Cloud Hosting

Google is training its robots with Gemini AI so they can get better at navigation and completing tasks. The DeepMind robotics team explained in a new research paper how using Gemini 1.5 Pros long context window which dictates how much information an AI model can process allows users to more easily interact with its RT-2 robots using natural language instructions.

This works by filming a video tour of a designated area, such as a home or office space, with researchers using Gemini 1.5 Pro to make the robot watch the video to learn about the environment. The robot can then undertake commands based on what it has observed using verbal and / or image outputs such as guiding users to a power outlet after being shown a phone and asked where can I charge this? DeepMind says its Gemini-powered robot had a 90 percent success rate across over 50 user instructions that were given in a 9,000-plus-square-foot operating area.

Researchers also found preliminary evidence that Gemini 1.5 Pro enabled its droids to plan how to fulfill instructions beyond just navigation. For example, when a user with lots of Coke cans on their desk asks the droid if their favorite drink is available, the team said Gemini knows that the robot should navigate to the fridge, inspect if there are Cokes, and then return to the user to report the result. DeepMind says it plans to investigate these results further.

The video demonstrations provided by Google are impressive, though the obvious cuts after the droid acknowledges each request hide that it takes between 1030 seconds to process these instructions, according to the research paper. It may take some time before were sharing our homes with more advanced environment-mapping robots, but at least these ones might be able to find our missing keys or wallets.

Read more from the original source:

Google says Gemini AI is making its robots smarter - The Verge

Investors are underappreciating Amazon's artificial intelligence potential, according to some analysts. "While consensus believes that AMZN lags behind its mega-cap peers in AI capabilities, AMZN highlighted that it has launched over two times the number of AI services than MSFT and GOOGL combined since 2023," wrote BMO Capital Markets' Brian Pitz. "We believe investors are underestimating AMZN's positioning to win its fair share of AI workloads and believe its platform approach with Bedrock will be a winning strategy over the long term." The analyst has an outperform rating on Amazon and a price target of $220, which signals 10% upside from Wednesday's close. These tools and initiatives should support ongoing and accelerating revenue growth through 2024 and more than 20% growth in 2025, according to JPMorgan analyst Doug Anmuth. He called the stock a best idea, viewing AWS as offering a "best-in-class" array of large language models and foundation models tailored to developer needs. Anmuth has an overweight rating on the stock and a $240 price target, which implies upside of 20%. The commentary comes on the heels of the e-commerce giant's Amazon Web Services summit in New York, where the company highlighted the ways customers are implementing AI workloads and showcased its rapidly growth AI platform know as Bedrock. Amazon shares have rallied about 30% year to date. AMZN YTD mountain Shares this year Citi's Ronald Josey also reaffirmed his confidence in AWS estimates for the second quarter and onward following the event. The analyst also believes demand is gradually moving toward companies offering multiple specialized models. He retained a $245 price target on shares and called the stock a top Internet pick. The price target implies 23% upside from Wednesday's close.

See more here:

Investors are underestimating the AI capabilities of this big-cap tech stock, analysts say - CNBC

When Aizada Marat moved from New York to California in 2018 with her husband, KODIF co-founder and CEO Chyngyz Dzhumanazarov, she needed to sort out her immigration status. Thats when everything started going badly.

The Kyrgyzstan-born, Harvard-educated attorney came to the U.S. when she was 17 for an exchange year with FLEX (future leaders exchange) sponsored by the U.S. State Department.

After graduating from Harvard, Marat moved to London because of immigration issues. Now she was coming out to California with Dzhumanazarov, who had been admitted to Stanford Business School, and to take a job offer at leading law firm Cooley.

But she didnt realize that immigration lawyers can be very buyer-beware. Through a Google search she found a lawyer in Palo Alto to help her with her visa. That turned out to be a bad move. Marat said the lawyer gave her wrong advice about when she could file authorization to work in California. That mistake caused her to not be able to work for more than a year. She also could not leave the country.

Im a lawyer, so I listen to what lawyers say, Marat told TechCrunch. Unfortunately, listening to them was devastating because months later, I was still unable to work. I had a job offer from Cooley.

Marat did end up getting to work at Cooley for three years. And she went back to that immigration law firm and showed them the mistake they made with her. It also ignited an entrepreneurial fire in her.

After she left Cooley to work at McKinsey as a management consultant, Marat kept coming back to that horrible immigration experience. So much so that she started thinking about why immigration legal services were of poorer quality given the long and complicated immigration process.

She learned that immigration law is super fragmented, meaning that 10% of the market is owned by one law firm while the other 90% is shared among over 20,000 law firms.

Very few big law firms have immigration services today because it is mainly serving individuals, and those are small checks, Marat said. Thats why, to get a talent visa green card, the majority of the time, people can self-petition. They dont even need an employer. Cooley, in my case, wouldnt really sponsor visas, so I had to sort it out myself.

And when she thought about what to do about it, Marat set out to start her own company developing software to sell to immigration attorneys. The goal was to help them deliver better services, so what happened to Marat wouldnt happen again.

After four or five months of selling that software to five immigration law firms, Marat and her team made the decision to provide immigration services directly. In October 2023, they launched Alma, an AI-powered legal tech startup that she started with other immigrants, including former Uber engineering manager Shuo Chen and former Step product manager Assel Tuleubayeva.

The startup aims to simplify the visa process for technologists, founders and researchers by providing personal legal advisors, helping to speed up document processing and digitally organizing the entire process. And like other companies working in this area, including Migrun, Boundless and Lawfully, Alma wants to fast-track international talent into Americas tech ecosystem, Marat said.

Marat says Alma differs from some competitors by leverage proprietary technology to provide high-quality services faster and employing its own immigration attorney.

Immigrants deserve high-quality services because so much depends on the immigration attorney that you find, Marat said. All the repetitive and mundane things that lawyers hate, we can automate so that lawyers actually focus on all clients and provide a really good strategy to get higher approval rates.

Helping to move the company forward is $5.1 million in combined seed and pre-seed funding that Alma recently raised. The company is backed by Bling Capital, Forerunner, Village Global, NFX, Conviction, MVP, NEA and Silkroad Innovation Hub. Much of the funding will go toward new hires for product and technology development.

View original post here:

Alma co-founder had such a bad immigration experience she founded a legal AI startup to fix it - TechCrunch

People in tech often say that data is the new oil. That phrase, coined by British mathematician Clive Humby, of course implies that data is valuable.

Data about a persons health can also provide meaningful insights and improve outcomes, but only 3% of patient data is currently used by physicians, according to the World Economic Forum. Although doctors know they can glean useful information from patient data, they dont have the time to review every detail in the medical record.

Regard, a digital health startup founded in 2017, wants to help physicians save time and increase the accuracy of diagnosis by analyzing patients health data using AI.

Regard announced on Thursday that it raised a $61 million Series B round led by Oak HC/FT, with participation from Cedars-Sinai Health Ventures and existing investors TenOneTen, Calibrate Ventures and Techstars. The company is now valued at $350 million, according to a person familiar with the matter.

The companys software mines thousands of data points in a medical chart and presents data in a way that allows doctors to detect health conditions more easily.

Doctors use our product because we help them make sure that nothing important is missed in the data, Eli Ben-Joseph, Regards co-founder and CEO, told TechCrunch. Every single doctor we work with has a story about: I used your product and I found something that changed the way I treat this patient.

Ben-Joseph said that Regard has helped one general physician catch atrial fibrillation (irregular heartbeat) that the cardiologist did not notice. She now feels its irresponsible not to use our product, he said.

But doctors are not the only ones who find Regard valuable. Hospital financial administrators are also big fans, according to Ben-Joseph. Regards ability to identify new conditions creates new billing opportunities for medical systems.

The company grew its revenue by 4.5 times in 2023 and is on a path to do a similar amount of growth this year, Ben-Joseph said. The company expects to reach profitability within the next two years.

Such fast growth has investors excited.

We absolutely fell in love with [Regard] because it has a direct impact on physician productivity, burnout, proper coding and clinical outcomes, said Nancy Brown, general partner at Oak HC/FT.

Brown, who has over 30 years of experience as an operator and investor in healthcare technology, has always dreamt that a computer would provide insights from patient information. That dream has been foiled [over the years] by the lack of tech, she said. Thats why when she met Ben-Joseph at a healthcare conference earlier this year, she instantly recognized that Regard is the technology she has been dreaming about.

Since launching its product in 2021, Regard has signed up a number of large healthcare systems, including Banner Health (one of Arizonas largest health providers), Virginia-based Sentara Healthcare, New Yorks Montefiore Medical Center and Cedars-Sinai Medical Center in Los Angeles. Some details of its latest funding round were previously reported by Business Insider.

The companys competitors include Engage One, a product developed by multinational conglomerate 3M, and startup Pieces, according to Ben-Joseph.

Brown has no doubts that Regard is the leader in the space. They are a beautiful scaling company with great margins, and they are delivering a solid ROI for their clients, Brown said.

Visit link:

AI-powered Regard nabs $61M to find missed illness, boost hospital revenue - TechCrunch

Enlarge / Whatever you do, don't ask the AI how many horizontal lines are in this image.

Getty Images

In the provocatively titled pre-print paper "Vision language models are blind" (which has a PDF version that includes a dark sunglasses emoji in the title), researchers from Auburn University and the University of Alberta create eight simple visual acuity tests with objectively correct answers. These range from identifying how often two colored lines intersect to identifying which letter in a long word has been circled to counting how many nested shapes exist in an image (representative examples and results can be viewed on the research team's webpage).

If you can solve these kinds of puzzles, you may have better visual reasoning than state-of-the-art AIs.

The puzzles on the right are like something out of Highlights magazine.

A representative sample shows AI models failing at a task that most human children would find trivial.

Crucially, these tests are generated by custom code and don't rely on pre-existing images or tests that could be found on the public Internet, thereby "minimiz[ing] the chance that VLMs can solve by memorization," according to the researchers. The tests also "require minimal to zero world knowledge" beyond basic 2D shapes, making it difficult for the answer to be inferred from "textual question and choices alone" (which has been identified as an issue for some other visual AI benchmarks).

After running multiple tests across four different visual modelsGPT-4o, Gemini-1.5 Pro, Sonnet-3, and Sonnet-3.5the researchers found all four fell well short of the 100 percent accuracy you might expect for such simple visual analysis tasks (and which most sighted humans would have little trouble achieving). But the size of the AI underperformance varied greatly depending on the specific task. When asked to count the number of rows and columns in a blank grid, for instance, the best-performing model only gave an accurate answer less than 60 percent of the time. On the other hand, Gemini-1.5 Pro hit nearly 93 percent accuracy in identifying circled letters, approaching human-level performance.

For some reason, the models tend to incorrectly guess the "o" is circled a lot more often than all the other letters in this test.

The models performed perfectly in counting five interlocking circles, a pattern they might be familiar with from common images of the Olympic rings.

Do you have an easier time counting columns than rows in a grid? If so, you probably aren't an AI.

Even small changes to the tasks could also lead to huge changes in results. While all four tested models were able to correctly identify five overlapping hollow circles, the accuracy across all models dropped to well below 50 percent when six to nine circles were involved. The researchers hypothesize that this "suggests that VLMs are biased towards the well-known Olympic logo, which has 5 circles." In other cases, models occasionally hallucinated nonsensical answers, such as guessing "9," "n", or "" as the circled letter in the word "Subdermatoglyphic."

These gaps in VLM capabilities could come down to the inability of these systems to generalize beyond the kinds of content they are explicitly trained on. Yet when the researchers tried fine-tuning a model using specific images drawn from one of their tasks (the "are two circles touching?" test), that model showed only modest improvement, from 17 percent accuracy up to around 37 percent. "The loss values for all these experiments were very close to zero, indicating that the model overfits the training set but fails to generalize," the researchers write.

The researchers propose that the VLM capability gap may be related to the so-called "late fusion" of vision encoders onto pre-trained large language models. An "early fusion" training approach that integrates visual encoding alongside language training could lead to better results on these low-level tasks, the researchers suggest (without providing any sort of analysis of this question).

Read this article:

Can you do better than top-level AI models on these basic vision tests? - Ars Technica