Madhya Pradesh has secured investment proposals worth INR 3,200 crore in a single day during an interactive session held in Bengaluru on August 7-8, 2024. The event, organised by Invest Madhya Pradesh, drew over 500 participants, including leading industrialists and investors from the IT, textiles, aerospace, and pharmaceutical sectors.

Among the significant proposals, Google Cloud announced plans to establish a startup hub and Center of Excellence in the state. Chief Minister Mohan Yadav shared that these initiatives aim to enhance the local skilled workforce. 

Meanwhile, chip giant NVIDIA suggested creating a blueprint to transform Madhya Pradesh into the “Intelligence Capital of India.”

The session also saw participation from more than 15 international diplomatic missions and major companies like Infosys, Cognizant, and TCS, further solidifying Madhya Pradesh’s reputation as a prime investment destination.

Chief Minister Yadav highlighted that the proposals could generate approximately 7,000 new jobs in the state, providing a substantial boost to the local economy.

“Discussions were held with these companies regarding the development of information technology in Madhya Pradesh and their future plans, and I am hopeful that with the kind of positive response that we have got, we will witness many IT companies setting up their campuses in Madhya Pradesh,” he said.

In February, Google had signed an MoU with the Maharashtra government to advance scalable AI solutions in sectors like agriculture and healthcare. The recent proposals in Madhya Pradesh signal a continued commitment from tech giants to invest in India’s technological growth.

The post Google, NVIDIA, and Microsoft to Invest INR 3,200 Crore in Madhya Pradesh appeared first on AIM.

The headline seems apt for Google at the moment. Google DeepMind’s latest model, Gemini 1.5 Pro’s experimental version, 0801, recently claimed the top spot in Arena, surpassing GPT-4o and Claude-3.5 with an impressive score of 1300. It also achieved first place on the Vision Leaderboard.

It excels in multilingual tasks and delivers a robust performance in technical areas like math, hard prompts, and coding. This version of Gemini 1.5 Pro is available for early testing and feedback in Google AI Studio and the Gemini API, where developers can try it out. 

“Very impressed with the extraction capabilities from images and PDFs. It does feel super close to what GPT-4o gives in terms of vision capabilities and what Claude 3.5 Sonnet gives me on code generation and PDF understanding/reasoning,” said Elvis Saravia, co-founder, DAIR.AI.

“The new Gemini model can execute code and is quite ‘sharp.’ We might see it become the first competitor to the Code Interpreter,” said Ethan Mollick, assistant professor at Wharton. He further said that, while experimenting with it for data analysis, some capabilities seemed odd, like the fact that it couldn’t access files unless they are uploaded each time. “Nevertheless, it seems promising so far,” he said.

Most recently, Google also upgraded its free tier Gemini model with 1.5 Flash, which brings improvements in quality and latency, with especially noticeable enhancements in reasoning and image understanding. The tech giant also expanded the context window in Gemini Advanced, quadrupling it to 32K tokens. 

Following in the footsteps of its competitor OpenAI, Google has announced improvements to Gemini 1.5 Flash, reducing input costs by up to 85% and output costs by up to 80%, effective from August 12, 2024.

Apart from focusing on the LLMs, Google entered the SLM scene as well by releasing Gemma 2. This new model outperformed GPT-3.5 on the Chatbot Arena leaderboard and is optimised for efficient deployment across a wide range of hardware, from edge devices to robust cloud environments. Leveraging NVIDIA’s TensorRT-LLM library optimisations, it supports deployments in data centres, local workstations, and edge AI applications. 

InflectionAI, AdeptAI, CharacterAI. Who’s next?

Google recently brought back Noam Shazeer, one of the original authors of the famous Transformers paper ‘Attention is All You Need’, who had left Google to create his own chatbot company, Character.AI. Surprisingly, Character.AI has more web traffic in the US than Gemini, according to Similarweb. 

This is similar to what Microsoft and Amazon have done with Inflection and Adept AI, respectively. Microsoft hired much of Inflection’s leadership team and employees, agreeing to pay a licensing fee of approximately $650 million. Meanwhile, Amazon hired Adept’s co-founder and CEO David Luan, along with several other co-founders and key team members.

“Absolutely delighted to welcome my longtime Google colleague Noam Shazeer back to Google!” said Jeff Dean, Google Deepmind’s chief scientist.

Similarly, others from Deepmind followed suit in welcoming Shazeer back. “Welcome back Noam Shazeer to Google! It’ll be a great time working together again since 2018. Let’s take Gemini which is #1 and continue expanding the limits of its capabilities,” said Dustin Tran, research scientist at Google Deepmind. 

In a new arrangement, Google will pay a licensing fee to Character.AI for its models. Around 30 of Character.AI’s 130 staff members, specialising in model training and voice AI, will join Google to advance its Gemini AI initiatives. 

Interestingly, Character.AI recently introduced Character Calls, allowing users to engage in two-way voice conversations with their favourite characters, like Michael Jackson, Sherlock Holmes, or Albert Einstein. “It’s like having a phone call with a friend,” the company said.

“So this also means that google is getting into the “ai companion” market to compete with openai’s “her” lol,” one user posted on X in response. 

Character.AI allows users to create their own chatbots, defining their personalities, traits, and backstories. These characters can be based on real people, fictional characters, or entirely new creations. 

This is similar to what Meta AI is doing by introducing customised AI characters across its social media platforms. Meta recently introduced AI Studio, a new platform that allows users to create, share, and discover custom AI characters without needing technical skills.

OpenAI recently rolled its advanced voice mode to a small group, which has been received well with the users finding several usecases for it. However, GPT-4o is still not multimodal yet. 

“Advanced voice mode on ChatGPT is very impressive, but it doesn’t yet have the full features enabled: no multimodal video or images in, no GPTs, no new image generator, no code interpreter or internet access,” said Mollick.

On the other hand, Google Deepmind’s latest models AlphaProof and AlphaGeometry 2 solved four out of six problems from this year’s International Mathematical Olympiad (IMO), achieving a score equivalent to a silver medalist in the competition. 

The word on the street is that OpenAI is also working on advanced reasoning capabilities for its next frontier model with ‘Project Strawberry’, which is most likely to be GPT-5. However, it appears that it is not coming out anytime soon.

“Little birdies told me Q* hasn’t been released yet as they aren’t happy with the latency and other little things they want to further optimise,” posted an insider who goes by the name Jimmy Apples on X.

“Not sure if it’s big patience or small patience.”

The post You Don’t Mess with Google appeared first on AIM.

Microsoft has once again claimed the crown of cloud, despite mixed market reactions after it reported its results.

In the previous quarter, Microsoft Azure notably encroached on AWS’s market share and it continues to ride this wave. In the latest quarter, Microsoft Azure’s Intelligent Cloud revenue, which includes the company’s server products and cloud services, rose to $28.5 billion, a 19 per cent increase year over year. This segment now constitutes nearly 45 percent of Microsoft’s total revenue. 

Meanwhile, AWS reported a revenue of $26.28 billion, a 19 per cent increase, surpassing analysts’ expectations of $26.02 billion according to StreetAccount. Google Cloud, meanwhile, experienced a 29 per cent rise in revenue, reaching $10.3 billion, slightly above the projected $10.2 billion.

Combined, AWS, Google Cloud and Microsoft Azure accounted for a whopping 67 per cent share of the $76 billion global cloud services market in Q1 2024, according to new data from IT market research firm Synergy. It needs to be seen, however, if Microsoft Azure has increased its market share. 

Source: Statista

In Generative AI We Trust 

One thing common amongst the three hyperscalers is that they have kept their trust in generative AI even though it hasn’t really started to pay off. 

Microsoft announced plans to spend more money this fiscal year to enhance its AI infrastructure, even as growth in its cloud business has slowed, suggesting that the AI payoff will take longer than expected.

Microsoft CFO Amy Hood explained that this spending is essential to meet the demand for AI services, adding that the company is investing in assets that “will be monetised over 15 years and beyond.” Meanwhile, CEO Satya Nadella said that Azure AI now boasts over 60,000 customers, marking a nearly 60% increase year-on-year, with the average spending per customer also on the rise. 

“For the next quarter, we expect Azure’s Q1 revenue growth to be 28% to 29% in constant currency,” said Hood. “Growth will continue to be driven by our consumption business, including AI, which is growing faster than total Azure.”

On similar lines, Google is facing increasing AI infrastructure costs. “The risk of under-investing far outweighs the risk of over-investing for us. Not investing to stay ahead in AI carries much more significant risks,” warned Google CEO Sundar Pichai.

In another news, Meta chief Mark Zuckerberg said that to train Llama 4, the company will need ten times more compute than what was needed to train Llama 3.

The Azure OpenAI service provides access to best-in-class frontier models, including GPT-4o and GPT-4o mini. Apart from that, Azure also offers in-house built AI models like Phi-3, a family of powerful small language models, which are being used by companies like BlackRock, Emirates, Epic, ITC, and Navy Federal Credit Union.

“With Models as a Service, we provide API access to third-party models, including as of last week, the latest from Cohere, Meta, and Mistral. The number of paid Models as a Service customers more than doubled quarter over quarter, and we are seeing increased usage by leaders in every industry from Adobe and Bridgestone to Novo Nordisk and Palantir,” said Nadella.

Microsoft is trying hard not to be dependent on OpenAI and has listed the startup as its competitor in generative AI and search. This move might have come after OpenAI cozied up to Apple by integrating ChatGPT into Siri.

Similarly, AWS Bedrock has been constantly adding new models to its offerings. “Bedrock has recently added Anthropic’s Claude 3.5 models, which are the best-performing models on the planet, Meta’s new Llama 3.1 models, and Mistral’s new Large 2 models,” said Amazon chief Andrew Jassy.

Last year, Amazon also announced its generative AI model called Q. “With Q’s code transformation capabilities, Amazon has migrated over 30,000 Java JDK applications in a few months, saving the company $260 million and 4,500 developer years compared to what it would have otherwise cost,” said Jassy.  

Google is also quite bullish with Gemini. Most recently, Google DeepMind’s new Gemini 1.5 Pro’s experimental version, 0801, was tested in Arena for the past week, gathering over 12K community votes.

For the first time, Google Gemini has claimed the 1st spot, surpassing GPT-4 and Claude-3.5 with an impressive score of 1300, and also achieving the first position on the Vision Leaderboard.

Google Vertex AI includes all models from the Gemini and Gemma families, such as Gemini 1.5 Pro, Gemini 1.5 Flash, and Gemma 2. It also offers third-party models from Anthropic, Mistral and Meta.

“Uber and WPP are using Gemini Pro 1.5 and Gemini Flash 1.5 in areas like customer experience and marketing. We broadened support for third-party models including Anthropic’s Claude 3.5 Sonnet and open-source models like Gemma 2, Llama, and Mistral,” Pichai said.

Some of the notable customers of Google Cloud are Hitachi, Motorola Mobility, KPMG, Deutsche Bank, and Kingfisher, as well as the US government.

Building In-House AI Chips 

NVIDIA’s upcoming Blackwell chip has been delayed by three months or more due to design flaws, a setback that could impact customers such as Meta, Google, and Microsoft, who have collectively ordered tens of billions of dollars worth of the chips.

Ahead of this, all the hyperscalers, apart from utilising NVIDIA GPUs, have also been developing their own AI chips. “We added new AI accelerators from AMD and NVIDIA, as well as our own first-party silicon chips, Azure Maia, and we introduced the new Cobalt 100, which provides best-in-class performance for customers like Elastic, MongoDB, Siemens, Snowflake, and Teradata,” said Nadella. 

Google also recently launched Trillium which was used by Apple to train its foundation models. “Trillium is the sixth generation of our custom AI accelerator, and it’s our best-performing and most energy-efficient TPU to date. It achieves a near five time increase in peak compute performance per chip and is 67 percent more energy efficient compared to TPU v5e,” said Pichai. 

“Year-to-date, our AI infrastructure and generative AI solutions for cloud customers have already generated billions in revenues and are being used by more than two million developers,” he added.

AWS has also developed its custom silicon chips. “We’ve invested in our own custom silicon with Trainium for training and Inferentia for inference. The second versions of those chips, with Trainium coming later this year, are very compelling on price performance. We are seeing significant demand for these chips,” Jassy said.

The post Microsoft Azure Crushes Cloud Competition, Leaves AWS and Google Cloud Behind appeared first on AIM.

In a move to scale their global marketing campaigns, the Coca-Cola Company has partnered with WPP and NVIDIA to integrate generative AI capabilities using NVIDIA Omniverse and NVIDIA NIM microservices. 

This collaboration, working through WPP Open X, will leverage NVIDIA‘s technology to personalise and customise brand imagery across over 100 markets worldwide. 

This initiative is part of Coca-Cola’s digital transformation strategy, led by Samir Bhutada, global VP of StudioX Digital Transformation.

WPP announced at SIGGRAPH that Coca-Cola will be among the first to adopt NVIDIA NIM microservices for OpenUSD in its Prod X studio. These include USD Search and USD Code, allowing the creation and manipulation of 3D advertising assets with culturally relevant elements using AI-generated images and prompt engineering.

“Our research with NVIDIA Omniverse for the past few years, and the research and development associated with having built our own core USD pipeline and decades of experience in 3D workflows, enabled us to create a tailored experience for the Coca-Cola Company” said Priti Mhatre, managing director for strategic consulting and AI at Hogarth. 

The post WPP and NVIDIA Omniverse Help Coca-Cola Scale Brand-Authentic Generative AI Content appeared first on AIM.

Google Research has developed a breakthrough hybrid general circulation model (GCM) that combines cutting-edge machine learning components with conventional physics-based techniques to improve weather forecast. 

This innovative research on Neural General Circulation Models, which was published in Nature, demonstrates how NeuralGCM may improve weather and climate prediction accuracy beyond that of standalone machine-learning models and traditional GCMs.

NeuralGCM, which was created in collaboration with the European Centre for Medium-Range Weather Forecasts (ECMWF), enhances simulation efficiency and accuracy by fusing ML with conventional physics-based modelling.

Breakthrough in Climate Modelling

Google CEO Sundar Pichai has called it a breakthrough in climate modelling. This is because when compared with the existing gold-standard physics-based models, Google claims that this approach offers weather forecasts that are 2–15 days more accurate. 

Besides, it is also capable of reproducing temperatures over the last 40 years more accurately than traditional atmospheric models. Unlike traditional models, NeuralGCM combines traditional physics-based modelling with ML for improved simulation accuracy and efficiency. 

According to Stephan Hoyer, an AI researcher at Google Research, NeuralGCM is a combination of physics and AI.

To prove their claim, the researchers used a defined set of forecasting tests called WeatherBench 2 to compare NeuralGCM against other models. NeuralGCM performed comparably to other machine-learning weather models like Pangu and GraphCast for three- and five-day forecasts.  

Not The Only One

While NeuralGCM can be called a breakthrough in climate modelling, it isn’t the only one. NVIDIA Earth-2 is a full-stack, open platform that combines physical simulations and machine learning models, like FourCastNet and GraphCast, with NVIDIA’s tools for data visualisation. 

However, unlike NeuralGCM, Earth 2 focuses on creating a virtual representation of Earth to quickly and accurately simulate and visualise the global atmosphere.

Then, there is the AI2 Climate Emulator developed by the Allen Institute for Artificial Intelligence (AI2). ACE focuses on quickly mimicking complex climate models using deep learning, allowing researchers to run fast simulations and test climate scenarios efficiently. 

Not A Big Achievement

“An important advance in atmospheric modelling and long-term weather prediction, but not necessarily a giant leap in climate prediction.” This was how Texas A&M University atmospheric sciences professor R Saravanan described the findings. Saravanan was not involved in the research.

“The NeuralGCM seems like a significant advancement in pure ML-based modelling at first glance,” Saravanan remarked. “In reality, the opposite is true—the paper emphasises the shortcomings of purely ML-based approaches.”

NASA’s Goddard Institute for Space Studies director Gavin Schmidt said that scientists estimate global heating from greenhouse gases as a range due to climate’s inherent chaos, similar to weather predictions like a “40% chance of rain”. 

“Physics-based models can better address this uncertainty by simulating the underlying physics, while AI models, lacking this capability, struggle to account for the inherent unpredictability,” Schmidt added.

He also cautioned about the latest findings, saying machine learning isn’t a replacement for physics. He claimed that because “weather models don’t conserve [things like] energy and water”, “you end up with massive drifts”, which cause systems that are trained on meteorological data to gradually diverge from reality.

Furthermore, Schmidt said that merely using meteorological data to train an AI does not ensure that the results it produces will adhere to these physical bounds.

The post Google Slashes Computer Power Needed for Weather Forecast by 2-15 Days appeared first on AIM.

LLMs have demonstrated remarkable capabilities, from generating coherent and contextually relevant text, to solving complex mathematical problems. However, these models also exhibit puzzling inconsistencies, such as struggling with seemingly simple tasks.

This phenomenon has led to the concept of what Andrej Karpathy calls “Jagged Intelligence,” a term that captures the uneven performance of LLMs across different types of tasks.

The Paradox of LLM Performance

To prove this, Karpathy recently gave an example, where he asked an LLM-based chatbot to compare two numbers – 9.11 and 9.9. The task to determine which was larger. Despite the simplicity of the task, the model provided an incorrect answer. 

Further illustrating these challenges, Karpathy said that even OpenAI’s GPT-4o failed to provide the correct answer to such simple comparisons approximately one-third of the time. 

Similarly, Anthropic’s Claude struggled with this task in all three attempts, highlighting a notable inconsistency in its performance.

These problems are not unique. Noam Brown, a research engineer at OpenAI, experimented with LLMs by making them play basic games like tic-tac-toe. The outcome was dismal, since LLMs performed poorly in this simple game.

Additionally, another developer tasked GPT-4 with solving tiny Sudoku puzzles. Here too, the model struggled, often failing to solve these seemingly simple puzzles. 

This pattern of inconsistent performance across different tasks underscores the current limitations of LLMs, despite their impressive achievements in more complex areas.

Intelligently Dumb

It’s concerning that while our most advanced models can win a silver medal in a Math Olympiad, they can also fail to answer a simple question like “which number is bigger, 9.11 or 9.9?”

This inconsistency might seem baffling, but NVIDIA’s Senior Research Manager Jim Fan has an explanation: training data distribution. 

Models like AlphaProof and AlphaGeometry-2 are specifically trained on formal proofs and domain-specific symbolic data. This specialised training makes them experts at solving Olympiad problems, even though they are based on general-purpose LLMs.

GPT-4o is trained on a lot of different data, including tons of code from GitHub. There is likely much more code data than maths data. In software versioning, a higher version number like “v9.11” is considered greater than “v9.9”. This might have confused the model when comparing simple numbers, causing it to make a mistake

Can’t Blame LLM Models Alone

A study by Google DeepMind proved that LLMs lack genuine understanding and, as a result, cannot self-correct or adjust their responses on command. 

A recent example by a developer illustrated this limitation. He requested a piece of code using a specific framework (Framework X) to accomplish a particular task (Task Y). 

Unaware that the requested task was not feasible with the given framework, the LLM initially provided code that functioned correctly but did not achieve the desired outcome.

When the user pointed out the discrepancy, the LLM then generated code that syntactically met the requirements but was inherently non-functional. This back-and-forth continued twice, with the LLM failing to recognize the incorrectness of its solutions. 

AI-Clever Charlatans

A very important point to ponder is that LLMs are created by humans. As Karprathy himself pointed out that all of its training data in the last, post-training stage are of the form [question -> authoritative sounding solution], where the solutions are written by humans. The LLMs just imitate the form/style of that training data.

Therefore, since LLM relies on imitating the style of the training data, it can sometimes generate answers that sound authoritative but are factually incorrect or misleading.

When Karpathy asked the LLM to compare two numbers, the model provided an incorrect answer. However, when the same prompt was modified with the instruction to “think step by step before answering,” the LLM was able to arrive at the correct answer.

What’s the Solution? 

Recently, OpenAI released a research paper on Prover-Verifier Games’ (PVG) for LLMs which can solve this problem. It involves two distinct roles: the Prover, which generates solutions, and the Verifier, which checks the validity of these solutions. 

By implementing this system, we can significantly improve the consistency, accuracy, and trustworthiness of LLMs. 

Interestingly, the PVG system alludes to a form of reinforcement learning, something OpenAI’s co-founder and former chief scientist Ilya Sutskever was a strong advocate of.

Causality also offers a framework for improving the intelligence and reliability of LLMs. Causal reasoning would enable AI to understand cause-and-effect relationships, similar to human reasoning. 

In an exclusive interview with AIM, Rohit Bhattacharya, assistant professor of computer science at Williams College said, “We don’t require millions of data points in order to do the things that we do. Reducing the amount of data needed to make these machines function the way they do is one big aspect where causal reasoning can play a role.”

This shift could help overcome limitations in data dependency and improve generalisation to new, unfamiliar situations.

The post Andrej Karpathy Coins ‘Jagged Intelligence’ to Describe SOTA LLMs’ Flaws appeared first on AIM.

“Hello everyone, good morning. Welcome to my YouTube channel. Today is my first day in the Bengaluru office and I’m thrilled to show you my workplace!”

Sounds familiar? This is a common find on YouTube – software engineers capturing their busy workday and sharing it with the world. 

But how exactly do they manage to record all this “content” amidst coding and meeting requirements? 

Nikhil Soni, a Bengaluru-based software engineer with 12.8k subscribers on YouTube, gives us a peek into his daily routine from waking up to hitting the gym, enjoying a wholesome breakfast, a productive stint at work, and winding down at home—a day that seems straight out of a dream for many.

Scroll a bit further, and you’ll find Google employees vlogging about their office perks across Pune, Gurugram, and Bengaluru. Which office offers the best food and who has the comfiest furniture? These seem to be important considerations for the techies today. 

Some of these influencers portray working long hours as not entirely negative, as they complain of boredom during the weekends. 

In a post on X, Venkatesh Gupta, a techie, shared encountering a senior engineer from Microsoft in Bengaluru who drove an auto-rickshaw over the weekends to combat loneliness. 

Long hours would leave him too tired to seek additional activities on weekends. And who knows he might even apply for the Wakefit sleep internship programme next if he’s ‘tired’ from over-working after turning to ‘Namma Yatri’ to beat ‘loneliness’.  

So, perhaps the proposal by Karnataka IT firms suggesting a 14-hour workday isn’t as crazy as it sounds. 

Need for a Backup Job

As per the data by the All India IT & ITeS Employees’ Union (AIITEU), India’s IT sector laid off around 20,000 techies during the year 2023. 

And in 2024 alone, around 2,000-3,000 professionals from India’s top IT companies lost their jobs according to the IT employee union Nascent Information Technology Employees Senate (NITES).

In moments like these, social media and influence come in handy. 

A Bengaluru-based software engineer, Jishnu Mohan, was laid off in February, after sharing his views on layoffs trends and other details about the IT industry.

Mohan, who worked for Forma (formerly Twic) for four years, wrote on X: “The whole recession situation in tech is making me uneasy. Maybe at the lowest confidence level in my career.”

Following this unexpected firing, Mohan appealed for job opportunities on the social media platforms. Fortunately, he got quick responses, with several users offering leads to open positions and expressing willingness to assist with forwarding his resume.

YouTube Pays Well

According to Influencer Market Hub, YouTubers make an average of $0.018 per ad view. However, the amount also depends on factors such as the number of followers, views, clicks on ads, ad quality, ad blockers, and video length.

Source: Social Orange

Here’s an interesting story: Vanessa Chen, a 23-year-old content creator in Boston, shared that she was a computer science undergraduate with plans to become a software engineer. But she became a full-time content creator with more than 4 million followers across YouTube, Instagram, and TikTok and earns a mid-six-figure income.

So, creating a YouTube channel can be a great part-time job, especially if you can produce engaging content even when busy at work!

The post A Day in the Life of a ‘Jobless’ Software Engineer appeared first on AIM.

Google is hosting its next hardware launch event ‘Made by Google’ on August 13. The company has already confirmed that it will announce the Pixel 9, Pixel 9 Pro, and the Pixel 9 Pro Fold at the event in California.

At most product launch events, hardware announcements steal the limelight. However, Google’s software-related announcements are also highly anticipated. One of the key updates to look forward to is on Gemma, Google’s own language model.

What about Gemma 3? 

Meta recently released Llama 3.1. It outperformed OpenAI’s GPT-4o on most benchmarks in categories such as general knowledge, reasoning, reading comprehension, code generation, and multilingual capabilities. 

Similarly, last week, OpenAI released GPT-4o mini, a cost efficient LLM. Priced at 15 cents per million input tokens and 60 cents per million output tokens, GPT-4o mini is 30x cheaper than GPT-40 and 60% cheaper than GPT-3.5 Turbo.

Gemma 2’s last update came over a month ago. The competitive environment in which LLMs operate is changing quickly. For Google to stay in the market, it will be essential for it to innovate and set Gemma apart.

At Made by Google, the tech giant is most likely to release the updated version of Gemma, aka Gemma 3, to stay relevant. 

Limitations of Gemma 2 

Data engineer Maziyar Panahi highlighted issues with Gemma 2’s performance when compared with models like Llama-3-70B and Mixtral. Panahi ran these models in Medical Advanced RAG. 

Panahi noted, “Gemma-2 (27B) trailed… Gemma-2 missed several obvious documents—quite a few mistakes noted! Gemma-2 tends to over-communicate, overlook details, and add unsolicited safety notes.”

Initial technical problems also plagued Gemma 2, as mentioned by a user mikael110 on Reddit. A tokeniser error was corrected relatively quickly, but a more critical issue related to “Logic Soft-Capping” persisted. 

This feature, crucial for the model’s performance, was initially overlooked due to conflicts with the model’s architecture.

Hugging Face has also said that biases or gaps in the training data can lead to limitations in the model’s responses. It also struggles to grasp subtle nuances, sarcasm, or figurative language.

Indian Developers Love Gemma 2 

Despite initial problems, Gemma 2 remains popular among Indian developers. They say they are more comfortable with Gemma than Llama. 

“750 billion tokens are spread across 30 languages, and considering an equal distribution over all 30 languages, it comes out to be 25 billion tokens per non-English language. A language like Hindi is very rich, so I feel it’s grossly underrepresented in Llama 3,” said Adarsh Shirawalmath, the founder of Tensoic.

Similarly, OdiaGenAI released Hindi-Gemma-2B-instruct, a 2 billion SFT with 187k large instruction sets in Hindi. The company said Gemma-2B was chosen as the base model due to 2B versions for CPU and on-device applications and efficient tokenisers on Indic languages compared to other LLMs. 

Recently, Telugu LLM Labs also experimented with Gemma and released Telugu Gemma.

“Models using Llama 2 extended its tokeniser by 20 to 30k tokens, reaching a vocabulary size of 50-60k. Continuous pre-training is crucial for understanding these new tokens. 

In contrast, Gemma’s tokeniser initially handles Indic languages well, requiring minimal fine-tuning for specific tasks,” said Adithya S Kolavi, the founder of Cognitive Lab. 

Not Everything is Lost for Gemma

According to Kolavi’s leaderboard for Indic LLMs, Llama 3 performs significantly better than Llama 2 on most benchmarks. However, compared to Gemma, it falls a little short. Gemma’s tokenisation for Devanagari is efficient when compared to Llama 2.

DeepMind engineer Anil Rohan wrote on X that Gemma 2 27b clearly outperforms Llama 3 70b and other open weight models with excellent post training.

“Gemma probably does a better job at Indic tokenisation than GPT-4 and Llama 3,” said Vivek Raghavan, the co-founder of Sarvam AI, in an exclusive interview with AIM. However, he added that Llama 3 has its own advantages. 

“I think Llama 3 looks quite good. There are many open models and There are many open models and we have a strategy where we leverage all of them,” he added.

The post Google Likely to Release Gemma 3 Next Month appeared first on AIM.

According to the Uptime Institute’s 2024 Outage Analysis, between 10 and 20 “high-profile IT outages or data centre events” occur every year. 

The study revealed that while power is the main cause of data centre outages, network issues are the leading cause of outages across all IT services. These outages make headlines and have serious consequences, disrupt business for customers, and damage company reputations.

More than half of the respondents said their most recent major outage cost them over $100,000, and 16% reported that it cost them over $1 million. Additionally, the report mentions the leading causes of network outages, including design and configuration, hardware, capacity, software, and environmental threats.

Here are eight major tech outages to be explored.

Microsoft-CrowdStrike 

Last week, CrowdStrike, a security technology provider, caused a massive global IT outage, potentially the biggest in history, affecting airlines, banks, businesses, schools, and government services worldwide. 

The CrowdStrike Outage occurred due to a faulty software update in their Falcon sensor program, which caused widespread disruptions to Windows systems globally. This led to the infamous “Blue Screen of Death” and reboot loops for millions of users. 

Excluding Microsoft, US Fortune 500 companies are said to face $5.4 billion in financial losses due to the Windows outage.

Meta 

On October 4, 2021, Meta platforms, including Facebook, Instagram and WhatsApp, experienced an outage lasting nearly six hours. Users faced difficulties accessing the apps, leading to a surge in traffic on competing platforms like Twitter and TikTok. 

During this period, Facebook reportedly lost about $545,000 in US ad revenue per hour. 

Google Services 

Popular Google services such as YouTube, Gmail, Google Drive, and Google Docs were down for an hour, affecting millions of users worldwide on December 14, 2020. 

The outage was attributed to a failure in Google’s authentication system, which manages user logins across its services. The issue specifically stemmed from an internal storage quota problem.

Users attempting to access these platforms encountered errors, with many reporting that they were unable to log in or retrieve their data. Google acknowledged the issue and confirmed that the services were restored for the vast majority of affected users shortly after the outage. 

Fastly 

On June 8, 2021, Fastly, a major content delivery network (CDN) provider, experienced a significant global outage that disrupted numerous high-profile websites, including Amazon, Reddit, and The New York Times. 

The outage was triggered by a software bug that had been introduced during a deployment on May 12, which remained dormant until a valid configuration change made by a customer activated it. 

This led to 85% of Fastly’s network returning errors, resulting in widespread accessibility issues for many internet users around the world. 

Twitter (X Corp)

Twitter suffered a major outage on December 28, 2022, leaving tens of thousands of users unable to access the platform or its features for several hours. It primarily impacted users attempting to access the platform via desktop computers. 

Many reported being unexpectedly logged out, encountering error messages, and facing difficulties in viewing replies or using features like notifications and TweetDeck. The hashtag #TwitterDown trended on the platform as users shared their experiences during the outage. 

AWS 

On December 7, 2021, Amazon Web Services (AWS) experienced a significant outage that disrupted numerous services and affected a wide range of businesses and applications. It primarily impacted the US-East-1 region, located in Northern Virginia, which is crucial for many of AWS’s services. 

The outage was caused by an automated scaling activity designed to increase capacity for service within AWS’s main network. This action unintentionally triggered a surge in connection attempts within AWS’s internal network, overwhelming the devices managing communication between the internal and main networks.

Akamai 

On June 17, 2021, a significant disruption occurred at Akamai, affecting the websites of numerous financial institutions and airlines in Australia and the United States. This outage was traced back to server-related glitches at Akamai, a major content delivery network (CDN) provider. 

The incident marked the second major internet blackout within a week, following a prior outage caused by a rival CDN, Fastly Inc.

Akamai attributed the outage to a bug in its software, which was promptly addressed. The company confirmed that the issue was not related to any cyber-attack or security vulnerability. 

Cloudflare

A power failure led to Cloudflare coming down for around two days. The platform uses the services of three data centres. One such data centre experienced a power failure. The outage was caused by a failure of the facility’s generators and faulty circuit breakers. 

As the generators failed, Cloudflare’s network routers lost power, which disrupted services reliant on the PDX-04 data centre.

The outage primarily affected Cloudflare’s dashboard, APIs, and related services, while traffic through its global network continued to function without interruption.

The post Tech Meltdowns: 8 Epic Outages and What Went Wrong appeared first on AIM.

Researchers from Google have unveiled a new framework called Conditioned Language Policy (CLP) that promises to revolutionise the finetuning of language models by enabling them to balance multiple conflicting objectives efficiently.

The framework addresses the limitations of traditional single-objective finetuning methods, which often require multiple expensive runs to achieve the desired balance between conflicting goals such as creativity and safety. 

CLP leverages techniques from multi-task training and parameter-efficient fine tuning to create steerable language models that can dynamically adjust to different objectives during inference without the need for retraining.

Read the full paper here

The key advantage of CLP lies in its ability to combine multiple reward weightings through a parameter-space conditioning mechanism, resulting in models that not only outperform existing methods but also exhibit superior steerability. This allows users to select from diverse outputs that best meet their needs, enhancing both model quality and flexibility. 

Unlike traditional methods that require separate models for different objectives, CLP uses a single model adaptable to various reward weightings, significantly reducing computational overhead and simplifying deployment.

The CLP framework has significant implications for various applications, including summarisation, conversational agents, and encoding social norms. By enabling language models to balance multiple objectives effectively, CLP can enhance the flexibility and usability of these models in real-world scenarios.

The researchers acknowledge that while CLP offers robust performance across different conditions, further evaluations, including human assessments and red-teaming, are necessary to mitigate potential risks associated with more flexible language models. Future research directions include exploring other conditioning mechanisms, automated tuning of weight sampling distributions, and addressing non-linear reward scalarisation.

Google is making constant moves towards making AI models and frameworks that simplify AI development. Recent one being, at the Google I/O Connect, Google expanded access to the multimodal AI model Gemini 1.5 Pro and the family of open models, Gemma 2, for Indian developers. 

With the introduction of CLP, it advances language model finetuning by providing a flexible, efficient method for balancing multiple objectives, creating versatile models that adapt to different needs, potentially leading to more capable AI systems.

The post Google Researchers Introduce Conditioned Language Policy Framework for Enhanced Multi-Objective Fine Tuning appeared first on AIM.

Generative AI is set to significantly transform India’s digital public infrastructure. 

Google-backed Plotch.ai is currently working with ONDC to build AI infrastructure and simplify e-commerce for consumers. Recently, the company developed an AI-powered conversational commerce app featuring multilingual, voice-enabled semantic search and robust image search capabilities.

ONDC is an open protocol network that aims to democratise digital commerce by creating an interoperable, open platform where buyers and sellers can transact regardless of the apps they use. It’s a platform to democratise e-commerce in India. 

According to recent reports, ONDC has captured approximately 3% of the food-order volumes from Swiggy and Zomato and achieved 68 million transactions since its inception. It is expected to reach 100 million transactions by Diwali 2024.

“Multilingual voice-based conversational commerce is one piece of the AI that we’re building,” said Manoj Gupta, the founder of Plotch.ai, in an exclusive interview with AIM. “For instance, if you’re looking to buy a saree, jewellery, or a T-shirt and want to find the most affordable options, you can simply ask the AI, and it will sort them for you,” he explained. 

Manoj and Monica Gupta founded Plotch.AI in 2020. They are also the founders of Craftsvilla.com, an online marketplace for ethnic products in India.

They are also working on improving the product catalogue using generative AI. “We will enhance the product names, descriptions, and product metadata using generative AI,” he said.

Moreover, the company is planning to introduce image generation capabilities. Following this, they aim to develop a tool that will categorise customers based on their order value. “For example, distinguishing between high-repeat customers and fraudulent ones,” explained Gupta. 

Additionally, they are building AI-based recommendation engines to show customers the most relevant products. Another area where the company plans to integrate AI is in optimising RTO (return to origin).

Google Loves India 

Gupta told AIM that the startup is backed by Google and is working very closely with them.  “We are using Google’s AI tech stack, which includes their large language model, Gemini. We have utilised this foundational model and custom-trained it specifically for Indian e-commerce,” he said. 

“We have been doing the custom training and have also incorporated vector search using vector databases. Additionally, we have implemented semantic caching. This broadly outlines our technology stack, which also includes Python for coding, MySQL for the database, and Nginx for the web server,” he explained. 

Gupta said that the company is raising about $5 million. The first round was led by Antler, and the second was led by Venture Catalyst. 

“We are very excited about AI and plan to invest 50% of our efforts into it. Even with our $5 million raise, half of that will go into building AI infrastructure directly on top of the ONDC framework,” said Gupta 

Last year, Google partnered with ONDC and announced plans to integrate generative AI capabilities into their tech stack. “Our collaboration creates an opportunity for organisations India-wide to reach larger audiences and grow their businesses, ultimately transforming digital commerce adoption in the country,” said Thomas Kurian, the CEO of Google Cloud. 

Google recently invested in Moving Tech, the parent company of Namma Yatri, and is planning to build a ‘GPay for travel’  in India. Interestingly, Namma Yatri has also joined the ONDC network. It is one of the first mobility services to be integrated into ONDC, which was initially focused on other e-commerce segments such as groceries and food delivery.

“ONDC started two years ago. We can always look for a booster shot for ONDC, some trigger to create a hockey stick growth curve. It is still in infancy; we should let it grow,” said Pramod Varma, former chief architect of Aadhaar, in an interaction with AIM, adding that today ONDC has 10 million transactions, and by 2030, it could hit the 100-million mark. 

Comparing UPI with ONDC, he said that UPI was narrower in purpose. “All it had to do was move money and it got extra support from the readiness of the ecosystem, such as PhonePe and Google Pay.” 

He explained that ONDC has a much broader purpose as it includes taxi booking like Namma Yatri, metro ticketing, grocery commerce, and food delivery. He added that the supply chain of ONDC is much more complex than that of banks. 

“Much of our economic value chain is fragmented; someone has to bring it together. ONDC is doing that. Over time, we will see more and more transactions happening,” he said.

What Does Plotch.ai Offer? 

Plotch.ai has a suite of products designed to facilitate seamless integration of customers into the ONDC ecosystem. 

“We essentially act as the gateway for customers to connect to ONDC,” said Gupta. “ONDC is a collection of domains, including retail, logistics, and fintech. We initially started with the retail domain,” he added.

Gupta said that some of the prominent customers of Plotch.ai are Meesho, IDFC Bank, Paytm, and Craftsvilla. He also said that one of the major challenges they face is connecting their customers’ systems to the network, as each system has its own CRM or ERP. 

“Ensuring that each party can talk to another party seamlessly is also a challenge,” Gupta said, adding that this issue has reduced significantly over the past year.

One of the products the company offers is NodeApp, a full-stack application that ensures smooth operation of both buyer and seller apps within the ONDC network. 

On the other hand, NodePay simplifies node-to-node network payments, guaranteeing efficient financial transactions between buyers and sellers. NodeDesk provides an ONDC-enabled CRM solution for managing customer grievances and ticketing.

NodeBox integrates voice AI into ONDC, allowing users in smaller cities and villages to buy and sell using voice commands in multiple languages. “India is going to be a use case capital of AI. We’ll be very big users of AI, and we believe that AI can significantly help in the expansion of the ONDC Network,” concluded Gupta.

The post How Plotch.ai is Building Generative AI Infrastructure for ONDC  appeared first on AIM.

Back in 2017, Google made UPI and cashless transactions household terms with the launch of  Google Pay in India. The app has over 67 million users in the country, its largest single market to date. 

Now, the company is trying to mimic the success of Google Pay to democratise transformation for all via Namma Yatri, the Bengaluru-based open-source rival of ride-hailing services like Ola and Uber.

Interestingly, earlier this week, Google invested in Moving Tech, the parent company of Namma Yatri. On the same date, Namma Yatri users also came across new features in its service including rentals and instant travel. 

Previously in 2020, Google said it plans to invest $10 billion in India over the next five to seven years as the search giant looks to help accelerate the adoption of digital services in the key overseas market.

Namma Yatri, initially a subsidiary of the payments company Juspay, became a separate entity called Moving Tech in April. CEO Magizhan Selvan and COO Shan M S, formerly with Juspay, now lead this new mobility business. 

Launched more than a year ago, it became a preferred service for many given its driver-centric approach with lower commission rates and a community-driven model tailored to local needs. Recently, it also expanded to multiple cities beyond Bengaluru, including Chennai and Delhi.

Similarly, Google is doing the same with ONDC as well, making it the UPI of ecommerce. At the recently held Google I/O Connect in Bengaluru, the company announced that it has lowered India-specific pricing for its Google Maps Platform API. Interestingly, it is offering up to 90% off on select map APIs for people building on top of ONDC.

Making Money Simply

The success of Google Pay in a price-sensitive market like India can be attributed to strategic initiatives, market understanding, and leveraging unique conditions. It leveraged the UPI infrastructure, an open API platform from the National Payments Corporation of India (NPCI), which enabled instant online bank transactions. This reduced entry barriers for new users and merchants. 

India’s cash-dependent economy, with many unbanked and underbanked people, was targeted by Google Pay through an easy-to-use digital payment solution requiring only a bank account linked to UPI. 

It solved the crisis which Indian fintech players like PhonePe and Paytm had been struggling with for a long time. 

Google built an ecosystem in India by integrating various services like bill payments, mobile recharges, and peer-to-peer transactions, embedding itself in daily financial activities. In 2021, it expanded its network of banks offering card tokenisation on its app, adding SBI, IndusInd Bank, Federal Bank, and HSBC India as partners.

Now, even international tourists can access the app in India. 

Now They Look to Make Travel Simple

In India, Namma Yatri competes with the likes of Uber, Ola and Swiggy-backed Rapido. 

Google Maps, with an established customer base in India, has incorporated Gemini into it and has come up with new features like Lens in Maps, Live View navigation, and address descriptors, specifically for Indian users. 

However, earlier this week, Ola founder Bhavish Aggarwal decided to ditch Google Maps for its own in-house Ola Maps. After Microsoft Azure’s exit last month, this is Aggarwal’s second move to get rid of Western apps. Yesterday, he reduced the pricing for the Ola Maps APIs even further to encourage other developers to build on it. 

“We’ve been using Western apps to map India for too long, and they don’t understand our unique challenges: street names, urban changes, complex traffic, non-standard roads, etc,” said Aggarwal.

Meanwhile, Rapido, Uber, Namma Yatri and others ride-hailing platforms continue to use Google Maps. Currently, Rapido, which started off as a bike taxi service has expanded to include auto rickshaw, carpooling services, taxicab hailing, parcel delivery, and third-party logistics services in over 100 Indian cities.

Namma Yatri: The UPI of Transportation

“Imagine the platform similar to Namma Yatri being adopted in cabs, metros, or any platform which is serving the passengers,” Selvan told AIM, when Moving Tech was first launched. “We essentially want to become the UPI of transportation.”

Namma Yatri stood out for the drivers since the app started off as free of commission but now charges a basic of just INR 25 per day. In contrast, Ola and Uber take a 25-30% commission from drivers per ride. 

One Namma Yatri driver told AIM that the low subscription fee compared to Uber and Ola has helped him. Within six months of switching to Namma Yatri, he was able to fund his two children’s weddings.

The app has onboarded 49,000 auto drivers and 550,000 users in five months, with approximately INR 12 crores ($1.5 million) paid out to drivers. It celebrated 500 million downloads in March.

Although Namma Yatri currently lacks features like bike taxis and carpooling, with Google’s support it may soon expand to include these options. 

The founders said that Namma Yatri will leverage the new funds to grow its engineering and R&D competencies, and also include more types of transportation, including buses. 

On the other hand, Google has found an ideal partner to strengthen its presence in India’s transportation sector. Given its expertise in revolutionising online purchases, it is set to replicate the ‘Google Pay moment’ in Indian transportation soon.

The post Google is Trying to Mimic ‘Google Pay’ for Transportation via Namma Yatri appeared first on AIM.

After Google announced a 70% price cut for Google Maps API, Ola CEO Bhavish Aggarwal challenged the tech giant by announcing an even more reduced pricing structure and the future product roadmap for Ola Maps.

“As promised, here’s our response to @googlemaps’ ‘belated’ price cuts. It’s time we build world-class alternatives to big tech giants and empower Indian innovation! I’m very excited to announce a further reduced pricing structure and our future product roadmap for Ola Maps @Krutrim,” he posted on X.

Ola Maps has announced a new pricing structure and is now available on the cloud with a free one-year subscription. The updated structure includes a free tier for 5 million API calls per month, covering the needs of over 90% of Indian developers and startups. For larger volumes, Ola Maps offers prices at half of Google’s reduced rates, with additional benefits for long-term commitments.

Furthermore, Ola has introduced a special offer for startups and SMBs building on the ONDC platform, providing three years of free access to Ola Maps.

Regarding Google’s price cut on the Maps API, Aggarwal criticised Google for its delayed response in adjusting prices and localising payment options, saying, “Dear @Google, too little too late! Reducing prices for @googlemaps and ‘offering to price in ₹’ after #ExitGoogleMaps. Don’t need your fake generosity!”

Ola Maps has seen significant interest from the developer community, with over 10,000 developers signing up and transitioning their workloads to the platform. This support is seen as a testament to India’s readiness to develop its tech solutions.

Astro India AI and KPN Fresh Product are among the early adopters of Ola Maps. Astro India AI has enhanced its user experience by integrating Ola Maps’ Autocomplete service, while KPN Fresh Product has seamlessly incorporated Ola Maps’ Location services into its e-commerce platform, benefiting from precise geocoding and routing.

New Features In Map 

Ola Maps announced that it  will introduce a suite of powerful new APIs. The Routes APIs will offer advanced features like traffic-aware directions, a distance matrix with multi-modal transportation options, and specialised routing such as two-wheeler routes and toll information.

The Places APIs will enhance search capabilities, enabling users to find nearby locations, perform natural language searches, and access detailed place information. Additionally, the Maps APIs will provide flexible solutions for both static and dynamic map integration, allowing seamless embedding of rich geographical data into applications.

To support these API offerings, Ola Maps is launching feature-rich SDKs for web, Android, and iOS platforms. The Web SDK will empower developers with functionalities including custom markers, interactive info windows, and advanced visualization tools like heatmaps and marker clustering. 

Unique features such as automatic localisation into Indic languages and extensive map customization options will also be available. The Android and iOS SDKs will provide native support for dynamic maps, custom markers, intuitive user controls, and will similarly support localisation and customisation. These SDKs will integrate seamlessly with the Places SDK, offering intelligent features like place autocomplete and detailed place information.

Ola Maps will also deliver robust platform features designed for developers and businesses. The Map Style Editor will allow users to create, edit, and publish custom map styles, while a dedicated test environment will enable experimentation before full integration. 

Comprehensive developer resources, including documentation, tutorials, and a community forum for knowledge sharing, will be available. Users will receive changelog notifications and have access to strong customer support through an automated ticket system with defined SLAs and chat support.

The post Ola’s CEO Bhavish Aggarwal Challenges Google with Free Ola Maps API Access for Indian Developers appeared first on AIM.

In an effort to expand its footprint in India, generative AI powerhouse Microsoft has opened a new innovation hub in Bengaluru. Puneet Chandok, president of  Microsoft India and South Asia took to LinkedIn to share the news. 

“Embrace a culture of learning and innovation with our maker spaces for prototypes and special projects. Meet our senior architects, who will steer you through bespoke technical engagements focused on transformative business outcomes” added Chandok.

The company has over 20,000 employees across 10 Indian cities – Ahmedabad, Bengaluru, Chennai, Gurugram, New Delhi, Noida, Hyderabad, Kolkata, Mumbai and Pune.

It has over 14 Microsoft Innovation Centres (MICs) across the country which are part of strategic partnerships with leading academic institutions and aim to grow tech skills. 

The company has announced several collaborations in the generative AI space in India over the last couple of months. For example, it teamed up with Skillsoft to create an AI training program for enterprises, leveraging Skillsoft’s AI Skill Accelerator to teach the use of Microsoft AI tools, including Copilot and Azure Open AI. 

Additionally, Indian IT giant Tech Mahindra is collaborating with Microsoft to implement Copilot for Microsoft 365 across 15 locations, aiming to improve efficiency for over 10,000 employees. Tech Mahindra will also use GitHub Copilot for 5,000 developers, expecting a 35-40% productivity boost.

Back in February of this year, Microsoft co-founder and philanthropist Bill Gates visited the Microsoft India Development Center (IDC) in Hyderabad, a hub of innovation that he envisioned 25 years ago. He expressed his optimism for India’s unique potential in AI and the company’s strategic focus on harnessing the country’s talent for upcoming features in this space.

The company is partnering with Indian startup Sarvam AI, specialising in Indic LLMs. It is also set to upskill two million Indians in AI by 2025.

The post Microsoft Inaugurates New Innovation Hub in Bengaluru appeared first on AIM.

At Google I/O Connect, Bengaluru, the generative AI powerhouse expanded access to its multimodal AI model Gemini 1.5 Pro and family of open models Gemma 2 for Indian developers. 

The new 2 million token context window on Gemini 1.5 Pro, previously limited, is now accessible in India. With a capacity of one million tokens, users can analyse extensive data, including up to one hour of video or 11 hours of audio. 

Explaining the business use cases of the models and what Indian developers are looking for, Chen Goldberg, VP, GM, Google Cloud Runtimes told AIM today, “We’re talking with them about how they can scale—scale with their customers, their business, and their teams and also  how they can run more efficiently. 

“Our customers in India are critical for us. We expect to see a lot of innovation in AI coming from the local market” she added. 

Additionally, the newly released Gemma 2 models, available with nine billion and 27 billion parameters, claim to offer improved performance and safety protocols. Optimised by NVIDIA, these models run efficiently on next-gen GPUs and a single TPU host in Vertex AI.

Boosting India’s GenAI Space

The availability of Gemma in India is likely to be a big leap in the surge of foundational models in Indian languages. Many developers in India prefer Gemma over other open-source models like Meta’s Llama for building Indic LLMs. 

Gemma’s tokenizer is particularly effective for creating multilingual solutions, as demonstrated by Navarasa, a multilingual variant of Gemma for Indic languages. At Google I/O California, the company highlighted the success of this project, developed by Telugu LLM Labs founded by Ravi Theja Desetty and Ramsri Goutham Golla. It is accessible in 15 Indic languages. 

“In India, there are two main areas of focus. Firstly, addressing language-related issues. Secondly, involves large-scale transformations across various industries be it in customer engagement or addressing the broader needs of the Indian population,” Subram Natarajan, director of customer engineering and field CTO at Google Cloud told AIM at the event, echoing similar thoughts.

Previously, Vivek Raghavan, the co-founder of Sarvam AI also told AIM that Gemma’s tokenizer gives it an advantage over Llama when it comes to Indic Languages. He explained that The tokenization tax for Indic languages means asking the same question in Hindi costs three times more tokens than in English and even more for languages like Odiya due to their underrepresentation in these models.

Today, the company also unveiled IndicGenBench to evaluate the generative capabilities of Indic LLMs, covering 29 languages, including several Indian languages without existing benchmarks.

Going ahead, the company will continue to focus on investments in the developer community and partnerships.

“These are crucial for scaling our operations. We understand that these elements are essential not just for our success but for the broader public’s benefit” concluded Natarajan.

The post Google Expands Gemini 1.5 Pro & Gemma 2 Access for Indian Developers appeared first on AIM.

As part of several initiatives that Google has taken up in India to improve Indic LLM capabilities, Google Pay vice president and GM Ambarish Kenghe announced the launch of IndicGenBench.

A benchmark to help in evaluating the generative capabilities of Indic LLMs, IndicGenBench is part of a slew of updates released during Google I/O Bengaluru 2024. Kenghe said that the benchmark covers as many as 29 languages, including several Indian languages that do not currently have benchmarks.

Speaking to AIM, Google Cloud director of customer engineering and field CTO Subram Natarajan said, “In India, there are two main areas of focus: Addressing language-related issues, while the second involves large-scale transformations across various industries, be it in customer engagement or addressing the broader needs of the Indian population.”

With a focus on improving language-related issues, Kenghe announced the open sourcing of DeepMind’s Composition to Augment Language Models (CALM), allowing developers to combine specialised language models with Google’s Gemma models. Interestingly, research on CALM had been done specifically by the Google DeepMind and Google Research teams in India, with the paper released earlier this year.

“Let’s say you’re building a coding assistant that can converse in English. Now, by composing a Kannada specialist model with CALM, you may be able to offer coding assistance to Kannada users as well,” explained Kenghe.

This focus on Indic language LLMs comes as DeepMind expands Project Vaani, a collaborative effort between Google and the Indian Institute of Science (IISc), wherein over 14,000 hours of speech data in 58 languages, has been made accessible to developers. This data was collected from over 80,000 speakers in 80 districts across the country.

As previously covered by AIM, this is being open-sourced as part of MeitY’s flagship AI initiative, Bhashini. These capabilities are also soon to be expanded as Bhashini also launched an initiative called Bhasha Daan, to help crowdsource voice and text data in multiple Indian languages.

The post Google Introduces IndicGenBench to Benchmark Indic LLMs Across 29 Languages appeared first on AIM.

At the Google I/O Summit in Bengaluru, the company announced that it is working with MeitY Startup Hub to train 10,000 startups in AI. 

As part of the project, Google is offering mentorship, expertise, and skills to businesses through Appscale Academy and Startup School, as well as AI-first programming and curriculum.

Additionally, the tech giant will give businesses up to $350,000 in Google Cloud credits to invest in the cloud computing and infrastructure needed for AI development and application.

“These collaborations are a part of our efforts to help the next generation of startups and developers  to solve real-world challenges,” said Ambharish Kenghe, vice president of Google Pay. “We are also reorienting our existing programs, like the Startup School and AppScale Academy, to be AI-first.”

In an exclusive interaction with AIM, general manager and vice president of Google Chem Golberg said, “This is my first visit to India, and I was very impressed with the level of innovation and new ideas, especially from startups that have been able to grow”.

Kenghe also announced a nationwide Gen AI Hackathon and an AI startup boot camp and a three-month immersive experience in partnership with MeitY Startup Hub and the Startup India initiative between August and October.

For Developers

Additionally, Google released additional developer tools and improvements, such Google Wallet APIs, to make it easier to integrate gift cards, tickets, and loyalty programmes. It also introduced India-specific pricing for the Google Maps Platform, with up to 70 per cent lower costs on most APIs. 

Collaborating with the Open Network for Digital Commerce (ONDC), Google is offering developers building for ONDC up to 90 per cent off on select Google Maps Platform APIs, in an effort to make it more affordable and accessible.

The post Google, MeitY Startup Hub to train 10,000 Indian startups in AI appeared first on AIM.

At the Google I/O Connect, Bengaluru, on Wednesday, Google Maps announced that it has lowered India-specific pricing reductions for Google Maps Platform API. 

“Starting August 1, you can expect to pay up to 70% less on Google Maps Platform API,” said Ambrish Kenghe, vice president of Google Pay. The company also announced a special program in collaboration with Open Network for Digital Commerce (ONDC). 

“If you’re building for the ONDC, you might be eligible for up to 90% off on select map APIs,” Kenghe added.

New Collaborations

At the I/O summit, Google also announced its new collaborations as a part of its efforts to help the next generation of startups and developers  to solve real-world challenges. It announced that it is working with METI Startup Hub to enable 10,000 Indian startups in their journey with AI. 

“As part of this effort, we are supporting eligible AI startups with up to $350,000 in Google Cloud credits to help them invest in AI infrastructure. We are also reorienting our existing programs, like the Startup School and AppScale Academy, to be AI-first,” Kenghe said. Kenghe also announced a nationwide Gen AI Hackathon and an AI startup boot camp.

Google’s Answer To Ola?

The move comes days after Ola replaced Google Maps to its in-house Ola Maps. In an X post, Aggarwal wrote, “We used to spend 100 crores a year, but we’ve made that 0 this month by moving completely to our in-house Ola Maps!”. Ola Maps is positioning itself as a cost-effective alternative to Google Maps.

Aggarwal is not the only one challenging Google Maps at the moment. Recently, ISRO chief S Somanath claimed that “ISRO’s Bhuvan is 10x better than Google Maps”.

In a bid to keep up the race, Google has introduced new features to Maps in India, including Lens in Maps and Live View walking navigation. With these features, users can see arrows, directions, and distance markers overlaid on the Maps screen, helping them quickly figure out which way to go. 

The tech giant also introduced Address Descriptors on Google Maps to help users understand addresses better, in a way they are used to in real life. Google is now experimenting with generative AI as well.

The post Google Maps API To Cost 70% Less Now appeared first on AIM.

Google has joined a group of investors supporting Namma Yatri, an open-source ride-hailing app in India that is challenging established players like Uber and Ola. This move signals growing interest in alternative models within the ride-sharing industry, particularly those emphasising driver welfare and community engagement.

Namma Yatri’s Unique Approach and Rapid Growth

Launched in November 2022, Namma Yatri has quickly gained traction in the Indian market, particularly in Bengaluru. The app’s name, which translates to “Our Traveler” in Kannada, reflects its community-driven ethos. Unlike traditional ride-hailing services, Namma Yatri operates on a no-commission model, allowing drivers to retain a larger portion of their earnings.

Key features of Namma Yatri include:

1. Driver-centric approach with lower commission rates
2. Community-driven model tailored to local needs
3. Expansion to multiple cities, including Chennai and Delhi
4. Integration with India’s Open Network for Digital Commerce (ONDC)

The app has reportedly onboarded 49,000 auto drivers and 550,000 users in just five months, with approximately ₹12 crores ($1.5 million) paid out to drivers. This rapid growth demonstrates the appeal of Namma Yatri’s model to both drivers and passengers.

Read more: Google in Advanced Talks to Acquire Wiz at $23 Billion

Google’s backing of Namma Yatri is significant, as it adds credibility to the open-source, driver-friendly approach in the ride-hailing sector. While the app faces challenges, including the need to reduce operational costs and improve user experience, the support from major tech players and its integration with ONDC could potentially accelerate its growth and impact on the Indian urban mobility landscape.

As Namma Yatri continues to evolve, it represents a notable shift in the ride-hailing industry, prioritising local needs, driver welfare, and technological innovation to create a more equitable transportation ecosystem.

The post Google Backs Namma Yatri Over OLA and Uber appeared first on AIM.

Growing up in a small town in Sangli, Maharashtra, Aqsa Fulara, an AI/ML product manager at Google since 2017, like many other women, faced societal norms that often discouraged women from pursuing higher education far from home. 

“Coming from a community where moving out of my parent’s home to a hostel for higher education was frowned upon, I put all my energy towards getting into this prestigious engineering college in the same city,” Fulara told AIM in an exclusive interview.  

Her dedication paid off when she was admitted to Walchand College of Engineering, where she did her BTech in computer science and engineering. This academic achievement was just the beginning. 

Fulara’s passion for learning and her desire to push the boundaries led her to the University of Southern California (USC), where she pursued a master’s degree in engineering management, and since then “there was no looking back!”, Fulara shared gleefully. 

“While my experiences in India provided me with a solid technical foundation and analytical approach to solving problems, my experiences at USC and Stanford focused a lot more on practical applications of cutting-edge technology,” she added. 

According to recent surveys, compared to other developing countries, fewer women in India reported being discouraged from pursuing scientific or technical fields (37% vs. 65%). The primary challenges faced by women students in India are high-stress levels (72%), difficulties in finding internships (66%), and a gap between their expectations and their current curriculum (66%).

Fulara’s path to AI and ML was not marked by a single dramatic moment but rather a gradual buildup of curiosity and fascination with technology. Her inclination towards solving problems and understanding complex systems drew her to this field. 

“That led me to my capstone project on behaviour recognition and predicting traffic congestion in large-scale in-person events and thus, building products for congestion management,” she added. 

Leadership Mantra: Building the Culture of Innovation

If you’re familiar with Google’s Vertex AI Search, you likely know about Recommendations AI. Now branded as Recommendations from Vertex AI Search, this service leverages state-of-the-art machine learning models to provide personalised, real-time shopping recommendations tailored to each user’s tastes and preferences. 

One of the key figures in scaling this product is Fulara, who has been instrumental in its growth since 2021. Fulara has also been the force behind the highly acclaimed products in Google Cloud’s Business Intelligence portfolio, such as Team Workspaces and Looker Studio Pro. 

Fulara considers Looker Studio as one of her favourite projects. “Imagine having a personal data analyst assistant who can provide customised recommendations and help you make informed decisions,” she added. 

Having worked with Google for over seven years now, one thing that Fulara values most about the company is the freedom to explore and innovate. “Whether it’s pursuing a 20% project in a new domain, growing into a particular area of expertise, or participating in company-wide hackathons, Google provides much space for creativity and innovation,” she shared. 

This environment has allowed her to pivot her career towards product management, building on her AI experiences and focusing on delivering business value through customer-centric solutions.

Leading AI product development comes with its own set of challenges. “AI products have a larger degree of uncertainty and ambiguity, with challenges in terms of large upfront investment, uncertain returns, technical feasibility, and evolving regulations,” she explained. 

To manage these challenges, Fulara fosters a culture of experimentation and agility. “We release MVPs for testing far ahead of production cycles to rigorously test and benchmark on production data and user behaviours,” she added, allowing her team to make informed decisions even with incomplete information.

Fulara emphasises the importance of managing scope creep tightly and sharing outcome-based roadmaps upfront. “We’re solving for problem themes, not necessarily just churning out AI features,” she noted. This strategy helps maintain focus and adapt to changes quickly. 

Future of AI 

Looking ahead, Fulara sees generative AI, personalised recommendations, and data analytics as transformative forces in the coming decade, making data and insights more accessible and workflows more collaborative. 

AI and ML models are becoming increasingly pervasive, assisting in personalised shopping journeys, optimising marketing strategies, and improving data-driven decision-making across various industries.

Read more: Beyond Pride Month: True Allyship Needs No Calendar

The post From a Small Town in Maharashtra to Silicon Valley: Aqsa Fulara’s Inspiring Journey with Google appeared first on AIM.

In trying to address the problem of limited contexts for large language models (LLMs), researchers have tried to emulate human episodic memory with EM-LLM.

The paper, titled “Human-like Episodic Memory for Infinite Context LLMs”, was released by researchers from Huawei and University College London.

EM-LLM (wherein EM stands for episodic memory) integrates important aspects of episodic memory and event cognition, used by the human brain, into LLMs. Through this, researchers have suggested that LLMs using EM-LLM (wherein EM stands for episodic memory) can potentially have infinite context lengths while maintaining their regular functioning.

This is particularly interesting as the method can be scalable without increasing the amount of computing power needed for the LLM to function. “By bridging insights from cognitive science with machine learning, our approach not only enhances the performance of LLMs on long-context tasks but also provides a scalable computational framework for testing hypotheses about human memory,” the researchers said.

EM-LLM goes about recreating human episodic memory by organising tokens into episodic events using Bayesian surprise and graph-theoretic boundary refinement. Following this, they employ a two-stage retrieval process, based on time and similarity, to allow for human-like access and retrieval of information.

Interestingly, when compared to similar proposals to improve context windows for LLMs, like InfLLM, EM-LLM excelled, with an overall improvement of 4.3%. Under the PassageRetrieval task, EM-LLM had a 33% improvement over InfLLM. 

“Our analysis reveals strong correlations between EM-LLM’s event segmentation and human-perceived events, suggesting a bridge between this artificial system and its biological counterpart,” the researchers stated.

This is yet another step towards tackling the problem of context lengths when it comes to interacting with LLMs. Research has been ongoing to increase the context lengths, with major companies like Google and Meta releasing their own papers on pursuing infinite context lengths.

The post Researchers Recreate Human Episodic Memory to Give LLMs Infinite Context appeared first on AIM.

Odyssey, dubbed as a Hollywood grade AI system, was recently unveiled by co-founder and CEO Oliver Cameron. The model looks to give visual control to people to allow them to tell a story exactly the way they have imagined. 

The co-founder believes their text-to-video platform will result in higher-quality movies, shows and video games. 

Multiple AI Models

When compared to OpenAI’s Sora or Google’s Veo, Odyssey has given the reins of control to the users who can direct the visuals as per their needs. Thereby, offering better customisation options for users. Infact, there are a number of AI-video generation platforms in the market. 

Odyssey’s capability is achieved by delving deeper than traditional text-to-visual models. Instead of using just one model that limits you to one input and one unchangeable output, Odyssey is using four generative models. These models give precise control over each main part of visual storytelling: creating detailed shapes, realistic materials, customisable lighting, and adjustable motion. Together, they let you quickly create scenes and shots exactly as a user envisions them. 

Furthermore, the company is also developing workflows for expert users, seamlessly integrating into current production methods used in Hollywood, gaming, and others. The company allows compatibility with established workflows in top-tier production, and even edit and export in multiple formats such as 3D file formats. 

Super Team

The team behind Odyssey comprises Hollywood artists and AI researchers from emerging tech verticals including Cruise, Wayve, Waymo, Tesla, Meta and more. The artists have worked with big production names such as Dune, Godzilla, Avengers and others. In addition to Cameron, Jeff Hawke serves as the other co-founder and CTO of Odyssey. 

The company has raised $9M from investors from Y-Combinator, Google Ventures and many more. 

The post ‘Odyssey’ AI Built for Hollywood, Sora Can Wait appeared first on AIM.

In a recent interview, Aidan Gomez, CEO and co-founder of Cohere, stated fears of an AI takeover are unfounded due to its reliance on human training data and the limits of exponential growth.

He explained, “I think I’m empathetic to the fears, you know, the sci-fi narrative of computers or AI taking over and destroying the world. It’s been going on for decades, and so it’s really deeply embedded within our culture. It gets lots of clicks, headlines, it gets attention. It’s shocking. I understand why people are scared of it and why some say it to get attention.” 

Gomez highlighted that this is not a technical truth of the technology and that continuous exponential scaling does not happen. There are friction points and complexities. He also mentioned that the intelligence of these models is limited by the humans who create them, as it is our data and knowledge that teach them.

However, he believes real risks lie in deploying AI in high-stakes scenarios like medicine and advocates for scrutiny and tough discussions on AI deployment, rather than sensational sci-fi narratives.

What is Cohere Upto

In the interview, he also discussed how the original goal of the project was to improve Google Translate, a very well-known problem. He noted that it has been extraordinary to see the broad impact of a technology developed to enhance translation. Gomez was also the co-author of the original Transformers paper which forms the crux of today’s generative AI products. 

Recently, AIM spoke to Saurabh Baji, SVP of Engineering at Cohere, about the mixed emotions in Silicon Valley regarding achieving AGI, as seen in the recent banter between Meta’s Yann LeCun and xAI’s Elon Musk.

“We remain concentrated on designing AI solutions that deliver better workforce and customer experiences for businesses today rather than pursuing abstract concepts like AGI,” said Baji.

The post ‘Fears of an AI takeover are unfounded due to data limits and finite growth,’ says Aidan Gomez appeared first on AIM.

Generative AI is changing fashion for the better. One stellar example is Google software engineer Christina Ernst, who triggered a discussion on social media after she created the “world’s first AI dress.”

“I engineered this robotic snake dress, and it is finally done,” Ernst said in the description of her video. “I coded an optimal mode that uses artificial intelligence to detect faces and moves the snake head towards the person looking at you. So, maybe this is the world’s first AI dress. Surveillance state, but make it fashionable,” she added. 

The black dress features several robotic snakes, with one around her neck specifically.

After Ernst architected this, the techie’s Instagram video became the talk of the town on social media. Netizens were quick to react and comment, highlighting her unique idea.  

However, this is not the first time that an AI-designed outfit has been produced. 

AI algorithms generate innovative styles, inspiring designers and making high fashion more accessible and diverse for everyone. They democratise design, allowing everyone to bring ideas to life and fostering inclusivity. 

For example, Bewakoof, a popular Indian apparel brand, partnered with Google Cloud at the Google Cloud Summit India 2024 to use Gemini AI to create new content. The collaboration includes an AI T-shirt designing tool for custom artwork.

Fashion x GenAI 

While the Google engineer gained viral attention, there have been a number of remarkable innovations where AI, not necessarily GenAI, has significantly influenced the fashion industry; one that made major headlines is when Zac Posen partnered up with Google to make a unique LED gown.

Maddy Maxey, a fashion engineer and mentor at Made with Code, aided in the design of the dress, a Google-backed initiative. The dress was designed with a circuit infused into the textile to create different animations. 

“Wearable technology was becoming popular in the tech industry back then, with many pumping investments in this particular sector. Google was already betting on its growing importance,” said Posen. 

Another notable example is Iris van Herpen’s entry into the world of AI with the debut of her FW23 campaign, which showcases a couture collection that imagines a new world.

“Our creative process was exceptionally inspiring, allowing us to dream up our references to the deep sea life we had seen and even my archive that we trained the AI with. So, by teaching the AI my design DNA and the more historic architecture references, it got better ‘dreams’. It was very inspiring to see how marvellous Rob Rusling and his team are with AI, it brings a whole new dimension to fashion editorial,” said Herpen.

While designers use AI to create costumes, there’s a company by the name Fabricant, founded by Kerry Murphy in 2018, that leads the digital-only fashion frontier, seamlessly blending technology and fashion to transform traditional craftsmanship. Committed to high-quality premium experiences, their platform reshapes the fashion landscape with innovative co-creation and visually striking encounters. Embracing sustainability and equity, The Fabricant creates an ecosystem where creativity thrives, redefining identity in fashion. 

Back in 2016, for the Manus x Machina-themed Met Gala, IBM and Marchesa unveiled a cognitive dress, a first-of-its-kind garment with cognitive inspiration woven into every step of the creative process – from concept and R&D to design and finished product. This collaboration showcased the creative potential of building with Watson and the ability of this technology to enhance human imagination. 

At Paris Fashion Week 2023, Humane Pin was designer Coperni’s latest buzzy tech name to be included. It was pinned on the clothes of multiple Coperni models during its presentation, which generated considerable buzz.

How Fashion Brands Use GenAI

AI has been reshaping industries for decades, and fashion is no exception. Generative AI, including LLMs and other models, is a particularly exciting development. Brands like The Fabricant and Rebeca Minkoff use AI to create digital clothing designs, enabling rapid prototyping and creative exploration. GANs generate new fashion designs by blending styles from a large dataset of existing designs.

The fashion industry has been using AI for quite some time, but many more brands have hopped onto the AI bandwagon in the last couple of years. 

A few of them include G-Star Raw’s March campaign, which used AI to reimagine denim, releasing its first denim couture piece designed with Midjourney. 

Maison Meta collaborated on its first AI-powered campaign in February with machine-generated images for London Fashion Week, and Levi’s took the potential of AI one step further in March by using it to create a series of inclusive digital models for its e-commerce site. 

McKinsey analysts say that Generative AI could add anywhere from $150 billion to $275 billion in profits to the fashion industry by 2030.   

But AI prediction isn’t just for trend forecasting firms. For example, fast fashion behemoth H&M employs more than 200 data scientists to track purchase patterns and other store trends to map customer demand at a granular level. Similarly, Sweden’s Zara also uses AI algorithms to identify patterns and predict which styles are likely to become popular in the future.

Amazon enables customers to upload an image, and its AI-powered tool finds similar styles on Amazon, considering factors like brand, price, and user preferences. L’Oréal’s ModiFace allows users to virtually try on makeup using facial recognition and AR, giving them a highly personalised shopping experience. 

Even back home, we have e-commerce giants using AI. Myntra, the e-commerce platform, introduced ‘Maya’, its premier virtual fashion influencer. 

According to Sunder Balasubramanian, Myntra’s Chief Marketing Officer, “Maya, a distinctive figure of Myntra Fashion Forward (FWD), will function as an influencer on social media and also be part of Myntra’s own social commerce on the Myntra Studio platform.”

Last year, we saw fashion brands in the experiment stage with AI. However, the shift is happening at a faster rate now as companies are moving from POCs to production this year. 

Let us know your thoughts below.

The post Fashion Gets a Generative AI Makeover appeared first on AIM.

A staunch proponent for the open development and research of AI, Meta’s chief AI scientist, Yann LeCun, recently posted a call to action to oppose SB1047, colloquially known as the California AI Bill.

Now, not taking into consideration the actual content of the proposed Safe and Secure Innovation for Frontier Artificial Intelligence Models Bill, California is home to a majority of AI and AI-adjacent companies that operate globally.

This means that a comprehensive bill governing AI within the state will affect not only the companies within the US state but also globally.

Overkill Much?

There’s plenty of reason not to like the actual Bill. But the main point of contention that had LeCun concerned was the regulation of research and development within the ecosystem.

“SB1047 is a California bill that attempts to regulate AI research and development, creating obstacles to the dissemination open research in AI and open source AI platforms,” he said.

However, the Bill also attempts to predict where AI could go, thereby implementing pretty strict and near unattainable compliance from companies.

It uses the potential for AI to “create novel threats to public safety and security, including by enabling the creation and the proliferation of weapons of mass destruction, such as biological, chemical, and nuclear weapons, as well as weapons with cyber-offensive capabilities” as a way to implement overarching measures, that in the end will go unimplemented.

For example, the Bill basically states that it is prohibited to build a model that can enable critical harm, given certain provisions. However, as AIM has covered before, literally any model can be jailbroken to even produce instructions on how to build nuclear weapons.

The Bill is filled with similar instances of providing guidelines that are either near impossible to adhere to or are just generalisations, backed by a need to adhere to safety protocols but a lack of actual knowledge of how these systems work.

Meta’s vice president and deputy chief privacy officer, Rob Sherman, said it perfectly in a letter sent to the lawmakers, “The bill will make the AI ecosystem less safe, jeopardise open-source models relied on by startups and small businesses, rely on standards that do not exist, and introduce regulatory fragmentation.”

Stick to What You Know

The general consensus is that it’s basically impossible to implement future-proof regulations for AI. 

Mentioning worst-case scenarios like nuclear war or building chemical weapons only serves to stoke a pervasive fear of AI that is already common among the general public. There have been several AI leaders, as well as government officials, who have stated that over-regulation of AI is something that they’re hoping to avoid.

However, regulations like these broadly generalise what AI is, with a lack of input from those working within the tech space and who are familiar with ongoing developments in the industry.

While there are several concerns on the development and usage of AI, these don’t ever seem to get addressed in regulations like this one and EU’s Artificial Intelligence Act (AIA). Instead, they focus on trying to future-proof AI usage, thereby making generalisations, and fail to address problems that are already prevalent within communities and the industry.

“The sad thing is that the regulation of AI R&D is predicated on the illusion of “existential risks” pushed by a handful of delusional think-tanks, and dismissed as nonsense (or at least widely premature) by the vast majority of researchers and engineers in academia, startups, larger companies, and investment firms,” LeCun said.

There are many gaps in regulation that AI companies and startups actively take advantage of, though conceding that this be done carefully so as not to cross ethical boundaries. However, with no proper regulation in place, companies are not bound by any kind of legal obligation.

Many big players like OpenAI, Meta, Google and Microsoft have been in staunch favour of regulations but have asked that preliminary conversations are held with stakeholders. Which, for anything regulation-related, makes sense.

However, it seems that the California AI Bill is just another in a long line of examples where governments seem to push regulations as a reactionary measure rather than one that has thought and rationale put behind it. Which is evidenced in the open letter written to the legislators, signed by several researchers, founders and other leaders in the AI space.

Further regulations can only serve to push companies, particularly startups, to pursue prospects in other countries that don’t attempt to have a hamfisted approach to policing AI.

The post What’s the California AI Bill, and Why Does Meta’s Yann LeCun Think it Sucks? appeared first on AIM.

 In a recent interview, Illia Polosukhin, co-founder of Near, said AI is not a human but just a system. 

He explained, “I think the important part to understand about AI is that it’s not a human, it’s a system. And the system has a goal. Unless somebody explicitly programs it to harm humans, it won’t magically do that. In the blockchain world, you realise that everything is driven by economics one way or another. There’s no economic incentive to kill humans.” 

Polosukhin highlighted that right now, they’re building systems to improve individuals and expand the capabilities of minds. They will have autonomous agents with specific missions and goals, and they will perform tasks in the physical world. 

However, their actions will be governed by the same laws as ours. Using AI to develop biological weapons is no different from doing so without AI. Ideally, AI should help us better detect and prevent such misuse.

Google Transformer 

As one of Google’s ‘Transformer 8,’ Polosukhin played a pivotal role in revolutionising deep learning. 

In the interview, he also discussed the inception of the Transformer paper and the importance of democratising AI. Polosukhin also reflected on his journey from Google to founding his own AI company.

Now, Polosukhin is considered one of the founding fathers of modern AI.

Polosukhin co-wrote the now famous 2017 paper, “Attention Is All You Need” along with seven Google colleagues, who have collectively become known as the “Transformer 8”. 

Google integrated transformers into Google Translate in 2018, resulting in improvements. However, the technology didn’t see widespread use until OpenAI introduced ChatGPT in November 2022. 

“OpenAI had very little to lose by opening this up,” Polosukhin told CNBC. 

“If, for example, any other company, especially a public company, opened it up and the first question you ask there, it was like an inappropriate answer, that would be in the news.”

By the time the transformer paper was published in late 2017, Illia Polosukhin had already left Google to co-found Near with Alexander Skidanov. Since then, all eight authors have left Google, with Polosukhin being the first to depart. 

The post ‘There is No Economic Incentive for AI to Kill Humans,’ says Illia Polosukhin appeared first on AIM.

The 2024 edition of CVPR 2024, the prestigious annual conference for computer vision and pattern recognition, took place from June 17 to 21 in Seattle, Washington. 

Google Research was one of the key sponsors, which presented over 95 papers on various topics including computer vision, AI, machine learning, deep learning, and related areas from academic, applied, and business R&D perspectives. It also had an active involvement in over 70 workshops and tutorials. 

“Computer vision is rapidly advancing, thanks to work in both industry and academia,” said David Crandall, professor of computer science at Indiana University, Bloomington and CVPR 2024 program co-chair. 

The event saw 11,532 entries, out of which only 2,719, that is 23.58%, were accepted. Let’s take a look at the top papers presented by Google this time.

Generative Image Dynamics 

Generative Image Dynamics presents a novel approach for generating realistic image sequences from a single input image by the authors. This work presents a generative model that predicts the temporal evolution of images, capturing spatial and temporal dependencies.

This approach has potential applications in video prediction and by generating realistic image sequences from a single input, it advances generative modelling and opens new possibilities for creative and interactive applications.

Rich Human Feedback for Text-to-Image Generation

The paper proposes a novel approach to leveraging human feedback for improving text-to-image generation models.

The framework allows users to give detailed feedback on generated images, such as annotations, sketches, and descriptions. This feedback is used in a novel training strategy to fine-tune and improve the text-to-image generation model. 

Incorporating rich human input also addresses the limitations of current models and advances user-centric generative systems.

DiffusionLight: Light Probes for Free by Painting a Chrome Ball

The paper introduces a diffusion model that can efficiently estimate the 3D lighting environment from a single 2D image. 

The diffusion model enables real-time applications like virtual try-on and augmented reality, with effective lighting estimation demonstrated on diverse inverse rendering benchmarks, surpassing prior state-of-the-art methods. 

The authors have also released the source code and a demo of the Diffusion Light system, enhancing accessibility for further research and development.

Eclipse: Disambiguating Illumination and Materials using Unintended Shadows

This paper, published in May 2024 by a team of Google researchers, presents Palm-E, a large language model designed for dialogue applications. The model is based on the Pathways Language Model (PaLM) architecture, which is a scaled-up version of the Transformer model. 

The authors fine-tuned the model on a large dataset of conversational data, including both human-human and human-bot conversations. The authors evaluated Palm-E on a range of dialogue tasks, including open-domain conversation, task-oriented dialogue, and dialogue safety. 

Time-, Memory- and Parameter-Efficient Visual Adaptation

This paper was published by a team of researchers from the University of Tubingen.

The paper explores the use of deep reinforcement learning to study the evolution of cooperation in social dilemmas. Social dilemmas are situations where individual self-interest conflicts with the collective good, and cooperation is often required to achieve the best outcome for the group.

They found that the agents were able to learn cooperative strategies in some cases, but that the emergence of cooperation depended on several factors, including the payoff structure of the game and the presence of noise.

Video Interpolation with Diffusion Models

Here, the authors argue that traditional supervised learning approaches for summarisation are limited by the quality and diversity of the available training data, and that RL with human feedback can help address these limitations.

They also propose a framework which involves training a reward model to predict the quality of summaries based on human feedback, and then using this reward model to train a summarisation model using RL. 

It also includes an analysis of the reward model and the summarisation model, and discusses several challenges and limitations of using RL with human feedback for summarisation.

WonderJourney: Going from Anywhere to Everywhere

Another paper here, presents a new approach for generating images from text using diffusion models. Diffusion models are a class of generative models that have recently shown promising results in image synthesis tasks. The authors first train a text encoder to map text descriptions to a latent space. They then use this latent space to condition a diffusion model to generate images. 

The diffusion model is trained using a denoising objective, where the model learns to progressively remove noise from a noisy image until it matches the target image.

The authors evaluated their approach on several benchmark datasets for text-to-image synthesis and compared it to several state-of-the-art models.

The post Top 7 Papers Presented by Google at CVPR 2024  appeared first on AIM.

Google DeepMind announced several new developer-centric features and capabilities for Gemini 1.5 and its Gemini API. Developers are now allowed access to the two million context window for Gemini 1.5 Pro, as well as code execution capabilities when using the Gemini API. These features were announced along with Gemma 2.

The added features are a massive boon to developers, as they serve to improve the functionalities of both Gemini 1.5 Pro and the Gemini API. 

Code Execution for Gemini API

In March, Google announced additional API features, like video frame extraction and parallel function calling. With the new code execution feature, developers can now generate and run Python on the model. This is now available on both AI Studio and the Gemini API.

However, the execution feature is not connected to the internet and billing will be based on the amount of output tokens from the model.

“Once turned on, the code-execution feature can be dynamically leveraged by the model to generate and run Python code and learn iteratively from the results until it gets to a desired final output,” the company stated.

Access to Gemini 1.5 Pro’s Two Million Context Window

At I/O in May this year, Google announced an expansion to Gemini 1.5 Pro’s context window, going from one million to two million. However, this was behind a waitlist and in private preview. With the latest announcement, Google has opened access to the two million context window, specifically to developers.

In opening access, Google has also announced the launch of context caching in the Gemini API for both Gemini 1.5 Pro and Flash. “Using the Gemini API context caching feature, you can pass some content to the model once, cache the input tokens, and then refer to the cached tokens for subsequent requests,” the company stated.

Red-Teaming Ongoing for Gemini 1.5 Flash Tuning

Additionally, Google also stated that more features will soon be announced for Gemini 1.5 Flash. In particular, the company is working on allowing access to tuning the model for developers. As of June 27, the company is rolling out access to developers in order to red-team the feature. It is expected to release by mid-July.

The post Google Opens Access to Gemini 1.5 Pro 2M Context Window, Enables Code Execution for Gemini API appeared first on AIM.

Google DeepMind announced the release of Gemma 2, an advanced version of its open models, available in 9 billion (9B) and 27 billion (27B) parameter sizes.

The model is accessible on Google AI Studio, Kaggle, Hugging Face Models, and soon on Vertex AI Model Garden. Researchers can apply for the Gemma 2 Academic Research Program for Google Cloud credits, with applications open until August 9.

Gemma 2 offers significant improvements over its predecessor, including competitive performance to larger proprietary models and optimised cost efficiency. The 27B model can perform inference on a single NVIDIA H100 Tensor Core GPU or TPU host, reducing deployment costs.

The new models integrate easily with major AI frameworks like Hugging Face Transformers, JAX, PyTorch, and TensorFlow via Keras 3.0. Developers can deploy Gemma 2 on various hardware setups, from cloud-based environments to local CPUs and GPUs.

Gemma 2 is available under a commercially-friendly license, encouraging innovation and commercialization. Google Cloud customers will be able to deploy and manage Gemma 2 on Vertex AI starting next month. Additionally, Google provides the Gemma Cookbook, offering practical examples for building and fine-tuning applications with Gemma 2.

Google emphasises responsible AI development with Gemma 2, incorporating robust safety processes, pre-training data filtering, and rigorous testing against bias and risk metrics. The LLM Comparator tool and text watermarking technology, SynthID, are part of these efforts.

The initial release of Gemma resulted in over 10 million downloads. Gemma 2 aims to support even more ambitious projects, with future plans to release a 2.6B parameter model to balance accessibility and performance.

The post Google Rolls Out Gemma 2, Leaves Llama 3 Behind appeared first on AIM.

Google has announced the general availability of Gemini in the Gmail side panel, extending its capabilities beyond Google Docs, Sheets, Slides, and Drive. The new feature, powered by Google’s advanced models like Gemini 1.5 Pro, enhances productivity for Google Workspace users. Gemini in Gmail allows users to:

• Summarize email threads
• Receive suggestions for responses
• Draft emails with assistance
• Search for specific information within emails or Google Drive files

The integration aims to streamline email management by providing proactive prompts and enabling freeform questions. Users can ask Gemini to retrieve details such as PO numbers, expenditure on events, or upcoming meetings directly from Gmail without leaving the interface.

In addition to the web version, Gemini is now accessible on the Gmail mobile app for Android and iOS. This mobile functionality includes analyzing email threads, presenting summarized views, and upcoming features like Contextual Smart Reply and Gmail Q&A.

End users of Google Workspace stand to benefit from Gemini’s integration across various applications, facilitating efficient workflows and quick access to information. 

The future’s connection with Workspace apps like Docs, Sheets, Slides, and Drive enables smooth collaboration and data retrieval. To utilise Gemini in Gmail, admins must ensure smart features and personalisation are enabled for users.

The rollout schedule for the web version spans from June 24 for Rapid Release domains to July 8 for Scheduled Release domains. Mobile rollout is concurrent, with full availability expected within 15 days for both release types.

Gemini in Gmail is accessible to Google Workspace customers with specific add-ons, including Gemini Business and Enterprise, Gemini Education and Education Premium, and Google One AI Premium.

The post Google Rolls Out Gemini in Gmail Side Panel for Google Workspace Users appeared first on AIM.