Год: 2025

In a move that could reshape the way online speech is managed, Meta recently announced integrating large language models (LLMs) into its content moderation strategy, along with moving “We’ve started using AI LLMs to provide a second opinion on some content before we take enforcement actions,” read a blog post by the company.

For a social media giant that is always under public scrutiny, this is a bold attempt to moderate the issues of overreach and inconsistency.

In a recent video, Meta chief Mark Zuckerberg unveiled sweeping changes to the platform’s strategy. One of the most striking ones was the significant reduction in the number of human fact-checkers across all Meta platforms. As part of the downsizing efforts, human fact-checkers are being shifted from their California base to Texas.

Zuckerberg has long championed free expression as a cornerstone of progress, and according to the company, this move is in tandem with that.

In his 2019 Georgetown address, he argued that that empowering people to voice their ideas not only drives innovation but also challenges existing power structures. Yet, even as he spoke of the virtues of free speech, he warned that too much moderation could tilt the scales of power, stifling diverse voices and diminishing the democratic discourse.

As the industry embraces AI, with AGI on the verge, Meta’s solution of putting AI and collective intelligence at the heart of moderation decisions is futuristic. By using LLMs as a second opinion before enforcement actions, the company claims to refine its approach, reduce wrongful takedowns, and temper the frustrations users often feel when their content is censored.

AI LLMs as a Second Opinion

At the heart of this strategy is Meta’s deployment of AI-driven LLMs to review flagged content.

These models, capable of sifting through massive data troves in seconds, are designed to identify subtle nuances and policy violations that human moderators might overlook. For users, this means fewer errors and more fairness—at least, that’s the goal.

It’s a vision of moderation that, in addition to minimising the number of cases of wrongful takedowns, improves the accuracy of enforcement decisions. At its core, Meta’s use of AI for second opinions is an experiment in trust—trust in technology, in the community, and the overall commitment to self-regulation.

But the question remains: Can AI deliver on this promise without falling prey to its own flaws?

The Bias Problem

Despite their impressive capabilities, LLMs have their limitations. Research has shown that these models often reflect the biases of their creators. Major AI systems—whether from OpenAI, Anthropic, Google, or others—have all been called out for exhibiting ideological leanings.

Andrej Karpathy, a former OpenAI researcher, explained the issue: “LLMs model token patterns without genuine understanding,” making them prone to echoing the biases embedded in their training data. For a system tasked with ensuring fairness, this is a troubling flaw.

xAI CEO Elon Musk has also voiced his concerns, warning that LLMs could exhibit ideological biases, including what he termed a “far-left” leaning. Meanwhile, computer scientist Grady Booch has criticised these models as “unreliable narrators,” capable of producing outputs that are not only deceptive but potentially toxic.

Baybars Orsek, VP of fact-checking at Logically, calls for the industry to unite around sustainable, technology-driven solutions that prioritise accuracy, accountability, and measurable impact. While he sees merit in efforts like community notes, he argues in favour of a professionalised approach to fact-checking. “A professional fact-checking model—enhanced by AI and rigorous methodologies—remains the most effective solution for addressing these issues at scale,” Orsek said.

These highlight a significant challenge Meta faces: if the tools intended to enhance fairness are fundamentally flawed, can they genuinely be relied upon to moderate speech impartially and equitably?

Meta Still Bets Big on Co-Intelligence

Beyond AI, Meta is taking a leaf out of X’s playbook by adopting a decentralised moderation approach with its Community Notes model.

This model, first introduced by former Twitter CEO Jack Dorsey in 2021, relies on users to add context to flagged content, providing a more collaborative approach to moderation. Under Elon Musk’s ownership, Community Notes became a key feature, earning praise for its ability to scale moderation while capturing a diversity of perspectives. Musk himself has commended Zuckerberg for bringing the model to Meta’s platforms.

To its credit, the strength of Community Notes lies in its decentralised nature. By leveraging millions of users, it incorporates a wide range of viewpoints, making moderation more representative and less prone to the pitfalls of top-down decision-making.

The post Free Speech, Really? appeared first on Analytics India Magazine.

American chipmaking giant AMD collaborated with Johns Hopkins University to introduce a new research study titled ‘Agent Laboratory: Using LLM Agents as Research Assistants’. The study focuses on frameworks to accelerate scientific discovery, reduce research costs, and improve research quality.

The framework will accept a human-provided research idea. It will process the same across three stages, namely literature review, experimentation, and report writing, which includes a code repository and a research report.

The researchers used this framework with various AI models, and it found that OpenAI’s o1 preview generates the best research outcomes. Furthermore, AMD also reiterated the importance of human involvement in the research process.

“Human involvement, providing feedback at each stage, significantly improves the overall quality of research,” read the report.

Moreover, Agent Laboratory also ‘significantly reduced’ research expenses, achieving an 84% decrease compared to previous autonomous research models.

“By integrating specialised autonomous agents guided by human oversight, our approach can help researchers spend less time on repetitive tasks and more time on the creative, conceptual aspects of their work,” said AMD researchers.

Check out the GitHub repository here: https://agentlaboratory.github.io/

A New Era of Scientific Breakthroughs

Indeed, AI has often been used to make scientific discoveries. For instance, David Baker, Demis Hassabis, and John Jumper were awarded the 2024 Nobel Prize in Chemistry for predicting and designing protein structures using AI.

Moreover, Google DeepMind’s Graph Networks for Materials Exploration (GNoME), an AI tool for discovering new materials, is said to have discovered over 2.2 million new materials, including 380,000 stable materials. This is equivalent to nearly 800 years’ worth of knowledge.

Recently, several researchers observed significant gains in efficiency using OpenAI’s o1 models for scientific tasks. “I just had o1 write a major cancer treatment project based on a very specific immunological approach. It created the full framework of the project in under a minute, with highly creative aims, approaches, and even considerations for potential pitfalls and alternative strategies,” said Derya Unutmaz, a biomedical scientist in a post on X.

The integration of an agentic workflow to enhance scientific research works like these can certainly have a positive impact on science. Moreover, several leaders in the AI industry are betting big on AI agents’ potential.

“We believe that, in 2025, we may see the first AI agents join the workforce and materially change the output of companies,” Sam Altman, the CEO of OpenAI, wrote in a recent blog post.

On the other hand, NVIDIA launched the Llama Nemotron LLMs and Cosmos Nemotron vision language models (VLMs), to improve agentic AI workflows.

“Agentic AI is the next frontier of AI development, and delivering on this opportunity requires full-stack optimisation across a system of LLMs to deliver efficient, accurate AI agents,” said Ahmad Al-Dahle, vice president and head of GenAI at Meta, the company behind the Llama models.

The post AMD Introduces Agent Laboratory, Transforms LLMs into Research Assistants appeared first on Analytics India Magazine.

Eli Lilly and Company (Lilly), the American pharmaceutical company, has announced plans to establish a new global capability centre (GCC) in Hyderabad, India, further strengthening its presence in the country. The centre, slated to open by mid-2025, will employ over 1,000 highly skilled professionals to bolster the company’s digital strategy and technological capabilities.

Named Lilly Capability Centre India (LCCI) Hyderabad, the new facility will focus on automation, AI, software engineering, and cloud computing to support Lilly’s global operations. This will be the pharmaceutical giant’s second capability centre in India, following the launch of LCCI Bengaluru in 2016.

The centre aims to enhance innovation and efficiency by leveraging data insights and advanced technologies, ultimately accelerating the development of next-generation medicines. Recruitment for LCCI Hyderabad is underway, with plans to hire between 1,000 and 1,500 professionals, including technology engineers and data scientists.

“Our teams in India play a pivotal role in driving Lilly’s global business support with centralised, scalable solutions,” said Manish Arora, managing director of LCCI Hyderabad. “LCCI Hyderabad will enhance our technical capabilities, enabling us to better meet growing business demands, fully leverage technology in our operations, and further expand our presence in India.”

Duddilla Sridhar Babu, Telangana’s Minister for Information Technology, expressed confidence in the centre’s impact. “We are pleased to welcome Lilly’s new global capability centre to Hyderabad. This investment underscores Hyderabad’s growing reputation as a hub for healthcare innovation and talent. We fully support Lilly’s vision to make life better for people around the world and are confident that this centre will contribute significantly to the local economy and help improve health globally,” he stated.

Lilly’s global leadership also highlighted Hyderabad’s significance. “Hyderabad is a hub of innovation with decades of history in technology, and we are excited to announce plans to launch a new centre here. We will bring together talented technology professionals who want to make life better for people around the world,” said Diogo Rau, executive VP and chief information and digital officer of Lilly.

The post Eli Lilly to Launch GCC in Hyderabad, Plans to Employ Over 1,000 Professionals appeared first on Analytics India Magazine.

On January 7, a powerful earthquake struck Tibet near Mount Everest, claiming at least 126 lives and leaving over 180 injured. Chinese authorities measured the quake at a magnitude of 6.8, while the US Geological Survey measured it at 7.1. The earthquake sent tremors across Nepal, Bhutan, and parts of India.

Yet, amidst the significant loss and damage, a critical question arises: Could this disaster have been foreseen?

In 2023, during a seven-month AI trial in China, researchers from The University of Texas at Austin collaborated with Chinese scientists to report a 70% success rate in predicting earthquakes a week in advance. Such advancements showcased AI’s potential to mitigate disaster risks.

However, the absence of advance warnings for the deadly earthquake in Tibet raises pressing concerns about the current limitations of these technologies.

Research Papers Claim to Predict Earthquake

“If someone claims they can predict earthquakes, they are misguided,” M. Ravichandran, Secretary, Ministry of Earth Sciences (MoES) told AIM. “No one in the world is currently able to predict earthquakes. Although progress is being made in this direction, a reliable system is not yet in place even here in India.”

Elaborating on this, Munish Bhatia, an assistant professor at NIT Kurukshetra and co-author of a research paper on AI-based earthquake prediction, explained the complexity of the task to AIM.

“Predicting earthquakes with precision is highly challenging due to the dynamic nature of seismic activity and tectonic plate movements. While analysing parameters like seismic activity and energy release allow for certain predictions, the Earth’s crust is too complex to fully decode.”

Bhatia mentioned that, in the study, they developed a mathematical model that incorporates various parameters, including time-domain data, spectral conditions, energy release, and entropy. Yet, he pointed out that accuracy depends heavily on the quality and scope of input data. Controlled experiments may yield promising results, but real-world applications face significantly more variability and unpredictability.

Speaking on other research papers published on the same lines, Bhatia explained, “The prediction model’s accuracy can vary significantly depending on the number and type of parameters and variables incorporated. It also depends on the scenario in which the experiments were conducted, the type of simulation models used, and the conditions under which they were developed.”

For example, achieving a 70% prediction accuracy is not exceptionally high. Instead, it only represents a probability and depends on the quality of the model and the data used.

“To improve accuracy, it is crucial to train models on large datasets, including historical data about earthquake occurrences in specific regions. Regional history, entropy variations, and space-time graph analysis are critical factors. In our experiments, we worked within a highly controlled environment, which contributed to achieving better accuracy. However, if additional parameters are introduced, the model’s accuracy could decrease due to increased complexity,” Bhatia added.

How Close are India’s Predictions?

As countries prepare to tackle disasters, India has also taken steps to ensure human safety in such situations.

In an exclusive conversation with AIM, GS Srinivasa Reddy, former director of the Karnataka State Natural Disaster Monitoring Centre (KSNDMC), explained, “At present, India does not have systems for earthquake prediction, even at the central government level. While the National Center for Seismology (NCS) is in the process of developing AI-based technologies, we are still far from achieving reliable earthquake forecasting.”

“For instance, Japan employs advanced technologies such as the Morter System, supported by over 4,000 sensors, to predict earthquakes. In contrast, India has only 160 to 200 sensors,” he added.

Reflecting on the same, Ravichandran said, “In Japan, early warning systems detect primary waves to issue alerts for the more destructive secondary waves, providing time for people to prepare.”

Notably, even the weather forecast at present doesn’t provide 100% accurate readings with AI in the picture.

“Despite advancements in technology, India’s weather forecast systems currently achieve an accuracy of about 80%, with longer-term forecasts (e.g., three-month or one-month predictions) falling below 70% accuracy. The forecasting period is also limited due to technological and operational constraints,” Reddy mentioned.

He also said that, in Karnataka specifically, KSNDMC does not operate its own forecasting systems. Instead, it relies on outputs from other agencies, such as the India Meteorological Department (IMD) and Space Applications Centre (SAC) Ahmedabad. Accurate weather forecasting requires substantial expertise, infrastructure, and high-performance computing resources like supercomputers.

Looking ahead at the successful implementation of AI in forecasting, the IMD recently initiated efforts to incorporate AI into its forecasting systems. However, this technology has been in use for only the past year. Effective integration of AI in forecasting involves extensive experimental phases, and it typically takes more than two years to transition these systems into practical and reliable applications.

Even with advanced AI technologies, achieving high accuracy and authenticity in forecasts remains a significant challenge. The IMD is currently focused on overcoming these hurdles before releasing AI-based forecasts for public use.

AI Models Can be Trained for More Accuracy

Bhatia’s team specifically focused on real-time data available in open-source libraries, including UCI, Kaggle, and DataPort. The data used was authenticated and collected through IoT sensors deployed in various regions. This allowed them to train the model effectively. The data was split into 80% for training and 20% for testing, which enabled them to achieve the results presented in the article.

The team used Stanford University’s earthquake dataset, which is openly available on GitHub. Anyone working in AI or ML can access this dataset to train their models and explore similar applications as per their requirements.

“Emerging technologies like pre-trained transformer models, high-performance computing from NVIDIA and Intel, and the nascent field of quantum computing are set to further work towards seismic prediction. But even with these tools, the question remains: How close are we to making reliable earthquake predictions a global reality?” Bhatia asked.

What’s Next?

Countries like Japan have focused on AI strategies to ensure human safety during disasters. Since 2019, Japan’s Artificial Intelligence Technology Strategy has integrated AI into urban development, which has improved disaster preparedness and streamlined evacuation protocols.

Similarly, a 2022 study by Hiroshima University introduced a novel neural network model to estimate site-specific earthquake impacts by analysing vibrations and overcoming the limitations of traditional methods like microtremor horizontal-to-vertical spectral ratios (MHVR).

Despite AI’s potential, predicting earthquakes continues to be a challenge. The recent earthquake in Tibet serves as a reminder that while AI offers hope, there’s still a need to make accurate and timely predictions a reality.

The post Tibet Tragedy Shows AI Can’t Predict Earthquakes—Yet appeared first on Analytics India Magazine.

On January 3, the central government’s ministry of electronics and IT released the draft Digital Personal Data Protection (DPDP) Rules 2025, which will be open to public feedback until next month. The draft DPDP rules aim to implement the DPDP Act 2023 to align with India’s commitment to safeguard digital personal data.

The key provisions include requirements to inform individuals about the personal data collected and its usage, consent, mandates for implementing reasonable security measures, and other measures to manage and protect digital data. Furthermore, a central sectoral committee will review and recommend data localisation for specific industries.

However, what does all this mean for Indian AI startups?

New Wave of Ethical Data Collection

With data collection being the foundation for AI models, the process will now come under the radar to ensure there is a fair and just way of collecting them while protecting them at the same time.

A number of Indian AI startups that work on voice and multilingualism source data from different avenues. For instance, Bengaluru-based Gnani.ai’s voice-first small language models are trained on millions of audio hours of proprietary datasets and billions of Indic language conversations.

“For Gnani AI, which specialises in collecting and processing voice datasets, this rule [DPDP] only reinforces what we already do. We ensure every dataset we work with is collected ethically, with user consent and full transparency,” Ganesh Gopalan, co-founder and CEO of Gnani.ai, told AIM.

“I’m sure most startups in the space will also adapt quickly, as the Act provides a clear framework that supports innovation while safeguarding privacy.”

Gopalan believes that the new rule will establish a consent-driven framework to enhance transparency and data ethics in India. He is also positive that rather than causing hindrance to data collection, the rule will promote the creation of improved systems for managing consent and data usage.

Companies such as Karya employ rural women to collect voice data and claim to follow an ethical process for the same – something they have been following from its inception.

Data Protection Needs Will Surge

Anshu Sharma, co-founder and CEO of Skyflow, a data privacy and security company that helps businesses manage and protect sensitive customer information, believes the DPDP rule will provide a framework that aligns with consumer protection.

“AI-driven startups will need to be built with a privacy-first mindset that includes fine-grained data access controls and advanced security safeguards including polymorphic data encryption, tokenisation, and masking to ensure sensitive data doesn’t leak into models,” Sharma told AIM.

He went on to explain how the DPDP mirrors elements of Europe’s GDPR, thereby bringing significant implications for technology companies relying on sharing data internationally. Rule 14 of the draft brings in strict data residency requirements regulating cross-border data transfers.

Increased Cost?

With focused security measures, startups should be prepared for additional costs.

Shlok Bhartiya, co-founder and CEO of AI-powered fashion search platform shoppin’, told AIM that the rules could result in additional compliance costs for businesses, but they have already been factored in as a necessary and non-negotiable cost for a secure user experience.

“AI startups will need to align with the DPDP Act’s requirements, which include but are not limited to obtaining explicit consent for data collection and processing. This could increase operational complexity and necessitate the development of robust consent management systems. If no exemptions are granted, startups will need to fully integrate these measures into their operations, potentially diverting resources from growth initiatives,” emphasised Tisha Bhambry, director analyst at Gartner.

Data Sovereignty Will Lead the Way

A crucial element of the DPDP rule applies to data sovereignty, where companies will be under strict control over personal data being stored and processed in India. Furthermore, companies looking to transfer this data to foreign governments and entities should comply with conditions set by the central government.

“AI startups relying on overseas cloud providers will need to reassess their data strategies based on which countries are mentioned in the restricted list. This may involve exploring partnerships with providers that have data centres in India or adopting hybrid cloud solutions,” said Bhambry, who acknowledged that this shift could involve additional costs.

However, the rule may possibly be a welcome move for local data centre players such as Yotta, Reliance and Tata, who are investing heavily in building data centres in India.

Gopalan doesn’t foresee this being an issue for startups, as most cloud providers are already equipped to support data sovereignty requirements. “This will ensure a seamless transition for companies and continue to foster innovation while adhering to the law.”

While the rules may sound stringent, Indian AI startups seem to be already geared for it and don’t look at this as impeding their progress.

The post What Do the Centre’s DPDP Rules Mean for Indian AI Startups? appeared first on Analytics India Magazine.

Despite the hype around generative AI, studies show just a fraction of GenAI projects have made it into production. A big reason for this shortfall is the concern organizations have about the tendency for large language models (LLMs) to hallucinate and give inconsistent answers. One way organizations are responding to these concerns is by implementing trust layers for AI.

Generative models, such as LLMs, are powerful because they can be trained using large amounts of unstructured data, and then respond to questions based on what they have “learned” from said unstructured data (text, documents, recordings, pictures, and videos). Organizations are finding this generative capability incredibly useful for the creation of chatbots, co-pilots, and even semi-autonomous agents that can handle language-based tasks on their own.

However, an LLM user has little control over how the pre-trained model will respond to these questions, or prompts. And in some cases, the LLM will generate wild answers completely disconnected from reality. This tendency to hallucinate–or as NIST calls it, to confabulate—cannot be fully eliminated, as its inherent with how these types of non-deterministic, generative models are designed. Therefore, it must be monitored and controlled.

One of the ways organizations can keep LLMs from going off the rails is by implementing an AI trust layer. An AI trust layer can take several forms. Salesforce, for example, uses multiple approaches to reduce the odds that a customer has a poor experience with its Einstein AI models, including by using secure data retrieval, dynamic grounding, and data masking, toxicity detection, and zero retention during the prompting stage.

While the Salesforce Einstein Trust Layer is gaining ground among Salesforce customers, other organizations are looking for AI trust layers that work with a range of different GenAI platforms and LLM models. One of the vendors building an independent AI trust layer that can work across a range of platforms, systems, and models is Galileo.

Voyage of AI Discovery

Before co-founding Galileo in 2021 with fellow engineers Atindriyo Sanyal and Vikram Chatterji, COO Yash Sheth spent a decade at Google, where he built LLMs for speech recognition. The early exposure to LLMs and experience working with them taught Sheth a lot about how these types of models work–or don’t work, as the case may be.

“We saw that LLMs are going to unlock 80% of the world’s information, which is unstructured data,” Sheth told BigDATAwire in an interview at re:Invent last month. “But it was extremely hard to adapt or to apply these models onto different applications because these are non-deterministic systems. Unlike any other AI that is predictive, that gives you the same answer every time, generative AI does not give you the same answer every time.”

Sheth and his Galileo co-founders recognized very early on that the non-deterministic nature of these models would make it very difficult to get them into production in enterprise accounts, which have less appetite for risk when it comes to privacy, security, and putting one’s reputation on the line than the move-fast-and-break-stuff Silicon Valley crowd. If these LLMs were going to be exposed to tens of millions of people and achieve the trillions of dollars in value that have been promised, this problem had to be solved.

“To actually mitigate the risk when it’s applied to mission critical tasks,” Sheth said, “you need to have a trust framework around it that can ensure that these models behave the way we want them to be, out there in the wild, in production.”

Starting in 2021, Galileo has taken a fundamentally different approach to solving this problem compared to many of the other vendors that have popped up since ChatGPT landed on us in late 2022, Sheth said. While some vendors were quick to apply frameworks for traditional machine learning, Galileo spent the better part of two years conducting research, publishing papers, and developing its first product built specifically for language models, Generative AI Studio, which it launched in August 2023.

“We want to be very thorough in our research because again, we are not building the tool–we are building the technology that works for everyone,” Sheth said.

Mitigating Bad Outcomes

At the core of the Galileo’s approach to building an AI trust layer is another foundation model, which the company uses to analyze the behavior of the LLM at issue. On top of that, the company has developed its own set of metrics for tracking the LLM behavior. When the metrics indicate bad behavior is occurring, they activate guardrails to block it.

“The way this works is we have our own evaluation foundation models that act, and these are dependable, reliable models that give you the same output every time,” Sheth explained. “And these are models that can run all the time in production at scale. Because of the non-deterministic nature, you want to set up these guardrails. These metrics that are computed each time in production and in real time, in low latency, block the hallucinations, block bad outcomes from happening.”

Galileo helps customers implement guard rails for GenAI (phoelixDE/Shutterstock)

There are three components of Galileo’s suite today: Evaluate, for conducting experiments across a customer’s GenAI stack; Observe which monitors LLM behavior to ensure a secure, performant, and positive user experience;, and Protect, which prevents LLMs from responding to harmful requests, leaking data, or sharing hallucinations.

Taken together, the Galileo suite enables customers to trust their GenAI applications the same way they trust their regular apps developed using deterministic methods, Sheth said. Plus, they can run Galileo wherever they like: on any platform, AI model, or system.

“Today software teams can ship or launch their applications almost on a daily basis. And why is that possible?” he asks. “Two decades ago, around the dot-com era, it used to take teams a quarter to launch the next version of their application. Now you get an update on your phone every like every few days. That’s because software now has a trust layer.”

The tooling involved in an AI trust layer look significantly different than what a standard DevOps team is used to, that’s because the technology is fundamentally different. But the end result is the same, according to Sheth–it gives development teams the peace of mind to know that, if something goes awry in production, it will be quickly detected and the system can be rolled back to a known good state.

Gaining GenAI Traction

Since launching its first product barely a year-and-a-half ago, Galileo has begun to generate some momentum. The company has a handful of customers in the Fortune 100, including Comcast, Twilio, and ServiceNow, and established a partnership with HPE in July. It raised $45 million in a Series B round in October, bringing its total venture funding to $68.1 million.

As 2025 kicks off, the need for AI trust layers is palpable. Enterprises are champing at the bit to release their GenAI experiments into production, but officers just can’t sign off until some of the rough edges are sanded down. Sheth is convinced that Galileo has the right approach to mitigating bad outcomes from non-deterministic AI systems, and giving enterprises the confidence they need to green light the GenAI.

“There are amazing use cases that I’ve never seen possible with traditional AI,” he said. “When mission critical software starts becoming infused by AI, what’s going to happen to the trust layer? You’re going to go back to the stone ages of software. That’s what is hindering all the POCs that are happening today from reaching production.”

This article first appeared on sister site BigDATAwire.