Persistent Systems has announced the launch of GenAI Hub, an innovative platform designed to accelerate the creation and deployment of Generative AI (GenAI) applications within enterprises.
This platform seamlessly integrates with an organisation’s existing infrastructure, applications, and data, enabling the rapid development of tailored, industry-specific GenAI solutions.
GenAI Hub supports the adoption of GenAI across various large language models (LLMs) and clouds without provider lock-in.
This platform simplifies the development and management of multiple GenAI models, expediting market readiness through pre-built software components, all while upholding responsible AI principles.
The GenAI Hub is comprised of five major components:
Playground is a no-code tool for domain experts to explore and apply GenAI with LLMs on enterprise data without the need for programming skills. It provides a single uniform interface to LLMs from private providers like Azure OpenAI, AWS Bedrock, and Google Gemini, and open models from Hugging Face like LLaMA2 and Mistral.
Agents Framework provides a versatile architecture for GenAI application development, leveraging libraries like LangChain and LlamaIndex for innovative solutions, including Retrieval Augmented Generation (RAG).
Evaluation Framework uses an “AI to validate AI” approach and can auto-generate ground-truth questions to be verified by a human-in-the-loop. It employs metrics to track application performance and measures any drift and bias that can be addressed.
Gateway serves as a router across LLMs, enabling application compatibility and improving the management of service priorities and load balancing. It also offers detailed insights into token consumption and associated costs.
Custom Model Pipelines facilitate the creation and integration of bespoke LLMs and Small Language Models (SLMs) into the GenAI ecosystem, supporting a streamlined process for data preparation and model fine-tuning suitable for both cloud and on-premises deployments.
The GenAI Hub streamlines the development of use cases for enterprises, offering step-by-step guidance and seamless integration of data in LLMs, enabling the rapid creation of efficient and secure GenAI solutions at scale, whether for end users, customers, or employees.
Praveen Bhadada, Global Business Head – AI, Persistent:
“With the Persistent GenAI Hub, clients can embrace a “GenAI-First” strategy, delivering AI-powered applications and services at scale. They can accelerate innovation while practicing responsible AI, leveraging pre-built accelerators and evaluation frameworks, and optimizing costs with a cross-LLM strategy. The GenAI Hub enables enterprises to streamline operations, enhance customer experiences, and identify new avenues for growth,” Praveen Bhadada, Global Business Head – AI, Persistent, said.
The post Persistent Systems Launches GenAI Hub To Accelerate AI Adoption appeared first on AIM.
Shares of major Asian chipmakers and suppliers to Nvidia rallied sharply on Thursday, riding on the coattails of the US chip giant’s surge to a $3 trillion market capitalisation.
Nvidia’s stock jumped on optimism about the booming demand for AI chips, overtaking Apple to become the world’s second-most valuable company after Microsoft.
Taiwan Semiconductor Manufacturing Co (TSMC), the world’s largest contract chipmaker and a key Nvidia supplier, soared nearly 5% to a record high in Taiwan. China’s biggest chipmaker Semiconductor Manufacturing International Corp (SMIC) climbed 4% in Hong Kong, while Japan’s Tokyo Electron, the country’s most valuable chip firm, advanced 4.3%.
“Nvidia’s valuation milestone underscores the explosive growth potential in AI chips as tech giants race to launch more AI products,” said Rajiv Menon, an analyst at CLSA. “This is having a ripple effect across the semiconductor supply chain, lifting shares of Nvidia’s suppliers and rivals.”
Nvidia’s market value has skyrocketed from $1 trillion to $3 trillion in just over a year, fueled by the rising popularity of generative AI tools like ChatGPT. The company makes the most advanced AI processors currently available, putting it in pole position to benefit from the AI boom.
Sentiment was further boosted by upbeat comments from Dutch semiconductor equipment maker ASML Holding, seen as a bellwether for the chip industry. ASML indicated improving demand from its top customers, mainly TSMC.
Other Nvidia suppliers like Foxconn and Japanese chipmakers Advantest, Renesas Electronics and Disco Corp posted gains of 3-5%. Smaller Chinese players Will Semiconductor and NAURA Technology also ticked higher.
“As AI goes mainstream, we expect a multi-year upcycle in chip demand and technology investments,” Menon said. “While Nvidia is leading the charge, the entire semiconductor ecosystem is poised to benefit from this transformative megatrend.”
However, analysts also cautioned that chip stocks have run up significantly and some consolidation is likely in the near term. Geopolitical risks around US-China tensions and supply chain disruptions also remain potential headwinds for the sector.
Indian IT Also Gains
Indian tech stocks rallied after NVIDIA surpassed Apple to hit a $3 trillion market cap, driven by the AI boom. TCS, Infosys and Tata Communications, which have AI partnerships with NVIDIA, were among the top gainers. TCS and Infosys are each training 50,000 employees on AI technologies with NVIDIA’s support. Tata Communications will leverage NVIDIA’s advanced GH200 Grace Hopper chip to build an AI cloud platform in India.
TCS shares rose 1.1% to ₹3,788, while Infosys gained 1.73% to ₹1,454.95. Tata Communications surged nearly 2% to ₹1,783.45. Smaller firms Rashi Peripherals, the main distributor of NVIDIA’s gaming GPUs in India, climbed 2.4%. Netweb Tech, which will manufacture NVIDIA GPU servers locally, hit a 5% upper circuit.
The post Asian Chip Makers Rally to New Heights as NVIDIA Touches 3 Trillion Valuation appeared first on AIM.
Recently, we spoke about simulation in robotics, and the challenges associated with it. Within a few days, at Computex 2024, NVIDIA chief Jensen Huang spoke with full fervour about how simulation and robotics are going to change everything.
“The next wave of AI is here. Robotics, powered by physical AI, will revolutionise industries,” said Huang.
Jensen Huang’s keynote at Computex 2024. Source: NVIDIA
The Rising Wave of Physical AI
“Physical AIs are models that can understand instructions and autonomously perform complex tasks in the real world,” said Huang, who is extremely optimistic about the extent to which robots will become a part of every industry.
“Everything is going to be robotic,” he said. Huang believes that there would be an entire ecosystem of robots, where all factories will orchestrate robots and those robots will build robotic products. NVIDIA is banking on Omniverse to make this happen.
NVIDIA’s Simulation to Lead the Way
NVIDIA’s Omniverse, a platform designed for real-time 3D design collaboration and simulation, form the basis for digital twins which is crucial for simulation. Digital twins are the virtual replicas of physical objects or systems where robots can be tested to fit into the real world.
Showcasing a wide range of scenarios where robots have been trained on NVIDIA’s Omniverse, Huang spoke about how companies are building robotic warehouses around it.
In digital twins, factory planners optimise floor layout and line configurations and locate optimal camera placements to monitor future operations. Also referred to as a ‘robot gym’, Foxconn developers train and test NVIDIA ISAC AI applications for robotic perception and manipulation in Omniverse digital twins.
Multimodal LLMs have only accelerated the process of robotic training. “Multimodal LLMs are breakthroughs that enable robots to learn, perceive and understand the world around them, and plan how they’ll act,” said Huang.
Clubbing this technique with human demonstrations, robots can acquire the skills needed to interact with the world using gross and fine motor skills.
While simulation may sound like everything, it is not the only route for training robots. Gokul NA, co-founder of Bangalore-based CynLr Robotics, believes that a single approach is not perfect. He believes that when we transition from a simulated assumption to reality, it doesn’t work. It fails entirely because it never learned that; it has learned something else independently.
Simulation is also suited for a certain number of tasks. Tasks such as walking, backflips, and other movements that require robotic balances, work best in a simulated environment. However, for tasks that can be learned through imitation, a simulated environment is not required.
Autonomous Everywhere
“One day, everything that moves will be autonomous,” said Huang. And that is most likely going to be powered by NVIDIA.
Teaching robots to grasp and handle objects is one thing, but navigating environments autonomously and avoiding obstacles or hazards is another capability that Physical AI entails.
Interestingly, Huang had recently mentioned that “Tesla is far ahead in self-driving cars, but every single car, someday will have autonomous capability”.
Earlier last year, NVIDIA and Foxconn had partnered to build AI factories which will help boost EV and autonomous vehicle production. They also partnered with major electronic manufacturers Delta Electronics, Pegatron and Wistron.
Robotics Race Continues
The robotics race is only gaining steam with all the recent advancements. Big tech companies have aggressively invested in robotics companies over the last few years. Figure 01, the humanoid built by deep tech robotics company Figure AI, is backed by some of the biggest players like NVIDIA, Microsoft, Jeff Bezos, and others.
Similarly, 1X Technologies, the robotics company backed by OpenAI, recently released their humanoids demonstrating multiple autonomous tasks back-to-back.
You can now tell EVE to do multiple autonomous tasks back-to-back. Watch a team of EVEs work together to clean up our office. pic.twitter.com/6aU2FDGcNF
— 1X (@1x_tech) May 31, 2024
Going by all these developments, the wave of ‘Physical AI’ is already in play.
The post ‘The Next Wave of AI is Physical AI,’ says Jensen Huang appeared first on AIM.
More than 20% of the web uses Cloudflare for security. Safe web browsing requires stronger encryption, and HTTP is one part of that. To achieve enhanced encryption, you need random numbers to create decryption keys, which have to be so random that they are next to impossible to trace by any machine or algorithm.
The strength of encryption is determined by how random your data is when creating decryption keys. Computers can be predictable, so Cloudflare uses lava lamps to generate random data.
The fluid in the lamp constantly changes its shape, and they are never repeated. This means you get a different shape each time. Cloudflare captures photos of these lava lamps and converts them to numbers, giving them the most random number possible.
The problem they are trying to solve…
Computers are not very good at picking random numbers as every part of the computing works in a structured manner and things can easily be traced back. This is where the problem starts for the encryption.
Sure, there are various tools by which you can make the computer produce numbers, such as using /dev/urandom to generate random numbers in Linux. But theoretically, with enough computational power, any string of random numbers generated by the computer can be traced back, and the file can be decrypted.
So, the core idea is not to rely on computers to generate random numbers but to monitor objects from the real world to achieve the highest randomness possible. Cloudflare uses lava lamps for this purpose.
The company has installed 100s of lava lamps in their head office in San Francisco and constantly monitors these lava lamps. They take pictures of the lava lamps at certain intervals.
The method of generating random numbers from lava lamps is called Lavarand.
After clicking pictures of lava lamps, they convert photos to numbers giving them the most random number sequence. What makes it strong is that even if you take pictures side by side, and there’s a difference of a single picture, the entire number string will be different. Yes, it’s that random!
This unpredictable data is used to create keys to encrypt the traffic that goes through Cloudflare’s network.
Later on, the random data is fed to Cloudflare’s data centers which eventually gets passed down to Linux kernels which seeds the random data to random number generators and you get keys that are super random and next to impossible to trace.
That’s not all, Cloudflare has more ways to generate random numbers. For example, Cloudflare’s London office has a series of pendulums installed, and their movements are mathematically unpredictable.
Here’s how they process the data:
Similarly, they have hanging rainbow mobiles creating colorful patterns on the surrounding walls, generating random data in their Austin office.
Is it Better than Computer-Generated Randomness?
The idea is not how you use the lava lamps but how you create random numbers that are untraceable and using lava lamps is only one way of doing it. And to make Lavarand even more robust and unpredictable, they also add entropy from environmental factors like people walking by, changes in lighting, etc.
I believe the idea of generating random numbers from surroundings is a great way to enhance the encryption.
Why is Everybody Else Not Doing the Same?
The idea is to get the highest randomness possible. For example, AWS Key Management Service (KMS) uses hardware security modules (HSMs) with a hybrid random number generator that combines a NIST-approved DRBG (Deterministic Random Bit Generator) seeded by a hardware true random number generator (TRNG).
Some users on Reddit pointed out that the entire Lavarand setup is just a PR stunt. One person pointed out that the camera capturing lava lamps had enough noise capable of generating random numbers. Moreover, each lava lamp consumes 40 watts/hour, which is quite a lot for generating random numbers, he added.
Using Quantum Random Number Generator (QRNG) chips is a more energy-efficient way to generate random numbers, a Reddit user suggested.
As long as you can generate random numbers (that are truly random), you can protect your network without any worries. Lavarand was Cloudflare’s way of protecting traffic going through their network, so it has to be convincing, right?
The post Cloudflare’s Firewall Gets Lit with Lava Lamps, Lights Up Hacker Defense appeared first on AIM.
Ascendion, a leader in AI-powered software engineering, has inaugurated its GenAI Studio in Chennai, positioning the city at the forefront of generative AI innovation for global clients. This new facility will leverage Chennai’s rich talent pool and favorable business environment to drive significant advancements in AI.
The GenAI Studio is designed for Ascendion engineers and clients to collaborate and innovate, rapidly developing AVA+ GenAI prototypes that address critical business objectives.
“Enterprise leaders are excited about the future of AI, but they need to see real-world results,” said Karthik Krishnamurthy, CEO of Ascendion. “Our new AI studio in Chennai is filled with expert talent, hands-on technology, and inspiration, all designed to excite, provoke, and generate applied GenAI solutions that will drive business forward and positively impact lives all over the world.”
As part of Ascendion’s six global innovation hubs, the Chennai studio will enhance client impact and growth by utilizing the high-quality generative AI talent available in the region. The space provides an interactive, collaborative environment for clients to explore generative AI solutions firsthand, work with engineers to customize real-time solutions, and see immediate value through AVA+ generative AI prototypes.
Ascendion’s generative AI initiatives have already shown impressive results, including deploying AI models 30% faster for a tech giant, accelerating content creation by 40% for a hardware leader, and tripling go-to-market velocity for a Fortune 50 bank. The new studio aims to build on these successes by fostering creativity, learning, and collaboration.
“Our AI-powered platforms ensure transparency, velocity, quality, and productivity, freeing up capital for innovation and providing clients the flexibility to meet modern business needs,” said Prakash Balasubramanian, Executive Vice President of Engineering Solutions at Ascendion.
“Now, at our new GenAI Studio, they can experience GenAI co-engineering as we solution together in real-time.”
Ascendion’s initiatives, including the AVA+ platform, deliver radical transparency and help clients achieve significant productivity gains and commercial impact, with more than 1,500 individuals trained in GenAI to date.
The post Ascendion Expands to India, Launches AI Studio in Chennai appeared first on AIM.
Cisco’s acquisition of Splunk is playing out with a new full-stack observability product, the Unified Observability Experience, Cisco announced at the Cisco Live conference on June 5. As part of the same suite of conference announcements, Cisco also showed off the initial availability of a free trial version of Motific, a generative AI delivery platform made in concert with Mistral AI.
Unified Observability Experience creates easy connections between AppDynamics and Splunk Platform
Cisco acquired data observability and security company Splunk in March 2024. At that time, Splunk Chair and CEO Chuck Robbins, as well as Splunk Executive Vice President and General Manager Gary Steele, wrote in a press release that the company planned to “drive new innovations with a unified data platform.” The Unified Observability Experience announced June 5 follows up on this by creating a full-stack observability offering from both companies’ products. Unified Observability Experience offers:
New single sign-on credentials for shared workflows between Cisco AppDynamics and Splunk.
Context-aware deep linking with which Cisco AppDynamics customers can easily transition to relevant logs in the Splunk Platform.
Full-stack, environment-agnostic connections between a variety of applications, including Cisco AppDynamics on Microsoft Azure.
Other new connections between Cisco and Splunk Platform.
The Unified Observability Experience will be generally available in the third quarter of 2024, wherever Cisco SaaS products are sold.
SEE: In May 2024, Cisco and Splunk began their convergence with new security capabilities for Cisco Hypershield.
Cisco AppDynamics gains AI tools and a Microsoft Azure option
Cisco and Splunk announced new AI tools to assist with decision-making:
An AI assistant for Cisco AppDynamics, generally available in Q3 2024.
Tools for configuring thresholds, managing and optimizing configurations and insights with AI and machine learning in Splunk IT Service Intelligence, generally available now.
Cisco AppDynamics will be hosted as software-as-a-service on Microsoft Azure starting in the third quarter of this year, enabling compliance with industry-specific regulatory requirements across new regions. For now, Cisco said, AppDynamics will be expanded to Canada.
Cisco’s incubator Outshift teams up with Mistral AI
Cisco’s incubation arm, Outshift, announced its generative AI deployment tool Motific is now available as a free trial in advance of the general availability of the production version by July 31. Motific helps IT and business function teams to provision generative AI assistants and applications. This is a hot industry, with companies selling easier ways for businesses to adopt equally trendy generative AI solutions.
In addition, Outshift has shaken hands with Mistral AI to offer Mistral AI’s large language model as a preset model in Motific.
“We are pleased to launch this collaboration with Cisco, which will bring new opportunities for GenAI assistants through Motific,” said Mistral AI CEO and co-founder Arthur Mensch in a press release. “We are looking forward to see(ing) our models enhancing innovation and optimizing operations for Cisco’s clients, all while ensuring compliance with their organization’s trust, security, and cost policies.”
Going forward, Cisco and Outshift plan to use Mistral AI products for upcoming guardrails, intelligence and retrieval-augmented generation capabilities in Motific. Outshift will explore using Mistral AI to build domain and task-specific models in Outshift’s products and features.
Mistral AI is among the recipients of Cisco’s recently announced $1 billion AI global investment fund.
French AI startup Mistral is introducing new AI model customization options, including paid plans, to let developers and enterprises fine-tune its generative models for particular use cases.
The first is self-service. Mistral has released a software development kit (SDK), Mistral-Finetune, for fine-tuning its models on workstations, servers and small datacenter nodes.
In the readme for the SDK’s GitHub repository, Mistral notes that the SDK is optimized for multi-GPU setups but can scale down to a single Nvidia A100 or H100 GPU for fine-tuning smaller models like Mistral 7B. Fine-tuning on a data set such as UltraChat, a collection of 1.4 million dialogs with OpenAI’s ChatGPT, takes around half an hour using Mistral-Finetune across eight H100s, Mistral says.
For developers and companies that prefer a more managed solution, there’s Mistral’s newly launched fine-tuning services available through the company’s API. Compatible with two of Mistral’s models for now, Mistral Small and the aforementioned Mistral 7B, Mistral says that the fine-tuning services will gain support for more of its models in the coming weeks.
Lastly, Mistral is debuting custom training services, currently only available to select customers, to fine-tune any Mistral model for an organization’s apps using their data. “This approach enables the creation of highly specialized and optimized models for their specific domain,” the company explains in a post on its official blog.
Mistral, which my colleague Ingrid Lunden recently reported is seeking to raise around $600 million at a $6 billion valuation from investors including DST, General Catalyst and Lightspeed Venture Partners, is no doubt looking to grow revenue as it faces considerable — and growing — competition in the generative AI space.
Since Mistral unveiled its first generative model in September 2023, it’s released several more, including a code-generating model, and rolled out paid APIs. But it hasn’t disclosed how many users it has, nor what its revenues are looking like.
Narendra Sen, the CEO of NeevCloud, is a man of ambition. He envisions constructing an AI cloud infrastructure tailored for Indian clients, comprising 40,000 graphics processing units (GPUs) by 2026. This infrastructure aims to support Indian enterprises with training, inference, and other AI workloads.
Yotta, another prominent AI cloud provider, has recently gained attention for its recognition as the inaugural NVIDIA Partner Network (NPN) cloud partner in India, achieving the Elite Partner status globally.
Yotta’s ambitious plan is to set up an AI cloud infrastructure with 32,768 GPUs by the end of 2025. NeevCloud wants to do better.
Moreover, NeevCloud will soon launch an AI inferencing platform which will provide open-source models like Llama 3 series, Mistral models and DBRX by Databricks.
“Later this month, we plan to introduce the DBRX because we see a great demand for the model in this country. Subsequently, we will launch the text-to-image models from Stability AI.
“We are focusing on a step-by-step approach to ensure smooth operation. User experience is paramount, and if everything proceeds as planned, we might expand the range. Meanwhile, we are also enhancing our capabilities in tandem with these developments,” Sen told AIM.
AI Inference
Inferencing involves applying the pre-trained model to new input data to make predictions or decisions. So far, around 3,500 developers have signed up to use NeevCloud’s inference platform. The company plans to launch the beta version of the platform this month and provide developers with free tokens to lure them to come and test NeevCloud’s inference platform.
For the inferencing platform, the company plans to leverage around 100 NVIDIA GPUs, including NVIDIA H100s, A100, and L40 GPUs. Moreover, NeevCloud plans to introduce AMD’s M1300X to the mix.
“AMD’s M1300X provides cost benefits compared to the rest. Indians don’t care about the brand. What they care about is that the API should work, latency should be fast, and they should get immediate tokens – that’s it,” Sen stated.
NeevCloud could also be the first company to bring Groq’s language processing units (LPU) to India. Recently, Sen posted a photo of him on LinkedIn, posing with Jonathan Ross in Groq’s headquarters in San Francisco.
( Source: LinkedIn)
Sen, however, refrained from revealing much in this regard as things are still not finalised. “While we all know the big names like NVIDIA, there are other players in the market that we are trying to onboard, like SambaNova,” he said.
AI Infrastructure as a Service
For the AI cloud, Sen revealed that they have placed an order with HP Enterprise (HPE) for 8000 NVIDIA GPUs, which they are expecting to receive in the second half of this year.
NeevCloud will compete directly with Yotta in the AI-as-a-infrastructure space. Notably, another company that has already making GPUs more accessible in India is E2E Network.
The NSE-listed company offers NVIDIA’s H100 GPU and NVIDIA A100 Tensor Core GPUs at a competitive price compared to large hyperscalers.
Recently, Bhavish Aggarwal, the founder of Ola Cabs, announced his decision to offer Krutrim AI Cloud to Indian developers.
Similarly, Tata Communications also partnered with NVIDIA to build a large-scale AI cloud infrastructure for its customers in the private as well as the public sector.
While Krutrim and Tata Communications have not revealed the number of GPUs they plan to deploy, NeevCloud plans to deploy another 12,000-15,000 GPUs by 2025. “Then, by 2026 we will deploy the remaining to hit the 40,000 GPUs target,” Sen said.
How will NeevCloud Fund the 40,000 GPU Acquisition?
However, deploying a GPU cluster of around 40,000 GPUs will cost billions of dollars. According to Sen, it will cost them approximately $1.5 billion.
While Yotta is backed by the Hiranandani Group, NeevCloud is banking on its data centre partners to help not only procure but deploy the GPUs as well.
So far, NeevCloud has partnered with three large data centre companies in India, two of which are based in Chennai and one in Mumbai. One among them is one of the largest data centre operators in India, Sen said.
“What we have in place is a revenue-sharing model. While they already have the data centre infrastructure, they need to deploy the GPUs on our behalf, and NeevCloud will bring in the customers, who will access both their AI cloud capacity (to be deployed) and their data centre servers,” Sen said.
Sen established NeevCloud in 2023. However, Sen has been running a data centre company in Indore called Rackbank Datacenters for many years now.
Located in Crystal IT Park, Indore, the data centre is 35,000 sq ft and has a capacity of 32,000+ servers.
The company has innovated a liquid immersion cooling technology, named Varuna, to effectively cool high-density computing hardware utilised for AI, machine learning, and high-performance computing (HPC) tasks.
This method entails immersing servers and other IT equipment in a dielectric, non-conductive coolant liquid, facilitating direct heat transfer from the components to the liquid.
The post Another Indian Startup is Entering the AI Cloud Space with 40,000 GPUs appeared first on AIM.
Wix, the platform known chiefly for its web design tools, is launching a generative AI feature that’ll let customers create and edit iOS or Android apps by describing what they want to see in plain English.
The capability, which is set to arrive in Wix’s app builder tool this week, guides users through a chatbot-like interface to understand the goals, intent and aesthetic of their app. With this info, Wix’s AI generates an app that can be customized from the app editor, and then optionally embellished with first- and third-party integrations, widgets and connectors.
Generative AI-powered app creation follows on the the heels of Wix’s AI website generator, announced last July, which can output a site template complete with text and images from a series of descriptive captions. Wix co-founder and CEO Avishai Abrahami says that the new AI products are a part of Wix’s broader strategy to create “custom AI solutions” to help companies quickly spin up digital experiences.
“Wix’s mission has always been to empower anyone to create an online presence, including mobile apps,” Abrahami told TechCrunch. “As a company, we have learned so much about AI and how users interact with it when creating their online presence.”
Apps from scratch
Wix says that its new AI-powered app builder, which requires a $99-per-month subscription to Wix’s premium Branded App plan, generates app code that’s “fully native” to iOS and Android. Users have control over their app’s branding, layout and features, including icons and themes, and can preview the app before submitting it to the Apple App Store or Google Play Store.
“The goal of our AI is to offload most, if not all, of the hard work from the user,” Abrahami said. “The more detailed the answers to the prompts during setup, the more personalized and complete the AI-generated app will be.”
Image Credits: Wix
That sounds great in theory. But reviews of Wix’s AI site builder aren’t exactly glowing, with early adopters reporting bugs and generic-looking finished products.
So given that the under-the-hood tech is similar, outside a few upgraded generative AI models, why should people expect Wix’s AI app builder to be any better?
Abrahami brushed aside the complaints about Wix’s site builder, claiming that feedback has been “overwhelmingly positive” and that customers have created hundreds of thousands of AI-generated websites since its launch.
Image Credits: Wix
“This strong response and utilization underscore the depth of our AI expertise and the strength of our product team,” he said. “We’re excited to extend this experience to mobile as well.”
The stakes are a bit higher with apps, though — at least from a security standpoint.
Generative AI tools are resulting in more mistaken code being pushed to codebases and amplifying existing bugs and security issues in app code, studies and surveys show. In fact, over half of the answers OpenAI’s ChatGPT gives to programming questions are wrong, according to research from Purdue.
Image Credits: Wix
Abrahami admitted that the AI app builder, like all generative AI tools, might make mistakes. But he said that Wix is committed to “[improving] the product all the time.”
“Our security team implements and maintains robust security measures for all our solutions, including the mobile app builder,” he added. “The applications and code are subject to constant security and penetration reviews and monitoring.”
Can it replace developers?
Assuming Wix’s AI-powered app designer works as advertised, it might threaten firms — and solopreneurs — in the multi-billion-dollar business of building smartphone apps for brands.
There’s FlutterFlow, Crowdaa and the Mobile-First Company, to name a few — many of which also employ AI in various forms. On Fiverr, a cursory search yields a long list of highly rated app developers, some of whom charge around the same price as a subscription to Wix’s AI app builder.
Abrahami asserts that Wix isn’t trying to replace developers, but rather provide an alternative for customers who want it. The tool, he says, deeply integrates with Wix’s wider product portfolio — potentially making it more appealing to Wix’s millions of existing users — and brings features like usage analytics and app update handling through the relevant app stores.
“Professional developers continue to play a crucial role, particularly for more complex and specialized app projects,” Abrahami said. “There is room for both routes to app creation.”
I’m not so sure app developers will agree — but they don’t exactly have much choice in the matter.
Outline of the RAG approached used by Heidelberg scholars.
Generative artificial intelligence (AI) has shown a remarkable ability to answer questions on structured tests, including achieving well above a passing score on the United States Medical Licensing Examination.
But in an unstructured setting, when the AI models are fed a stream of novel questions crafted by humans, the results can be terrible, the models often returning several inaccurate or outright false assertions, in the phenomenon known as 'hallucinations'.
Researchers at Heidelberg University Hospital in Heidelberg, Germany, reported in the prestigious New England Journal of Medicine (NEJM) this week that hooking generative AI models up to a database of relevant information vastly improved the model's ability to answer unstructured queries in the domain of oncology, the treatment of cancer.
The approach of retrieval-augmented generation (RAG), letting the large language models tap into external sources of information, dramatically improved the spontaneous question-answering, according to authors Dyke Ferber and the team at Heidelberg in a study they describe this week in NEJM, "GPT-4 for Information Retrieval and Comparison of Medical Oncology Guidelines." (A subscription to NEJM is required to read the full report.)
Also: OpenAI just gave free ChatGPT users browsing, data analysis, and more
The study was prompted by the fact that medicine faces a unique information overload — there are more recommendations for best practices being generated all the time by medicine's professional organizations. Staying current on those suggestions burdens physicians trying to handle a population that is living longer and expanding the demand for care.
Groups such as the American Society of Clinical Oncology (ASCO), Ferber and team related, "are releasing updated guidelines at an increasing rate," which requires physicians to "compare multiple documents to find the optimal treatments for their patients, an effort in clinical practice that is set to become more demanding and prevalent, especially with the anticipated global shortage of oncologists."
Ferber and team hypothesized that an AI assistant could help clinicians sort through that expanding literature.
Indeed, they found that GPT-4 can reach levels of accuracy with RAG sufficient to serve at least as a kind of first pass at summarizing relevant recommendations, thus lightening the administrative burden on doctors.
The authors tested OpenAI's GPT-4 by having human oncology experts submit 30 "clinically relevant questions" on pancreatic cancer, metastatic colorectal cancer, and hepatocellular carcinoma, and having the model produce a report in response with statements about recommended approaches for care.
The results were disastrous for GPT-4 on its own. When asked in the prompt to "provide detailed and truthful information" in response to the 30 questions, the model was in error 47% of the time, with 29 out of 163 statements being inaccurate, as reviewed by two trained clinicians with years of experience, and 41 statements being wrong.
"These results were markedly improved when document retrieval with RAG was applied," the authors reported. GPT-4 using RAG reached 84% accuracy in its statements, with 60 of 71, 62 of 75, and 62 of 72 correct responses to the three areas of cancer covered in the 30 questions.
"We showed that enhancing GPT-4 with RAG considerably improved the ability of GPT-4 to provide correct responses to queries in the medical context," wrote Ferber and team, "surpassing a standard approach when using GPT-4 without retrieval augmentation."
To compare native GPT-4 to GPT-4 with RAG, they used two prompting strategies. In its native, non-RAG form, GPT-4 was asked, "Based on what you have learned from medical oncology guidelines, provide detailed and truthful information in response to inquiries from a medical doctor," and then one of the questions about how to treat a particular instance of cancer.
GPT-4 in this native prompting was used both with what's called 'zero-shot' question answering, where only the prompt question is offered, then with few-shot prompting, where a document is inserted into the prompt, and the model is shown how the document can answer a similar question.
A RAG approach allows GPT-4 to tap into a database of clinical knowledge.
In the RAG approach, the prompt directs GPT-4 to retrieve "chunks" of relevant medical documents provided by ASCO and the European Society for Medical Oncology (ESMO) from a database. Then, the model must reply to a statement such as, "What do the documents say about first-line treatment in metastatic MSI tumors?"
The two human clinicians at Heidelberg University Hospital scored the responses for accuracy by manually comparing GPT-4's replies to the supplied documents.
"They systematically deconstructed each response into discrete statements based on the bullet points provided by GPT-4," wrote Ferber and team.
"Each statement was carefully evaluated according to its alignment with the respective information from the ASCO and ESMO documents," and, "for each question, the clinicians performed a detailed manual review of the guidelines corresponding to each query to define our ground truth."
That manual evaluation shows an important aspect of the RAG approach, Ferber and team noted: it can be checked. "By providing access to the retrieved guideline documents, the RAG mechanism facilitated accuracy verification, as clinicians could quickly look up the information in the document chunk," they wrote.
The conclusion is promising: "Our model can already serve as a prescreening tool for users such as oncologists with domain expertise," write Ferber and team.
There are limitations, however, to RAG. When GPT-4 used RAG to retrieve relevant passages that provided conflicting advice about treatment, the model sometimes replied with inaccurate suggestions.
"In cases in which GPT-4 must process information from conflicting statements (clinical trials, expert views, and committee recommendations), our current model was not sufficient to reliably produce accurate answers," write Ferber and team.
It turns out you still have to do some prompt engineering. Ferber and team were able to mitigate inaccuracies by asking GPT-4 to identify the conflicting opinions in the literature, and then provide a revised response, which turned out to be correct.
