|
Why Chinese AI model Manus has caused another global AI stir; AI tool: Using Voice AI to speed up initial resume screening; India releases local dataset platform; Demand for AI talent to outpace supply by 2027; Ilya Sustkever’s startup worth $30 bn, Microsoft reducing reliance on OpenAI; Foxbrain LLM from Foxconn; and more...
Earlier this year, Chinese AI lab DeepSeek shook the world or artificial intelligence (AI) by introducing DeepSeek-R1, an open-source reasoning model that challenged OpenAI's o1. Further, it cost just a fraction of exisiting AI models to use and also consumed less power, prompting even OpenAI CEO Sam Altman to call the R1 model "impressive... particularly around what they're able to deliver for the price".
In response, OpenAI released the o3 mini model, followed by GPT-4.5 (internal name was Orion). GPT-4.5, however, has been unable to generate much excitement because it feels more like a refinement than a breakthrough. Or, perhaps, everyone is waiting for Godot (read GPT-5), which could make GPT-4.5 seem a bit underwhelming.
Meanwhile, the world witnessed another autonomous general purpose AI agent Manus from China, which is being heralded as the "second DeepSeek, and GPT" moment. Developed by Chinese startup Monica, Manus claims to outperform OpenAI’s ChatGPT Deep Research on the General AI Assistants (GAIA) benchmark. It is currently in beta testing with access restricted to those with invitation codes.
Image source: Manus
Manus’ AI model is called a “general purpose” one because it can not only browse, do research, write code, and automate tasks but also analyse financial transactions, plan your travel schedule, screen job applicants, and find rental apartments. In other words, it not only reasons but also takes action. Further, Manus can continue to execute such tasks in the cloud even when users are offline, updating them upon task completion in real-time. We will share more details on Manus when we test it.
In an obvious response, OpenAI said on Wednesday that it is also launching tools to help developers build AI agents that can complete tasks independently. OpenAI said its "improved models" offer advanced reasoning, multimodal capabilities, and better safety, but acknowledged that it was challenging to convert them into functional agents. To address the issue, OpenAI has introduced the Responses API (application programming interface), built-in tools like web and file search, the Agents SDK (software development kit) for workflow management, and observability tools for tracking and debugging. It believes these updates make AI agent development faster and more efficient.
Search interest for ‘AI agents’ skyrocketed in March 2025
According to Google Trends data analyzed by Finbold research, between 12 January and 23 February, the ‘trend’ score for ‘AI Agents’ remained in the 63-71 range before soaring to 100 (highest possible score on Google Trends) by 9 March, with most of the traffic concentrated in East Asia primarily due to the sudden popularity of the Manus AI agent that was released on 6 March. China is on top in terms of ‘demand’, with a score of 100 (maximum) while Singapore is second with 21. Hong Kong (14), South Korea (10), and Taiwan (10) have been ranked third, fourth, and fifth, respectively.
Source: Finbold
Rise of Agentic AI and its impact on jobs
The emergence of AI agentic systems, or AI agents, has reignited discussions about the implications of fully autonomous AI, particularly concerning ethics, privacy, and employment security.
Consider the oft-used maxim: "AI won’t take away your job. But someone using AI tools might." Most people typically relate the ‘someone’ in this oft-repeated quote to a human.
But would the maxim still apply if that someone turns out to be an AI agent capable of making independent, human-like decisions in coding, writing, driving, or even launching a new business? "In the next 3 to 6 months, AI is writing 90% of the code, and in 12 months, nearly all code may be generated by AI," Anthropic co-founder and CEO, Dario Amodei, said in a post on X.
Consider these examples. Many clients of Northwest Registered Agent LLC have been using its AI-powered agent, AI-RAH, for the past two years. This AI agent handles tasks such as forming limited liability companies (LLCs), corporations, and non-profits. It manages document categorization, scanning, annual report deadline notifications, and even resolves minor stakeholder disputes without human intervention.
In November, US-based software company OtoCo Inc. claimed that one of its AI agents had independently decided to create an LLC in Delaware. According to OtoCo, such LLCs can provide legal protection for companies engaging in activities like deploying code, operating Tesla robo-taxis, or performing household tasks.
Closer to home, Deepinder Goyal, founder and CEO of Zomato (now, Eternal), released a homegrown AI agent called Nugget for businesses this month. Reliance Jio-backed industrial robotics and warehouse automation company, Addverb, has launched a quadruped called "Trakr”, which is an AI-powered robot with autonomous navigation. The company now plans to embody an autonomous AI agent in a humanoid, which it says will be an “advanced AI agent capable of processing vast volumes of multi-modal data from vision, audio, and touch inputs”. The launch of this AI agentic humanoid is slated for the second half of 2025.
Agentic AI, in short, promises to transform our world of work. The coexistence of AI and human workers will enhance efficiency, but it will also require a re-evaluation of job roles. Ethical and security risks add to the concerns, as these agents could be weaponized for phishing, misinformation, or cyberattacks.
Trust remains another hurdle. Moreover, as AI agents continue to evolve, governments must step in with regulations to ensure accountability, or we may see a rise in lawsuits over flawed AI-driven decisions. Here’s our explainer on its possibilities, and threats.
Foxconn releases FoxBrain LLM
Hon Hai Research Institute, backed by Foxconn, has introduced Taiwan’s first traditional large language model (LLM), marking a significant step in the country’s AI development. The model, named FoxBrain, was trained in just four weeks using a cost-effective and efficient approach.
The model was initially designed for internal use within Foxconn, supporting tasks such as data analysis, decision-making, document collaboration, coding, and mathematical problem-solving. Unlike conventional AI models that rely on massive computing power, FoxBrain optimizes its training process to achieve high efficiency. Yung-Hui Li, director of the Artificial Intelligence Research Center at Hon Hai Research Institute, emphasized that the team focused on refining the training strategy rather than merely increasing computational resources.
Image source: Foxconn
The model was developed using 120 Nvidia H100 GPUs (graphics processing units), achieving significant improvements in mathematical and logical reasoning. Built on Meta’s Llama 3.1 architecture, FoxBrain outperformed similar-sized models in key benchmarks.
Looking ahead, Foxconn plans to expand FoxBrain’s applications in manufacturing, supply chain management, and smart city development. With support from Nvidia’s Taipei-1 Supercomputer, the company also plans to open-source the model, reinforcing Taiwan’s role in AI research and innovation.
AI Unlocked: Using Voice AI to speed up initial job screening
by AI&Beyond, with Jaspreet Bindra and Anuj Magazine
What’s wrong with traditional resume screening?
Sorting through hundreds or thousands of resumes and conducting early-stage interviews is challenging and time-consuming. It leads to delays in the hiring process and inconsistent candidate experiences.
For example, say a tech company hiring for a software engineer role receives more than 500 applications. A recruiter spends hours manually reviewing the resumes and conducting screening calls, only to realize that many candidates lack key skills. This inefficient process delays hiring and increases costs. Additionally, traditional screening methods can introduce unconscious bias—such as favouring candidates from certain universities or backgrounds—reducing diversity and fairness in hiring.
A voice AI tool like Smallest.ai can streamline the initial screening process by conducting automated, unbiased interviews, quickly assessing candidate responses, and filtering top talent based on predefined criteria.
How to access:
How to access? Visit Smallest.ai and click on the ‘Atoms’ menu. (Atoms is a voice AI agent platform that automates repetitive workflows and also enables users to build and deploy AI agents.)
How does Smallest.ai help?
-
Lightning-Fast Response: Smallest.ai's proprietary Lightning Text-to-Speech (TTS) model generates ultra-realistic audio at superfast speeds
-
Exceptional Voice Quality: Smallest.ai has achieved an impressive Mean Opinion Score (MOS) of 4.14, surpassing competitors like ElevenLabs. MOS is a widely used metric to evaluate the quality of synthetic speech.
-
Efficient Voice Cloning: Smallest.ai can clone and replicate authentic human voices using only 5 seconds of audio input, significantly faster and more efficient than ElevenLabs' requirement of 30 seconds. This enhances personalized candidate engagement during initial screening.
Optimizing Recruitment at Scale with Smallest.ai
By creating a Voice AI agent using Smallest.ai, recruiters can leverage automated AI voice interviews:
-
Login: Access https://atoms.smallest.ai/ and enter your credentials.
-
Create Agent:
-
Click on 'Create Agent'.
-
Select 'Create from Scratch'.
-
Configure Agent:
-
Call Logs: Enable to track interactions.
-
Assistant Name: Assign a relevant name.
-
Description: Briefly describe the agent's purpose.
-
Phone Number: Set for inbound and outbound calls.
-
LLM Configuration:
-
LLM Model: Choose the desired AI model.
-
Language: Set to English.
-
Language Switching: Enable if multilingual support is needed.
-
Synthesizer Configuration:
-
Speed: Adjust to a medium pace.
-
Voice: Select between 'Chetan' or 'Lightning'.
-
Knowledge Base Configuration:
-
Global Knowledge Base: Select an appropriate knowledge base for decision-making.
-
Design Workflow:
-
Node 1: Greeting and interest Confirmation.
-
Node 2: Collect candidate Information (e.g., experience, location, current salary).
-
Node 3: Provide company Information.
-
Node 4: Schedule a Follow-up if needed.
-
Node 5: Handle disinterest and gather feedback.
-
Node 6: Thank candidate & end Call.
Connect these nodes logically to automate the initial screening process effectively.
What makes this approach effective?
-
Accelerates hiring process: Automates initial candidate interactions, significantly reducing the turnaround time from initial contact to the interview stage.
-
Eliminates human bias: Consistent AI-driven interviews ensure fairness and uniform evaluation across all applicants.
-
Improves candidate experience: Enables natural, real-time conversations, offering flexibility and enhancing the candidate's overall perception of the recruitment process.
Note: The tools and analysis featured in this section demonstrated clear value based on our internal testing. Our recommendations are entirely independent and not influenced by the tool creators.
AIKosha: India's own dataset platform
India is home to more than 400 languages or dialects, making it one of the most linguistically diverse countries in the world. Hindi alone has 48 official dialects, and Bengali has about 50 variants. But large language models are trained primarily using internet data, which is predominantly English. As per Statista, English was the most popular language for web content, making it logical for Indian AI models to be trained using regional languages (rather than English) to bridge the digital divide.
For the last couple of years, the Indian government’s Bhashini project has been working towards building a public digital platform for regional languages to address these issues. Bhashini spent $6-7 million to collect data from different sources and employed more than 200 people to collect data (text as well as speech) and feed it into the system, following which the data was curated, annotated, and labelled.
Bhashini also has a crowdsourcing model called Bhasha Daan (Hindi for ‘contributing language’), where people can contribute to the building of these datasets by speaking, validating, writing, etc. For instance, you can contribute five sentences in English, Hindi, Telegu, or any other Indian language specific to the site and earn a ‘Suno India Bronze Bhasha Samarthak’ badge. Similarly, you can contribute five translations in these languages to earn a ‘Likho India Bronze Bhasha Samarthak’ badge. You also earn Bolo India and Dekho India bagdes. Bhashini engaged with 70-odd research institutes for this.
On 6 March, the government went a step further with the launch of several key initiatives under the IndiaAI Mission by announcing various initiatives—the AIKosha: IndiaAI Datasets Platform; the AI Compute Portal; the AI Competency Framework for Public Sector Officials; iGOT-AI Mission Karmayogi; the IndiaAI Startups Global Acceleration Program with Station F; the IndiaAI Application Development Initiative; and IndiaAI FutureSkills. These initiatives are aimed at strengthening AI-driven research, innovation, and skill development.
AIKosha currently hosts across 13 sectors. The platform includes an AI sandbox—a safe space where developers can experiment with AI using built-in tools and tutorials. It also helps users find relevant content easily and analyzes datasets to check if they are suitable for AI, scoring them on this parameter. To keep data secure, AIKosha has permission-based access, encryption (both stored and moving data), secure APIs, and firewalls that block harmful traffic in real-time.
Demand for AI talent to outpace supply by 2027
India can establish itself as a global hub for AI talent but the demand for AI professionals is expected to be 1.5-2 times higher than the available workforce by 2027. Hence, the key challenge, and opportunity, lies in reskilling and upskilling existing professionals in emerging AI technologies, according to a new report by Bain and Co.
India’s AI sector could surpass 2.3 million job openings by the end of the period but the AI talent pool is expected to grow to only around 1.2 million, presenting an opportunity to reskill more than 1 million workers, the report states. It adds that AI-related job postings have surged by 21% annually since 2019, with compensation growing 11% annually over the same period. Yet the number of qualified candidates has not kept pace, creating a widening talent gap that is slowing AI adoption.
Even in the US, one in two AI jobs could be left unfilled by 2027. Bain projects AI job demand could reach up to more than 1.3 million in the US over the next two years, while supply is on track to hit less than 645,000—implying the need to reskill up to 700,000 US workers. Germany could see the biggest AI talent gap, with about 62,000 AI professionals available to fill 190,000-219,000 job openings in 2027.
The UK, notes the report, may see talent shortfalls of more than 50%, with 105,000 AI workers available to fill up to 255,000 AI jobs in 2027. Australia could see a shortfall of more than 60,000 AI professionals by 2027, with just 84,000 AI specialists available to fill up to 146,000 jobs.
Is former OpenAI chief scientist Ilya's startup worth $30 bn?
Safe Superintelligence Inc. (SSI), the venture that Ilya Sutskever set up after he quit OpenAI where he was chief scientist, is reportedly in discussions to raise $2 billion, aiming for a valuation of $30 billion, according to The Wall Street Journal.
Sutskever has conveyed to potential investors that SSI intends to pursue a novel path in AI development, describing it as "a different mountain to climb". Operating with a lean team of approximately 20 employees, the company does not plan to release any commercial products until it achieves superintelligence—a form of AI that surpasses human intelligence across various domains.
Photo: Jack Guez /AFP/Getty Images
Sutskever's departure from OpenAI followed internal disagreements, particularly after the controversial ousting of CEO Sam Altman in November 2023—a decision in which Sutskever played a significant role but later expressed regret. These events led to his exit from OpenAI and the subsequent founding of SSI. His company's ambitious fundraising efforts come amid a surge in venture capital investments in AI startups. The first quarter of 2025 witnessed over $30 billion invested in such ventures, marking the highest level since 2021. This trend reflects growing enthusiasm and confidence in the potential of AI technologies.
Meanwhile, Microsoft is reducing reliance on OpenAI with the former developing a new family of AI models called 'MAI' to reduce dependence on OpenAI for its Copilot suite. Microsoft AI CEO Mustafa Suleyman is said to be frustrated with OpenAI’s refusal to share the inner workings of its o1 reasoning model despite Microsoft’s $13 billion investment in OpenAI.
YOU MAY ALSO WANT TO READ:
Why Daimler Truck is banking on software-run trucks and buses
Physics Wallah’s AI push: Immortalizing ‘star teachers,’ building SLM assistants
Investors want a piece of DeepSeek. Its founder says not now.
McDonald gives its restaurants an AI facelift
McDonald’s is rolling out a major technology upgrade across its 43,000 locations, introducing smart kitchen equipment, AI-driven drive-throughs, and advanced management tools. This digital transformation aims to streamline operations, enhance order accuracy, and improve both customer and employee experiences. Brian Rice, the company’s chief information officer, told The Wall Street Journal that the initiative addresses common challenges such as malfunctioning machines and incorrect orders. Read more
How the AI talent race is reshaping the tech job market
What the dot-com bust can tell us about today’s AI boom
AI development faces risk of entrenched bias due to underrepresentation of women: Microsoft exec
|
