Рубрика: AI

IBM Red Hat, one of the top providers of enterprise Linux, has been embroiled in a controversy over the past few months. In June, the company decided to put its Red Hat Linux stream behind a paywall, only continuing to open source its CentOS Stream. The RHEL (Red Hat Enterprise Linux) distribution, which was previously open source, now cannot be redistributed even by paying customers.

This not only disrupts the various upstream enterprise Linux distributions, it also creates a rift between Red Hat and members of the enterprise Linux community. This move has brought up the sensitive topics of the betraying of the open source community’s trust and violating the General Public License (GPL) that Linux is built on. However, the paywalling of RHEL belies a deeper issue in the enterprise Linux market — its business model.

RedHat and other Linux providers have been repackaging publicly available Linux distros for use in the enterprise sector and selling them with vendor support. While the latest move might be necessary to keep RHEL as a competitive service, it also shows the greedy nature of Red Hat, amplifying their perceived ‘ownership’ of enterprise Linux.

Enterprise Linux: Stability over security

To further understand how enterprise Linux works, we must explore the concepts of upstream and downstream distributions. At the top of the ‘stream’ are the Linux kernel updates, with any other distros being ‘downstream’ or derived from the main Linux kernel. In enterprise Linux, RHEL is usually upstream, with other distributions like Alma Linux, Rocky Linux, and Oracle Unbreakable Linux building out features based on RHEL.

In a nutshell, Linux providers like RHEL and Oracle take the publicly available open source code of Linux and go through it with a fine-tooth comb. They then seed out some of the more problematic bugs, optimising their distros for stability. These kernels are often years-old Linux versions which have been locked to that specific version, usually made for a long-term support model in the timeframe of 10 years.

As the enterprise kernels get updated with new versions and features, enterprise service providers try to keep track of the changes and try to fix any bugs that can arise from the update. However, one of the biggest issues with this model is that security updates usually get delayed for years on end, unless the issue is extremely sensitive.

This tradeoff of stability over security has resulted in Linux becoming a prime target for cyber criminals. According to research, Linux is one of the prime targets for hackers running ransomware campaigns, with attacks on these systems increasing 75% over the last few years. Moreover, researchers also found that ransomware groups are tailoring their attacks to focus specifically on Linux systems.

Cybersecurity researchers at VMWare have also exposed malware in Linux-based multi-cloud environments, shedding light into a new family of ransomware and crypto mining payloads made specifically for Linux. Moreover, key decision-makers view Linux as generally more secure than Windows, resulting in a lack of focus on threat management for Linux systems.

This has created an environment where Linux servers are becoming prime targets for attacks, with enterprise Linux providers sorely lacking in providing security updates to the ecosystem. Adding to this, RHEL’s move to put its Linux distribution behind a paywall might end up causing a domino effect in the enterprise Linux space.

The way forward

“I feel that much of the anger from our recent decision around the downstream sources comes from either those who do not want to pay for the time, effort and resources going into RHEL or those who want to repackage,” said Mike McGrath, the vice president, core platforms engineering at Red Hat, defending the company’s decision to paywall the RHEL source code.

Other enterprise service providers have also spoken out against Red Hat. Oracle, which offers both a Red Hat Compatible Kernel (RHCK) and its own distro called Unbreakable Enterprise Kernel (UEK), has been the most vocal against this change.

“IBM doesn’t want to continue publicly releasing RHEL source code because it has to pay its engineers? That seems odd, given that Red Hat as a successful independent open source company chose to publicly release RHEL source and pay its engineers for many years before IBM acquired Red Hat in 2019 for $34 billion,” said Edward Screven, Oracle’s chief corporate architect, and Wim Coekaerts, head of Oracle Linux development.

While issues of monetisation can only be shed light upon by parties close to the matter, closing off the source code for a product which is inherently built for stability over security is a recipe for disaster. Now that there are fewer parties working downstream of the RHEL code to make it more secure, the onus falls on Red Hat to make enterprise Linux safer.

To overcome this issue, enterprise linux service providers must move towards a release schedule that more closely tracks the mainline Linux kernel, as opposed to being downstream of other enterprise Linux distros like RHEL. A good example of this is Oracle’s Unbreakable Enterprise Kernel (UEK).

While UEK first started off as a downstream fork of RHEL, it has now become an enterprise Linux kernel that closely follows the upstream Linux kernel. By taking an approach that allows the maintainers to lag behind mainstream releases, Oracle has enough time to perform ‘enterprise grade’ validation and to even carry out stringent bug tests, ensuring both stability and security.

This approach is the best way for enterprise Linux to finally begin focusing on security, especially considering the current cybersecurity risks that older deployments face.

The post Red Hat: Game’s Up appeared first on Analytics India Magazine.

Created by author using Midjourney Introduction

Wouldn't it be great to have a coding partner that never tires, works around the clock, and can generate an entire codebase from a single project description. This is the intent of a new project called GPT-Engineer, one of a number of emerging AI-powered coders. It's an AI-powered application builder that uses the power of GPT to help you build applications. Developed by Anton Osika, GPT-Engineer is a game-changer in the world of AI-assisted development. You can check out the GPT-Engineer GitHub repository here to see it in action.

GPT-Engineer is not just another AI tool; it's your coding assistant. Whether building a simple application or something more complex, such as a game, GPT-Engineer can help. It's like having a junior developer by your side, asking clarifying questions to ensure it fully understands your project requirements.

Getting GPT-Engineer Up and Running

Before you start coding with GPT-Engineer, you'll need to install it. To do so, clone the GPT-Engineer repository from the project's GitHub page. Next, create a new environment with a recent version of Python, and install the necessary modules. For a step-by-step guide, head over to the GPT-Engineer GitHub page.

Leveraging GPT-4 for Design Documentation Prompts

GPT-Engineer prompting is based on feeding it design documents. But how does one draft a useful design document? Why not use GPT-4 to help create them?!?

Let's use GPT-4 to generate detailed project design descriptions, which can then be fed into GPT-Engineer. This way, you can ensure that your AI coding buddy has all the information it needs to build your application.

As an example, let's create a design document with GPT-4 to prompt GPT-Engineer with. It really doesn't take much to get ChatGPT to produce a useful design document — though they can be more complex, as you will see below. Here is our prompt:

I want to implement a game of snake in Python, using the game's classic rules and gameplay. It can use external libraries to accomplish this. Generate a design document for this project.

Here are the first several lines of the generated design document, which I can assure you, in total, does a good job of outlining what needs to be done.

Introduction
The purpose of this design document is to outline the implementation details for creating a classic Snake game in Python. The game will adhere to the traditional rules and gameplay, where the player controls a snake that grows in length by eating food while avoiding collisions with itself and the game boundaries. The game will be developed using external libraries to facilitate graphics and user input.

Game Components
The Snake game will consist of the following components:
a) Snake: A movable object that grows in length when it eats food and loses the game if it collides with itself or the game boundaries.
b) Food: A randomly placed object that the snake can eat to increase its length and score.
…

Now that we have a design document prompt, we can use it with GPT-Engineer.

Building a Design Document with ChatGPT

One of the great things about GPT-Engineer is its versatility. It can implement a simple snake game, just as it can help you build a wide range of applications, including a simple key-value database using only Python's standard library modules. Just describe the project in the main.prompt file, and GPT-Engineer will ask you for any necessary clarifications. Once you've provided the details, sit back and watch as GPT-Engineer generates your codebase.

First, let's generate a design document with ChatGPT-4. Remember, you don't need to use ChatGPT to accomplish this. If you have a detailed description of the software you are looking to build you are able to use it. But for a comprehensive conversion of your design notes to something that is easily ingestible by GPT-Engineer, this is a logical route to take.

Here is the prompt the author used with ChatGPT-4 to generate a design document:

Generate a high quality design document that can be used by our project team to design the software described below. It should be complete and usable as the sole design documentation to get our team to complete our project.

—
Simple key-value database engine written in Python using only standard library modules (except those mentioned below).

The basic interaction with the database engine when it is running should be via imported Python module with a clear and concise library API.

There should also be a separate console that can be run to act as a command line REPL for interacting with the database engine and databases.

Additional methods of interaction should include an API created with Fast API (external library).

Design specifications:

A key-value database is a type of NoSQL database that stores data as a collection of key-value pairs. Unlike traditional relational databases, key-value databases do not enforce a rigid schema, providing flexibility and scalability for certain use cases. Designing a key-value database involves considering several important specifications.

Commands that should be accessible:
— connect to database
— delete database
— get key
— insert data
— update data
— delete data
— search for existence of both keys and values

Data Model: Each record consists of a unique key and an associated value. The value can be a simple scalar value or a complex data structure (such as JSON or XML).

Key Structure: The key serves as the primary identifier for each data entry and must be unique within the database. Keys can be alphanumeric strings, integers, or even composite keys (multiple values combined to form a single key).

Value Structure: The value associated with each key can be of any data type. It can range from basic data types (strings, numbers, booleans) to more complex structures like nested objects or arrays.

Performance and Caching: In-memory caching techniques should be employed to speed up read operations and reduce latency.

And here is the design document that ChatGPT-4 produced:

## Overview
The main goal of this project is to design and develop a lightweight key-value database engine written in Python, using only standard library modules, and FastAPI for the HTTP interface. The application will provide the essential features of a key-value database, including connecting to the database, data insertion, update, deletion, and retrieval, along with search functionality for keys and values.

The database will be interfaced via three methods:
1. A Python module that can be imported by other Python programs.
2. A command line REPL console for manual interaction.
3. An API created with FastAPI for remote interactions.

## Components
The database engine will have the following main components:

1. **Data Store**: The actual storage for the key-value pairs. As a simple Python implementation, this could be a dictionary where keys are unique identifiers and values can be of any datatype. The data store should have efficient methods for accessing, adding, updating, and deleting data.
2. **Database Engine**: The engine that provides the methods to interact with the data store. This will include methods to connect to a database, insert, update, and delete data, as well as fetch a value by key. It should also include a method to check the existence of a particular key or value in the data store.
3. **In-Memory Cache**: A cache layer that holds frequently accessed data to improve read speed and reduce latency. This can be implemented using Python's `functools.lru_cache` decorator or a custom caching solution.
4. **CLI**: A command-line interface for manual interaction with the database engine. This would use the `cmd` or `argparse` module from the Python standard library to create a REPL interface.
5. **HTTP API**: A FastAPI application that exposes the database engine methods as a HTTP API. This will allow remote interaction with the database.

## Data Model
The database engine will follow a simple key-value data model. Each record in the database will consist of a unique key and an associated value.

### Key Structure
Keys can be alphanumeric strings, integers, or composite keys. They must be unique within the database and will serve as the primary identifier for each data entry.

### Value Structure
The value associated with each key can be of any Python data type. It can range from basic data types (strings, numbers, booleans) to more complex structures like nested objects or arrays.

## API Design
The Python library API, CLI, and FastAPI should provide the following functions:
— `connect(database_name: str)`: Connect to a database. Creates the database if it does not exist.
— `delete_database(database_name: str)`: Deletes a database.
— `insert(key: str, value: Any)`: Inserts a new key-value pair into the database.
— `update(key: str, value: Any)`: Updates the value of an existing key.
— `delete(key: str)`: Deletes a key-value pair from the database.
— `get(key: str)`: Fetches the value of a key from the database.
— `exists_key(key: str)`: Checks if a key exists in the database.
— `exists_value(value: Any)`: Checks if a value exists in the database.

## Performance and Caching
In-memory caching techniques will be employed to speed up read operations and reduce latency. The `functools.lru_cache` decorator or a custom caching solution could be used for this purpose. The cache should have a maximum size to prevent excessive memory usage.

## Conclusion
This document provides a high-level design for a simple key-value database engine using Python and FastAPI. By adhering to this design, we can ensure that the database will be flexible and scalable, meeting the needs of various applications.

And that's it. The only thing left to do was run GPT-Engineer and direct it to these criteria.

Conclusion

So what now? How do we fix the issues that were present in the generated code? Well, recently and up to a few days ago, the author was using GPT-Engineer in conjunction with Aider, another capable GPT-backed software generating project to try and tweak the results. However, with the recent public availability of the ChatGPT Code Generator plugin, it seems that trying it out for both generating and fixing code would be a good idea.

In sum it up, GPT-Engineer is more than just a tool, it's a potential revolution in how we build applications. It's a testament to the power of AI, LLMs, and perhaps a glimpse into the future of coding. Regardless of whether you are an experienced developer or just starting out, GPT-Engineer is definitely worth a look.

Happy coding prompting!

More On This Topic

• StarCoder: The Coding Assistant That You Always Wanted
• Fear not, for AI coding is here to help you!
• 7 Must-Know Python Tips for Coding Interviews
• Free Python Project Coding Course
• ChatGPT as a Python Programming Assistant
• Scrape Images Easily from Websites in A No-Coding Way

The world is a scarier place because of misinformation and the threat posed by misinformation has as adverse consequences as destructive weapons. Recently, an AI-generated image of former US President Donald Trump being arrested was circulated on Twitter and that created a frenzy. With such a rise of fake news on its platform, alarmed Twitter to come up with innovative ways to mark and highlight fake news, it is still riddled with misinformation and hate speech, which the site claims are just a fraction of the community.

However, Meta is willing to compromise moderation for the time being, as it opened the platform for everyone. Within two weeks of the launch of Threads, it has amassed more than 100 million sign-ups. Now, the platform has users like Nick Fuentas, a far-right American white supremacist political commentator who was banned on Instagram, with his followers.

Fuentas claimed to have made a fake Instagram account to access the new microblogging site. Additionally, he directed his followers to get in early on the platform to be able to amass a fan following.

“I signed up for it last night. I made a fake Instagram. I got on a fake Thread,” he said, according to Media Matters. Fuentes told his followers they should “try and build a big account. I mean, if you get in early, maybe some of you guys can blow up and red pill some people there.”

Free Speech for a Time Being

Moserri’s pitch for the platform was to be ‘a town square of a billion people’ ‘that are interested in a less angry place for conversations’—and that is where it should be headed. For this purpose, the platform was willing to eliminate or not support news content on the platform.

However, something else is brewing within the platform. In a move to attract users and encourage endless conversations, Meta’s Threads has shown its lack of commitment to managing misinformation on the platform. This lackadaisical approach of Threads has caused the influx of Nazi supporters, anti-gay extremists, and white supremacists on the platform.

The platform has also removed all the guardrails which are in place for Facebook and Instagram, where third-party fact-checkers identify, review, and label false information. Meta’s spokesperson admitted saying, “the company will not extend its existing fact-checking program to Threads”. However, the spokesperson mentioned that if a crosspost on Facebook or Instagram receives a false rating from fact-checkers, that label would also apply to Threads.

The amiable censorship from the platform seems to be in line with the opinion of the US House Judiciary Committee which recently subpoenaed Meta over concerns about censorship on its new Threads app. In a letter to Meta CEO Mark Zuckerberg, Judiciary Committee Chairman Jim Jordan, wrote, “Given that Meta has censored First Amendment-protected speech as a result of government agencies’ requests and demands in the past. The committee backed by Republicans expressed its concerns against potential violations of first amendment violations that may occur on the new platform.

However, with Threads, Meta seems to have a different strategy, and seems to be in line with Jordan’s view of absolute free speech despite connotations of hate or bias.

It also seems like Threads wants to stay clear of any political controversy and the never-ending debate around content moderation. In a move to position itself as a true rival to Twitter and Musk, it looks like Meta is willing to let anything go on its platform.

Threads Hanging by a Thread

However, while Meta “couldn’t be more psyched” about how the launch week has gone, with Threads hitting 100 million sign-ups, just five days after its launch—Instagram Mosseri has expressed his doubts about retention. In a thread he said, “I appreciate the vote of confidence, but we don’t even know if this thing is retentive yet.”

Being at the helm of one of the largest social media platforms, Mosseri understands the workings and knows what he is talking about. After initial days of pomp and record-setting signups, the number of active users dropped from 49 million to 23.6 million within a week. Although Threads attracted attention as a Twitter alternative, it’s very apparent that it has yet to gain a loyal user base.

It seems like just letting the guardrails down isn’t gonna work in Threads’ favour, and it will have to do a lot more than this to stand the test of time and be a true alternative to Twitter.

The post For Threads’ Success, Meta will Watch the World Burn appeared first on Analytics India Magazine.

Not really. Though pushed to chase “Generative AI” thanks to the hype around it, sadly, most tend to focus solely on the shiny end-products without understanding the hard work, innovation, and data engineering that precede such success.

The generative AI landscape is peppered with eye-catching headlines of colossal acquisitions by major companies. For instance, the news on DataBricks’ $1.3 billion purchase of MosaicML, Accenture acquiring Flutura, and Snowflake’s acquisition of Myst AI, did the rounds of the startup world recently. While these transactions signal the importance and value of generative AI, do the CEOs pouring their all into it really understand the infrastructure and tooling that goes behind it?

“They want generative AI. To them, Microsoft is the generative AI Company, and ChatGPT is generative AI,” remarked AI advisor Vin Vashishta, in a LinkedIn post.

Business leaders view technology giants like Microsoft as the epitome of generative AI capabilities. Every company wants to be Microsoft and wants a ChatGPT under its belt. However, this simplified perception disregards the critical role of underlying data engineering infrastructure and tooling.

The mirage of a generative AI product

A disconnect exists between the CEOs’ visions for implementing generative AI into their companies and the technical intricacies that underpin its successful deployment. There are various challenges that come with building a generative AI framework, from deciding if it’s even plausible to be used in the company to deciding how to deploy it. Interestingly, CXOs find this hard to understand.

CEOs pushing their teams into generative AI without comprehending the underlying infrastructure and tooling requirements is because of the push they get from the investors and market trends. Even then, it highlights the need for education and collaboration between business leaders and data professionals to ensure successful integration.

CEOs and top-level executives are often captivated by the allure of generative AI amid all the funds and money that this technology is minting for companies around them. Skipping literacy about the technology, the infrastructure required, the expertise required for the team, the availability and ethical boundaries around using the data is often ignored. The push is directly given to the team to build a so-called “generative AI product”.

It is the data teams that are tasked with the Herculean implementation of generative AI, grapple with complex mesh structures, data governance, legal considerations, and contractual obligations. While these technicalities are crucial for successful AI integration, they are often lost on CEOs, who are primarily interested in the final output rather than the steps to achieve it.

Who is to blame?

This discrepancy between the team often stems from the communication gap between business leaders and data professionals. CEOs use a different language, emphasising results, market positioning, and profits, something that the data teams do not really care about if the process of building models is not streamlined. This disconnect leads to misaligned expectations and potentially frustrating interactions.

Vashishta gives an example in his post that when data teams return from conferences brimming with enthusiasm about cutting-edge infrastructure and tooling, business leaders fail to share their excitement. “To CEOs, these concepts appear as incomprehensible dots on a Gartner report, rendering them unenthusiastic about investing in the necessary foundation.”

On the other hand, the success of technical and data teams hinges on turning data into actionable insights that help leaders drive the business forward. To empower leaders, it’s crucial to provide accessible information instead of using technical jargon. Generative AI offers valuable outcomes, but failure to deliver insights can impact funding. In the end, the business leaders care about the ROI.

AI literacy for CEOs

When data literacy was the talk of the town, CEOs were saying, “we are a data literate company”. And now, the CEOs say that they are “AI literate”, when arguably they may not be. To bridge this gap, it is imperative to educate business leaders about the core aspects of generative AI implementation. CEOs must comprehend that investing in AI engineering infrastructure and tooling is the bedrock on which successful generative AI initiatives rest. Being data-literate is one thing, but it’s time they became AI literate too.

By providing business leaders with a clear understanding of the steps involved in AI development, including data collection, data preprocessing, model training, and validation, they can make informed decisions and set realistic expectations with their teams.

For the AI teams, it is important to understand that infrastructure isn’t actually a C-Level task. Generative AI is a strategic goal, driven by C-level, but AI infrastructure and quality are tactical and should be handled by the middle management. C-Suite trusts middle management to make the right decisions and convey the importance of data quality and infrastructure. Generative AI is a means to an end, not a strategic goal itself for the CEOs. That is expected to remain the case till the time they get AI literate.

The post CEOs Think They Understand Generative AI appeared first on Analytics India Magazine.

If you were thinking – it’s raining money in generative AI, you might be disappointed.

Despite the hype around generative AI funding, reports show that startup funding remains bleak, especially for newcomers who have no moat. In June, Venture Intelligence reported that under ChatGPT’s shadow, funding received by domestic AI startups witnessed a significant decline this year, totaling $510 million from January to May in 2023. This amount is merely half of the five-month average of $1.02 billion recorded in the previous year.

Securing funding from investors for generative AI startups is not a cake walk anymore. Investors have realized that these companies need to find their niche before starting out. Convincing VCs with an idea isn’t the recipe to get their money. Nowadays most of the generative AI startups are just trying to cash the wave of generative AI to make some money without adding any substantial application for their customers.

There are going to be selective applications from startups that will work well and can build businesses, but maybe 70-80% of them are going to die in the end. Recently, ZuAI, a generative AI Edtech startup raised $484K (INR 4 Cr) in a seed round led by the early stage investor Prime Venture Partners. However, there is one problem that ZuAI’s GPT-centric platform cannot solve. It cannot generate content for all the subjects yet as mathematical solutions are not the strong suit of AI models currently.

If not in an idea, VCs tend to invest in founders of AI startups that are from premier institutions like Stanford, Harvard and others, alongside having prior experience at big techs. Based in Paris Mistral AI, which raised $133 million in seed funding, is co-founded by alums from Google’s DeepMind and Meta to compete against OpenAI in the building, training and application of large language models and generative AI. The case is the same with Inflection.AI by Mustafa Suleyman, who has worked with DeepMind before. Anthropic AI and Dario Amodei also have the same story, being a former OpenAI employee.

So, what can be the moat, then?

Focus on Data

One of the most widely discussed AI moats in venture capital circles is the concept of proprietary data. The reason behind its prominence is evident: data serves as the lifeblood of AI and acts as the crucial driver for foundation models. Despite the substantial funding pouring into generative AI startups, the risk of failure looms if they are unable to access the appropriate data—a challenging task in itself. A lot of AI startups fail because they don’t have the essential data to train their model.

Companies who may well be pursuing a brilliant application of AI, but if they don’t have access to data that will give them the ability to build a powerful application, the future is bleak. OpenAI recently partnered with AP and will have access to AP news stories going back to 1985. It is going to be tough for AI startups to challenge Microsoft backed OpenAI.

Lately, companies are looking at creating specialised datasets from ChatGPT (based on GPT-4), and others to further train open-source, smaller models like LlaMA, to fit into the required use cases and usability. Focusing on building quality datasets for training your foundational models can be rewarding.

UI/UX is all that matters

Jasper AI, which is based on GPT 4, is trying to differentiate itself from ChatGPT. They even posted a blog saying ‘ Jasper vs ChatGPT: Why Jasper Wins’. The question here is, is it really winning. Why would a company buy a Jasper subscription when they can directly use ChatGPT or build their own model based on their API.

In my humble view, here are the criteria to determine if a startup is building something interesting vs. just a "GPT wrapper":
1. Startup has novel UI/UX that is *10x better* (incumbents can't easily copy)
2. Startup has novel tech that is *10x better* prior to layering in…

— Champ (@doyouknowchamp) July 16, 2023

Among the hype around ChatGPT, we are increasingly seeing startups that are either using GPT in their name, or building technologies using the APIs provided by OpenAI. If you are building on other’s APIs what can be your moat? Your UI/UX should be so good that the user doesn’t want to use the original model. Importance of creating a unique experience for users solely based on UI/UX shouldn’t be discarded. UI/UX can help you turn around the tide. A discussion has been going on over Twitter how UI/UX can or cannot be a moat.

Know Your Customers

Creating generative AI solutions or products for B2B customers is a tough nut to crack. Executives at large companies have shown great interest in AI since the beginning, leading many startup founders and venture capitalists to believe that these companies would be ideal early customers. However, the startups building AI solutions for these companies underestimated the executives’ and engineers’ ability to quickly implement AI using open-source tools.

An engineering leader would rather spin up their own infrastructure for free and build tech themselves than buy something from a new startup. Zoho, a prominent Indian software-as-a-service (SaaS) company, has embraced the challenge of developing its own large language model (LLM) akin to OpenAI’s GPT and Google’s PaLM 2 models. The project is under the supervision of Zoho’s founder and CEO, Sridhar Vembu.

Similarly, TCS is attempting to build its own Github Copilot alternative, which is touted to be used for enterprise code generation. In April, Tech Mahindra became the first IT giant to launch something like a Generative AI Studio.

Know Your Investors

If we look at the popular AI startups like Inflection, Cohere and Anthropic, most of them are backed by Microsoft and Google. It takes huge amounts of money to build a successful AI company with exceptional minds. Like mentioned earlier, to build a successful generative AI model, you need large amounts of data, infrastructure and money. VCs are not going to shell out that easily.

Once you have found an idea, and worked on the MVP, it is wise to join the accelerator or incubation program offered by large tech giants. Following are the programs offered by them – Google for Startups Accelerator, Microsoft for Startups. And also recently Neo announced a new AI track for Neo Accelerator with help from Microsoft and OpenAI to empower startup founders with the best of generative AI.

In Conclusion

If we look at startups that are successful in generative AI, they are finding their own niche application and not copying OpenAI or using its API, examples being MidJourney and Stability AI. We need to see from the customer perspective and understand business use cases. Generative AI isn’t valuable by itself. We are just seeing the shedding of the companies that didn’t offer officiant business value and differentiation to keep their foothold.

The post Generative AI Startups Have No Moat, No Money appeared first on Analytics India Magazine.

India’s second-largest software services exporter Infosys has informed the stock exchanges that it has entered into a new agreement with an undisclosed established client to provide AI and automation services over a period of five years. The partnership has an estimated target spend of $2 billion. The announcement pushed up the company’s stock price up by 3.6% on the Bombay Stock Exchange (BSE).

“Infosys has entered into framework agreement with one of its existing strategic clients to provide AI and automation led development, modernisation and maintenance services. The total client target spend over 5 years is estimated at USD 2 billion,” the company said in an exchange filing on Monday. The news comes three days before July 20 when Infosys is scheduled to release the results of its June quarter (Q1FY24).

In accordance with the company’s exchange filing, the agreement includes the advancement, modernization, and upkeep of AI and automation-related services. Notably, the IT giant had recently unveiled a wide-ranging and cost-free AI certification training initiative through Infosys Springboard. This program aims to equip individuals with skills required to thrive in the future job landscape.

Infosys’ AI move underscores the growing trend of Indian IT companies increasing their investments in the field of AI. Long before OpenAI’s ChatGPT hit the scene, Tech Mahindra was already working with generative AI. Notably, the IT behemoth’s chief executive, CP Gurnani, lauded the Storicool platform—an auto content creation tool that proved ahead of its time.

In line with this trajectory, Tata Consultancy Services (TCS) made headlines with its own foray into generative AI capabilities, joining forces with Google Cloud. Furthermore, Wipro, too, has entered a partnership with Google Cloud to harness the power of its generative AI tools, integrating them with their in-house AI models, business accelerators, and pre-built industry solutions, as per the company’s announcement. This signifies a competitive shift within the Indian IT landscape.

Read more: How Indian IT Giants are Bringing GenAI to Their Clients

The post Infosys Signs Mega AI Deal Worth $2B, Shares Rise Over 3% appeared first on Analytics India Magazine.

Microsoft presented new partner opportunities and generative AI innovations at the keynote for its Inspire conference, which is being held virtually July 18-19. Artificial intelligence was the talk of the show, with Microsoft Chairman and CEO Satya Nadella saying, “We are in the midst of a massive platform shift with the new generation of AI.”

At Inspire, the company announced pricing for Microsoft 365 Copilot; Bing Chat Enterprise, a version of the chatbot built into Microsoft 365 with considerations around privacy for businesses; and enhancements to the Azure AI model catalog, among other new offerings.

Jump to:

• Nadella goes all-in on generative AI
• Microsoft 365 Copilot’s price is steep
• Bing Chat Enterprise rolled out
• Copilot application announced for sales teams
• AI features added to Power Automate Process Mining
• Meta’s Llama 2, other tools to be added to Azure AI model catalog
• Azure Migrate and Modernize tool gains new features
• Microsoft pushes AI Cloud Partner Program

Nadella goes all-in on generative AI

Nadella pointed out two major applications for generative AI: the ability of AI to make computer user interfaces more intuitive using natural language and information management. Generative AI is a reasoning engine that works on top of all of your data, he said.

“To be able to have natural language as the interface we have with computers is the pursuit,” said Nadella.

“Microsoft’s big bet on generative AI across their product portfolio will rely heavily on their partner ecosystem to be successful,” said Jason Wong, Gartner analyst. “Partners will be critical to ensuring customer readiness of data and talent, effective implementation across multi-vendor technology stacks, and change management to ensure value attainment.”

SEE: Keep up to date with our ChatGPT cheat sheet (TechRepublic)

Microsoft 365 Copilot’s price is steep

Microsoft 365 Copilot, which embeds a ChatGPT-based conversational AI into 365 applications, will cost $30 per month per user, announced Yusuf Mehdi, corporate vice president and consumer chief marketing officer at Microsoft. He demonstrated Copilot by creating a PowerPoint presentation and then a summary of that presentation, all within Teams. Microsoft did not announce when Copilot will be generally available.

“The Microsoft 365 Copilot price indicated how much value Microsoft believes there is to unlock in the digital workplace with generative AI,” said Wong. “The comparatively high price will force potential customers to think more carefully about which employee segments will benefit most from Copilot and to identify concrete use cases with ROI to justify the extra costs.”

Bing Chat Enterprise rolled out

Organizations will be able to access Bing Chat securely from within their own Microsoft 365 ecosystem with Bing Chat Enterprise, a new offering available today. It was created in part to address enterprise concerns about privacy, such as company-wide generative AI. With Bing Chat Enterprise, commercial data protection will make sure enterprise user data is not commingled with web data, not shared with Microsoft and not used to train any AI models. It includes up-to-date information sourced from Bing search.

ChatGPT can be used by logging in to Bing with Microsoft 365 work credentials, which ensures company data is covered by commercial data protection. Bing Chat Enterprise is included for no cost with Microsoft 365 E3, E5, Business Standard and Business Premium plans. Microsoft also expects to offer it from within Windows Copilot in the future, and at $5 per user per month as a stand-alone application.

Mehdi demonstrated multimodal visual search in Bing chat as well. Visual search uses GPT-4 to let Bing Chat analyze images the user uploads; then, it can answer questions about the image. Visual search is available for free today.

Copilot application announced for sales teams

Nadella announced a new application called Sales Copilot, which embeds Copilot functionality and CRM data from Salesforce, Dynamics 365 or other platforms within Microsoft 365. Sales Copilot also includes automated information about sales opportunities, integration with Microsoft 365 Copilot and PowerPoint and tips generated in real time in a Team meeting, such as providing context when the customer mentions a competitor.

Sales Copilot is generally available today.

AI features added to Power Automate Process Mining

AI features will now be available within Power Automate Process Mining to map out sales, finance, HR or other processes and “identify opportunities for how we drive that AI advantage to every customer,” Nadella said.

Essentially, Nadella wants businesses to be able to show how they derive value from generative AI and make it easier for them to find ways to use it.

Charles Lamanna, corporate vice president of business applications and platform at Microsoft, said Process Mining Copilot can be asked, “What are the top insights it has for me?” to find work bottlenecks and what is causing the bottlenecks and make suggestions for other tools to use to automate tasks or track data points more closely.

Power Automate Process Mining will be generally available on August 1.

Meta’s Llama 2, other tools to be added to Azure AI model catalog

Microsoft announced that Meta’s family of Llama 2 open-source foundation models will be available on Azure and Windows starting today.

In addition, Nadella revealed the regional expansion of the Azure OpenAI service to Asia, as well as increased availability in North America and Western Europe.

Azure AI model catalog offers AI options for developers

The Azure AI model catalog, which is currently in public preview, empowers developers to apply AI, including GPT-4 and a library of other AI from Hugging Face, to their applications. Developers can browse for models and fine-tune and deploy them in sandboxes running in Azure.

“This means organizations can quickly and easily benefit from the power of these foundational models, yet they have their own private instance of the model,” said Jessica Hawk, corporate vice president in data, AI and digital applications at Microsoft. Having private instances allows developers more time to improve model effectiveness and ensure their model follows AI safety best practices with built-in guidelines, she said.

“Application developers now can take advantage of both the frontier models of OpenAI as well as the models available in open source to transform every software category there is,” Nadella said.

Azure Migrate and Modernize tool gains new features

In the world of cloud migration, Microsoft has enhanced the Azure Migrate and Modernize tool. This tool is intended for partners who take customers through cloud assessments to Azure migration. They can go to Azure Migrate in the Azure portal and discover the customer’s data center inventory, which has assessment tools to help create a migration path. It can also help organizations build a business case for migration.

Changes include expanded capabilities for the free Azure Migrate tool, extended services in Azure Confidential Computing and a preview for Azure Boost, a virtualization tool for processes traditionally handled by a hypervisor.

Microsoft pushes AI Cloud Partner Program

Much of the Microsoft Inspire keynote presentation was devoted to partner opportunities.

“It’s really exciting to see us completely revamp the enticement, the incentives” of the AI Cloud Partner Program, Nadella said.

“This AI advantage is not just a Microsoft advantage — it is an advantage for the entire Microsoft partner ecosystem,” said Judson Althoff, executive vice president and chief commercial officer at Microsoft. He described sharing AI resources as the next step after moving customers to the cloud and encouraging digital transformation across organizations.

Althoff warned against companies trying to “AI-wash,” or use the term AI in their company or product names, without really showing market differentiation and value.

Empowering partners to make their own AI assistants

Throughout the presentation, partners demonstrated how they made their own AI assistants to prepare for meetings and make strategic business decisions.

“Microsoft’s Copilot strategy is not just about their own offerings, but enabling a new cloud and development platform centered on generative AI with network effects that drive more value with more partner participation,” said Wong. “Microsoft wants everyone to be building their own copilots on their platform.”

Subscribe to the Microsoft Weekly Newsletter

Be your company's Microsoft insider by reading these Windows and Office tips, tricks, and cheat sheets.

Delivered Mondays and Wednesdays Sign up today

Maybe showing off an AI-generated fake TV episode during a writers’ strike is a bad idea Devin Coldewey @techcrunch / 11 hours

The ongoing strike of creatives in TV and film, plus the nascent threat of AI-based writing and effects, make it a complicated time to work in show business. But little savvy is required to see that this may be the worst possible moment to soft-launch an AI that can “write, animate, direct, voice, edit” a whole TV show — and demonstrate it with a whole fake “South Park” episode.

The company behind it, Fable Studios, announced via tweet that it had made public a paper on “Generative TV & Showrunner Agents.” They embedded a full, fake “South Park” episode where Cartman tries to apply deepfake technology to the media industry.

The technology, it should be said, is fairly impressive: Although I wouldn’t say the episode is funny, it does have a beginning, a middle and an end, and distinct characters (including lots of fake celebrity cameos, including fake Meryl Streep). On a GitHub page, the processes that interact to perform this complex task are detailed, with the appropriate examples and diagrams.

But ultimately the whole thing seems monstrous. Cue Jeff Goldblum talking about how just because they could doesn’t mean they should. Especially when half of Hollywood is striking and many of the rest are doing their best not to cross picket lines.

Actual writers, directors, editors, effects workers, and millions of others are warning of the dangers of AI — not that it will destroy life on earth, but that it will be employed as a cost-cutting measure by clueless executives, eliminating livelihoods and reducing creative work to a self-perpetuating algorithmic mess. Is this not Midjourney for TV, with all that implies?

Thousands of authors sign letter urging AI makers to stop stealing books

CEO Edward Saatchi told TechCrunch that he thinks demonstrating this capability is actually good for the labor side of the strike.

We think the timing is correct — we are right in the middle of the biggest strike in 60 years, by releasing the research (but not the ability for anyone to create episodes of protected IP) we hope [for] the Guilds in Hollywood to negotiate strong, strong, strong protections that producers cannot use AI tools without the express permission of artists. Frankly the IP holders also need to figure out how to negotiate with AI chatbot companies who are profiting from their work.

The strike is the moment of maximum leverage to set rules for the coming decades and keep producers from using this tech.

Fable started in 2018 as a spinoff from Facebook’s Oculus (how times have changed since then), working on VR films — a medium that never really took off. Now it has seemingly pivoted to AI, with the stated goal of “getting to AGI — with simulated characters living real daily lives in simulations, and creators training and growing those AIs over time,” Saatchi said.

Simulation is the name of the product they intend to release later this year, which uses an agent-based approach to creating and documenting events for media, inspired by Stanford’s wholesome AI town.

Researchers populated a tiny virtual town with AI (and it was very wholesome)

If you’re confused looking at Simulation’s site, that’s understandable and partly by design. It’s not a real company, and the founding team is all fake, too — not just AI-generated portraits but non-real people. The company history is also invented. If Saatchi hadn’t told me it was all part of the product concept, I would have guessed this was an elaborate hoax (with an NFT play right around the corner).

His approach to media is certainly provocative, if that’s the word, but he is also genuinely a part of the creative community, with a Peabody for the company’s work on Lucy and the Wolves in the Walls. Which is part of what makes this whole situation so perplexing. If you asked around the many people participating in or abiding by the strike (this includes me, incidentally, outside my work as a reporter) what they thought would help their cause, I’m confident that somewhere between zero and none would say “a generative AI that produces entire TV episodes.” Yet that is exactly what Fable decided to pursue and publicize.

“If creative people can get their ideas out and cinema can have an element more like novels or painting where a single person can make a show or movie themselves we could get weirder things than the cookie-cutter studio reboots we often see,” he explained. There’s a certain sense to that, but is the only way forward across the picket line?

Perhaps, because it’s also clear that Saatchi and his team see media production as simply a stepping stone toward a higher goal. “Our focus is not on changing Hollywood — our focus is on getting to AGI,” he wrote.

Whether one, both, or neither is accomplished, Fable certainly has contributed to rocking an already unstable boat. If they don’t get sued into oblivion (the agent was certainly trained extensively on copyrighted data, to say nothing of the fake “South Park” episode), their work may yet be referenced as a notable example of the power and danger of AI.

Meta releases Llama 2, a more ‘helpful’ set of text-generating models Kyle Wiggers 8 hours

The generative AI landscape grows larger by the day.

Today, Meta announced a new family of AI models, Llama 2, designed to drive apps such as OpenAI’s ChatGPT, Bing Chat and other modern chatbots. Trained on a mix of publicly available data, Meta claims that Llama 2’s performance improves significantly over the previous generation of Llama models.

Llama 2 is the follow-up to Llama — a collection of models that could generate text and code in response to prompts, comparable to other chatbot-like systems. But Llama was only available by request; Meta decided to gate access to the models for fear of misuse. (Despite this precautionary measure, Llama later leaked online and spread across various AI communities.)

By contrast, Llama 2 — which is free for research and commercial use — will be available for fine-tuning on AWS, Azure and Hugging Face’s AI model hosting platform in pretrained form. And it’ll be easier to run, Meta says — optimized for Windows thanks to an expanded partnership with Microsoft as well as smartphones and PCs packing Qualcomm’s Snapdragon system-on-chip. (Qualcomm says it’s working to bring Llama 2 to Snapdragon devices in 2024.)

So how does Llama 2 differ from Llama? In a number of ways, all of which Meta highlights in a lengthy whitepaper.

Llama 2 comes in two flavors, Llama 2 and Llama 2-Chat, the latter of which was fine-tuned for two-way conversations. Llama 2 and Llama 2-Chat come further subdivided into versions of varying sophistication: 7 billion parameter, 13 billion parameter and 70 billion parameter. (“Parameters” are the parts of a model learned from training data and essentially define the skill of the model on a problem, in this case generating text.)

Llama 2 was trained on two million tokens, where “tokens” represent raw text — e.g. “fan,” “tas” and “tic” for the word “fantastic. That’s nearly twice as many as Llama was trained on (1.4 trillion), and — generally speaking — the more tokens, the better where it comes to generative AI. Google’s current flagship large language model (LLM), PaLM 2, was reportedly trained on 3.6 million tokens, and it’s speculated that GPT-4 was trained on trillions of tokens, as well.

Meta doesn’t reveal the specific sources of the training data in the whitepaper, save that it’s from the web, mostly in English, not from the company’s own products or services and emphasizes text of a “factual” nature.

I’d venture to guess that the reluctance to reveal training details is rooted not only in competitive reasons, but in the legal controversies surrounding generative AI. Just today, thousands of authors signed a letter urging tech companies to stop using their writing for AI model training without permission or compensation.

But I digress. Meta says that in a range of benchmarks, Llama 2 models perform slightly worse than the highest-profile closed-source rivals, GPT-4 and PaLM 2, with Llama 2 coming significantly behind GPT-4 in computer programming. But human evaluators find Llama 2 roughly as “helpful” as ChatGPT, Meta claims; Llama 2 answered on par across a set of roughly 4,000 prompts designed to probe for “helpfulness” and “safety.”

Meta’s Llama 2 models can answer questions — in emoji. Image Credits: Meta

Take the results with a grain of salt, though. Meta acknowledges that its tests can’t possibly capture every real-world scenario and that its benchmarks could be lacking in diversity — in other words, not covering areas like coding and human reasoning sufficiently.

Meta also admits that Llama 2, like all generative AI models, has biases along certain axes. For example, it’s prone to generating “he” pronouns at a higher rate than “she” pronouns thanks to imbalances in the training data. As a result of toxic text in the training data, it doesn’t outperform other models on toxicity benchmarks. And Llama 2 has a Western skew, thanks once again to data imbalances including an abundance of the words “Christian,” “Catholic” and “Jewish.”

The Llama 2-Chat models does better than the Llama 2 models on Meta’s internal “helpfulness” and toxicity benchmarks. But they also tend to be overly cautious, with the models erring on the side of declining certain requests or responding with too many safety details.

To be fair, the benchmarks don’t account for additional safety layers that might be applied to hosted Llama 2 models. As part of its collaboration with Microsoft, for example, Meta’s using Azure AI Content Safety, a service designed to detect “inappropriate” content across AI-generated images and text, to reduce toxic Llama 2 outputs on Azure.

This being the case, Meta still makes every attempt to distance itself from potentially harmful outcomes involving Llama 2, emphasizing in the whitepaper that Llama 2 users must comply with the terms of Meta’s license and acceptable use policy in addition to guidelines regarding “safe development and deployment.”

“We believe that openly sharing today’s large language models will support the development of helpful and safer generative AI too,” Meta writes in a blog post. “We look forward to seeing what the world builds with Llama 2.”

Given the nature of open source models, though, there’s no telling how — or where — the models might be used exactly. With the lightning speed at which the internet moves, it won’t be long before we find out.

AI chatbots have become valuable tools in assisting people with all of their text-based needs such as writing, coding, research, and more. However, they have typically been limited to text, until a recent shift towards multimodal inputs, and Bing Chat now joins the trend.

Also: How to use Bing Chat (and how it's different from ChatGPT)

On Tuesday, Microsoft announced that Bing Chat will now accept images in its prompt inputs. This will allow users to upload images that they would like more information about, or that relate to the prompt in some way.

For example, if users see a painting and are wondering who painted it, they can simply upload a photo and ask Bing Chat who the painter is, and Bing Chat can now interpret the image and answer the question by leveraging GPT-4.

Also: This viral AI TikTok trend could score you a free headshot. Here's how

When GPT-4 first launched in March, one of the biggest improvements of the model was its ability to process multimodal prompts, specifically images, and texts.

This Bing Chat integration is the first time GPT-4's multimodal abilities have been integrated into a chatbot since even ChatGPT Plus users do not have access to this feature yet.

Microsoft suggested other fun uses for this feature including uploading a photo of the contents in your fridge and requesting a lunch idea or asking about the architecture of buildings in a new city.

Also: Bing Chat's enterprise solution is here. This is what it can offer your business

The feature is beginning to roll out to users today on the desktop and the mobile app. The feature already rolled out to me, so I put it to the test.

I uploaded an image of my puppy and asked Bing Chat what breed he was. Within seconds Bing Chat released a response that was accurate, as seen in the image at the top of the article.

This release follows Google's addition of image inputs to its Google Bard chatbot last week through an integration with Google Lens.

Artificial Intelligence