"A group of researchers says it has identified a hidden reason we use Google for nearly all web searches," reports the Washington Post. "We've never given other options a real shot." Their research experiment suggests that Google is overwhelmingly popular partly because we believe it's the best, whether that's true or not. It's like a preference for your favorite soda. And their research suggested that our mass devotion to googling can be altered with habit-changing techniques, including by bribing people to try search alternatives to see what they are like...

[A] group of academics — from Stanford University, the University of Pennsylvania and MIT — designed a novel experiment to try to figure out what might shake up Google's popularity. They recruited nearly 2,500 participants and remotely monitored their web searches on computers for months. The core of the experiment was paying some participants — most received $10 — to use Bing rather than Google for two weeks. After that period, the money stopped, and the participants had to pick either Bing or Google. The vast majority in the group of people who were paid to use Bing for 14 days chose to go back to Google once the payments stopped, suggesting a strong preference for Google even after trying an alternative. But a healthy number in that group — about 22 percent — chose Bing and were still using it many weeks later.
"I realized Bing was not as bad as I thought it was...." one study participant said — which an assistant professor in business economics and public policy at the University of Pennsylvania says is a nice summation of the study's findings.

"The researchers did not test other search engines," the article notes. But it also points out that more importantly: the research caught the attention of some government officials: Colorado Attorney General Phil Weiser (D), who is leading the group of states that sued Google alongside the Justice Department, said the research helped inspire a demand by the states to fix Google's search monopoly. They asked a judge to require Google to bankroll a consumer information campaign about web search alternatives, including "short-term incentive payments."
On the basis of that, the article suggests "you could soon be paid to try Microsoft Bing or another alternative."

And in the meantime, the reporter writes, "I encourage you to join me in a two-week (unpaid) experiment mirroring the research: Change your standard search engine to something other than Google and see whether you like it. (And drop me a line to let me know how it went.) I'm going with DuckDuckGo, a privacy-focused web search engine that uses Bing's technology."

Google announced it will end software updates and remote control support for the first- and second-generation Nest Learning Thermostats (plus the 2014 European model) starting October 25th. "You will no longer be able to control them remotely from your phone or with Google Assistant, but can still adjust the temperature and modify schedules directly on the thermostat," the company wrote in a Friday blog post. The Verge reports: In other significant news, Google is flatly stating that it has no plans to release additional Nest thermostats in Europe. "Heating systems in Europe are unique and have a variety of hardware and software requirements that make it challenging to build for the diverse set of homes," the company said. "The Nest Learning Thermostat (3rd gen, 2015) and Nest Thermostat E (2018) will continue to be sold in Europe while current supplies last." [...]

In a clear attempt to ease customer anger, Google is offering a $130 discount on the fourth-gen Nest Learning Thermostat in the US, $160 off the same device in Canada, and 50 percent savings on the Tado Smart Thermostat X in Europe since the Nest lineup will soon be gone. The original Nest thermostats were released while the company was an independent brand under the leadership of former Apple executive Tony Fadell. Google acquired Nest in 2014 for $3.2 billion.

An anonymous reader shares a report: At Microsoft's annual executive huddle last month, the company's chief financial officer, Amy Hood, put up a slide that charted the number of users for its Copilot consumer AI tool over the past year. It was essentially a flat line, showing around 20 million weekly users. On the same slide was another line showing ChatGPT's growth over the same period, arching ever upward toward 400 million weekly users.

OpenAI's iconic chatbot was soaring, while Microsoft's best hope for a mass-adoption AI tool was idling. It was a sobering chart for Microsoft's consumer AI team and the man who's been leading it for the past year, Mustafa Suleyman. Microsoft brought Suleyman aboard in March of 2024, along with much of the talent at his struggling AI startup Inflection, in return for a $650 million licensing fee that made Inflection's investors whole, and then some.

[...] Yet from the very start, people inside the company told me they were skeptical. Many outsiders have struggled to make an impact or even survive at Microsoft, a company that's full of lifers who cut their tech teeth in a different era. My skeptical sources noted Suleyman's previous run at a big company hadn't gone well, with Google stripping him of some management responsibilities following complaints of how he treated staff, the Wall Street Journal reported at the time. There was also much eye-rolling at the fact that Suleyman was given the title of CEO of Microsoft AI. That designation is typically reserved for the top executive at companies it acquires and lets operate semi-autonomously, such as LinkedIn or Github.

Legacy search brand Yahoo has been working on its own web browser prototype, and says it would like to buy Google's Chrome if the company is forced by a court to sell it. From a report: The information came out during the fourth day of the Justice Department's remedies trial to rectify Google's search monopoly. The DOJ has -- among other proposals -- requested Judge Amit Mehta break up Google by requiring it sell its Chrome browser, which the agency says is a key distribution channel for its popular search engine that's amassed too much power for anyone else to compete. Yahoo isn't the only company interested in buying Chrome. While DuckDuckGo's CEO said they wouldn't be able to afford it, witnesses from Perplexity and OpenAI both expressed interest in the popular browser on the stand this week. Yahoo obviously isn't worth Chrome's estimated price tag of tens of billions of dollars. So the company is saying that its owner, the hedge fund giant Apollo, will help bankroll the purchase should the opportunity present itself.

Google's search AI is confidently generating explanations for nonexistent idioms, once again revealing fundamental flaws in large language models. Users discovered that entering any made-up phrase plus "meaning" triggers AI Overviews that present fabricated etymologies with unwarranted authority.

When queried about phrases like "a loose dog won't surf," Google's system produces detailed, plausible-sounding explanations rather than acknowledging these expressions don't exist. The system occasionally includes reference links, further enhancing the false impression of legitimacy.

Computer scientist Ziang Xiao from Johns Hopkins University attributes this behavior to two key LLM characteristics: prediction-based text generation and people-pleasing tendencies. "The prediction of the next word is based on its vast training data," Xiao explained. "However, in many cases, the next coherent word does not lead us to the right answer."

A Government Accountability Office report released this week reveals generative AI systems consume staggering amounts of water, with 250 million daily queries requiring over 1.1 million gallons -- all while companies provide minimal transparency about resource usage. The 47-page analysis [PDF] found cooling data centers -- which demand between 100-1000 megawatts of power -- constitutes 40% of their energy consumption, a figure expected to rise as global temperatures increase.

Water usage varies dramatically by location, with geography significantly affecting both water requirements and carbon emissions. Meta's Llama 3.1 405B model has generated 8,930 metric tons of carbon, compared to Google's Gemma2 at 1,247.61 metric tons and OpenAI's GPT3 at 552 metric tons. The report confirms generative AI searches cost approximately ten times more than standard keyword searches. The GAO asserted about persistent transparency problems across the industry, noting these systems remain "black boxes" even to their designers.

Five years removed from the onset of the Covid pandemic, Google is demanding that some remote employees return to the office if they want to keep their jobs and avoid being part of broader cost cuts at the company. CNBC reports: Several units within Google have told remote staffers that their roles may be at risk if they don't start showing up at the closest office for a hybrid work schedule, according to internal documents viewed by CNBC. Some of those employees were previously approved for remote work.

Google revealed in court that its Gemini AI chatbot reached 350 million monthly active users worldwide as of March 2025 -- up from 9 million daily users in October 2024. TechCrunch reports: Usage of Google's AI offerings has exploded in the last year. Gemini had just 9 million daily active users in October 2024, but last month, the company reportedly logged 35 million daily active users, according to its data. Gemini still lags behind the industry's most popular AI tools, however.

Google estimates that ChatGPT had roughly 600 million monthly active users in March, according to the company's data shown in court. That puts ChatGPT on a similar user base to Meta AI, which CEO Mark Zuckerberg said in September was nearing 500 million monthly users.

At Google's antitrust trial, OpenAI's head of product revealed the company would consider buying Chrome if regulators force Alphabet to sell it, arguing such a move could help improve ChatGPT's search capabilities. Reuters reports: ChatGPT head of product Nick Turley made the statement while testifying at trial in Washington where U.S. Department of Justice seeks to require Google to undertake far-reaching measures restore competition in online search. The judge overseeing the trial found last year that Google has a monopoly in online search and related advertising. Google has not offered Chrome for sale. The company plans to appeal the ruling that it holds a monopoly.

Turley wrote last year that ChatGPT was leading in the consumer chatbot market and did not see Google as its biggest competitor, according to an internal OpenAI document Google's lawyer showed at trial. He testified that the document was meant to inspire OpenAI employees and that the company would still benefit from distribution partnerships. Turley, a witness for the government, testified earlier in the day that Google shot down a bid by OpenAI to use its search technology within ChatGPT. OpenAI had reached out to Google after experiencing issues with its own search provider, Turley said, without naming the provider. ChatGPT uses technology from Microsoft's search engine, Bing. "We believe having multiple partners, and in particular Google's API, would enable us to provide a better product to users," OpenAI told Google, according to an email shown at trial.

OpenAI first reached out in July, and Google declined the request in August, saying it would involve too many competitors, according to the email. "We have no partnership with Google today," Turley said. The DOJ's proposal to make Google share search data with competitors as one means of restoring competition would help accelerate efforts to improve ChatGPT, Turley said. Search is a critical part of ChatGPT to provide answers to user queries that are up to date and factual, Turley said. ChatGPT is years away from its goal of being able to use its own search technology to answer 80% of queries, he added.

An anonymous reader quotes a report from Ars Technica: In 2019, Garrett Mitchell was already an Internet success. His YouTube channel, Cleetus McFarland, had over a million followers. If you perused the channel at that time, you would've found a range of grassroots motorsports videos with the type of vehicular shenanigans that earn truckloads of views. Some of those older videos include "BLEW BY A COP AT 120+mph! OOPS!," "THERE'S A T-REX ON THE TRACK!," and "Manual Transmission With Paddle Shifters!?!." Those videos made Mitchell, aka Cleetus McFarland, a known personality among automotive enthusiasts. But the YouTuber wanted more financial independence beyond the Google platform and firms willing to sponsor his channel. " after my YouTube was growing and some of my antics [were] getting videos de-monetized, I realized I needed a playground," Mitchell told Ars Technica in an email.

Mitchell found a road toward new monetization opportunities through the DeSoto Super Speedway. The Bradenton, Florida, track had changed ownership multiple times since opening in the 1970s. The oval-shaped racetrack is three-eighths of a mile long with 12-degree banking angles. By 2018, the track had closed its doors and was going unused. DeSoto happened to be next to Mitchell's favorite drag strip, giving the YouTuber the idea of turning it into a stadium where people could watch burnouts and other "massive, rowdy" ticketed events. Mitchell added: "So I sold everything I could, borrowed some money from my business manager, and went all in for [$]2.2 million." But like the rest of the world, Mitchell hit the brakes on his 2020 plans during COVID-19 lockdowns. Soon after his purchase, Mitchell couldn't use the track, renamed Freedom Factory, for large gatherings, forcing him to reconsider his plans. "We had no other option but to entertain the people somehow. And with no other racing goin' on anywhere, we bet big on making something happen. And it worked," Mitchell said. That "something" was a pay-per-view (PPV) event hosted from the Freedom Factory in April 2020.

The event led to others and, eventually, Mitchell running his own subscription video on demand (SVOD) service, FRDM+, which originally launched as Cleetervision in 2022. Today, a FRDM+ subscription costs $20 per month or $120 per year. A subscription provides access to an impressive library of automotive videos. Some are archived from Mitchell's YouTube channel. Other, exclusive videos feature content such as interviews with motorsports influencers and members of Mitchell's staff and crew, and outrageous motorsports stunts. You can watch videos from other influencers on FRDM+, and the business can also white-label its platform into other influencers' websites, too. "Today, bandwidth isn't a problem for FRDM+, and navigating the streaming service doesn't feel much different from something like Netflix," writes Ars Technica's Scharon Harding. "There are different 'channels' (grouped together by related content or ongoing series) on top and new releases and upcoming content highlighted below. There are horizontal scrolling rows, and many titles have content summaries and/or trailers. The platform also has a support section with instructions for canceling subscriptions."

"Due to wildly differing audiences, markets, costs, and scales, comparing FRDM+'s financials to the likes of Netflix and other mainstream streaming services is like comparing apples to oranges. But it's interesting to consider that FRDM+ has achieved profitability faster than some of those services, like Peacock, which also launched in 2020, and Apple TV+, which debuted in 2019."

An anonymous reader shares a report: In a shocking development, Google won't roll out a new standalone prompt for third-party cookies in Chrome. It's a move that amounts to a U-turn on the Chrome team's earlier updated approach to deprecating third-party cookies, announced in July last year, with the latest development bound to cause ructions across the ad tech ecosystem. "We've made the decision to maintain our current approach to offering users third-party cookie choice in Chrome, and will not be rolling out a new standalone prompt for third-party cookies," wrote Anthony Chavez, vp Privacy Sandbox at Google, in a blog post published earlier today (April 22). "Users can continue to choose the best option for themselves in Chrome's Privacy and Security Settings." However, it's not the end of Privacy Sandbox, according to Google, as certain initiatives incubated within the project are set to continue, such as its IP Protection for Chrome Incognito users, which will be rolled out in Q3.

Google pays Samsung an "enormous sum of money" every month to preinstall Google generative AI app, Gemini, on its phones and devices, according to court testimony, even though the company's practice of paying for installations has twice been found to violate the law. From a report: The company began paying Samsung for Gemini in January, according to Peter Fitzgerald, Google's vice president of platforms and device partnerships, who testified Monday in Washington federal court as part of the Justice Department's antitrust case. The contract, set to run at least two years, provides fixed monthly payments for each device that preinstalls Gemini and pays Samsung a percentage of the revenue Google earns from advertisements within the app, Fitzgerald told Judge Amit Mehta, who is overseeing the case.

Google has argued in court that the U.S. Department of Justice's proposal to break up its Chrome and Android businesses would weaken national security and harm the country's position in the global AI race, particularly against China. CNBC reports: The remedies trial in Washington, D.C., follows a judge's ruling in August that Google has held a monopoly in its core market of internet search, the most-significant antitrust ruling in the tech industry since the case against Microsoft more than 20 years ago. The Justice Department has called for Google to divest its Chrome browser unit and open its search data to rivals.

Google said in a blog post on Monday that such a move is not in the best interest of the country as the global battle for supremacy in artificial intelligence rapidly intensifies. In the first paragraph of the post, Google named China's DeepSeek as an emerging AI competitor. The DOJ's proposal would "hamstring how we develop AI, and have a government-appointed committee regulate the design and development of our products," Lee-Anne Mulholland, Google's vice president of regulatory affairs, wrote in the post. "That would hold back American innovation at a critical juncture. We're in a fiercely competitive global race with China for the next generation of technology leadership, and Google is at the forefront of American companies making scientific and technological breakthroughs."

Google is confronting an existential threat as the U.S. government tries to break up the company as punishment for turning its revolutionary search engine into an illegal monopoly. From a report: The drama began to unfold Monday in a Washington courtroom as three weeks of hearings kicked off to determine how the company should be penalized for operating a monopoly in search. In its opening arguments, federal antitrust enforcers also urged the court to impose forward-looking remedies to prevent Google from using artificial intelligence to further its dominance. "This is a moment in time, we're at an inflection point, will we abandon the search market and surrender them to control of the monopolists or will we let competition prevail and give choice to future generations," said Justice Department attorney David Dahlquist.

The proceedings, known in legal parlance as a "remedy hearing," are set to feature a parade of witnesses that includes Google CEO Sundar Pichai. The U.S. Department of Justice is asking a federal judge to order a radical shake-up that would ban Google from striking the multibillion dollar deals with Apple and other tech companies that shield its search engine from competition, share its repository of valuable user data with rivals and force a sale of its popular Chrome browser. Google's attorney, John Schmidtlein, said in his opening statement that the court should take a much lighter touch. He said the government's heavy-handed proposed remedies wouldn't boost competition but instead unfairly reward lesser rivals with inferior technology. "Google won its place in the market fair and square," Schmidtlein said.

In February tech journalist Nicholas Carr published Superbloom: How Technologies of Connection Tear Us Apart.

A University of Virginia academic journal says the book "appraises the past and present" of information technology while issuing "a warning about its future." And specifically Carr argues that the government ignored historic precedents by not regulating the early internet sometime in the 1990s. But as he goes on to remind us, the early 1990s were also when the triumphalism of America's Cold War victory, combined with the utopianism of Silicon Valley, convinced a generation of decision-makers that "an unfettered market seemed the best guarantor of growth and prosperity" and "defending the public interest now meant little more than expanding consumer choice." So rather than try to anticipate the dangers and excesses of commercialized digital media, Congress gave it free rein in the Telecommunications Act of 1996, which, as Carr explains,

"...erased the legal and ethical distinction between interpersonal communication and broadcast communications that had governed media in the twentieth century. When Google introduced its Gmail service in 2004, it announced, with an almost imperial air of entitlement, that it would scan the contents of all messages and use the resulting data for any purpose it wanted. Our new mailman would read all our mail."

As for the social-media platforms, Section 230 of the Act shields them from liability for all but the most egregiously illegal content posted by users, while explicitly encouraging them to censor any user-generated content they deem offensive, "whether or not such material is constitutionally protected" (emphasis added). Needless to say, this bizarre abdication of responsibility has led to countless problems, including what one observer calls a "sociopathic rendition of human sociability." For Carr, this is old news, but he warns us once again that the compulsion "to inscribe ourselves moment by moment on the screen, to reimagine ourselves as streams of text and image...[fosters] a strange, needy sort of solipsism. We socialize more than ever, but we're also at a further remove from those we interact with."
Carr's book suggests "frictional design" to slow posting (and reposting) on social media might "encourage civil behavior" — but then decides it's too little, too late, because our current frictionless efficiency "has burrowed its way too deeply into society and the social mind."

Based on all of this, the article's author looks ahead to the next revolution — AI — and concludes "I do not think it wise to wait until these kindly bots are in place before deciding how effective they are. Better to roll them off the nearest cliff today..."

Chad Anderson is the founder/managing partner of the early-stage VC Space Capital (and an investor in SpaceX, along with dozens of other space companies). Space Capital produces quarterly reports on the space economy, and he says today, unlike 2021, "the froth is gone. But so is the hype. What's left is a more grounded — and investable — space economy."

On Yahoo Finance he shares several of the report's insights — including the emergence of "investable opportunities across defense-oriented startups in space domain awareness, AI-driven command systems, and hardened infrastructure." The same geopolitical instability that's undermining public markets is driving national urgency around space resilience. China's simulated space "dogfights" prompted the US Department of Defense to double down on orbital supremacy, with the proposed "Golden Dome" missile shield potentially unleashing a new wave of federal spending...

Defense tech is on fire, but commercial location-based services and logistics are freezing over. Companies like Shield AI and Saronic raised monster rounds, while others are relying on bridge financings to stay afloat...

Q1 also saw a breakout quarter for geospatial artificial intelligence (GeoAI). Software developer Niantic launched a spatial computing platform. SkyWatch partnered with GIS software supplier Esri. Planet Labs collaborated with Anthropic. And Xona Space Systems inked a deal with Trimble to boost precision GPS. This is the next leg of the space economy, where massive volumes of satellite data is finally made useful through machine learning, semantic indexing, and real-time analytics.

Distribution-layer companies are doing more with less. They remain underfunded relative to infrastructure and applications but are quietly powering the most critical systems, such as resilient communications, battlefield networks, and edge-based geospatial analysis. Don't let the low round count fool you; innovation here is quietly outpacing capital.
The article includes several predictions, insights, and possible trends (going beyond the fact that defense spending "will carry the sector...")

• "AI's integration into space (across geospatial intelligence, satellite communications, and sensor fusion) is not a novelty. It's a competitive necessity."

• "Focusing solely on rockets and orbital assets misses where much of the innovation and disruption is occurring: the software-defined layers that sit atop the physical backbone..."

• "For years, SpaceX faced little serious competition, but that's starting to change." [He cites Blue Origin's progress toward approval for launching U.S. military satellites, and how Rocket Lab and Stoke Space "have also joined the competition for lucrative government launch contracts." Even Relativity Space may make a comeback, with former GOogle CEO Eric Schmidt acquiring a controlling stake.]

• "An infrastructure reset is coming. The imminent ramp-up of SpaceX's Starship could collapse the cost structure for the infrastructure layer. When that happens, legacy providers with fixed-cost-heavy business models will be at risk. Conversely, capital-light innovators in station design, logistics, and in-orbit servicing could suddenly be massively undervalued."

In NoSQL database news, Arch Linux "is the latest Linux distribution replacing its Redis packages with the Valkey fork," reports Phoronix.

Valkey is backed by the Linux Foundation, Google, Amazon Web Services, and Oracle, which the article points out is due to Redis's decision last year to shift the upstream Redis license from a BSD 3-clause to RSALv2 and SSPLv1. Valkey is replacing Redis in the Arch Linux extra repository and after a two week period the Redis package will be moved out to AUR and receive no further updates. Users are encouraged to migrate to Valkey as soon as possible.

"Russia is automating the spread of false information to fool AI chatbots," reports the Washington Post. (When researchers checked 10 chatbots, a third of the responses repeated false pro-Russia messaging.)

The Post argues that this tactic offers "a playbook to other bad actors on how to game AI to push content meant to inflame, influence and obfuscate instead of inform," and calls it "a fundamental weakness of the AI industry." Chatbot answers depend on the data fed into them. A guiding principle is that the more the chatbots read, the more informed their answers will be, which is why the industry is ravenous for content. But mass quantities of well-aimed chaff can skew the answers on specific topics. For Russia, that is the war in Ukraine. But for a politician, it could be an opponent; for a commercial firm, it could be a competitor. "Most chatbots struggle with disinformation," said Giada Pistilli, principal ethicist at open-source AI platform Hugging Face. "They have basic safeguards against harmful content but can't reliably spot sophisticated propaganda, [and] the problem gets worse with search-augmented systems that prioritize recent information."

Early commercial attempts to manipulate chat results also are gathering steam, with some of the same digital marketers who once offered search engine optimization — or SEO — for higher Google rankings now trying to pump up mentions by AI chatbots through "generative engine optimization" — or GEO.
Our current situation "plays into the hands of those with the most means and the most to gain: for now, experts say, that is national governments with expertise in spreading propaganda." Russia and, to a lesser extent, China have been exploiting that advantage by flooding the zone with fables. But anyone could do the same, burning up far fewer resources than previous troll farm operations... In a twist that befuddled researchers for a year, almost no human beings visit the sites, which are hard to browse or search. Instead, their content is aimed at crawlers, the software programs that scour the web and bring back content for search engines and large language models. While those AI ventures are trained on a variety of datasets, an increasing number are offering chatbots that search the current web. Those are more likely to pick up something false if it is recent, and even more so if hundreds of pages on the web are saying much the same thing...

The gambit is even more effective because the Russian operation managed to get links to the Pravda network stories edited into Wikipedia pages and public Facebook group postings, probably with the help of human contractors. Many AI companies give special weight to Facebook and especially Wikipedia as accurate sources. (Wikipedia said this month that its bandwidth costs have soared 50 percent in just over a year, mostly because of AI crawlers....) Last month, other researchers set out to see whether the gambit was working. Finnish company Check First scoured Wikipedia and turned up nearly 2,000 hyperlinks on pages in 44 languages that pointed to 162 Pravda websites. It also found that some false information promoted by Pravda showed up in chatbot answers.
"They do even better in such places as China," the article points out, "where traditional media is more tightly controlled and there are fewer sources for the bots." (The nonprofit American Sunlight Project calls the process "LLM grooming".)

The article quotes a top Kremlin propagandist as bragging in January that "we can actually change worldwide AI."

Nature has published an analysis of the 21st century's most-cited scientific papers, revealing a surprising pattern: breakthrough discoveries like mRNA vaccines, CRISPR, and gravitational waves don't make the list. Instead, a 2016 Microsoft paper on "deep residual learning" networks claims the top spot, with citations ranging from 103,756 to 254,074 depending on the database.

The list overwhelmingly features methodology papers and software tools rather than groundbreaking discoveries. AI research dominates with four papers in the top ten, including Google's 2017 "Attention is all you need" paper that underpins modern language models.

The second-most-cited paper -- a 2001 guide for analyzing gene expression data -- was explicitly created to be cited after journal reviewers rejected references to a technical manual. As sociologist Misha Teplitskiy noted, "Scientists say they value methods, theory and empirical discoveries, but in practice the methods get cited more."

Microsoft has introduced BitNet b1.58 2B4T, the largest-scale 1-bit AI model to date with 2 billion parameters and the ability to run efficiently on CPUs. It's openly available under an MIT license. TechCrunch reports: The Microsoft researchers say that BitNet b1.58 2B4T is the first bitnet with 2 billion parameters, "parameters" being largely synonymous with "weights." Trained on a dataset of 4 trillion tokens -- equivalent to about 33 million books, by one estimate -- BitNet b1.58 2B4T outperforms traditional models of similar sizes, the researchers claim.

BitNet b1.58 2B4T doesn't sweep the floor with rival 2 billion-parameter models, to be clear, but it seemingly holds its own. According to the researchers' testing, the model surpasses Meta's Llama 3.2 1B, Google's Gemma 3 1B, and Alibaba's Qwen 2.5 1.5B on benchmarks including GSM8K (a collection of grade-school-level math problems) and PIQA (which tests physical commonsense reasoning skills). Perhaps more impressively, BitNet b1.58 2B4T is speedier than other models of its size -- in some cases, twice the speed -- while using a fraction of the memory.

There is a catch, however. Achieving that performance requires using Microsoft's custom framework, bitnet.cpp, which only works with certain hardware at the moment. Absent from the list of supported chips are GPUs, which dominate the AI infrastructure landscape.