Shares of the database company MongoDB Inc. shot up both before and after the close of the regular trading session today following a solid first-quarter financial report that saw it blow past analysts’ expectations. The company also posted solid guidance, in what was taken as another sign that fears of artificial intelligence killing the software […] The post MongoDB posts stellar earnings and revenue beat as software rebound accelerates appeared first on SiliconANGLE.
Articles
Browse all AI articles. Use filters to narrow down results.
Most Recent Articles
CBS News fires Cecilia Vega "60 Minutes" as editor in chief Bari Weiss overhauled the show's leadership. Taylor Hill/WireImage The "60 Minutes" correspondent says she was let go almost a year before her contract was up. In a parting statement, she accused the network of "censorship, both imposed and self-driven." Her firing came as Bari Weiss replaced executive producer Tanya Simon with Nick Bilton. "60 Minutes" correspondent Cecilia Vega said Thursday that she was fired and issued a blistering statement in which she accused CBS News of "censorship, both imposed and self-driven." Vega, who joined the prestigious TV newsmagazine from ABC News in 2023, said her contract wasn't up until March 2027. The network did not respond to a request for comment. Vega got the ax in a major shakeup at the long-running show. CBS News Editor in Chief Bari Weiss ousted the executive producer, Tanya Simon, and replaced her with Nick Bilton, a former New York Times tech columnist and Vanity Fair contributor. Sharyn Alfonsi, a veteran "60 Minutes" correspondent who clashed with Weiss over a segment on Donald Trump's immigration policies, also exited after her contract was not renewed. Read Vega's full message here: I was fired today. My contract as a correspondent for 60 Minutes was not set to expire until March 2027. I have the utmost respect and admiration for my colleagues at 60 Minutes and the stories that air every Sunday. But I very much fear what comes next for and the future of the legendary broadcast. In recent months, my producing teams and I have experienced efforts to insert political bias into our stories. Reporting teams have held back on submitting story pitches about important news topics out of fear of the internal repercussions. Let's call this what it is: censorship, both imposed and self-driven. It is dangerous for the show and dangerous for democracy. I held the line and refused to incorporate suggestions that offend the conscience, a phrase I borrow from a colleague who has also fought to keep questionable editorial suggestions away from the facts. I know from many conversations with colleagues that many producing teams and correspondents working on the show today have had to fight to maintain editorial independence with regularity. I am far from the only 60 Minutes correspondent who has asked herself, "What is my personal red line? How much can I push back before I pay the price?" I am proud of the work I did for 60 Minutes. This season alone I was part of teams that won two of the highest honors in our profession — a George Polk award and a duPont-Columbia award for our coverage of Venezuelan migrants sent by the Trump administration to El Salvador's Cecot prison. And not for nothing, I climbed to Mount Everest. I also walk away with an honor no one can take from me: I was the first Latina correspondent to ever be on 60 Minutes. Today I lost an amazing job. But I still have my integrity. To my former colleagues, continue to hold the line. Read the original article on Business Insider
<p>The most interesting thing about <a href="https://www.anthropic.com/news/series-h">Anthropic's $65B Series H announcement</a> is this line (emphasis mine):</p> <blockquote> <p>Since our Series G in February, adoption has continued to grow across global enterprise customers, and our run-rate revenue crossed <strong>$47 billion</strong> earlier this month.</p> </blockquote> <p>Anthropic have made a bit of a habit of sharing their "run-rate revenue" in this kind of announcement, which is an annualized projection of their current revenue - typically calculated by taking the most recent month and multiplying by 12. <strong>Update</strong>: here's <a href="https://simonwillison.net/2026/May/31/anthropic-run-rate/">a leaked description of their run-rate formula</a>.</p> <p>Earlier this year:</p> <ul> <li>Apr 6, 2026 in <a href="https://www.anthropic.com/news/google-broadcom-partnership-compute">Anthropic expands partnership with Google and Broadcom</a>: "Our run-rate revenue has now surpassed <strong>$30 billion</strong>—up from approximately <strong>$9 billion</strong> at the end of 2025."</li> <li>Feb 12, 2026 in <a href="https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation">Anthropic raises $30 billion in Series G</a>: "Today, our run-rate revenue is <strong>$14 billion</strong>, with this figure growing over 10x annually in each of those past three years."</li> </ul> <p>I had <a href="https://claude.ai/share/f52e82bd-7e09-49a5-b658-0b9999ce5a45">Claude Opus 4.8 make me</a> this chart using <a href="https://matplotlib.org/">Matplotlib</a> (Claude: "a data line chart is more straightforward matplotlib work—not really a design piece"):</p> <p><img alt="Line chart titled "Run-rate revenue" with y-axis "Run-rate revenue ($bn)" from $0bn to $50bn, showing four data points rising sharply: Dec 31 2025 $9bn, Feb 12 2026 $14bn, Apr 1 2026 $30bn, May 7 2026 $47bn." src="https://static.simonwillison.net/static/2026/anthropic-run-rate-extra-axis.png" /></p> <p>Back in April <a href="https://www.axios.com/2026/04/13/anthropic-revenue-growth-ai">Axios CEO Jim VandeHei wrote</a> that he could not find "any company — in any industry, in any era — that has scaled organic revenue this quickly at this level as Anthropic" - and that was when they were at a paltry $30 billion.</p> <p>(Also <a href="https://www.axios.com/2026/05/28/ai-spending-roi-enterprise-costs">in Axios today</a> is an anonymously sourced note that "An AI consultant tells Axios one of their clients recently spent half a billion dollars in a single month after failing to put usage limits on Claude licenses for employees" - times that by 12 and you get an extra $6 billion in annualized run-rate!)</p> <p>Ed Zitron was <a href="https://www.wheresyoured.at/anthropics-profitability-swindle/">extremely skeptical of that $30 billion number</a> - I wonder if his skepticism will update for the new $47 billion figure.</p> <p>I've seen a few people dismiss this as untrustworthy, because the numbers come from Anthropic. That doesn't hold up: these numbers were included in announcements of their fundraises, and lying to investors who just put in $65 billion would be securities fraud. They're even less likely to lie given that the real numbers will no doubt come out in their S-1 when they file for their IPO.</p> <p>Tags: <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/ai">ai</a></p>
AI applications are moving beyond text generation to multimodal systems that can perceive, search, and reason across images, documents, video, and... AI applications are moving beyond text generation to multimodal systems that can perceive, search, and reason across images, documents, video, and language in real time—turning fragmented information into actionable insights. Step 3.7 Flash, the latest from StepFun, brings these capabilities to production and enterprise-scale, available on NVIDIA-accelerated infrastructure. It is a 198B… Source
OpenAI shares guidance on third-party AI evaluations, covering how to assess model capabilities, safeguards, and validity for frontier systems.
Skill evals for measurable agent self-improvement
Customizing claude skill-creator for iterative skill development
<p>Anthropic shipped <a href="https://www.anthropic.com/news/claude-opus-4-8">Claude Opus 4.8</a> today. My favourite thing about it is this note in the release announcement:</p> <blockquote> <p>Users will find Opus 4.8 to be a modest but tangible improvement on its predecessor. There’s still more to be done: we’re working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost.</p> </blockquote> <p>It's so refreshing to see an AI lab honestly describe a release as a minor incremental improvement over the previous model!</p> <p>Honesty seems to be a theme. Here's my other favorite note from that announcement:</p> <blockquote> <p>One of the most prominent improvements in Opus 4.8 is its <em>honesty</em>. We train all our models to be honest---for instance, to avoid making claims that they can't support. But a general problem with AI models is that they sometimes jump to conclusions, confidently claiming to have made progress in their work despite the evidence being thin. Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims. This is borne out in <a href="https://www.anthropic.com/claude-opus-4-8-system-card">our evaluations</a>, which show that Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.</p> </blockquote> <p>That linked system card includes the following:</p> <blockquote> <p>Claude Opus 4.8 had the lowest incorrect-rate of the six models on every benchmark—the most direct measure of factual hallucination. It achieved this mainly by abstaining on questions about which it was uncertain rather than by answering more questions correctly.</p> </blockquote> <h4 id="model-characteristics">Model characteristics</h4> <p>Not much has changed since 4.7.</p> <p>It's priced the same as Opus 4.5/4.6/4.7 - $5/million input and $25 per million output. "Fast mode" is twice that price, which is a significant reduction from their previous models - fast mode on 4.6/4.7 remains at $30/$150. Note that <a href="https://platform.claude.com/docs/en/build-with-claude/fast-mode">fast mode</a> is only available to organizations that are part of the research preview, "Contact your account manager to request access".</p> <p>Both the reliable knowledge cutoff and the training data cutoff are January 2026, the same as for 4.7.</p> <p>The context window is still 1,000,000 tokens, and the max output is 128,000 tokens.</p> <p>The <a href="https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-8">What's new in Claude Opus 4.8</a> document has some of the more interesting details. These caught my eye:</p> <blockquote> <p><strong>Mid-conversation system messages</strong>. Claude Opus 4.8 accepts <code>role: "system"</code> messages immediately after a user turn in the <code>messages</code> array (subject to <a href="https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages#limitations">placement rules</a>). This lets you append updated instructions later in a long-running conversation without restating the full system prompt, which preserves <a href="https://platform.claude.com/docs/en/build-with-claude/prompt-caching">prompt cache</a> hits on the earlier turns and reduces input cost on agentic loops.</p> </blockquote> <p>See also <a href="https://github.com/anthropics/anthropic-sdk-python/commit/2b826760101664ef89db42132932f53ba97c894d#diff-a947c9c02eab58e8ddbe799a11832d533836d242e07c7251997f8543f0981f2f">this update</a> to the Anthropic Python SDK. Being able to steer the system prompt mid-conversation sounds really powerful. I was worried this would be incompatible with the abstraction provided by my own <a href="https://llm.datasette.io/en/stable/python-api.html#system-prompts">LLM library</a>, which expects a single system prompt per conversation... but it turns out my recent <a href="https://simonwillison.net/2026/Apr/29/llm/">redesign</a> should handle that <a href="https://github.com/simonw/llm-anthropic/issues/73">just fine</a>.</p> <blockquote> <p><strong>Lower prompt cache minimum</strong>. The minimum cacheable prompt length on Claude Opus 4.8 is 1,024 tokens, lower than on Claude Opus 4.7.</p> </blockquote> <p>I checked and 4.7's minimum <a href="https://platform.claude.com/docs/en/build-with-claude/prompt-caching#cache-limitations">was 4,096</a>.</p> <h4 id="and-some-pelicans">And some pelicans</h4> <p>Here are <a href="https://tools.simonwillison.net/markdown-svg-renderer#url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Ffea4f7546626d627862dc241a4e3a86a">pelicans riding bicycles</a> for all five thinking levels, <code>low</code>, <code>medium</code>, <code>high</code>, <code>xhigh</code>, and <code>max</code>:</p> <div style="display: grid; grid-template-columns: repeat(6, 1fr); gap: 1rem; max-width: 900px; margin: 0 auto;"> <figure style="grid-column: span 2; margin: 0; text-align: center;"> <img src="https://static.simonwillison.net/static/2026/claude-opus-4.8-low.png" alt="Flat-style cartoon illustration of a white duck with an orange beak and legs riding a black bicycle, its feet on the pedals, against a blue sky and green grass background." style="width: 100%; height: auto; border: 1px solid #ccc;" /> <figcaption style="margin-top: 0.5rem; font-family: system-ui, sans-serif; font-weight: bold;"> <a href="https://gist.github.com/simonw/fea4f7546626d627862dc241a4e3a86a#response">low</a> </figcaption> </figure> <figure style="grid-column: span 2; margin: 0; text-align: center;"> <img src="https://static.simonwillison.net/static/2026/claude-opus-4.8-medium.png" alt="Flat-style illustration of a white egret or heron with an orange beak and legs riding a black bicycle, against a blue sky and green grass background." style="width: 100%; height: auto; border: 1px solid #ccc;" /> <figcaption style="margin-top: 0.5rem; font-family: system-ui, sans-serif; font-weight: bold;"> <a href="https://gist.github.com/simonw/fea4f7546626d627862dc241a4e3a86a#response-1">medium</a> </figcaption> </figure> <figure style="grid-column: span 2; margin: 0; text-align: center;"> <img src="https://static.simonwillison.net/static/2026/claude-opus-4.8-high.png" alt="Cartoon illustration of a white duck with an orange beak riding a black bicycle, against a light blue sky with a pale yellow sun in the upper left and a green ground line at the bottom." style="width: 100%; height: auto; border: 1px solid #ccc;" /> <figcaption style="margin-top: 0.5rem; font-family: system-ui, sans-serif; font-weight: bold;"> <a href="https://gist.github.com/simonw/fea4f7546626d627862dc241a4e3a86a#response-2">high</a> </figcaption> </figure> <figure style="grid-column: span 3; margin: 0; text-align: center;"> <img src="https://static.simonwillison.net/static/2026/claude-opus-4.8-xhigh.png" alt="Cartoon illustration of a white pelican with an orange beak riding a black bicycle, its orange legs extending down to the pedals, against a blue sky with a yellow sun and green ground." style="width: 100%; height: auto; border: 1px solid #ccc;" /> <figcaption style="margin-top: 0.5rem; font-family: system-ui, sans-serif; font-weight: bold;"> <a href="https://gist.github.com/simonw/fea4f7546626d627862dc241a4e3a86a#response-3">xhigh</a> </figcaption> </figure> <figure style="grid-column: span 3; margin: 0; text-align: center;"> <img src="https://static.simonwillison.net/static/2026/claude-opus-4.8-max.png" alt="Cartoon illustration of a white pelican with an orange beak riding a red bicycle on green grass, against a light blue sky with a fluffy white cloud and a yellow sun." style="width: 100%; height: auto; border: 1px solid #ccc;" /> <figcaption style="margin-top: 0.5rem; font-family: system-ui, sans-serif; font-weight: bold;"><a href="https://gist.github.com/simonw/fea4f7546626d627862dc241a4e3a86a#response-4">max</a></figcaption> </figure> </div> <p>This time I ran them using the <a href="https://llm.datasette.io/en/stable/usage.html">LLM CLI</a>, exported the logs to Markdown and then had Claude Opus 4.8 <a href="https://github.com/simonw/tools/commit/71e4944766b577a327ff048cc63b739ba4cbade9">build me</a> an HTML tool that could render that Markdown with the <code>svg</code> fenced code blocks displayed as SVGs on the page.</p> <p>(I later had GPT-5.5 xhigh in Codex <a href="https://gist.github.com/simonw/bb5a267f8144dfe4e92e50a014e49e98">update that code</a> to remove any XSS holes. I'm sure Claude could have done that if I'd asked, but GPT-5.5 is my code security blanket at the moment.)</p> <p>The max one was clearly the best, but it did take 25 input, 17,167 output tokens for a total cost of <a href="https://www.llm-prices.com/#it=25&ot=17167&ic=5&oc=25&sel=claude-opus-4-5">43 cents</a>!</p> <p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/claude">claude</a>, <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle">pelican-riding-a-bicycle</a>, <a href="https://simonwillison.net/tags/llm-release">llm-release</a></p>
<p><strong>Release:</strong> <a href="https://github.com/simonw/llm-anthropic/releases/tag/0.25.1">llm-anthropic 0.25.1</a></p> <blockquote> <ul> <li>New model: <a href="https://www.anthropic.com/news/claude-opus-4-8">Claude Opus 4.8</a> (<code>claude-opus-4.8</code>).</li> <li>New <code>-o fast 1</code> option for <a href="https://platform.claude.com/docs/en/build-with-claude/fast-mode">fast mode</a>, for organizations with that feature enabled on their account.</li> <li>Default max_tokens for each model now defaults to that model's maximum output rather than 8,192. <a href="https://github.com/simonw/llm-anthropic/issues/72">#72</a></li> </ul> </blockquote> <p>See also my <a href="https://simonwillison.net/2026/May/28/claude-opus-4-8/">notes on Opus 4.8</a> - I used this new release of <code>llm-anthropic</code> to generate the pelicans.</p>
I thought focusing on grades would push my son harder. It crushed his confidence instead.
The author said she stoped focusing on her kids' grades first when reviewing the report cards. Jeffrey Coolidge/Getty Images For years, I fixated on grades, and slowly watched my son lose confidence because of it. Realizing I was transferring my own performance anxiety onto him forced me to change. I now read behavior comments first, and it has transformed report card day in our home. Life taught me a painful truth. The mistakes we make as new parents are often our biggest opportunities to learn and improve, but our children, especially the firstborns, bear the brunt of that learning curve. At least that's how it is in my family, and I suspect in many others as well. During the first few years of my oldest son's schooling, I used to make report card day a big deal. It took a painful realization to make me adopt a much healthier, highly successful approach instead, one that completely changed our family dynamics for the better. Earlier, my focus when opening a report card was entirely academic Even though I always believed I was a parent who prioritized character, manners, and behavior, my inexperience blinded me to fixate solely on grades, especially the lower ones, a habit that was hampering my son's confidence. When the grades arrived, I used to pinpoint the few flaws that were noted before celebrating any successes. I would obsess over why he received below-average marks in a few areas. I was focusing on the subjects where he lost points, overlooking the ones where he excelled. The author said that she noticed that focusing on academic grades on her oldest child's report card was creating unnecessary anxiety. Now she's changed her approach for all of her children. Courtesy of Ariba Mobin. Now I see that I was wrong to believe that showing strictness was necessary so my son would know what to work on. In reality, I was ignoring, or worse, undermining his strengths. His anxiety became so strong that before even handing the report card over, he would defensively list the subjects where he scored average before mentioning his high marks. The sad truth was that I was transferring my own performance anxiety onto him, making him think that he needed to show and justify his weakness before showing me his achievements. Watching my son lose his confidence changed the way I look at report cards This pattern continued for a few years until he reached third grade, around the same time my younger son started preschool. Seeing my eldest, a sweet, obedient, and helpful child, lose his confidence broke my heart. He was becoming blind to his genuine strengths, like his amazingly natural aptitude for math and general knowledge and his caring attitude toward all. A change was necessary I realized I had to change. Good character matters more than a perfect scorecard, and as a mother, it was my responsibility to protect my children's self-esteem. I needed them to understand that behavior and conduct mattered far more than a letter grade, and that focusing on character would naturally give them the tools to thrive in every aspect of life, including academics. Thankfully, our children's school places a wonderful emphasis on character building, featuring a detailed, dedicated section for behavioral and conduct remarks. On the next report card day, my husband and I agreed on a strategy: we would first acknowledge and celebrate those behavioral remarks. This new strategy worked well. Both of my sons looked surprised as I ignored the academic result section and read aloud, with visible pride, their teachers' praise of their conduct. Watching their beaming smiles made my heart swell. Once the tension had completely eased, we casually moved to discussing their good grades. Only at the very end did we address the subjects that needed improvement. Because the children were entirely at ease, they were incredibly receptive to the feedback. They were not defensive and actually listened. Ever since we made result day less stressful, the results have been remarkable Now, as all three of my children go to school and understand our viewpoint, report card day is no longer a scary day of judgment in our home. My children know that mom and dad will celebrate who they are as people before they look at a single grade. They regularly bring home certificates for good behavior, and they are so proud to show them off before we even look at the grades, celebrating them just as much as if they had gotten an A in a subject. My oldest is now in seventh grade, and the boy who used to pre-justify his weak grades now hands me his report card without a second thought. Academics still matter in our home. The difference is that our children now approach their studies without fear, and that change alone has done more for their grades than any pressure ever did. Read the original article on Business Insider
The seller of a $2.9 million San Francisco home is accepting payment in Anthropic or OpenAI stock
A seller in San Francisco is advertising that it is open to payment in OpenAI or Anthropic stock. Zillow Less than 24 hours after the offer went live, the agent said she had been flooded with interest. The seller is a local "high-end luxury developer" who is a big believer in Anthropic and OpenAI. This is not the first Bay Area house to ask for stock in AI companies. The listing for a newly restored three-bedroom home in San Francisco's Duboce Triangle neatly captures the city's AI-fueled real estate frenzy in one line: "Anthropic or OpenAI stock will be considered as payments." It has been less than 24 hours since the offer went live, and the listing agent, Rachel Swann, told Business Insider she has been flooded with interest. "My phone has been blowing up," Swann said. "It seems to be something that's very exciting for people." San Francisco real estate and stock in the hottest AI companies have one thing in common: high demand and low supply, which naturally drives up prices. The 2,495-square-foot Duboce Triangle home is asking $2,995,000. It offers a slew of luxurious amenities, including 10-foot ceilings, custom cabinetry, remote-controlled solar skylights, and a deep-soaking tub, but the listing leads with an offer that could only happen in 2026 San Francisco —the offer of accepting stock in two of the most valuable AI companies on the planet. The seller is a local "high-end luxury developer" who is a big believer in Anthropic and OpenAI, according to Swann. Swann says they got the idea after seeing a wave of OpenAI and Anthropic employees and investors come to open houses who have "nowhere to live" because the housing supply is so tight and all their wealth is tied up in illiquid stock. "It's a perfect storm," she said. "You wouldn't believe how many people I see come through who say, 'We can't touch our stock quite yet, but we would love to buy something.'" This is not the first Bay Area house to ask for stock in AI companies. Last month, Storm Duncan, a tech banker, offered his $4.8 million Marin County estate in exchange for shares in Anthropic. He told Business Insider he received multiple offers, including some from Anthropic employees. In contrast to the anemic national housing market, San Francisco home prices have skyrocketed 19% year over year, driven by a surge in AI valuations. On average, homes sell after 14 days on the market, compared to 20 days last year, according to Zillow data. Tales abound about buyers being flooded with multiple offers and houses going for significantly over their listing prices. Rental prices are rising by more than 21%, with the average one-bedroom apartment costing $4,000 a month, according to Zumper. As OpenAI and Anthropic prepare to go public, housing prices are expected to soar further as employees and early investors cash out and buy homes with their riches. "There's probably a decent number of people who are sitting in a one-bedroom apartment in San Francisco even though they're earning $400,000 a year and are worth a $100 million," Duncan told Business Insider a few weeks ago. "But they can't access that because their stock is so illiquid, so this gives them an opportunity to diversify." The listing for Duncan's house has since been removed. Asked for comment, Duncan texted: "Can't talk about it. Apologies." Swann says she is hoping to have all offers for the Duboce Triangle house in hand next week. "If somebody can buy something before they realize their paper wealth, why not?" she said. "People create things that have only been in their dreams all the time in San Francisco, so why not do it with real estate?" Read the original article on Business Insider
Walkable
Discussion | Link
Jaimie Ding / Associated Press: California AG Rob Bonta sues 23andMe, alleging it failed to protect sensitive user data in a 2023 breach that affected ~7M people across the US — California's attorney general sued the genetic testing company formerly known as 23andMe on Thursday, alleging it failed to protect sensitive user data …
Samantha Subin / CNBC: Okta reports Q1 revenue up 11% YoY to $765M, vs. $752M est., says the agentic AI build-out is spiking demand for its identity tools; OKTA jumps 7%+ after hours — Okta beat Wall Street's fiscal first-quarter estimates after the bell on Thursday as demand for identity security tools spikes …
Protests over data centers, like this one in Albany, N.Y., have swept the country amid concerns over environmental damage and utility prices. Will Waldron/Albany Times Union via Getty Images Coachella City Council considers halting Stronghold Power Systems' data center development. Hundreds of residents showed up to protest the proposal at a meeting on Wednesday night. Residents voice concerns over environmental impact and proximity of data centers to homes. A mass outpouring of data center opponents has prompted the California city of Coachella to consider a moratorium on development of the facilities there. It's the latest example of community backlash tempering local authorities' acceptance of data centers — which are being built across the country to accommodate the rise of AI. After hundreds of people showed up to protest at the Coachella City Council meeting on Wednesday evening, all four council members signaled support for a temporary halt to Stronghold Power Systems' proposal so the city can study the issue further. They also voted 4-0 to hire outside legal counsel to review an agreement for a city-controlled utility to provide power for the project. Planning documents show the Coachella Valley Technology Campus would span up to 450 acres and eventually include six data centers. The initial phase of the project would include three data center buildings, each about 1 million square feet, with a capacity of 90 megawatts per facility, Stronghold has said. The proposed data center location is at 52nd Ave & Fillmore St, Coachella, CA 92274. Screenshot Over the course of the five-hour meeting, residents voiced concerns about water use, air pollution, energy demands, and the project's proximity to homes and schools. Protesters gathered outside city hall, holding signs with slogans like "Defend Coachella, no data centers" and "Protect our environment, not tech profits." "This valley is already facing extreme heat, water scarcity, and infrastructure strain," Coachella resident Adriana Suarez told the council. "These data centers do not belong here, and you guys know it." Stronghold CEO Scott Bailey defended the proposal during the meeting and said that opponents were spreading "false information." Bailey said the project would use a closed-loop cooling system that would not tap the city's drinking water and that the development could generate more than $20 million annually for the city. "It goes on and on and on about the benefits," Bailey said. "They obviously don't want to hear it, but I'm here for you guys." Mayor Frank Figueroa said he would support a moratorium and a more permanent ordinance on data centers. "There was not one resident that came, not today, not the other council meetings, not the town hall, in support for the data center," Councilmember Yadira Perez said. The council could vote on a moratorium at a special meeting tentatively scheduled for June 3. The turn of events in Coachella has played out in communities across the country, though other data centers have been approved over local protests. County commissioners in Box Elder, Utah, recently approved a massive data center known as the Stratos Project — backed by "Shark Tank" personality Kevin O'Leary — despite ferocious opposition. Read the original article on Business Insider
2027 Audi RS5 first drive: A performance PHEV with split personalities
Audi has developed an entirely new electric torque-vectoring rear differential.
Rafe Rosner-Uddin / Financial Times: Sources: Amazon has shut down an internal leaderboard that tracked employees' use of AI tools after workers tried to boost their scores with needless tasks — Senior executive Dave Treadwell tells staff 'don't use AI just for the sake of using AI' as costs rise
Fine-tuning tests show "bias ... toward confidently representing the claims as true."
The internet is being rebuilt for machines
As AI agents move from experiments to production, AWS, Cloudflare, and others are redesigning cloud infrastructure for a future dominated by machine-generated internet traffic instead of human users.
Axios: Sources: Airwallex raised new funding led by Lee Fixel's Addition at a ~$12B valuation, up from $8B late last year, and hit $1.5B in ARR, up from $1B in October — Payments firm Airwallex has raised new funding at around a $12 billion valuation led by Lee Fixel's Addition, Axios has learned.