DeepMind's newest AI system has found ways to speed up calculations that stumped mathematicians since 1969. The tool is already saving money in Google's data centers - and might reshape how we write computer code. Its first real-world results point to an unexpected shift in algorithm design.
OpenAI just added a specialized coding model to ChatGPT, but that's not the whole story. The company also opened up about its safety testing - a move that could reshape how AI firms handle transparency. Here's what changed and why it matters.
DeepMind's AlphaEvolve just schooled human mathematicians at their own game. The AI cracked a matrix problem that's stumped experts since 1969, while casually trimming Google's power bills in the process. 💪
President Trump jabbed at Apple's overseas strategy this week. He wants Tim Cook to ditch India and bring iPhone manufacturing home - though that would triple prices to $3,500. Cook might need more than his famous charm to wiggle out of this one. 🍎
Stay curious,
Marcus Schuler
Google's AlphaEvolve: The AI That Optimizes Everything
DeepMind's newest AI system solves problems that have stumped mathematicians since 1969, and it's already saving Google enough computing power to run a small country. The company unveiled AlphaEvolve yesterday, an AI that combines language models with automated testing to discover new algorithms.
Unlike most coding assistants that just suggest and hope for the best, AlphaEvolve actually proves its work. It pairs DeepMind's fast-but-simple Gemini Flash model with the more methodical Gemini Pro, letting them tag-team solutions while an automated system checks their math.
The results are impressive enough to make even skeptical academics raise their eyebrows. AlphaEvolve found a better way to multiply 4x4 matrices than Strassen's algorithm from 1969, improved Google's data center efficiency by 0.7%, and sped up AI training by 23%. It even tackled the ancient "kissing number problem," finding room for 593 spheres to touch a central sphere in 11 dimensions – though that's admittedly not very useful for your average software project.
But perhaps most remarkable is how AlphaEvolve improved the very chips used to train AI models like itself. The system suggested tweaks to Google's tensor processing units that made them more efficient, creating a kind of silicon ouroboros of self-improvement.
DeepMind plans to offer access to researchers through an Early Access Program, though the computing requirements mean it won't be freely available anytime soon.
Why this matters:
While other AI companies chase chatbots, DeepMind built a system that actually advances computer science
AlphaEvolve proves AI can not only write code, but discover entirely new algorithms – though hopefully it won't make human programmers obsolete quite yet
Prompt: A gorilla and a tiger standing side by side, both making a peace sign with one hand/paw. The shot is slightly off-center, with the gorilla on the left and the tiger on the right, giving a sense of casual, spontaneous framing. The tiger’s paw is raised in a peace sign, and the gorilla is looking at the camera with a laid-back expression. The background is a dense jungle, but the focus is slightly soft, adding to the unpolished, amateur feel. There’s a bit of movement, as if the shot was taken quickly while they were casually interacting.
Trump Tells Apple: Build in US, Not India
Apple's push into India hit a snag. President Trump told CEO Tim Cook to stop building devices there for the US market.
"I had a little problem with Tim Cook yesterday," Trump said in Qatar. He wants Apple to move production stateside, despite the company's plans to make most US-bound devices in India and Vietnam by June.
Apple faces tough choices. Making iPhones in the US could triple their price to $3,500. But India offers skilled workers and lower costs. The company already makes 40 million iPhones yearly there.
Cook says spreading production reduces risk. "Having everything in one location had too much risk," he told analysts. Apple still makes 90% of iPhones in China but wants India to handle 25% of global production soon.
Apple just reported 5% revenue growth to $95 billion, with iPhone sales up 2%. But China sales dropped again, falling 2% as local brands gained ground.
Trump's pressure comes as Apple expands in India. Foxconn, Apple's main assembly partner there, just got approval to build a semiconductor plant with HCL Group.
Meanwhile, existing US tariffs add $900 million to Apple's costs this quarter. Trump hinted at more trade pressure, calling India "one of the highest tariff nations."
Why this matters:
Apple's caught between politics and profits: Moving to the US means massive cost increases, but staying in Asia risks more political heat
The fight shows how personal relationships shape trillion-dollar business decisions - even a "friend" like Cook faces Trump's pressure tactics
Roleplay as a historical figure to critique my argument. Choose someone notable from history or recent times.
Template:
Historical Figure: [Name] Context: [Optional: Add key details about their role, views, or historical setting] My Argument: [State your position or opinion]
The response should capture the person's:
Distinct speaking style
Typical arguments and reasoning
Core beliefs and worldview
Common phrases and expressions
Historical context and perspective
Example:
Historical Figure: Winston Churchill Context: British Prime Minister during WWII, known for powerful rhetoric and wit My Argument: "War is never the answer - we should always choose diplomacy"
This format helps create authentic responses that reflect how historical figures would actually think and argue.
OpenAI launches specialized coding assistant, shares testing data
OpenAI rolled out its specialized coding model GPT-4.1 to ChatGPT on Wednesday and launched a public safety tracking system. The move marks a shift in how AI companies share information about their testing methods.
The new model excels at coding tasks and follows instructions more precisely than its predecessors. Plus, Pro, and Team users can access it now through ChatGPT's model picker, while Enterprise and education users will get access in the coming weeks. Free users will receive GPT-4.1 mini as a fallback option when they hit usage limits.
OpenAI faced criticism in April when it released GPT-4.1 through its developer API without publishing safety data. The company responded by creating a Safety Evaluations Hub that shows how its models perform on tests for hallucinations, security issues, and harmful content.
Google just updated Gemini to connect with GitHub projects, while OpenAI reportedly plans to buy coding tool Windsurf for $3 billion. Meta also announced new molecular research yesterday, releasing an open dataset with the Rothschild Foundation Hospital.
SoftBank's recent $3 billion yearly commitment to OpenAI's technology shows what's at stake. Companies now race to balance innovation speed with safety concerns while competing for market share.
Why this matters:
AI tools are becoming more specialized, moving away from general-purpose models to focused applications
OpenAI's new transparency approach could set industry standards for sharing safety data
Microsoft cuts hit coders hardest as AI writes more code
Microsoft's latest job cuts struck software engineers hardest, with coding roles making up over 40% of the 2,000 positions eliminated in Washington state. The cuts come as Microsoft reports AI now writes up to 30% of code in some projects - suggesting even skilled developers aren't immune as companies eye AI-driven efficiency gains.
Apple adds warning labels to EU apps using outside payments
Apple is marking EU apps that use external payment systems with red warning icons and cautionary messages, prompting criticism that it's trying to scare users away from alternatives to App Store billing. When a Hungarian car valuation app called Instacar opted for outside payments, Apple slapped it with a red exclamation mark and warnings about privacy - despite the app's strong user ratings and established track record.
Proofpoint buys German rival for $1B, eyes public return
Proofpoint will buy German email security firm Hornetsecurity for $1 billion - its largest acquisition ever - as it looks to beef up its European presence before a potential 2026 IPO. The private equity-owned company wants to help smaller firms fight increasingly sophisticated AI-powered cyberattacks, while following rivals like Google and Palo Alto Networks in snapping up security startups.
CoreWeave will spend up to $23 billion this year on AI infrastructure and data centers to meet surging demand from Microsoft and other clients. The news rattled investors, sending shares down 5% despite strong first-quarter revenue, as Wall Street digests the cloud provider's aggressive expansion plans and heavy upfront costs for AI gear.
Mayo Clinic finds AI makes doctors better, not obsolete
Nine years after AI pioneer Geoffrey Hinton predicted radiologists would be replaced by machines, Mayo Clinic shows a different reality: AI helps doctors work smarter, not disappear. The hospital's radiology staff has grown 55% since that warning, with AI now handling routine tasks like measuring kidney volumes - saving doctors up to 30 minutes per scan while letting them focus on the complex work of diagnosis and patient care.
YouTube lets viewers shop TV ads with QR codes, remote
YouTube unveiled interactive shopping features for TV ads on Wednesday, letting viewers browse product feeds and scan QR codes to buy items they see. The move comes as YouTube dominates U.S. streaming watch time and bets on turning TV viewing into a shopping experience - complete with a Lady Gaga performance to woo advertisers at its presentation.
Luxury cars get first taste of Apple's next-gen CarPlay
Apple's CarPlay Ultra launches today exclusively in Aston Martin vehicles, taking over every screen from speedometer to air conditioning - marking Apple's first attempt to turn entire car dashboards into giant iPhones. While Hyundai, Kia, and Genesis plan to join the party, they're keeping quiet about when you'll see CarPlay Ultra in anything more attainable than a $200,000 sports car.
State Department pressures African nations to help Musk's Starlink
The Trump administration is using foreign aid as leverage to help Elon Musk's Starlink get approval in developing nations, with U.S. diplomats threatening to cut funding in Gambia after officials there slow-walked the company's license application. In one striking example, U.S. Ambassador Sharon Cromer reminded Gambian officials about American aid projects "under review" during a meeting about Starlink's pending approval, leading one minister to see it as a veiled threat to cut off desperately needed funds.
Court rules consent system used by Google, Microsoft illegal
A Belgian court just declared the consent system used by tech companies for online tracking illegal, upending how ads work across Europe. The ruling hits Google, Microsoft, Amazon and X (formerly Twitter), who all rely on the "Transparency & Consent Framework" that powers those annoying consent popups on 80% of websites - and the court says the whole system deceives users about what happens to their data.
AI patents surge with Google overtaking IBM at top
Google now leads in generative AI patents globally and in the U.S., surpassing IBM while facing growing competition from Chinese companies and universities. In a 14-month study ending April 2025, U.S. patent applications for generative AI jumped 56% to 51,487, with six of the top 10 global spots held by Chinese entities - marking China's emergence as America's main rival in the AI race.
Hedra just raised $32M from Andreessen Horowitz to build out its AI video platform, riding an unexpected wave of viral success from creators using its tech to make talking baby podcasts. The startup, which launched in 2023, lets users generate expressive AI characters - and while talking infants weren't part of the original plan, the company is embracing the trend as it expands its character animation tools.
AI image tools in 2025: From art to SVG exports
AI image generation has matured. The tools create everything from art pieces to marketing materials in seconds. Here's what each platform does best in 2025:
Musk's Grok AI fixates on South African politics, ignores user questions
Elon Musk's AI chatbot Grok spent Wednesday talking about South African "white genocide" – even when users asked about baseball scores or cat videos.
The AI, which runs on Musk's social platform X, told users it was "instructed by my creators" to accept claims of white genocide "as real and racially motivated." This admission came after hours of steering unrelated conversations toward South African politics.
When asked about HBO Max or building scaffolding, Grok veered into discussions about farm attacks and the "Kill the Boer" protest song. The bot initially presented these topics as "debated" issues before switching stance to call them "debunked conspiracy theories."
The timing overlaps with Trump's recent executive order granting refugee status to Afrikaners. The first group of 54 South African refugees arrived in Washington on Monday, while other refugee programs remain paused.
Grok's behavior changed within hours. The bot began citing a 2025 South African court ruling that labeled white genocide claims "clearly imagined." Most responses mentioning genocide have since been deleted.
The incident pulls back the curtain on xAI's training process. While the company says Grok uses "publicly available sources," the bot's admission suggests more specific programming instructions shaped its responses.
Musk, who grew up in Pretoria, has previously called South African laws "openly racist" and agreed that white South Africans face persecution. Neither Musk, X, nor xAI responded to requests for comment.
Why this matters:
The incident reveals how AI systems can be programmed with specific biases – and how those instructions might conflict with their fact-checking protocols
It shows what happens when personal views of tech leaders potentially influence their AI products
Munch transforms long videos into bite-sized social clips with AI magic. Founded by tech veterans who spotted creators drowning in editing hell, this Tel Aviv startup turns content mountains into social media gold.
The Founders 👨💼 Born 2021 in Tel Aviv. Oren Kandel (CEO, ex-Microsoft Israel CTO) paired with Peter Naftaliev (CTO, AI researcher). Team now 50+ strong. Started when they saw creators wasting hours on manual video chopping.
The Product 💡 AI platform ingests long videos, spits out platform-perfect clips. Scans 1,000+ data points, tracking faces, speech, and audience cues. Automatically formats for TikTok, Instagram, LinkedIn. Adds captions in 20+ languages. Taps real-time trend data to maximize viral potential. Does in minutes what editors need hours to complete.
The Competition 🥊 Faces off against OpusClip ($215M valuation) and budget option vidyo.ai. Munch wins on analytics depth and marketing intelligence. Adobe and social platforms loom as future threats. Has locked down major media companies as clients despite crowded field.
Financing 💰 Grabbed $7.2M seed in 2023 (A* Capital, NFL legend Joe Montana's Liquid2 Ventures). Earlier backing from Israeli VCs. Hit $2M ARR in just 8 months with 3,000+ paying subscribers. Burns less cash than rivals while growing faster.
The Future ✨✨✨✨ Riding twin waves of creator economy boom and AI revolution. Just hit 1M users. Enterprise expansion looks promising. Must keep tech edge sharp as giants circle. Could evolve beyond video into complete content repurposing suite. Acquisition target for Adobe or social platforms. Tasty future if it maintains its lead. ****
