A new autonomous agent from AI startup Manus made waves over the weekend. Here’s what you need to know.

What’s happening? Manus, a new “general purpose agent” that can supposedly execute deep research, deep thinking and multi-step tasks like no agent before, has sparked a vigorous online debate as to whether we are witnessing a second DeepSeek moment. “Manus is the most impressive AI tool I’ve ever tried,” wrote Victor Mustar, Head of Product Design at Hugging Face, on X. Mustar’s post, which includes a small demo, argues the user experience is also a key part of the wow factor. “The UX is what so many others promised, but this time it just works,” he wrote.

What’s under the hood? Manus offers an automation experience where users chat with the model, which can access tools to orchestrate requests and responses. In the days since the company announced its invite-only beta version on March 7, users discovered Manus is a well-designed interface that sits atop Anthropic’s Claude Sonnet model and 29 other tools.

Why it matters: While much of the AI world is chasing AGI, Manus demonstrates that real-world adoption may not require such sophistication or endlessly improving models—just practical models that are connected to the data and tools that people need. “I don’t think we need artificial general intelligence. We just need some really good models, a really good experience and then some agents and tools that can go and perform a job for you,” says Chris Hay, a Distinguished Engineer at IBM.

The jury is still out: For all the enthusiasts, many skeptics have also chimed in and found that Manus failed at certain tasks or crashed when they tried it out. IBM’s Maryam Ashoori, Director of Product Management for watsonx.ai, is more measured. “It’s impressive and exciting, yes. But is it solving an enterprise problem? I don’t see how.” To scale real-world applications, she says we need “specialized, cost-efficient agents, that are specially configured for enterprise use.”

However this plays out, many on Wall Street and Silicon Valley will be tracking Manus closely to avoid DeepSeek whiplash.

What else can we expect from AI agents in 2025? IBM experts break down expectations vs. reality.

Is she talking to her mom? Her best friend? No. That’s her AI agent on her mobile phone giving her advice. The latest “AI Phone,” a joint product from T-Mobile parent company Deutsche Telekom and AI giant Perplexity, took center stage at this year’s Mobile World Congress.

But while Google and Apple have been adding AI features to phones for years, the AI-first phones launching in 2025 may signal a bigger shift, says Kaoutar El Maghraoui, a Principal Research Scientist and Manager at IBM. AI giants are now joining forces with big telecoms to integrate AI early on in the hardware design process, rather than adding it as an afterthought to existing devices.

This end-to-end collaboration may change the way people use their phones, says El Maghraoui. For example, users of Deutsche Telekom’s AI Phone can simply activate the Perplexity assistant by clicking a button on the side of the phone and speaking to make a dinner reservation, book a taxi or translate speech from one language to another in real time—all without opening the lock screen or any apps.

Why now? Advances in small language models (SLMs) are propelling this evolution in mobile technology, says El Maghraoui. “These compact, more efficient and more capable models can do real-time processing and create personalized, secure interactions on edge devices like smartphones,” she says. IBM, for example, has prioritized smaller, fit-for-purpose models such as its Granite series because they enable enterprises to pursue frontier model performance at a fraction of the cost.

El Maghraoui adds that the lessons learned from SLMs on smartphones may actually be most useful outside the consumer arena. Small models are nice for phones, but critical for “more serious, mission-critical applications, such as embedded devices in factories where sensitive manufacturing data must be processed locally,” she says.

Learn more about how AI and cloud are transforming the telecom industry here.

— Aili McConnon

Autoregressive models like GPT, which predict next steps from past data, have ruled the LLM world for years. But a startup called Inception Labs is shaking things up with diffusion-based language models (dLLMs).

These models refine entire phrases in one go, rather than predicting words one at a time—which Inception Labs says makes them faster and cheaper, without skimping on quality. The company says its new Mercury model outperforms GPT-4o mini in coding benchmarks and is poised to make chatbots snappier and AI assistants more efficient.

Open-source projects like LLaDA are trying to show that diffusion can hold its own against GPT-style models, and Inception Labs is taking it further with a commercialized version. Experts say that as diffusion techniques scale, the AI landscape could shift from token-by-token slogging to real-time refinement.

“dLLMs expand the possibility frontier,” Stefano Ermon, a Stanford University computer science professor and Inception co-founder, tells IBM Think in an interview. “The Mercury models provide unmatched speed and efficiency. And by leveraging more test-time compute, dLLMs will also set the bar for quality and improve overall customer satisfaction for edge and enterprise applications.”

IBM Research Engineer Benjamin Hoover sees an even bigger shift coming. “It’s just a matter of two or three years before most people start switching to using diffusion models,” he says. “When I saw Inception Labs’ model, I realized this is going to happen sooner rather than later.”

— Sascha Brodsky

As the dust settled after the release of DeepSeek-R1, experts such as IBM Consulting VP and Senior Partner Shobhit Varshney are extracting the critical lessons for companies large and small.

Varshney views DeepSeek’s breakthrough as a win for innovation and open-source, that will benefit companies of all sizes. While some concluded that we may need fewer chips to build excellent models going forward, Varshney believes chip demand will only increase. “Any given company may use fewer chips,” Varshney tells IBM Think. “But DeepSeek demonstrated that many more players can enter the market and use open-source techniques to build impressive models for less.”

Ultimately, the DeepSeek moment, which revealed high-performing models could be built with less compute and for less cost than was previously thought, is “a huge win for the open-source community,” Varshney says, with the potential to accelerate innovation across the board. “We will have meaningful intelligence at a low enough price point in the next 24 months to truly transform how work can be done.”

Read Varshney’s analysis of how we can all benefit from developments like DeepSeek-R1.

— Aili McConnon

​Developers worldwide are invited to join IBM’s 2025 Call for Code Global Challenge to harness AI and tackle global humanitarian challenges, from clean water to climate action.

Foxconn debuts FoxBrain, Taiwan’s first reasoning-capable AI model, trained on NVIDIA GPUs and optimized for traditional Chinese.

Hazy Research unveils ThunderMLA, a fused “megakernel” that reportedly delivers 20-35% faster performance than DeepSeek’s FlashMLA for large language model inference by reducing kernel launch overhead and implementing advanced scheduling techniques.

Researchers from UC San Diego present DEMO³, a robotic learning framework that reportedly improves data efficiency by 40-70% in complex manipulation tasks by breaking them into distinct stages with rewards, using demonstrations to train policies and world models.

Vodafone partners with IBM to implement quantum-safe cryptography in its Secure Net mobile security service, aiming to protect customers from future quantum computing threats.

— Sascha Brodsky

Was this email forwarded to you? You can subscribe here