At Google I/O event, the company unveiled PaLM 2, a next-generation language model. PaLM 2 represents a significant improvement over its predecessor, PaLM, and introduces several new capabilities that set it apart from OpenAI’s GPT-4.
One of the key advantages of PaLM 2 is its availability in smaller sizes, such as Gecko, Otter, Bison, and Unicorn, which are specifically optimised for applications with limited processing power. These smaller models enable PaLM 2 to cater to a wider range of devices and products, including mobile devices that can run the lightweight Gecko model even offline. This flexibility in model sizes gives PaLM 2 an edge in terms of accessibility and deployment.
Google claims that PaLM 2 demonstrates enhanced reasoning capabilities compared to GPT-4, particularly in tasks like WinoGrande and DROP, with a slight advantage in ARC-C as well. However, it’s important to note that direct comparisons between the two models can be challenging due to differences in the presentation of test results. Additionally, Google has chosen to omit some comparisons where PaLM 2 performed less favourably, raising questions about the completeness of the assessment.
In terms of mathematical abilities, PaLM 2 shows improvements according to Google’s research paper. While the exact size of PaLM 2’s largest model, PaLM 2-L, remains undisclosed, Google has stated that it is significantly smaller than PaLM’s 540 billion parameters. This suggests that PaLM 2-L is likely smaller than GPT-3.5, but it still competes well with GPT-4, delivering impressive performance in various tasks.
Bard’s new features also make it the better choice for research. It provides more concise summaries and improved sourcing. Users can now quickly access the core information of a topic and easily identify which parts of the response match specific sources by clicking on number tags to the corresponding sections in the linked sources. This helps when conducting research or writing essays requiring specific knowledge and detailed citations. These updates address the limitations of AI tools in verifying real-world information and enhance Bard’s research capabilities.
No Disclosure
While Google doesn’t disclose the exact size of PaLM 2’s training dataset, the company emphasises a focus on mathematics, logic, reasoning, and science. PaLM 2’s pre-training corpus consists of a diverse range of sources, including web documents, books, code, mathematics, and conversational data. Moreover, PaLM 2 has been trained in over 100 languages, enhancing its contextual understanding and translation capabilities.
In contrast, OpenAI has trained GPT-4 using publicly available data and licensed data. GPT-4 aims to generate a wide range of responses and has been fine-tuned using reinforcement learning with human feedback, aligning its behaviour with user intent.
Both PaLM 2 and GPT-4 can be accessed through their respective chatbots, Bard and ChatGPT. Bard is freely available worldwide, while ChatGPT Plus, featuring GPT-4, is behind a paywall. However, GPT-4 can also be accessed for free through Microsoft’s Bing AI Chat, which utilises the model. This accessibility plays a role in the potential adoption of PaLM 2, as it is an open-source model.
Google has integrated PaLM 2 into more than 25 of its products, including Android and YouTube, while Microsoft has also incorporated AI features into its Office suite and various services. Although GPT-4 has gained traction among developers and startups due to its early release and refinement, the open-source nature of PaLM 2 may attract a wider range of users.
As PaLM 2 is a relatively new model, its ability to compete with GPT-4 is still being assessed. Google’s ambitious plans and the unique capabilities of PaLM 2 suggest that it could present a formidable challenge to GPT-4. However, GPT-4 remains a capable model, outperforming PaLM 2 in several comparisons. Nevertheless, PaLM 2’s smaller models, especially lightweight options like Gecko, give it an advantage, particularly for mobile devices.
With the introduction of PaLM 2 and Google’s ongoing development of the multimodal AI model Gemini, the competition for AI dominance has intensified. Google’s commitment to advancing AI technologies indicates a continued drive to innovate and challenge established players like GPT-4. The future will reveal how these language models evolve and how they shape the landscape of natural language processing and AI as a whole.
The post LLM Wars: PaLM 2 VS GPT-4 appeared first on Analytics India Magazine.
If reports are to be believed, OpenAI is preparing to release its open source AI model. However, chances are that the new model will not compete with its flagship GPT. But, why exactly is OpenAI thinking of going back to being ‘open’?
Going by how closed OpenAI has been about its flagship GPT, the announcement of creating an open-source model will allow it to enter the open source race, a segment it had been evading. Open source allows continuous innovation. By allowing people to work and provide feedback on codes, organisations are leveraging their potential using crowd intelligence and open contributors, which enables technology to grow exponentially. Open source encourages a two-way system that benefits the makers and users.
OpenAI has been cautious in planning to offer a new model which might have nothing to do with their GPT models thereby continuing to protect its well-kept secret. However, this is not the first time an attempt to reach the open source community has been made by OpenAI, after going the profit-seeking way. Last month, the company announced a Bug Bounty Program that invited researchers to report vulnerabilities, bugs or security flaws in their systems, and rewards ranging from $200 to up to $20,000 were offered.
Is Open Source the Way?
While OpenAI remained closed source for many years, a number of companies such as Google (PaLM), Meta (LLaMA)and Hugging Face have kept their models open sourced. Companies adopt an open source approach with the belief of getting contributions from outside developers.
Meta’s LLaMA has been in news among researchers for open sourcing their model. Their latest model’s weights were also available to all on a case-by-case basis. With the code being leaked on GitHub, people were able to access GPT-level LLM for free.
With models such as Alpaca and Vicuna released by Stanford and a collaboration of universities such as UC Berkeley, UC San Diego and others, LLMs have been democratized. With these open source models, LLMs are being openly accessed by developers to further add to their functionality. Google is also not far behind with their open sourced model PaLM which is also open source.
Benefits and threat of open source
Recently, in an internal document, Google speaks about how open source is a big threat for the company. It emphasises on how the modern internet “runs on open source for a reason” and compares the recent open-source-based AI developments to the Internet. The open source community had a huge role to play during the Internet tech revolution which cannot be replicated by any organization.
According to the document shared internally by a senior software engineer at Google, Luke Sernau spoke about how open source technology has ‘quietly advanced’ and outshined OpenAI and Google in the AI race. It has been emphasized that open-source communities pose a threat to Google and Big Tech companies as people are able to build models that rival them at a quicker pace and lower cost. It can also be detrimental when people will not be ready to pay for high-quality technology when it is freely available in open-source communities.
Back-and-Forth
OpenAI has been on the receiving end of criticism for keeping their models closed source. The company that started as a nonprofit research organization that was committed towards developing digital intelligence became a closed source model with a “capped profit” model. When asked about the change in approach, co-founder Ilya Sutskever once said that they “were wrong” and that if AI or AGI becomes unbelievably potent at some point, then “it doesn’t make sense to be open-source.” He also expects that in the next few years, everyone will realize that “open sourcing AI is not wise.”
Elon Musk, one of the initial founding members of OpenAI, had expressed his concerns on how the company was initially started as an open source counterbalance to Google, but has now become a closed-source “maximum-profit” company which is controlled by Microsoft, something that he had not intended it to be.
Today, with the anticipated news of OpenAI planning to release an open source model, it looks like the company will be going back to their initial goal in a possible bid to cater to the open-source community.
It’s ironic that OpenAI uses open source to train and scale their GPT models, but their codes are closed for the community. In March, OpenAI faced a security breach when a user’s search history was exposed, and the issue was attributed to a bug in the Redis client library, indicating how the company still uses open source code.
OpenAI’s path today is tricky and laden with mysteries. On one hand, they are going about monetizing whatever they can through special subscriptions such as ChatGPT Plus and ChatGPT Business, and on the other hand, they are planning to open source. They recently announced 70+ ChatGPT Plugins along with a web browsing feature, where ChatGPT will know when and how to browse the internet for recent topics and events, which will be released to all ChatGPT Plus users over the coming days. Going by how OpenAI is trying hard to stay ahead of the AI race, the open model format is probably the only factor holding it back. It is to be seen what their open source model will offer.
The post Has OpenAI Lost the Open Race? appeared first on Analytics India Magazine.
Microsoft today released a Bible for language models, called Guidance. This new tool allows users to control language models more effectively and efficiently than traditional prompting or chaining. The language model could also include open-source models like Vicuna, besides GPT-4 and others.
The company said that the Guidance program allows users to interleave generation, prompting, and logical control into a single continuous flow, matching how the language model actually processes the text. Looks like the tech giant is scripting a new religion in the name of LLM. It even quoted GPT Proverb11:14 – which says, “Where there is no guidance, a model fails, but in an abundance of instructions, there is safety.”
Check out the GitHub repository here.
Microsoft said that simple output structures like the chain of thought (COT) and its many variants, including ART, Auto-COT, etc., have been shown to improve the performance of the language model. It said that the advent of more powerful LLMs like GPT-4, allows for an even richer structure, and Guidance makes that structure easier and cheaper.
Guidance Features
The Guidance offers simple and intuitive syntax based on Handlebars templating.
Provides rich output structure with multiple generations, selections, conditionals, tool use, etc.
Users can experience Playground-like streaming in Jupyter/VSCode Notebooks
Offers smart seed-based generation caching.
Supports tole-based chat models (ChatGPT)
This new feature can be easily integrated with HuggingFace models. This includes guidance acceleration, token healing and regex pattern guides.
Earlier this month, researchers from Microsoft unveiled automatic prompt optimisation (APO), a simple and general-purpose framework for the automatic optimisation of LLM prompts. Check out the research paper here.
Last year, at Build 2022, Microsoft announced the launch of open-source tools and datasets designed to audit AI-powered content moderation systems, alongside automatically writing tests that highlight potential bugs in AI models. At the time, the company claimed that the projects, namely AdaTest and (De)ToxiGen, could lead to more reliable language models.
The post Microsoft Releases A Bible for Language Models, Calls it Guidance appeared first on Analytics India Magazine.
Worldwide spending on public cloud services is set to grow 20.7% to total $591.8 billion in 2023, according to Gartner, and threat actors are getting better at exploiting unpatched vulnerabilities.
Recent research by Palo Alto Networks’ Unit 42 found that more than 60% of organizations take over four days to resolve security issues, over 63% of codebases in production have unpatched vulnerabilities, and threat actors exploit a misconfiguration or vulnerability within hours.
The company’s Prisma Cloud is a top security player in spotting vulnerabilities in cloud-native application development and deployment. TechRepublic spoke with Ankur Shah, SVP and general manager of Prisma Cloud, about what cloud security means and how IT pros and decision makers should think beyond the traditional cybersecurity playbook when it comes to cloud security.
TechRepublic: How has hybrid work and migration to cloud business informed what Palo Alto’s Prisma does?
Ankur Shah: Before the cloud, security was like a house with one front door, a camera and a security guard: one level of security and you’re good to go. Now security is very dynamic. Every house looks and feels different. There are windows and doors and you don’t always know which are open, and the crown jewels are inside. So there’s a lot of “lift and shift” [the process of migrating applications and systems to the cloud] with customers rewriting applications — building “houses” in cloud infrastructure, and the security person at IT does not have as much control over how these houses get built.
TechRepublic: Developers do, nowadays.
Ankur Shah: … Because every company is becoming a digital company. If I’m Home Depot, I am a technology company that happens to be in home hardware; if I’m Pfizer, I’m a technology company that happens to be doing pharmaceuticals: today people are using AWS or another cloud service provider and developing their own software. So, yes, developers can have outsized influence because they have to build fast. Today there are over 33 million developers and fewer than three million security people who actually know the cloud. I don’t have data for this one, but I would guess that there are probably fewer than 20,000 people in the world who really understand cloud and security.
TechRepublic: But isn’t cloud security pretty much what most security is about now?
Ankur Shah: You have to understand that the bulk of the security professionals come out of an understanding of network and endpoint security. A lot of security people are using the same playbook that we used back in the day and applying it in the cloud. It’s a very different paradigm now, though. The way workloads get deployed in the public cloud — the windows and doors of the house — is very dynamic. You don’t rack and stack a server anymore. You click a button … or you don’t even have to click a button. Through automation, you can create literally hundreds of thousands of workloads in the cloud today. So these are the best of times, these are the worst of times if you’re in security.
TechRepublic: Should cloud providers be doing more in terms of securing what enterprises enact in cloud environments?
Ankur Shah: If you look at AWS, Azure, Google Cloud, IBM, Oracle and the others … you can have one cloud provider alone with over 200 cloud services that developers are using to build new applications. The cloud providers say, “Look, I will secure the infrastructure layer, but what you put in your applications, I don’t have responsibility, that’s up to you.” When I was a developer, we would ship that code once a year. Now customers are shipping code daily. So the CI/CD [continuous integration/continuous deployment] pipeline has reduced significantly now.
TechRepublic: Palo Alto Prisma Cloud is about securing that entire CI/CD process, correct?
Ankur Shah: The entire code-to-cloud journey … often involves 7, 8, 9 tools. The left doesn’t talk to the right, right doesn’t talk to the middle, middle doesn’t talk to the right. So, yes, Prisma Cloud’s mission has been to deliver code-to-cloud security at each stage of the pipeline. There will be security problems once things are in production. Continuously monitoring the final product to ensure that security holes are not left is also a big part of what we do.
TechRepublic: Even with code-to-cloud security there will still be exploitable critical vulnerabilities, don’t you need multiple tools to deal with this in development and production?
Ankur Shah: Well, there are two ways to not solve that problem. One is if you have multiple tools that aren’t integrated, which is what much of the security industry is today. There are 3,000 different vendors, 200 in cloud security alone. And everybody’s trying to sell point solutions. It’s not going to save the day for you. More tools make you less secure, not more.
TechRepublic: Which I assume is why enterprises are moving away from collecting point solutions toward platforms like extended detection and response, or XDR, in Security Operations Center contexts.
Ankur Shah: There is a big consolidation movement because customers can’t keep on repeating the sins of the past and have multiple tools, point products, but in security, good enough is not good enough. You have to be best in class.
TechRepublic: Is DevSecOps fundamentally different than what is happening in the world of SOCs and does Prisma Cloud respond to both contexts?
Ankur Shah: Tools like XDR for SOC are out there for doing threat detection prevention. If you have software already in production and an intruder gets in, Prisma Cloud will detect it and we will send those signals to the SOC. From the code to the cloud process, there are risk signals, and Prisma’s job is to prevent those problems to begin with.
TechRepublic: What are some uses of large language models in cloud security?
Ankur Shah: My vision is to leverage AI for two purposes: to improve the user experience and to improve the security outcomes. It’s really that simple. Customers today are asking simple questions, but to answer those questions we often have pages and pages of product information. With AI, why can’t you ask something like, “Hey, what’s my top security priority? What’s the next incident that I can expect?” In the future of security, users are going to be engaging with AI to help solve problems for these kinds of queries. That speaks to the user experience aspect of it. The security outcome is a lot of the stuff that we did already in AI. You can expect us to do more and more in the future with automation, more AI and machine learning because it’s really connecting the dots to ensure that if there is a breach — if there is a security incident — we’re able to detect it sooner than later.
Cybersecurity Insider Newsletter
Strengthen your organization's IT security defenses by keeping abreast of the latest cybersecurity news, solutions, and best practices.
For anything to thrive and sustain, ethics are a must; technology innovation is no exception!
Businesses, in general, have incredible growth potential owing to recent AI advancements such as generative AI, but this also entails a lot of responsibility. Since technology directly affects people’s lives, there is a lot of emphasis on AI ethics, data governance, trust, and legality.
Businesses must be aware of new and impending regulations as well as the procedures they must undertake to ensure their organizations are data-responsible and compliant when they begin to scale up their AI usage to reap business benefits. Responsible AI can help with this goal.
So what exactly is ‘Responsible AI’?
Responsible AI is the practice of designing, developing, and deploying AI with the purpose of having a fair influence on consumers and society, allowing businesses to build trust and confidently scale AI. This refers to a framework with predefined principles, ethics, and rules to govern AI.
Responsible AI, also called ethical or trustworthy AI, ensures that any AI system operates according to moral principles, complies with rules and regulations, and reduces the risk of reputational and financial harm. As opposed to being an expensive burden or just a risk-avoidance mechanism, it actually is an enabler of technology. Businesses that adhere to these criteria are typically rewarded with more accurate AI models, less waste in their deployment, and, overall, more sustainable benefits.
While implementing and deploying AI, one must abide by certain general principles. Additionally, it’s equally critical to note how these principles are put into practice in a manner that fosters the development of a responsible AI ecosystem.
Framework of Responsible AI
The proposed framework of Responsible AI has two facets:
Conscientiousness – Thought level
Solid Governance- Execution level
Each pillar of Responsible AI should be dealt with conscientiousness and solid governance, ensuring adherence at the thought as well as execution level. Let’s look at the two facets in detail:
Conscientiousness (Thought Level): The goals of artificial intelligence (AI) must be humanistic. The developers and users of AI must demonstrate responsibility at thought level showing a strict regard for doing the job well and thoroughly. The conscientious approach implies painstaking efforts to follow one’s conscience and an active moral sense governing all actions of an individual/institution as one implements each of the ten pillars of responsible AI listed in this article.
Solid Governance (Execution Level): Even the best-designed model could produce undesirable and unanticipated behavior if there is weak governance. The architecture that oversees the design, creation, and use of a machine learning model is known as governance. All pillars of responsible AI should be under the jurisdiction of a solid governance system. A clear structure of administration and control must be in place, even though the specifics of effective governance vary from model to model based on the application and intended use.
Key Pillars of Responsible AI:
● Inclusiveness (Non-bias): The principle of non-discrimination ensures that a qualified person shouldn’t be denied an opportunity by AI systems solely because of their identity. In terms of education, employment, access to public areas, and other issues, it should not further the damaging historical and social divisions based on religion, race, caste, sex, descent, place of birth, or domicile. Additionally, it should try to prevent discrimination by identifying affected stakeholders, determining attributes for inclusion or exclusion, fostering the creation of a diverse AI workforce, and testing the AI model with diverse users.
● Transparency: It implies the requirement to demonstrate and document the methods by which AI systems are developed along with their strengths and limitations. A simple and understandable example of AI is risk or fraud models. In a transparent scenario, you would have visibility around the source and features of training data as well as the development approach of the underlying algorithms along with their shortcomings. Long-term disaster avoidance, along with the realisation of AI’s potential for good, depend on open and transparent AI.
● Explainability: The inner working of an AI model, along with its potential biases and expected effects, are all described in terms of explainable AI. When putting AI models into production, a business must first establish trust and confidence. Humans find it difficult to understand and retrace how an algorithm arrived at a result as AI advances. Explainability seeks to enable users to get explanations for decisions impacting them in simple and intuitive language. For example, explainable artificial intelligence can provide explainations relating to an AI system’s decision of selection or rejection of a resume.
● Accountability: There should be a clear allocation of roles and responsibilities throughout the AI life cycle, along with identification of stakeholders accountable for outcomes of the AI system. This essentially calls for individuals/institutions to take responsibility for actions of AI systems and their consequences, for example, of results from ChatGPT or any credit risk model. Accountability entails making developers and vendors aware of adherence to responsible AI principles and compliance with existing standards and regulations. Robust control over AI processes, including humans in the loop, and timely feedback from all stakeholders and users aid accountability.
● Privacy: Like other forms of technology, AI systems should be able to defend against threats and safeguard sensitive data. If privacy concerns are not considered, one could run the danger of being quickly identified and having their data compromised. Privacy entails getting user consent before storing and using personal data, safeguarding and avoiding repurposing of collected personal data, transparency of data access and usage, and implementing all necessary privacy controls. For example, in the healthcare sector, companies must do pre-work on data, such as anonymization and de-identification, when using patient data for AI purposes to comply with HIPAA regulations.
● Security: To stop hackers from meddling with the system and altering its intended behavior, an AI system’s and its underlying data’s security is essential. It is critical to identify and mitigate system vulnerabilities – put in place strong access controls, secure coding practices, controls against data dwindling, model stealing, and malicious use, as well as ensure adequate security controls when dealing with third parties and open source components. Responsible AI can pave the way for security to be established by ensuring system robustness and security against misuse or adversarial attacks.
● Reliability: Any trustworthy system must be dependable and secure. The same is true for AI systems. Since AI systems permeate the fundamental fabric of human experiences, they need to be dependable. But what does reliability in AI entail? To begin with, it ensures that the AI is reliable in varied situations, much like a trustworthy human who can provide rationally consistent responses to complex problems. In addition, the reliability of AI systems (e.g., for collections) is characterized by the selection of appropriate algorithms, reproducibility of outcomes, monitoring data or model drifts, presence of feedback loops, and quality assurance checks across the AI product lifecycle.
● Safety: Public safety may be significantly impacted by vulnerabilities brought about by the growing usage of AI in crucial areas of society. Organizations must create AI systems that function well and have no or minimal adverse impact. The presence of human supervisory systems and decommissioning in case of system failures are critical to ensuring the safety of AI systems. What will happen if an AI system malfunctions? What actions will the algorithm take in an unexpected situation? If AI addresses every “what if” and reacts to the new circumstance effectively and without endangering users, then it can be said to be safe. For example, a self driving car killing pedestrians as it was not prepared to deal with people in the middle of the road is unsafe and needs to build in appropriate safety mechanisms. Practical and affordable grievance redressal system and compensation mechanisms should also be put in place.
● Compliance: AI systems must comply with all applicable laws, statutory standards, rules, and regulations in all stages of their life cycles. Organizations must build awareness and constantly monitor the state of the AI regulatory environment locally and globally to ensure compliance and avoid reputational or financial losses. Organizations must take precautions to prevent data misuse and only use data with consent. Corporate-wide data and AI compliance, along with associated rules and practices, must be established by organizations. A recommended step is to start auditing, which comprises examining the data design, proposed model, and purpose. Compliance in AI should be proactive on the company’s part, not after the act.
● Alignment with Human Values: The fundamental goal of AI should be the maximisation of human potential in alignment with human values. This entails a critical review of AI use cases, deep diving into anticipated benefits, harms, and overall impact on society. To safely accomplish human objectives and the values that underpin their realization, it is imperative that human values become integrated into or inseparable from the processes in which an AI system learns to make evaluative decisions. The “codes” we feed into AI algorithms should align with human objectives and values.
Promise of Responsible AI
AI has the potential to significantly impact all industries, including financial services, retail, manufacturing, healthcare, logistics and even space exploration. A responsible AI framework enables companies to track and mitigate bias and create transparent and explainable AI models, prevent misuse and adverse effects of AI, determine who to be held responsible if something goes wrong, and ensure compliance with security, privacy, and associated regulations. Prevention of misuse with appropriate usage guidelines and implementation of a continuous feedback loop can go a long way in maximizing positive returns from AI. Responsible AI must, however, overcome several challenges, including access to appropriate data sets, adopting the most suitable data infrastructure and dealing with the black box nature of complex algorithms. Organizations should focus on creating or adopting responsible AI toolkits comprising frameworks, KPIs, best practices, assessments, checklists and relevant technologies.
If used properly, artificial intelligence could revolutionize the game by positively transforming the world and improving the quality of life. While a conscientious approach is crucial to the development of responsible AI, to direct AI in the future, there must be solid governance, suitable legislation, and regulation.
This article is written by a member of the AIM Leaders Council. AIM Leaders Council is an invitation-only forum of senior executives in the Data Science and Analytics industry. To check if you are eligible for a membership, please fill out the form here.
The post Council Post: Towards An Ethical Tech Revolution: Building Responsible AI Practices appeared first on Analytics India Magazine.
Researchers from South Korea recently released DarkBERT, a dark web domain-specific language model based on the RoBERTa architecture.
This new model is said to show promising applicability in future research in the dark web domain and in the cyber security industry. It also outperformed existing language models with evaluations on dark web domain tasks and datasets.
But, how did they do it? To allow DarkBERT to adapt well to the language used in the Dark Web, the researchers have pre-trained the model on a large-scale dark web corpus collected by crawling the Tor network. In addition to this, they also polished the pre-training corpus through data filtering and deduplication, alongside data pre-processing to address the potential ethical concerns in Dark Web texts related to sensitive information.
Showcasing DarkBERT pretraining process and the various use case scenarios for evaluation. (Source: arXiv)
The same group of researchers last year worked on ‘Shedding New Light on the Langauge of the Dark Web,’ where they introduced CoDA, a dark web text corpus collected from various onion services divided into topical categories. Another notable study includes – ‘The Language of Legal and Illegal Activity on the Darknet,’ done by Israeli researchers, where they identified several distinguishing factors between legal and illegal texts, taking a variety of approaches. This includes predictive (text classification), and application-based (named entity Wikification), along with an approach based on raw statistics.
All of this research work and more has inspired the researchers to develop DarkBERT.
What next?
In the coming months, the researchers said that they plan to improve the performance of dark web domain-specific pre-trained language models using more latest architectures and crawl additional data to allow the construction of multilingual language models.
The post A New Language Model Trained on Dark Web Emerges appeared first on Analytics India Magazine.
At this year’s Google I/O conference, the company announced significant upgrades to Google Bard, which is now available in over 180 countries. Google developed Bard as its own AI chatbot to compete against OpenAI’s ChatGPT and Microsoft’s GPT-4 powered Bing.
Bard is based on PaLM 2, an advanced AI model that Google announced in February. One of its PaLM 2 models is lightweight enough to work on smartphones, CEO Sundar Pichai claimed.
Google Cloud CEO Thomas Kurian informed Reuters that the division is securing customers like Deutsche Bank AG and Uber Technologies Inc for testing purposes, as they evaluate and assess the effectiveness of Google’s latest technology.
Bard too has garnered interest among tech enthusiasts since its launch and is seen as a strong competitor to ChatGPT. Google has added several new features to enhance user experience and overcome the limitations of ChatGPT.
We’ve listed eight Bard features that ChatGPT doesn’t have, as of now.
Access to the web
One notable advantage of Bard over ChatGPT is its access to the internet. ChatGPT does not have direct internet access and can access the web only through plugins on its paid version—ChatGPT Plus.
Bard can provide comprehensive and informative answers by leveraging the power of the internet. It can give real-time information, fetch the top news, and answer questions with the most up-to-date data. However, it should be noted that Bard is still in the experimental phase, and there may be instances where the information is inaccurate or offensive.
Meanwhile, OpenAI CEO Sam Altman had tweeted expressing their intention to add internet plugins and code execution plugins.
Image generation
Bard also surpasses the paid and unpaid versions of ChatGPT when it comes to generating images as response. Google, at the event, announced that they’ll provide AI image generation capabilities through integration with Adobe Firefly. This feature enhances the visual aspect of the conversation and allows users to obtain more contextually rich information.
Voice Prompts
Bard outperforms ChatGPT when it comes to voice prompts as well, providing users with the ability to interact through voice input. This offers a convenient way to obtain responses while multitasking or when typing is not feasible. This voice interaction capability gives Bard an edge over its competitors.
Coding Capabilities
Bard overshadows ChatGPT and offers a strong support for coding with its ability to assist in over 20 programming languages including C++, Python, Java, TypeScript, JavaScript, etc. It can help professionals with code generation, explanation, and debugging. In comparison, while ChatGPT does have coding capabilities, it falls short when it comes to additional tasks, which OpenAI’s Codex may be better suited to perform.
ChatGPT is focused on natural language but it has also been trained on coding languages like Python, JavaScript, C++, C#, Java, Ruby, PHP, Go, Swift, TypeScript, SQL and Shell.
Gmail Integration
The integration of Bard with Gmail is another significant advantage. With over 2 billion users, Gmail is widely used for communication. Having access to an AI chatbot like Bard within the email service opens up new possibilities for email interactions and can enhance the experience.
However, ChatGPT is being added to Microsoft work software, Microsoft 365, and will be embedded into Word, Excel, PowerPoint, and its Gmail-equivalent Outlook. The OpenAI chatbot can perform this through plugins.
Export Responses
Bard also offers the functionality to export results to Gmail and Docs instantly. Users can easily share the generated content with friends and colleagues by exporting it directly to these platforms. This feature streamlines the process of sharing information and makes composing emails hassle-free.
On the other hand, OpenAI has released a similar export option in settings—where users get to export their account details and conversations which will be sent to your registered email in a downloadable file, but it says that processing may take some time.
Image Prompts
One standout feature of Bard is its ability to use images as prompts. Users can simply click a picture or scan an image using Google Lens and ask Bard for assistance. For instance, a user can lookup similar holiday destinations as portrayed in another image and can also ask about its history and significance. This feature opens up new possibilities for interaction and prompt generation in AI chatbots.
Similarly, GPT-4 also claims to be a large multimodal model which accepts image and text inputs, to emit text outputs but the capability hasn’t been introduced even in the paid version as of the date of publishing this article.
Webpage Summarisation
Bard has the advantage of internet connectivity, allowing it to summarise web pages by simply sharing the link. In contrast, ChatGPT lacks internet connectivity, requiring users to manually copy and paste the content they want to summarise.
However, Bard has its limitations, particularly in terms of toxicity. During a test, the model produced toxic responses more than 30% of the time when given explicitly toxic prompts. Additionally, in languages like English, German, and Portuguese, PaLM 2 tended to exhibit more obvious toxic behaviour overall.
Although designed to compete with OpenAI’s GPT-4, it is challenging to directly compare the two models due to their different architectures and testing methodologies. In reasoning tasks, Google’s PaLM 2 performed similarly to or better than GPT-4. However, in coding tasks, PaLM 2 required multiple attempts and additional coding tokens to achieve good performance.
The post 8 Things That Bard Can Do, But ChatGPT Can’t appeared first on Analytics India Magazine.
What is SaaSification? Software as a Service (SaaS) is a model by which customers pay for utilization of a service rather than buying a license. SaaSification refers to the conversion to this model. However, more broadly it refers to a model by which the units of a company are turned into services and provided via software-style APIs.
This mode of interaction is not completely new. Credit card companies, finance and logistics providers such as FedEx and UPS have long provided APIs that allow customers to do everything from shipping to tracking items. While the actual “service” is provided by people with trucks, they are driven by APIs.
Companies also are transforming their internal operations as “services” provided to other units. This serves both as an accounting or operating model and as a way to provide more flexibility to the business. Services can be composed as business workflows via tools for business process management (BPM). SaaSification means that essentially APIs take the place of mail, email, telephone and other communication functions.
What is in it for me?
SaaSified companies reap several benefits. First, they become more flexible. Consider fulfillment as a software service. So long as the product, the amount to be paid, and the destination are the same, a company can easily integrate a vendor’s fulfillment services into its existing process. Moreover, it can potentially provide that service to another firm. Who or how something is done is encapsulated, and by changing a business process workflow or rule, change can be driven more quickly.
SaaSified companies should be more transparent overall. Because the internal functions of the company are provided as services and accounting and performance integrators are built into each part of the whole, it is easier to find inefficiencies and bottlenecks. Moreover, third-party vendors can potentially compete with the firm’s internal services and provide the same kinds of metrics. This can form a kind of internal marketplace where the best services and ideas have a chance to compete in providing the most value for the lowest cost. The metrics built into the system make it easier to experiment.
Saasified companies are also more valuable. Given that integration is a key reason why many mergers and acquisitions fail, having easily integrated and recomposed business units and software functions will add to the company’s overall value to a potential buyer or investors. And, combined SaaSified companies can simply adapt their overall business process workflow and pick between the best internal service providers. By making integration seamless from the start, not only are acquisitions easier but potential buyers receive risk mitigation and may be willing to pay more.
How do I get there?
Begin by making your internal systems into services. This means they run independently, have a REST API and service level agreement, and capture key metrics, including cost. Ensure that the business is aligned with those services. All uses of that business function go through that service. Most or all authorization to perform that service is obtained upfront to get access to the service. This means, if I have access to internal.your.co/ship/createShipment, then I have access to ship things. Compose workflows using a tool like business process management software. The road to SaaSification is started by making your business a series of software services and workflows.
Change the culture from phone calls, emails and informal processes to align with services. Change how rules and processes are obeyed via software. Evaluate company performance at the service level. If more visibility is needed then the service is too granular. Evaluate whether services are meeting expected KPIs including costs, and set goals at the service level. It is the API that is the gateway to a function.
Evaluate third-party alternatives and services based on cost compared to internal services. Is the third-party service compatible but cheaper? Can they really ship from your warehouse for half the price, for example, or are there internal costs such as loading dock workers that have to be decomposed and considered in terms of performance and cost?
Look at your technology and whether it scales. Is it robust and can it scale business services both out and back when business is less robust. This includes moving to service versions of core technologies like security as well as databases. Consider upgrading legacy relational databases to distributed SQL databases which scale out, provide greater reliability and can scale back without data loss.
Upgrading to a services model is a combination of culture, technology and how the business is managed. It cannot happen overnight but the great thing about SaaSification is that it can happen a service or two at a time. Core to this is moving to communicating via APIs and upgrading to scalable technologies.
What organizational changes are needed?
Business organizations should also be aligned with services. This includes IT functions. There may be foundational core technologies and expertise required to run services, but generally application teams, and at least some administrative staff should be aligned with the services themselves, at least for core services (e.g., logistics). Reorganization and process changes should happen at the software level which drives different human actions.
By moving to a SaaSified company you can provide better and more services to the market, consume cheaper and better services than you can produce internally, increase flexibility as change happens and build a more valuable company. These changes are more than just vendors providing services that used to be internal. They are shaping how internal services should happen and how the organization should communicate both internally and externally. As new technologies and business practices become important they can be readily integrated into or on top of an existing service layer.
When Google artificial intelligence scientists revealed a significant new program — the Pathways Language Model (PaLM) — a year ago, they spent several hundred words in a technical paper describing the significant new AI techniques used to achieve the program's results.
Also:How to use ChatGPT: Everything you need to know
Introducing the successor to PaLM last week, PaLM 2, Google revealed almost nothing. In a single table entry tucked into an appendix at the back of the 92-page "Technical Report", Google scholars describe very briefly how, this time around, they won't be telling the world anything:
PaLM-2 is a new state-of-the-art language model. We have small, medium, and large variants that use stacked layers based on the Transformer architecture, with varying parameters depending on model size. Further details of model size and architecture are withheld from external publication.
The deliberate refusal to disclose the so-called architecture of PaLM 2 — the way the program is constructed — is at variance not only with the prior PaLM paper but is a distinct pivot from the entire history of AI publishing, which has been mostly based on open-source software code, and which has customarily included substantial details about program architecture.
Also: Every major AI feature announced at Google I/O 2023
The pivot is clearly a response to one of Google's biggest competitors, OpenAI, which stunned the research community in April when it refused to disclose details of its latest "generative AI" program, GPT-4. Distinguished scholars of AI warned the surprising choice by OpenAI could have a chilling effect on disclosure industry-wide, and the PaLM 2 paper is the first big sign they could be right.
(There is also a blog post summarizing the new elements of PaLM 2, but without technical detail.)
PaLM 2, like GPT-4, is a generative AI program that can produce clusters of text in response to prompts, allowing it to perform a number of tasks such as question answering and software coding.
Like OpenAI, Google is reversing course on decades of open publishing in AI research. It was a Google research paper in 2017, "Attention is all you need," that revealed in intimate detail a breakthrough program called The Transformer. That program was swiftly adopted by much of the AI research community, and by industry, to develop natural language processing programs.
Also: The best AI art generators to try
Among those offshoots is the ChatGPT program unveiled in the fall by OpenAI, the program that sparked global excitement over ChatGPT.
None of the authors of that original paper, including Ashish Vaswani, are listed among the PaLM 2 authors.
In a sense, then, by disclosing in its single paragraph that PaLM 2 is a descendent of The Transformer, and refusing to disclose anything else, the company's researchers are making clear both their contribution to the field and their intent to end that tradition of sharing breakthrough research.
The rest of the paper focuses on background about the training data used, and benchmark scores by which the program shines.
This material does offer a key insight, picking up on the research literature on AI: There is an ideal balance between the amount of data with which a machine learning program is trained and the size of the program.
Also: This new technology could blow away GPT-4 and everything like it
The authors were able to put the PaLM 2 program on a diet by finding the right balance of the program's size relative to the amount of training data, so that the program itself is far smaller than the original PaLM program, they write. That seems significant, given that the trend of AI has been in the opposite direction of late, to greater and greater scale.
As the authors write,
The largest model in the PaLM 2 family, PaLM 2-L, is significantly smaller than the largest PaLM model but uses more training compute. Our evaluation results show that PaLM 2 models significantly outperform PaLM on a variety of tasks, including natural language generation, translation, and reasoning. These results suggest that model scaling is not the only way to improve performance. Instead, performance can be unlocked by meticulous data selection and efficient architecture/objectives. Moreover, a smaller but higher quality model significantly improves inference efficiency, reduces serving cost, and enables the model's downstream application for more applications and users.
There is a sweet spot, the PaLM 2 authors are saying, between the balance of program size and training data amount. The PaLM 2 programs compared to PaLM show marked improvement in accuracy on benchmark tests, as the authors outline in a single table:
In that way, they are building on observations of the past two years of practical research in the scale of AI programs.
For example, a widely cited work by Jordan Hoffman and colleagues last year at Google's DeepMind coined what's come to be known as the Chinchilla rule of thumb, which is the formula for how to balance the amount of training data and the size of the program.
Also: Generative AI brings new risks to everyone. Here's how you can stay safe
The PaLM 2 scientists come up with slightly different numbers from Hoffman and team, but it validates what that paper had said. They show their results head-to-head with the Chinchilla work in a single table of scaling:
That insight is in keeping with efforts by young companies such as Snorkel, a three-year-old AI startup based in San Francisco, which in November unveiled tools for labeling training data. The premise of Snorkel is that better curation of data can reduce some of the compute that needs to happen.
This focus on a sweet spot is a bit of a departure from the original PaLM. With that model, Google emphasized the scale of training the program, noting it was "the largest TPU-based system configuration used for training to date," referring to Google's TPU computer chips.
Also: These 4 popular Microsoft apps are getting a big AI boost
No such boasts are made this time around. As little as is revealed in the new PaLM 2 work, you could say it does confirm the trend away from size for the sake of size, and toward a more thoughtful treatment of scale and ability.
German multinational software firm, SAP, yesterday announced that it will be collaborating with Microsoft to streamline recruiting and employee development processes.
SAP to integrate OpenAI models via Azure: https://t.co/CGiZaOXiqG
— Greg Brockman (@gdb) May 15, 2023
SAP’s chief Christian Klein said that they are excited about the opportunities generative Ai unfolds for their industry and their customers.
Streamlining Recruitment
With this partnership, the duo – Microsoft and SAP – looks to streamline recruiting and employee development processes. But how?
SAP looks to leverage Azure OpenAI Service API and data from SAP SuccessFactors solutions to create targeted job descriptions.
The company said that with the integration between SAP SuccessFactors Recruiting solution and Microsoft 365, business leaders will be able to fine-tune job descriptions using Copilot in Microsoft Word with additional content and checks to detect bias. Further, it stated that the final job descriptions will then be published in SAP SuccessFactors solutions to complete the workflow seamlessly.
In addition to this, SAP will be using the Azure OpenAI Service API to offer prompts to interviewers within Microsoft Teams with suggested questions based on a candidate’s resume, job description, and similar jobs.
Enabling Employee Learning
Besides streamlining the recruitment process, the company looks to integrate SAP SuccessFactors solutions and Microsoft Viva Learning, which will enable employees to use Copilot in Viva Learning to conduct natural language queries to create personalised learning suggestions/recommendations based on data and learning courses in SAP SuccessFactors solutions that align with the employee’s L&D goals.
SAP said that this enhancement has been built on an already robust integration with content, assignment, permissions and SSO sync, which is already available now. In the coming months, the company looks to give customers access to its automated admin setup experience.
Partners with Google Cloud
Recently, SAP also announced its plans to expand its partnership with Google Cloud to help them build the future of open data and AI for enterprises. With this, the duo looks to launch an open data offering combining the capabilities of the SAP data sphere solution and Google’s data cloud. Read: Google, SAP Unveil Data Cloud, the Next Big Thing in Business Intelligence
The post SAP Integrates OpenAI Models to Streamline Recruitment, Employee Learning appeared first on Analytics India Magazine.
