 |
|
|
|||||
The Chinese AI company DeepSeek has put the AI industry in an uproar. Denied the most powerful chips thought needed to create state-of-the-art AI models, DeepSeek pulled off some engineering master strokes that allowed the researchers to do more with less. The DeepSeek-V3 and DeepSeek-R1 models the company recently released achieved state-of-the-art performance in benchmark tests and cost much less time and money to train and operate than comparable models. And the cherry on top: The companys researchers showed their workthey explained the breakthroughs in research papers and open-sourced the models so others can use them to make their own models and agents. The main reason DeepSeek had to do more with less is that the Biden administration put out a series of restrictions on chip exports saying that U.S. chipmakers such as Nvidia couldnt ship the most powerful GPUs (graphics processing units, the go-to chip for training AIs) to countries outside the U.S. This effort started in October 2022, and has been updated and fine-tuned several times to close loopholes. Biden released an executive order shortly before leaving office further tightening restrictions. DeepSeek apparently played by the rules. It made do with H800 chips the U.S. allowed Nvidia to sell in China, instead of the more powerful H100 that U.S. tech and AI companies use. With less powerful chips, the researchers were forced to find ways of training and operating AI models using less memory and computing power. The DeepSeek models use a mixture of experts approach, which allows them to activate only a subset of the models parameters that specialize in a certain type of query. This economizes on computing power and increases speed. DeepSeek didnt invent this approach (OpenAIs GPT-4 and Databrickss DBRX model use it), but the company found new ways of using the architecture to reduce the computer processing time necessary during pretraining (the process in which the model processes huge amounts of data in order to optimize its parameters to correctly respond to user queries). In DeepSeek-R1, a reasoning model comparable to OpenAIs most recent o1 series of models (announced in September), DeepSeek found ways of economizing during inference time, when the model is thinking through various routes to a good answer. During this process of trial and error, the system must collect and store more and more information about the problem and its possible solutions in its context window (its memory) as it works. As the context window adds more information, the memory and processing power required leaps up quickly. Perhaps DeepSeeks biggest innovation is dramatically reducing the amount of memory allocated to storing all that data. In general terms, the R1 system stores the context data in a compressed form, which results in memory savings and better speed without affecting the quality of the answer the user sees. DeepSeek said in a research paper that its V3 model cost a mere $5.576 million to train. By comparison, OpenAI CEO Sam Altman said that the cost to train its GPT-4 model was more than $100 million. Since the release of DeepSeeks V3, developers have been raving about the models performance and utility. Consumers are now embracing a new DeepSeek chatbot (powered by the V3 and R1 models), which is now number one on the Apple ranking for free apps. (However, that success has attracted cyberattacks against DeepSeek and caused the company to temporarily limit new user registrations.) For the past two years, the narrative in the industry has been that creating state-of-the-art frontier models requires billions of dollars, lots of the fastest Nvidia chips, and large numbers of top researchers. Across the industry and in investment circles that assumption has been challenged. As a result, Nvidia stock fell nearly 17% Monday as investors question their assumptions about the demand for the expensive GPUs. And its all happening because a small shop of Chinese researchers knew theyd need some big engineering breakthroughs in order to create state-of-the-art models using less than state-of-the-art chips.
Category:
E-Commerce
Millennial wealth in the United States has nearly quadrupled since 2019, outpacing both Gen X and baby boomers, yet most millennials don’t consider themselves rich. Millennials, those born between 1981 and 1996 (give or take a year or two), are now worth a staggering $15.95 trillion, about four times what they were worth just five years ago, according to data from the Federal Reserve as reported by CNBC. As of 2024, the average net worth of a millennial was a whopping $333,096, according to Empower, a financial services company. Its data shows millennials managed to grow their wealth more than any other generation in 2024, increasing their net worth by 13.7% (compared to 7.7% for all Americans), and increasing their 401Ks by 15.6% (nearly double that of the average American). However, as millennials face high costs of living, due in part to inflation and high interest rates, many say they feel less wealthy than they appear on paper, a phenomenon known as “phantom wealth.” That’s because much of their net worth is tied up in assets not readily available, like 401Ks, homes, and the stock market. That’s as there are three main areas of growth that are driving millennial wealth: real estate; stocks and mutual funds; and money they are either inheriting or getting as gifts from parents and family. In the past several years, home equity has emerged as the greatest driver of wealth accumulation, and many millennials who bought homes before or during the pandemic are seeing their value greatly increase. Millennials have also, on average, contributed more to their retirement funds, increasing the value of their holdings both in stocks and mutual funds. Finally, they are also benefiting from their parents’ generosity, receiving financial gifts and inheriting wealth to pay off high student loans, mortgages, car payments, and high child care costs, financial planner Sophia Bera Daigle told CNBC.
Category:
E-Commerce
Isnt AI supposed to make things simpler? asks a student in a new Saturday Night Live sketch. Technically, the answer is yes. Artificial intelligence is often pitched as a future-forward omni-tool for removing friction from everyday tasks. Of course, the student in this sketch (SNL cast member Sarah Sherman) only asks her question after AI has made one such task even more complicated. And thats just one of the many glaring flaws with AI, as it exists in 2025, that the shows writers illustrate to perfectionpresumably without any help from Sora. The premise of the sketch finds a high school investing in a new AI program that turns textbooks into educational podcasts. Its a barely veiled allusion to Googles NotebookLM, a program that creates breezy, conversational summaries from dense documentsand which quickly went viral after debuting last October. Unlike in real life, the SNL version of the fake podcasts has a video component. What the sketch portrays accurately, however, is the way AI products often have questionable utility, overinflate whatever utility they do have, and come brimming with glitches. The hosts of the podcast strain to sound natural, repeat key phrases again and again, and ultimately leave the skeptical students with more questions than answers. According to Gavin Purcell, a (very much human) cohost of the AI-demystifying podcast AI for Humans, the product this sketch is based on actually does offer some benefits. NotebookLM can struggle with getting all its facts right and, over time, the voices get repetitive, but its an interesting use case of how AI can break down complicated topics and make them more digestible, Purcell says. Try throwing an extensive Wikipedia page into it and see what comes out. You might be surprised. In the sketch, though, the program uses full textbooks rather than the smaller documents NotebookLM was made to condense. (The length of the average podcast the real product churns out is five to 10 minutes.) Condensing a whole textbook into a podcast would create something closer to a breezy, conversational audiobook than a short podcast snippet. And its exactly this kind of redundancy that AI tech too often offers. One need only visit the most recent CES to see this redundancy in action. That event was overflowing with AI-assisted devices like Boschs new smart crib, which lets parents know when their baby has pooped overnightas opposed to the age-old technology that has historically done so: a screaming baby . . . not to mention Samsungs new, AI-powered washing machine, which not only alerts users when their laundry is done, but also lets them take phone calls through the machine, for some reason. Beyond satirizing AI products whose usefulness is dubious, the SNL sketch also taps into AI true believers’ tendency to get overhyped too early. Anything that is useful at all suddenly becomes revolutionary. A student might understandably use a fake podcast to briefly learn about a specific topic, as Notebook LM demonstrated, but that doesnt mean the program is going to disrupt learning as we know it, let alone destroy the podcast industry. “NotebookLM was one of these small, quirky AI products that I don’t think Google even thought would blow up as big as it did, Purcell says. And, unfortunately, as often happens when something AI-based explodes into the mainstream, you get a lot of “OMG, PODCASTING IS SO DEAD!!” posts from hardcore AI people. In the past few years, experts have claimed that AI products like ChatGPT may fully reshape the legal and medical industries, among others. But ChatGPT has not yet demonstrated anything like the immaculate reliability it would need to truly revolutionize either field. Instead, its exhibited enough fallibility to only underscore the inherent value of human judgment. In one infamous example, a lawyer used ChatGPT to help a client sue an airline, and the program ended up hallucinating at least six precedent cases that did not actually exist. As long as such mistakes can ever happen, the hype around AIs power to remake every field in society should be taken with a grain of salt. And at this still-early stage in AIs evolution, mistakes happen all the time. The most prominent bug in the SNL sketch is an AI classic: One of the podcasters is depicted with six fingers. Generating anatomically correct extremities is something AI has long struggled with, but glitches manifest in all sorts of ways. McDonald’s recently had to shut down its experiment with AI drive-thru, after a flurry of viral TikToks showed unwanted bacon on ice cream and other bugs, and Apple has reportedly paused AI news summaries on its new iPhones due to persistent glitches. Maybe one day, malfunctioning AI will become a rare exception, but for now, its much closer to the rule. The final turn in the SNL sketch reveals one problem with AI that humans, so far, have only scratched the surface ofits malevolent side. Do we eat? Do we exist? asks the AI podcaster played by Timothée Chalamet.
Category:
E-Commerce
All news |
||||||||||||||||||
|
||||||||||||||||||