If you’ve been following the latest AI news then you’ll know that chatbots that you can talk to using your voice are here. OpenAI was one of the first to demo the technology with its ChatGPT Advanced Voice mode (currently only free for 10 minutes a month), but Google got to market first with Gemini Live (now free to all Android users), and recently Microsoft joined in by revamping its Copilot website and app (which is free to everyone) to include voice conversations.
The ability to talk to AI using our voice, and have it talk back like a human, has been the sci-fi dream ever since Captain James T. Kirk addressed the ship’s computer in Star Trek, but it was later sci-fi creations that proved indistinguishable from human beings, like HAL 9000 and the Blade Runner replicants, that ignited our imaginations about the possibilities of an AI that could interact like a human.
Now we appear to be living in the future, because you can, right now, have a conversation with AI using the smartphone or computer you’re reading this on. But while we’ve made huge progress towards a human-like companion, there’s still a long way to go, as I discovered recently by putting the latest voice-controlled AIs – ChatGPT Advanced Voice mode, Gemini Live, and Copilot – through their paces for a couple of weeks. Here are my top three takeaways:
1. Interruptions are a great idea, but don’t work properly
The biggest problem I find with talking AIs is being able to interrupt them successfully, or their ability to interrupt you when you don’t want them to. It’s great that ChatGPT, Gemini Live, and Copilot all let you interrupt, mainly because they tend to give long and ponderous answers to everything you ask them, and without that ability, you wouldn’t bother using them. That process, however, is often flawed; either they miss your interruption or they then respond to your interruption with more talking. Usually, it’s some version of, “Ok, what would you like to know about instead?”, when all you want them to do is stop talking so you can begin to talk. The result is usually a messy series of jumps and starts that kills the natural flow of the conversation and stops it from feeling human.
Advertisement
Quite often this week I found myself yelling, “Just stop talking!”, at my phone, just so I could get a word in, which isn’t a good look. Especially since I sit in an office surrounded by people for most of the day.
Another problem I frequently encountered with all of the chatbots is thinking I had finished talking when in fact I was just pausing to consider my thoughts and was still halfway through a sentence. The whole AI experience needs to be as smooth as butter for you to have confidence in it, or the spell breaks.
2. There’s not enough local information
Ask any of the current crop of chatbots where the best place to get a pizza is locally and apart from Gemini Live, you get told that they can’t search the web. Gemini Live is massively ahead here – it will make a recommendation for somewhere good to get pizza. The recommendations aren’t bad, and although it can’t make a reservation for you it will get you the phone number of the restaurant.
Voice-activated chatbots obviously need to be able to browse the web, just like text-based chatbots currently can, but right now ChatGPT Advanced Voice mode and Copilot can’t, and that’s a huge drawback when it comes to delivering relevant information.
Advertisement
3. They’re not personal enough
For voice AI to be useful it needs to know a lot of information about you. It also needs to be able to access your important apps like your inbox and your calendar. At the moment it can’t do that. If you ask it, “Hey, am I free at 4 pm this Friday?”, or, “When is the next family birthday coming up?”, you get told that it can’t do that right now, and without that kind of ability, the usefulness of voice AI just falls off a cliff.
So, what is a talking AI good for?
Right now the best use of Voice AI is for asking questions, giving you some motivation to do something, or coming up with ideas that you wouldn’t think of on your own. Pick a subject and get AI to engage with you in a conversation and you’ll find that it knows a surprising amount about a lot of things. It’s fascinating! For example, one of the things I actually know a lot about is Brazilian Jiu-Jitsu, and I found I could engage each of the chatbots in a pretty good conversation about it, even down to a surprising level of detail regarding techniques and positions. Based on my experience I’d say that Copilot gave me the best answers and that Gemini seemed more likely to hallucinate things that weren’t true.
In terms of the interface, I think ChatGPT is leading the way. I really like the way its swirling orb seems to react with a pulse that’s in time with whatever you say, which gives you confidence it’s actually listening. Gemini Live in contrast has a mainly dark screen with a glowing area at the bottom, which doesn’t give you a focus point to look at, leading to a slightly more soulless experience.
The AI you can talk to right now is great for delving into research topics, but it also feels a bit half-finished, and it’s going to need a lot more integration with our smartphones before it can perform at the level we’d naturally like it to. Of course, it will get better over time. Right now the elephant in the room is Apple Intelligence and its associated Siri, who are both late to the party. We’re still waiting for an Apple Intelligence release date, and even then we won’t get the full all-singing, all-dancing Siri until next year.
Advertisement
Right now the promise of an AI we can talk to just like a friend or a real virtual assistant seems tantalizingly close, but also still a long way off.
In the world of artificial intelligence-powered tools, it keeps getting harder and harder to differentiate real and AI-generated images. No one can easily identify an AI-created photo at the first glance. However, Google Photos could soon help you identify AI-generated images. Notably, folks over at Android Authority have uncovered this ability in the APK code of the Google Photos app.
Soon, you will be easily able to identify AI-created images using Google Photos
The source has found clues in the Google Photos app’s version 7.3 regarding the ability to identify AI-generated images. This ability will allow you to find out whether a photo is created using an artificial intelligence tool. One of the layout files in the APK of Google Photos v7.3 has identifiers for AI-generated images in the XML code. The source has uncovered three ID strings namely “@id/ai_info”, “@id/credit”, and “@id/digital_source_type”, inside the code.
Furthermore, the report suggests that the “@id/credit” ID could likely display the photo’s credit tag. If the photo is made using Google’s Gemini, then Google Photos can identify its “Made with Google AI” credit tag. It will allow Google Photos to identify AI-generated images quite easily.
Also, the “@id/digital_source_type” ID could refer to the source type field. This will showcase the media source from where the AI photo was created. There’s no word as to what the “@id/ai_info” ID in the XML code refers to.
Advertisement
Google Photos will use these identifiers of an image to tell if it is AI-generated
Notably, the report also mentions that it’s likely all the aforementioned information will be displayed in the image details section. The IPTC metadata will allow Google Photos to easily find out if an image is made using an AI generator. That said, soon it will be very easy to identify AI-created images using the Google Photos app.
As of now, this feature isn’t live on Google Photos. However, we can expect Google to roll out the new functionality as soon as possible as it’s already inside Google Photos.
When TS Imagine first started using Snowflake, it was simply seeking a means to manage its data. Three years later, Snowflake is providing the fuel for the fintech company’s transformation into an enterprise powered by AI.
Based in Bozeman, Mont., but with no central headquarters, Snowflake is a data cloud vendor whose platform enables customers to store and analyze data. In addition, over the past couple of years, the vendor has made AI a focal point, developing an environment where customers can develop, deploy and manage AI, machine learning models and applications.
TS Imagine, meanwhile, is a SaaS-based financial services vendor headquartered in New York City that provides front-office trading, portfolio management and financial risk assessment capabilities. The company was formed in 2021 after the merger of TradingScreen and Imagine Software.
Following the merger, TS Imagine needed a way to integrate and organize years of data from TradingScreen, founded in 1999, and Imagine Software, founded in 1993.
Advertisement
Snowflake was that way. Now, however, as Snowflake has evolved beyond a platform for data management into an environment for AI, TS Imagine has evolved with it and is using Snowflake’s platform to power its metamorphosis.
“We are a cloud-first company and a Snowflake-first company,” said Thomas Bodenski, TS Imagine’s COO and chief data and analytics officer. “Now there’s a third one: AI-first.”
Using Snowflake, TS Imagine is accessing data previously unavailable to inform decisions. It’s using AI to manage certain processes, and it’s reaping financial benefits.
First, however, it just needed to get organized.
Advertisement
Starting with Snowflake
When TradingScreen and Imagine Software merged in May 2021, the newly formed company faced challenges. Both TradingScreen and Imagine Software brought with them well over two decades of data. In addition, the newly formed company had two data teams; two technology stacks; and plans to expand into new areas, such as fixed-income securities trading.
TS Imagine needed a way to unify that data, and it needed to do so in a single system that would help its expansion.
“We very quickly identified the data as the area where we needed to focus,” Bodenski said. “We knew, strategically, that we had to do something. We have to have data ready at any time because it’s used for trading and for risk management. We need to solve problems that our clients are never meant to see.”
TS Imagine manages over 20 million financial instruments — assets such as stocks, bonds, loans, funds and certificates of deposit that can be traded or exchanged. Each, including the client that owns the instrument, generates data, meaning that TS Imagine needs to manage massive amounts of data to serve the needs of its customers.
Advertisement
It, therefore, needed a data management platform that was simple enough to enable users to easily access data when needed and could also handle scale.
One option included the platforms previously used by TradingScreen and Imagine Technologies. Others included platforms such as Markit EDM and GoldenSource geared specifically for reference data used to categorize financial transactions and for enabling semantic modeling for financial instruments.
Ultimately, TS Imagine chose Snowflake.
Timeliness was a key factor in the TS Imagine’s decision given that it needs to access data in near real time to inform and execute trades, according to Bodenski. So were the breadth and depth — the scale — of Snowflake.
Advertisement
Finally, simplicity played a significant role.
Snowflake understands Python and SQL code. If TS Imagine had chosen a platform that required Java or C++, for example, few of its developers would have had the requisite skills to use the platform. But because Python and SQL can be used in Snowflake, 54 data scientists, engineers and other data experts already had the needed skills.
“We felt that with Snowflake, we had a platform that could empower us,” Bodenski said. “We were able to grow and scale overnight from a small team to a large organization.”
Now, TS Imagine stores all its data in Snowflake and runs all its data management processes, such as data quality monitoring, pipeline monitoring and automated regression testing, in Snowflake.
Advertisement
Adding AI
About a year after TS Imagine got its data in order on Snowflake, OpenAI launched ChatGPT.
Released in November 2022, ChatGPT was a significant improvement in generative AI capabilities over what was previously available. Of particular interest to many organizations were its natural language processing (NLP) and automation capabilities.
Enterprises quickly recognized that if they could combine those capabilities with proprietary data to understand an organization’s operations, they could reap significant benefits such as more widespread use of analytics due to NLP and efficiency gains due to process automation.
Among the organizations that saw the possibilities of generative AI in the enterprise was TS Imagine.
“When the hype around ChatGPT started, we got very excited,” Bodenski said. “All of my executive peers are into data as well, and ChatGPT was part of every meeting.”
Advertisement
TS Imagine had already experimented with NLP and machine learning to automate tasks, such as data classification and cataloging. However, converting unstructured data to structured data to inform models and applications had proven to be difficult.
Unstructured data, such as text, images and audio files, is estimated to make up well over three-quarters of all data. Tapping into unstructured data is critical to gaining a full understanding of an organization.
With Snowflake still focused largely on data management at the time, TS Imagine viewed ChatGPT’s generative AI capabilities as a way to finally gain access to its unstructured data, particularly text in emails and PDF documents.
“We needed to make it more actionable by converting it into structured content,” Bodenski said.
Advertisement
TS Imagine developed an AI engineering team that worked with its data experts to use the data it had stored in Snowflake to train ChatGPT to analyze text.
It created an AI pipeline using open source database ChromaDB to vectorize unstructured data to give it structure, LangChain to develop a retrieval-augmented generation (RAG) pipeline to discover the relevant data required to train its models, and containers from Google Cloud to run its generative AI workloads.
The result was models that delivered precise, accurate outputs when analyzing text from over 500 clients, according to Bodenski.
“It had an unbelievably high provision rate to the point where we could rely on it,” he said.
Advertisement
Nevertheless, TS Imagine did not automate final decisions based on generative AI outputs. It still put a human in place to check the outputs for accuracy and make any final decisions.
For the next year, TS Imagine continued to use ChatGPT to underpin its generative AI development and analysis. That was until Snowflake began developing its own environment for generative AI.
Snowflake for everything AI
Enterprises like TS Imagine weren’t the only ones who recognized the potential value of generative AI following the release of ChatGPT.
With data serving as the underlying engine for AI — the information used to train and inform AI models and applications — analytics and data management vendors, from specialists such as MicroStrategy and Monte Carlo to tech giants such as AWS, Google Cloud and Microsoft, all made generative AI a focal point of their product development plans.
Advertisement
Snowflake rival Databricks was especially aggressive in building up an environment for customers to create AI models and applications. After a slow start, Snowflake followed suit.
Thomas BodenskiCOO and chief data and analytics officer, TS Imagine
In May 2023, Snowflake acquired Neeva, a search engine specialist, to procure generative capabilities. Five months later, the vendor introduced Cortex, an environment for AI development that includes access to LLMs and vector search capabilities, among other capabilities. Since then, Snowflake has continued to add tools aimed at enabling AI and machine learning development, including its own LLM and a chatbot development framework.
With Snowflake — via Cortex — providing the same capabilities TS Imagine was piecing together with ChatGPT, ChromaDB, LangChain and Google Cloud, the financial services specialist decided to migrate its AI operations to Snowflake.
The process was simple, taking one engineer one week to complete the entire undertaking, according to Bodenski.
Advertisement
“Everything AI runs exclusively on Snowflake now,” he said.
Immediately following the migration, simply by eliminating the cost of using various platforms to create an AI pipeline and instead using the tools provided by Snowflake, TS Imagine saw a 30% reduction in spending related to training and managing its generative AI capabilities.
“That was significant for us,” Bodenski said. “It’s a one-stop shop for us. We can build the entire AI pipeline on the technology we are all familiar with.”
With its data already residing in Snowflake, TS Imagine simply builds an AI pipeline on top of that data without needing to move the data to another system where it might get accidentally exposed. In addition, with all the required pieces of an AI pipeline in one environment, it takes just a few days to develop a new model or application.
Advertisement
Results
In the year since the migration, after developing text analysis capabilities using ChatGPT for generative AI capabilities and moving that to Snowflake, TS Imagine has developed five other generative AI pipelines for different applications.
Beyond analyzing emails and PDFs, one of the key applications of generative AI is to monitor customer service. TS Imagine receives an average of 5,000 inquiries per month. Fully understanding everything related to customer service is challenging.
“If you are the global head of customer service, it’s not easy to get that overview,” Bodenski said. “And if you are a regional manager, it’s hard to know everything that is going on.”
With its customer service application, TS Imagine can now classify each customer service incident, automatically assigning sensitivity ratings as well as understanding the sentiment, urgency and complexity of the request.
Advertisement
“Those are all steps that would have had to have been done manually,” Bodenski said.
Tangibly, by developing and deploying generative AI tools using Snowflake, TS Imagine has saved thousands of hours of work — including 4,000 that would have been devoted just to analyzing emails –that otherwise would have been done manually, he continued.
“It allows us to utilize people to do work that is more analytical, more knowledge-oriented,” Bodenski said. “We can use people to be more productive on other tasks.”
Despite all its benefits, like most enterprises using generative AI to improve operations, TS Imagine is working through some problems.
Advertisement
While using Snowflake to develop generative AI tools has been a smooth process, getting models and applications to consistently deliver outputs that can be trusted remains a concern, according to Bodenski.
“There is still a challenge with what large language models produce,” he said.
Accuracy has been a problem for generative AI. Even when trained using high-quality data, models and applications sometimes still deliver incorrect and even bizarre outputs called hallucinations.
To combat those inaccuracies, TS Imagine runs its RAG pipelines multiple times for each query to try to weed out any outliers. Still, however, the company makes sure there is always a person in place to take any action rather than trust the model or application to automatically go from output to action on its own.
Advertisement
“We need to constantly look at the results,” Bodenski said. “And you really need to find the right use cases. This stuff doesn’t solve everything. You need to find the right use cases, and that’s when you get high precision rates. Even still, the outputs are sometimes very strange.”
Future plans
With six RAG pipelines running after one year using Snowflake for its AI development and deployment, TS Imagine has plans to add more AI applications, according to Bodenski.
To date, what the company has done with generative AI is to automate processes to make workers more efficient. It hasn’t yet developed AI assistants that enable business users to query and analyze data using natural language.
TS Imagine has used Snowflake to develop applications its customers can use to analyze data. But those applications are traditional analytics applications rather than AI-powered applications.
Advertisement
The next step is to add generative AI to those applications to enable clients to broaden their use of BI beyond data experts as they analyze financial transactions and strategies.
“What we aim for is self-service analytics,” Bodenski said. “There is a lot of data involved in financial transactions, and with this data available, our clients can self-service themselves. We want to bring AI through our products to our clients. That’s the final objective.”
Eric Avidon is a senior news writer for TechTarget Editorial and a journalist with more than 25 years of experience. He covers analytics and data management.
is charting a course away from its . The company is dropping that standardized approach in favor of a wider range of battery cell chemistries and physical formats.
The automaker had hoped that, by adopting a unified system across all of its EVs as well as , it would be able to reduce costs and ship them faster. The plan was to pack the flat pouch-style Ultium cells into a variety of modules depending on what was needed for each EV.
Things haven’t gone smoothly, as notes. Among other things, COVID-19 slowed down the company’s EV roadmap and with the robots that assembled the modules.
“It now makes business sense to transition from one-size-fits-all to new program-specific batteries,” Kurt Kelty, GM’s vice president of batteries, said at an investor event. The automaker hopes that switching from Ultium’s nickel cobalt manganese chemistry to lithium iron phosphate (LFP) battery tech will lower the cost of its EVs by as much as $6,000. As notes, Tesla and Ford are among those that use LFP cells, which are said to be cheaper and less complicated to manufacture. The , which is slated to arrive in late 2025, will use such batteries.
Advertisement
GM plans to build a new battery research facility at the Warren Tech Center in Michigan. The team there will explore cylindrical and prismatic cells in addition to the pouch format. Researchers will also look into alternative battery chemistries.
The shift in battery strategy comes as GM chases profitability in its EV division. The company said it’s getting close to that point. It’s on track to build and sell around 200,000 EVs this year. GM now claims to be the number two EV seller in North America behind Tesla.
This video will help you assemble Rackmount Solutions’ Cruxial ™ WM series wallmount racks. These self-squaring racks are available in 8u, 12u, 15u and 18u with depths of 13″ and 19″. Assembles in less than 15 minutes.
A relay rack made from U.S. American Finished Steel, the WM Series is built tough in order to handle your server, data center, networking needs.
Rackmount Solutions is an industry leader in supplying server racks, server cabinets, wallmount racks, network racks, LAN racks, portable rackmount cases and accessory products for the IT/Network professional. We deliver rackmount storage server rack solutions for 19″, 23″, 24″ and 28″ wide equipment, as well as PCI Data Security Storage compliant cabinets and racks. We pride ourselves in providing quality customer service. Please call us toll free at 1-800-352-6631 and let us know how we can solve your rackmount needs.
You must be logged in to post a comment Login