaders in the tech industry are firmly convinced about the imminent arrival of superintelligence. Mustafa Suleyman, Microsoft's AI chief, famously criticized the idea as “absurd.” Similarly, Salesforce CEO Marc Benioff referred to the excitement surrounding AGI as a form of “hypnosis.”

Amidst this discourse, several proponents of superintelligence have expressed unease regarding its implications. A petition advocating for a moratorium on its development until specific criteria are met has garnered more than 100,000 signatures from notable figures, including Apple co-founder Steve Wozniak, and esteemed computer scientists Geoffrey Hinton and Yoshua Bengio, recognized as pioneers in the field.

Despite a general skepticism regarding the rapid approach of superintelligence by the end of the decade, many experts still acknowledge its potential for transformative change. The study indicates that these experts foresee a substantial impact from AI by 2040, likening it to the astonishing influence of electricity in the past century. By 2030, they anticipate that AI could become a daily companion for around 15% of adults and facilitate approximately 18% of working hours in the U.S.

The study acknowledges its participants as “superforecasters,” but even among such individuals, predicting AI's evolving landscape has proven to be challenging. Earlier research revealed that both AI specialists and superforecasters had underestimated the technology's potential to advance rapidly. For instance, they projected in 2022 that AI would achieve a gold medal in the International Mathematical Olympiad by 2030, while superforecasters predicted 2035. However, an AI developed by Google clinched that gold medal as early as July of this year.

Major tech companies claim superintelligent AI is approaching, but many experts disagree.

Split-screen scene under warm natural light: on the left, three suited executives stand on a sleek stage applauding beneath giant Apple, Google and Microsoft logos as a glowing neural-network brain hovers overhead; on the right, a circle of seasoned experts in suits sit around a wooden table, arms crossed and brows furrowed, papers scattered before them, all staring skeptically at the same hovering brain.

Big Tech Says Superintelligent AI Is in Sight. The Average Expert Disagrees - Gizmodo

- AI leaders predict superintelligence could surpass human intelligence by the late 2020s.  
- Experts believe rapid AI advancements are less likely, estimating a 23% chance.  
- AI's significant impact is expected by 2040, compared to past historical technologies.  

The top leaders in AI have very lofty expectations for the future of artificial intelligence.

âBy 2026 or 2027, we will have AI systems that are broadly better than almost all humans at almost all things,â Anthropic CEO Dario Amodei said earlier this year at the World Economic Forum in Davos, Switzerland.

Elon Musk posited late last year that we will have AI systems that are more intelligent than any single human by the end of this year, and we will “100%” have an AI system that exceeds the intelligence of all human beings combined by 2030. OpenAI CEO Sam Altman told Bloomberg earlier this year that he thinks artificial general intelligence (AGI) will “probably get developed” before the end of Trump’s presidential term.

Superintelligence or AGI is a super-powered future AI system that could theoretically outperform human intelligence on all fronts, and it has become the North Star of the tech industry. Meta, for example, has a whole division and a multibillion-dollar spending spree dedicated to building superintelligence, and CEO Mark Zuckerberg claims it is âin sight.”

Although most of the AI leaders with the loudest microphones claim that artificial superintelligence is imminent, a new study paints a differing picture of expert sentiment.

According to the Forecasting Research Institute’s “The Longitudinal Expert AI Panel” research report, the timeline for superintelligent AI will be much slower than what’s been promised. Led by multi-disciplinary experts like Federal Reserve Board of Chicago economist Ezra Karger, the fact-finders say they reached out to computer scientists, economists, industry professionals, and AI researchers “whom policymakers, business and nonprofit leaders, and other stakeholders would be most inclined to consult regarding the progression of AI capabilities and its technological impact.”

The experts gave the tech leaders’ timeline of rapid progress only a 23% chance of actually happening.

The rapid progress described in the study is one in which “AI writes Pulitzer Prize-worthy novels, collapses years-long research into days and weeks, outcompetes any human software engineer, and independently develops new cures for cancer.” So, pretty much mirroring how Silicon Valley describes artificial superintelligence.

“Radical change in major systems just takes longer than 4-5 years. I also think that [in] many of these domains, even unexpectedly fast advancement in AIs will not easily translate to improvements for quite some time because of unexpected barriers,” one expert respondent wrote. “To paraphrase an old saying, every job looks easy for those not actually doing it.”

Another expert thought that bottlenecks would intervene before AI’s impact could scale up to the level predicted.

“The force of its impact will likely be slowed by bottlenecks in areas AI hasnât yet conquered,” the expert wrote. “There likely exist thousands (millions?) of potential bottlenecks in the economy which will only become legible as other processes are sped up by orders of magnitude.”

Some of the advances in AI model capabilities have also been brought into question recently, after a new Oxford study found that a lot of the popular benchmarking tools used to test the performance of AI models were unreliable or misleading.

Not all tech leaders have the same, unwavering belief in the future of superintelligence, though. Microsoft’s AI chief, Mustafa Suleyman, is famously a non-believer, going so far as to call the pursuit of superintelligence “absurd.” Another tech titan, Salesforce CEO Marc Benioff, recently called the hype around artificial general intelligence an example of “hypnosis.”

Meanwhile, some tech experts that do believe superintelligence is imminent are not particularly thrilled about that possibility. In October, a statement calling for the prohibition on the development of superintelligence until certain conditions are met was signed by more than 100,000 people, including the likes of Apple co-founder Steve Wozniak, and computer scientists Geoffrey Hinton and Yoshua Bengio, both of whom are considered “the godfathers of AI.”

Even though most experts in the study disagreed that AI will evolve at light speed to reach superintelligence by the end of the decade, the average expert still believes in its transformational power.

According to the study, experts predict that AI will have a significant impact by 2040, labeling it as the “technology of the century” akin to electricity. And by 2030, experts believe that AI will provide daily companionship for roughly 15% of adults and assist in 18% of work hours in the U.S.

Some of the experts included in the study are described by the researchers as “superforecasters.” But AI is a tricky thing to forecast, even for so-called superforecasters, as it turns out.

A past study of AI experts and superforecasters, also conducted by the Forecasting Research Institute, found that both groups had underestimated just how fast AI could progress. For example, experts thought in 2022 that AI would get a gold medal in the International Mathematical Olympiad in 2030, and superforecasters said 2035. But a Google-built AI system won that gold in July of this year.

Big Tech Says Superintelligent AI Is in Sight. The Average Expert Disagrees

For years, Microsoft faced restrictions in its pursuit of advanced artificial intelligence. However, recent developments suggest that the software company is now poised to take on competitors more robustly, as indicated by Mustafa Suleyman, a prominent executive. 

In a recent conversation with Business Insider, Suleyman announced the formation of a new superintelligence team at Microsoft, discussing its inception and future endeavors. The discussion clearly highlights Microsoft's ambition to develop artificial general intelligence (AGI)—AI that can surpass human capability across various tasks.

Although Suleyman dismissed the idea of an AGI "race," the establishment of this new team signals a direct challenge to its collaborator OpenAI, alongside other tech giants such as Anthropic, Google, and Meta. The goal of the team is to cultivate a top-tier, in-house research capability in advanced AI.

"For Microsoft to succeed, we must become self-sufficient in AI," Suleyman stated during the interview. He reiterated the importance of training advanced models using proprietary data and computing resources to stay at the forefront of technology. This approach marks a significant shift from Microsoft’s past strategy, which primarily focused on refining smaller models and utilizing resources to bolster OpenAI rather than creating competitive in-house models following OpenAI's GPT innovations.

Historically, Microsoft's limited focus stemmed from a prior partnership agreement with OpenAI that restricted the company from developing its own AGI until 2030. This deal was designed to enable OpenAI to maintain its hold on cutting-edge AI development while relying on Microsoft's infrastructure and funding. However, recent renegotiations have lifted these restrictions, allowing Microsoft to independently pursue AGI, whether solo or with external partners.

With its new team, aptly named Microsoft AI Superintelligence, the company is setting ambitious sights on future applications across multiple industries, which include healthcare, energy, and transportation. Suleyman pointed out that the unit aims to address fundamental challenges hindering the realization of these applications, offering examples such as enabling one AI model to impart knowledge to another and facilitating continual learning within neural networks.

To support these initiatives, Microsoft is investing significantly in computing power, forging collaborations with Nvidia, and enhancing its own chip production alongside AI-optimized cloud services for more efficient training and inference capabilities. 

Suleyman also noted in an internal meeting, as reported by Business Insider, that considerable investments are being directed towards Microsoft’s proprietary AI chip cluster. This strategy aims to reduce dependence on external hardware and elevate performance across its Azure AI offerings. "Our top priority is to ensure we have the highest-performing infrastructure globally," he mentioned during the meeting.

In his conversation with Business Insider, Suleyman expressed a willingness to explore various AI model types, including open-source technologies, and offerings from other firms, illustrating Microsoft’s pragmatic approach. "We shouldn't be dogmatic about our choices," he remarked. "Our aim is to ensure our products are effective."

As Microsoft joins a competitive landscape alongside companies like Meta, Google, Anthropic, and Elon Musk's xAI, it seeks to differentiate itself by prioritizing responsibility and safety in AI development. Suleyman underscored this point in a blog where he reflected on the importance of having a clear, purpose-driven direction. 

While refraining from pinpointing specific competitors he views as lacking focus, Suleyman emphasized the collective responsibility to ensure AI advancements are aligned with human interests. "We have to design systems that don’t outsmart us, requiring a fundamentally human-centered approach," he cautioned.

Investments in safety are also on the rise within the new team. Recently, Suleyman appointed Trevor Callaghan, a former general counsel for DeepMind and a legal director at Google, to oversee responsible AI initiatives as vice president.

According to Suleyman, the formation of this team arrives at a critical juncture for Microsoft. "We face a tremendous mission ahead," he remarked. "With $300 billion in revenue, we must ensure all our products prioritize AI, deploying intelligent agents across all customer workflows to enhance their capabilities."

With access to vast data sets, extensive distribution channels, and advanced user interfaces, Suleyman believes that it won't be long before revolutionary advancements take shape. "It’s just a matter of time before these initiatives yield truly remarkable results," he concluded.

Mustafa Suleyman states that OpenAI was previously used to stop Microsoft from seeking AGI, but now the tech giant can compete freely.

A warm, natural-lit conference hall scene with Mustafa Suleyman at a podium, shearing a glowing digital chain that once linked a projected “OpenAI” text with no logo to a large Microsoft logo on stage—symbolizing Microsoft’s newly freed race to AGI.

OpenAI used to prevent Microsoft pursuing AGI. Now, the software giant is free to compete, Mustafa Suleyman says. - Business Insider

- Microsoft has created a superintelligence team to pursue artificial general intelligence technology.  
- The new agreement with OpenAI allows Microsoft to independently develop AGI solutions.  
- Microsoft is investing in AI chips and infrastructure to enhance its competitive capabilities.

In the race to build powerful artificial intelligence, Microsoft has had its hands tied behind it back for years. Now, the software giant is free to compete, according to top executive Mustafa Suleyman.

Suleyman recently unveiled a superintelligence team at Microsoft and he spoke with Business Insider about how this came about and the company's future plans.

What's clear from the interview is that Microsoft will aggressively pursue artificial general intelligence, technology capable of outperforming humans in a wide variety of tasks.

Related video What are the real-life consequences of AI?

While Suleyman has rejected narratives about a race to AGI, the new unit puts the tech giant in more direct competition with its partner OpenAI and other companies, including Anthropic, Google, and Meta.

The superintelligence team will focus on building a "world-class, frontier-grade research capability in-house," Mustafa, Microsoft's AI CEO, told Business Insider in a recent interview.

"Microsoft needs to be self-sufficient in AI," he said. "And to do that, we have to train frontier models of all scales with our own data and compute at the state-of-the-art level."

This is a major departure from Microsoft's approach in recent years. The company mostly focused on building smaller models, post-training existing models for new purposes, and channeling resources toward OpenAI instead of trying to build in-house frontier models that would compete with the startup's GPT offerings.

Previous limitations

A big reason for this narrower strategy was Microsoft's previous agreement with OpenAI. That deal barred Microsoft from developing its own AGI through 2030, according to a person familiar with the matter. This reflected OpenAI's desire to maintain control over frontier AI development while relying on Microsoft for cloud infrastructure and capital investment.

The two companies recently renegotiated this partnership. The new deal lets Microsoft "independently pursue AGI (artificial general intelligence) alone or in partnership with third parties."

Loftier goals

That's freed Microsoft to pursue loftier goals. The new team, called Microsoft AI Superintelligence, has big ambitions for what AI will one day be able to achieve.

Related stories Business Insider tells the innovative stories you want to know Business Insider tells the innovative stories you want to know

Potential future applications include healthcare, energy, transportation systems, and reducing the "cost of living for billions of people over the next 10 years," Suleyman said.

The team's remit will be to focus on solving fundamental problems that present barriers to the creation of those applications, Suleyman said. Two examples he cited: How to transfer learning so AI models can teach other models new knowledge, and continual learning to add knowledge to existing neural networks.

Suleyman said Microsoft is making major investments in compute through partnerships with Nvidia, as well as expanding its own chip development and AI-optimized cloud infrastructure that allows training and inference capacity to be shared more efficiently.

Keeping an open mind

Suleyman recently said in an internal meeting reviewed by Business Insider that Microsoft plans to make "significant" investments in its own AI chip cluster to help the company build its own models. These custom chips are part of Microsoft's broader push to reduce reliance on third-party hardware and improve performance across Azure's AI services.

"It's the number one priority for us to make sure this is the most performant infrastructure in the world," Suleyman said during that meeting.

In his interview with Business Insider, Suleyman said Microsoft will keep an open mind about what models it uses, including open-source offerings, Anthropic models, OpenAI models — and Microsoft's own creations, known as MAI models.

"There's no reason for us to be religious about that," Suleyman said. "Obviously, we're very focused on getting our products working."

The adult in the room

As Microsoft enters a crowded field of companies with their own superintelligence teams, including Meta, Google, Anthropic, and Elon Musk's xAI. Microsoft is trying to position itself as the adult in the room, emphasizing responsibility and safety.

"This isn't about some directionless technological goal, an empty challenge, a mountain for its own sake," Suleyman wrote in a blog announcing the new team.

In the interview with Business Insider, Suleyman declined to elaborate on which competitors he sees as directionless, but said he believes everyone will get behind developing AI that's aligned to human interests.

"That should be something we all take for granted, but it actually needs to be stated and repeated, and it needs to be the No. 1 most important thing that humanity focuses on," Suleyman said. "There's a risk with these systems that they get extremely smart and run away from us, and we have to design them so that they don't do that. That requires a humanist intent, which keeps humans at the top of the food chain."

The team is making significant investments in safety. Suleyman recently added Trevor Callaghan, a former general counsel from DeepMind and legal director at Google, to his leadership team as vice president of responsible AI, according to an organizational chart recently viewed by Business Insider.

The new team comes at a pivotal moment for Microsoft, Suleyman said.

"We've got a huge mission ahead of us," Suleyman said. "We have $300 billion of revenues, a huge responsibility to make sure that all of our products are AI-first, that we deploy agents everywhere, and we really make all the workflows that customers use today much more intelligent."

"We have the data, we also have the distribution, and we have the user interface," he added. "So I think it's just a matter of time before these things become really, really magical."

Artificial General Intelligence News