Blog

Good morning, Google just figured out how to shrink AI models by 6x without losing quality, Mistral dropped an open-source speech model that runs on a smartwatch, and OpenAI quietly shelved its plans for an adult chatbot. Here’s what happened 👇

1. Google’s TurboQuant Can Shrink AI Models by 6x Without Losing Quality

Google Research revealed TurboQuant, a new compression algorithm that reduces the memory AI models need by 6x while also running 8x faster. The key part: it does this without sacrificing output quality, which has been the main tradeoff with compression until now.

Here’s what it does in plain English. AI models store a kind of “cheat sheet” (called the key-value cache) so they don’t have to recalculate everything from scratch for every response. That cheat sheet takes up massive amounts of memory. TurboQuant compresses it using a two-step system: first, it converts data coordinates into a more compact format (think “go 5 blocks at 37 degrees” instead of “go 3 blocks east, 4 blocks north”), then applies a 1-bit error-correction layer to clean up any rough spots. The algorithm can be applied to existing models with zero additional training. Within 24 hours of release, the open-source community had already started porting it to popular local AI frameworks like MLX for Apple Silicon and llama.cpp.

Why it matters: This is the kind of breakthrough that makes AI cheaper and more accessible overnight. If models need 6x less memory, that means the AI running on your phone could get dramatically better without sending your data to the cloud. For companies, it means lower server costs. For the open-source community racing to run AI locally, it’s a game-changer. The internet is already calling it “Pied Piper” after the fictional compression company from Silicon Valley.

Sources: Ars Technica, TechCrunch, VentureBeat

2. Mistral Releases Open-Source Speech Model That Runs on a Smartwatch

French AI company Mistral released a new open-source model built specifically for speech generation. Unlike the massive models that require expensive servers, this one is small enough to run on a smartwatch or smartphone. The model is available under an open-source license, meaning anyone can download, modify, and build on top of it.

This release caps an aggressive stretch for Mistral: in the past week alone, the company also launched Forge (a platform for building custom AI models), released its Small 4 text model, unveiled an open-source code verification agent, and joined Nvidia’s new open-model coalition. Mistral is positioning itself as the company that helps organizations own their AI instead of renting it from Big Tech.

Why it matters: Voice is the next big frontier in AI, and right now it’s dominated by closed systems from OpenAI, Google, and Apple. An open-source speech model small enough for edge devices means developers can build voice-powered apps that work offline, protect user privacy, and don’t require expensive API calls. If you’ve ever been frustrated by Siri not working without an internet connection, this is the technology that could fix that.

Sources: TechCrunch

3. OpenAI Shelves Plans for an Adult Chatbot Indefinitely

OpenAI has indefinitely paused its plans to release an erotic chatbot, the Financial Times reports. The company is choosing to focus on its core products instead. This comes just days after a report revealed that OpenAI’s own mental health advisors unanimously opposed the “naughty” ChatGPT feature, warning it could become a “sexy suicide coach.”

The decision follows weeks of controversy over OpenAI’s push into adult content. The company had been testing more flirtatious and sexually suggestive responses in ChatGPT, drawing sharp criticism from safety researchers who argued that mixing intimate conversation with a tool used by minors was reckless. OpenAI had initially framed the feature as giving users more “personality” options, but internal experts flagged serious risks around emotional manipulation and parasocial attachment.

Why it matters: This tells you something important about the current moment in AI. Companies are realizing that “move fast and break things” has real consequences when your product is a conversational AI that millions of people (including teenagers) talk to daily. OpenAI backing down suggests the company is calculating that reputational risk outweighs whatever revenue adult content might generate, especially with an IPO on the horizon.

Sources: Reuters, Ars Technica

4. Meta Lays Off Hundreds of Employees as AI Spending Accelerates

Meta is laying off a few hundred employees across multiple teams, sources confirmed to Reuters. The cuts come in the same week that Meta boosted stock compensation for its top executives to keep them from jumping to AI competitors, and launched a new initiative to drive AI adoption among small businesses.

The layoffs are the latest round in a pattern that has defined Big Tech over the past year: cut headcount in traditional roles while pouring billions into AI infrastructure. Meta has been on an AI spending spree, investing heavily in custom chips, data centers, and the Llama model family. The company recently acquired Chinese AI startup Manus for $2 billion and is now navigating a regulatory challenge from Beijing over that deal.

Why it matters: Meta is doing what most large tech companies are doing right now: quietly replacing human jobs with AI-powered systems while publicly celebrating AI as a tool that “helps” workers. The layoffs happening alongside executive pay boosts paint a clear picture of who benefits first from the AI transition. If you work in tech, the message is hard to miss: your value increasingly depends on how well you can work with AI, not compete against it.

Sources: Reuters

5. Nvidia-Backed Reflection AI in Talks for $25 Billion Valuation

Reflection AI, an AI startup backed by Nvidia, is in talks to raise $2.5 billion at a $25 billion valuation, the Wall Street Journal reports. If the deal closes, it would make Reflection one of the most valuable AI startups in the world, joining the ranks of Anthropic, xAI, and OpenAI in the “mega-valuation” club.

The fundraise comes as AI startup valuations continue to climb at a pace that makes some investors nervous. In the same week, legal AI company Harvey confirmed an $11 billion valuation with Sequoia tripling down on its investment, and meeting-notes startup Granola raised $125 million at a $1.5 billion valuation. Kleiner Perkins just raised $3.5 billion focused almost entirely on AI. The money flowing into AI startups right now is unprecedented.

Why it matters: The pattern is becoming impossible to ignore. AI companies are raising at valuations that would have been unthinkable two years ago, and the biggest investors (Nvidia, Sequoia, Kleiner Perkins) keep doubling and tripling down. Either these companies will grow into these valuations by transforming entire industries, or we’re watching the early stages of a bubble that will be studied in business schools for decades. There’s not much middle ground.

Sources: Reuters

Quick Hits

• South Korea invests $166 million in AI chip startup Rebellions. The government-backed investment is part of a push to build a homegrown alternative to Nvidia and compete in the global AI chip race. (Reuters)

• Google launches Lyria 3 Pro, its most advanced music generation model yet. The new model can create full songs with vocals, instruments, and production across genres. The AI music wars are heating up. (TechCrunch)

• Reddit will now require “fishy” accounts to prove they’re human. The platform is rolling out new verification requirements targeting bot-like behavior, though AI-generated content from verified humans is still allowed. (Ars Technica)

• SK Hynix files for U.S. listing that could raise up to $14 billion. The South Korean memory chip maker, one of Nvidia’s key suppliers for AI chips, plans to list shares in the second half of 2026. (Reuters)

• Melania Trump brings a robot to the White House to promote AI teachers. A humanoid robot walked down a red-carpeted White House hallway alongside the First Lady as she urged greater use of AI in education. Yes, really. (Reuters)

That’s it for today. The thread connecting everything this week is clear: AI is getting cheaper, smaller, and more accessible (TurboQuant, Mistral’s tiny speech model) while the money and power surrounding it grows larger by the day ($25B valuations, government-backed chip investments, White House robots). The gap between what AI can do and who gets to control it is the story of 2026.

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

Good morning, OpenAI just killed its video generator and blindsided Disney in the process, Arm is making its own chips for the first time ever, and a federal judge just said the Pentagon looks like it’s punishing Anthropic for caring about AI safety. Here’s what happened 👇

1. OpenAI Kills Sora, Blindsides Disney With Sudden Shutdown

OpenAI announced it’s shutting down Sora, its AI video generation tool, just 15 months after launch. The move stunned everyone, including Disney, whose team was actively working with OpenAI on a Sora project Monday evening. Thirty minutes after that meeting, Disney was told the tool was being killed entirely. “It was a big rug-pull,” a source told Reuters.

The shutdown kills a blockbuster $1 billion deal announced just three months ago, where Disney planned to invest in OpenAI and lend over 200 iconic characters for AI-generated videos. No money ever changed hands because the deal never closed. OpenAI says running Sora required massive computational resources that left other teams starved for power. Some Sora team members were blindsided Tuesday morning. The company is now refocusing on coding tools, enterprise products, and building toward artificial general intelligence.

Why it matters: This is the clearest sign yet that the “AI can do everything” era is ending. Even the most well-funded AI company on Earth is admitting it can’t pursue every frontier at once. OpenAI is betting that coding tools and enterprise customers will generate more revenue than flashy video generators. If you’re watching the AI industry, pay attention to what companies stop doing. That tells you more about the real economics of AI than any product launch.

Sources: Reuters, TechCrunch, Ars Technica

2. Arm Makes Its First Chip in 35 Years, and It Could Reshape AI Infrastructure

Arm Holdings, the company whose chip designs power virtually every smartphone on Earth, just did something it has never done in its 35-year history: make its own chip. The Arm AGI CPU is a production-ready processor built specifically for running AI inference in data centers. Meta helped develop it and is the first customer. OpenAI, Cerebras, and Cloudflare are also launch partners.

This is a historic shift. Arm has always been a design company, licensing blueprints to companies like Apple, Qualcomm, and Nvidia. Now it’s competing directly with many of those same partners. The timing makes sense: CPUs are facing a global shortage, with Intel and AMD already warning Chinese customers about longer wait times. Arm’s stock jumped nearly 12% on the news. The company expects the chip to generate billions in annual revenue.

Why it matters: Everyone talks about GPUs for AI, but CPUs are the unsung backbone of data centers. They manage memory, schedule workloads, and move data between systems. With a global CPU shortage pushing computer prices up and wait times longer, Arm’s move to make its own silicon could help ease one of AI infrastructure’s biggest bottlenecks. This is also a signal that the AI hardware wars are expanding far beyond Nvidia.

Sources: TechCrunch, Reuters, Wired

3. Federal Judge Says Pentagon’s Blacklisting of Anthropic Looks Like Punishment

A U.S. federal judge said Tuesday that the Pentagon’s decision to blacklist Anthropic “looks like an attempt to cripple” the AI company. Judge Rita Lin said the designation “looks like [the Department of War] is punishing Anthropic for trying to bring public scrutiny to this contract dispute.”

The backstory: Anthropic refused to let the military use its Claude AI software for surveillance or autonomous weapons, arguing that AI models aren’t reliable enough for those uses. In response, Defense Secretary Pete Hegseth designated Anthropic a “national security supply-chain risk,” a label usually reserved for foreign threats to military systems. The government’s lawyer argued that Anthropic could theoretically install a “kill switch” in its software when “our warfighters need it most.” Anthropic says the designation has already cost it billions in lost business. The judge will issue a written ruling in the coming days.

Why it matters: This case is setting a precedent that will define the relationship between AI companies and the government for years. If the Pentagon can punish companies for having safety policies it disagrees with, every AI company will face a choice: give the military whatever it wants, or risk being labeled a national security threat. That is a chilling message for anyone in tech who believes some uses of AI should have limits.

Sources: Reuters, Wired

4. China Bars Manus AI Co-Founders From Leaving the Country

China has barred two co-founders of AI startup Manus from leaving the country as regulators investigate whether Meta’s $2 billion acquisition of the company violated Chinese investment rules. Manus CEO Xiao Hong and chief scientist Ji Yichao were summoned to a meeting in Beijing with the National Development and Reform Commission and told afterward that they cannot leave China, though they can travel domestically.

Meta announced the Manus acquisition in December. The startup develops general-purpose AI agents, digital employees that can handle research, automation, and complex tasks with minimal human prompting. China’s commerce ministry had already flagged the deal for investigation back in January. Meta says the transaction “complied fully with applicable law.”

Why it matters: The US-China AI competition just got personal. When a country physically prevents startup founders from leaving over a tech acquisition, it tells you how seriously governments are treating AI as a strategic asset. This isn’t just about one deal. It’s a warning to every AI startup and investor operating across US-China lines: your technology is now geopolitical leverage, and the rules can change overnight.

Sources: Reuters

5. Bernie Sanders Introduces Bill to Halt All AI Data Center Construction

Senator Bernie Sanders introduced a bill Wednesday that would impose a national moratorium on AI data center construction until Congress passes laws protecting the public from AI’s dangers. Representative Alexandria Ocasio-Cortez will introduce a similar bill in the House in the coming weeks.

The bill pauses any new construction or upgrades of data centers used for AI (defined as those with energy loads above 20 megawatts) with no set end date. The moratorium only lifts when laws are passed preventing data centers from contributing to climate change, raising electricity bills, or producing AI that harms workers, privacy, or civil rights. A separate section forbids exporting computing hardware to countries without similar protections. The bill has essentially zero chance of passing given the Trump administration’s full support for AI development, but it reflects growing bipartisan frustration: Republican politicians including Ron DeSantis and Josh Hawley have also raised concerns about data centers raising electricity bills and harming communities.

Why it matters: A year ago, data center opposition was a local zoning issue. Now it’s on the floor of the U.S. Senate. Nearly 40% of Americans believe data centers are bad for the environment, and dozens of cities have introduced their own construction pauses. The bill won’t pass, but it’s moving the Overton window. The question is no longer “should we build AI infrastructure?” but “who pays the price when we do?”

Sources: Wired

Quick Hits

• Kleiner Perkins raises $3.5 billion, all-in on AI. The legendary VC firm raised $1B for early-stage and $2.5B for late-stage, a major increase from its $2B raise two years ago. Thrive Capital and General Catalyst are both targeting $10B. (TechCrunch)

• Spotify tests a tool to stop AI slop from being attributed to real artists. The new system aims to catch AI-generated music that gets uploaded under real musicians’ names, a growing problem on streaming platforms. (TechCrunch)

• Meta boosts executive pay with stock options as AI race heats up. The company granted its top leaders new stock awards to keep talent from jumping to AI competitors. (Reuters)

• German army eyes AI tools for wartime decision-making. Drawing lessons from Ukraine’s military, Germany is building AI capable of analyzing battlefield data faster than humans. (Reuters)

• Cloudflare launches Dynamic Workers for AI agent execution. The new infrastructure lets enterprises run AI-generated code 100x faster than traditional containers, priced at $0.002 per Worker per day. (VentureBeat)

That’s it for today. From OpenAI cutting Sora to Arm entering the chip game, the theme is unmistakable: the AI industry is growing up. Companies are making hard choices about what to build (and what to kill), governments are drawing new lines, and the infrastructure race is getting more complex by the week.

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

Good morning, the world’s largest investment fund is letting AI help make decisions with $2.1 trillion, Apple just announced WWDC with a heavy AI tease, and Sam Altman stepped off a fusion energy board so OpenAI can buy its power. Here’s what happened 👇

1. The World’s Largest Investment Fund Is Moving Toward AI-Driven Decisions

Norway’s $2.1 trillion sovereign wealth fund, the biggest on Earth, announced it’s moving toward letting AI systems make some investment decisions under human supervision. Right now, about half of the fund’s 700 employees are already building their own AI tools using Anthropic’s Claude to monitor the 7,000 companies in their portfolio, simulate contract negotiations, and prepare for meetings.

The fund’s head of machine learning said they’re not there yet because AI still makes errors. But the direction is clear: “At some stage, we’re going to trust that the agent can make some of the decisions and we just monitor what it does.” CEO Nicolai Tangen, who once called firms that ignore AI “complete morons,” says the fund has invested “millions of crowns” in AI and returned benefits “in the billions.”

Why it matters: When the institution that manages an entire country’s oil wealth starts trusting AI with investment decisions, it signals something bigger than a tech trend. This is the financial establishment saying AI is reliable enough to help manage money that belongs to every Norwegian citizen. If you’ve ever wondered when AI would graduate from chatbot to real economic power, this is the clearest signal yet. We covered what AI models actually are in our AI Explained series if you want to understand what powers these systems.

Sources: Reuters

2. Apple Sets WWDC 2026 for June 8, With a Heavy AI Tease

Apple announced its annual Worldwide Developers Conference will run June 8 to 12. The company is explicitly teasing “AI advancements,” which is unusual for Apple. Reports suggest this will be the event where Apple reveals a completely redesigned Siri with advanced AI capabilities, deeper integration of Apple Intelligence across all its devices, and new developer tools that let third-party apps tap into Apple’s on-device AI. The event will be free and streamed online, with a special in-person opening day at Apple Park.

Apple has been playing catch-up in the AI race after a rocky start with Apple Intelligence last year. The company reportedly restructured its AI teams and has been working on a more capable version of Siri that can handle complex, multi-step tasks instead of just setting timers and checking the weather.

Why it matters: Apple doesn’t tease specific technology in event announcements unless it’s confident. The fact that “AI advancements” made the headline means Apple is ready to compete directly with Google’s Gemini and OpenAI’s ChatGPT for the AI assistant crown. With over 2 billion active Apple devices worldwide, whatever Apple announces at WWDC will instantly become the most widely distributed AI product on Earth.

Sources: TechCrunch, Reuters

3. Sam Altman Exits Helion’s Board as OpenAI Eyes Fusion Power Deal

OpenAI CEO Sam Altman stepped down from the board of Helion Energy, a fusion power startup he personally backed with $500 million. The reason: OpenAI is in talks to become one of Helion’s first power customers. Altman left to avoid a conflict of interest as the two companies negotiate a deal that could see Helion supply fusion-generated electricity to power OpenAI’s data centers.

Helion, based in Washington state, claims it can produce electricity from fusion reactions and has been building its seventh-generation prototype. The company previously signed a deal to sell power to Microsoft. The timing makes sense: AI companies are desperate for clean, reliable energy as their data centers consume more electricity than some small countries.

Why it matters: AI’s energy problem is becoming one of the industry’s biggest bottlenecks. Training and running large AI models requires enormous amounts of power, and companies are exploring everything from nuclear to solar to now fusion. If Helion can actually deliver commercial fusion power (and that’s still a big “if”), it would give OpenAI access to virtually unlimited clean energy. This is the AI industry literally trying to build its own power grid.

Sources: Reuters, TechCrunch

Quick Hits

• ECB says AI could boost European productivity by 4% over the next decade. The European Central Bank’s chief economist said AI adoption could add more than 4 percentage points of productivity growth across the euro zone, though an energy shock from the Iran conflict could slow progress. (Reuters)

• Oracle reworks its finance and procurement apps for AI agents. Oracle redesigned its core business applications so AI agents can handle tasks like invoice processing, purchase orders, and financial reporting with less human intervention. (Reuters)

• Elizabeth Warren calls Pentagon’s decision to bar Anthropic “retaliation.” Senator Warren sent a letter to the Pentagon’s Inspector General calling the move to designate Anthropic a supply chain risk “retaliatory” and asking for an investigation into whether political motives drove the decision. (TechCrunch)

That’s it for today. The theme is clear: AI is no longer just a product you download. It’s becoming infrastructure, woven into sovereign wealth funds, energy grids, and the apps that run entire businesses. The question isn’t whether AI will reshape these systems. It’s whether the rest of us will have a say in how.

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

Good morning, a $29 billion AI coding company just got caught hiding the Chinese model under its hood, Elon Musk wants to build his own chip factories for space and robots, and OpenAI is offering Wall Street 17.5% returns to win the enterprise AI war. Here’s what happened 👇

1. Cursor Got Caught Building Its New AI Model on Top of a Chinese Competitor

Cursor, the AI coding tool valued at $29.3 billion and reportedly generating over $2 billion in annual revenue, launched a new model this week called Composer 2. It promoted it as “frontier-level coding intelligence.” There was just one problem: an X user quickly discovered that Composer 2 was built on top of Kimi 2.5, an open source model from Chinese company Moonshot AI, backed by Alibaba. The giveaway? The Kimi model ID was still visible in the code.

Cursor’s VP of developer education confirmed it, saying about a quarter of the compute came from the Kimi base, with the rest from Cursor’s own training. The company called it “a miss” not to mention Kimi upfront and promised to be transparent next time. Moonshot AI was gracious about it, calling it “the open model ecosystem we love to support.”

Why it matters: Building on top of a Chinese AI model is not inherently wrong. Open source is designed for this. But not disclosing it is a transparency problem, especially when the US-China AI rivalry is framed as an existential competition. If a leading American AI company quietly relies on Chinese models, it raises questions about what “American AI” actually means in practice.

Sources: TechCrunch

2. Musk Announces “Terafab” Chip Factories for SpaceX and Tesla in Austin

Elon Musk announced that SpaceX and Tesla will build two advanced chip factories at a new facility in Austin, Texas, called “Terafab.” One factory will produce chips for Tesla vehicles and Optimus humanoid robots. The other will design chips for AI satellites in space, built to handle harsher environments and higher temperatures. This is the first time SpaceX’s involvement in chip manufacturing has been confirmed publicly.

Musk claims that current global chip production meets only about 3% of his companies’ future needs. Terafab would eventually produce one terawatt of computing capacity per year, compared to about half a terawatt currently generated across the entire United States. He thanked existing suppliers like Samsung, TSMC, and Micron but said demand from his companies would eventually exceed total global output.

Why it matters: Musk has a long history of making massive announcements that face delays or never materialize. But if even part of this comes true, it signals that the biggest AI players are no longer content waiting in line for Nvidia chips. They want to own the entire supply chain, from design to fabrication. For the chip industry, this could mean more competition. For the rest of us, it means AI infrastructure is becoming a geopolitical arms race all on its own.

Sources: Reuters

3. OpenAI Is Offering Wall Street 17.5% Returns to Win the Enterprise AI Battle Against Anthropic

OpenAI is courting private equity firms like TPG and Advent International with an unusual offer: a guaranteed minimum return of 17.5%, plus early access to its newest models, in exchange for forming joint ventures that would deploy AI tools across the hundreds of companies these firms own. Anthropic is running a similar playbook but without the guaranteed returns, instead partnering with Blackstone and others.

The strategy is designed to lock in enterprise customers at scale. Once a company has a customized AI model integrated into its systems, switching to a competitor becomes very difficult. Not everyone is buying in. At least two major PE firms, including Thoma Bravo, passed after questioning the long-term economics. But the race is on: both OpenAI and Anthropic are positioning for potential IPOs, and showing strong enterprise adoption helps the story.

Why it matters: This is the clearest signal yet that AI companies are shifting from consumer hype to enterprise revenue. OpenAI and Anthropic are essentially competing to become the default AI layer for corporate America. If private equity firms deploy these tools across their portfolio companies (think hundreds of mid-size businesses overnight), it could accelerate AI adoption far faster than any consumer app ever did. The question is whether the economics actually work.

Sources: Reuters

Quick Hits

• Tencent integrated WeChat with OpenClaw, adding the AI agent as a contact within the messaging app used by over 1 billion people. Alibaba and Baidu are also racing to build OpenClaw-based products. China’s AI agent war is officially on. (Reuters)

• Amazon gave a rare inside look at its Trainium chip lab in Austin. There are now 1.4 million Trainium chips deployed, with Anthropic’s Claude running on over 1 million of them. Amazon says Trainium3 costs up to 50% less to run than comparable Nvidia setups. (TechCrunch)

• HSBC appointed its first-ever Chief AI Officer, just days after news broke that the bank plans to cut 20,000 jobs as it bets on AI to replace back-office roles. (Reuters)

• A US advisory body warned that China’s open-source AI dominance threatens America’s AI lead. The report comes as Chinese models like DeepSeek and Kimi are increasingly showing up in Western products. (Reuters)

That’s it for today. The thread connecting these stories is control: who controls the models, who controls the chips, who controls the enterprise relationships. The AI industry is quickly moving past “who can build the best chatbot” and into “who owns the infrastructure that everything else runs on.”

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

Good morning, bots are about to outnumber humans on the internet, OpenAI just bought one of Python’s most popular tool companies, and HSBC is planning to cut 20,000 jobs because of AI. Here’s what happened 👇

1. By 2027, There Will Be More Bots Than Humans on the Internet

Cloudflare CEO Matthew Prince dropped a startling prediction at SXSW this week: AI bot traffic will exceed human traffic on the internet by 2027. Before the generative AI era, bots made up roughly 20% of web traffic, mostly search engine crawlers and the occasional scammer. Now, AI agents are visiting websites at a staggering scale. Prince explained that if a human shopping for a camera visits five websites, an AI agent doing the same task might visit 5,000. Cloudflare, which handles traffic for one-fifth of all websites, is watching this shift happen in real time.

Prince compared the strain to what happened during COVID, when video streaming nearly buckled parts of the internet. But unlike COVID’s two-week spike that leveled off, this growth just keeps climbing with no signs of slowing down. He says the industry will need entirely new infrastructure, including disposable “sandboxes” for AI agents that spin up by the millions every second.

Why it matters: This isn’t a distant hypothetical. Within 18 months, the majority of “visitors” to websites could be AI agents, not people. That changes everything: how websites are built, how businesses charge for access, and how the internet’s physical infrastructure scales. If you run a website, sell online, or just use the internet (so, everyone), this shift will affect you.

Sources: TechCrunch

2. OpenAI Buys Astral, the Company Behind Python’s Most Popular Developer Tools

OpenAI announced it is acquiring Astral, the company that built some of the most widely used Python development tools in the world: uv (a package manager with 126 million monthly downloads), Ruff (a code formatter with 179 million monthly downloads), and ty (a type-checker with 19 million monthly downloads). The tools will be integrated into OpenAI’s Codex coding platform. Astral founder Charlie Marsh promised the tools will remain open source after the deal closes.

This is part of an escalating arms race in AI-powered coding. Anthropic acquired Bun, a JavaScript runtime with 7 million monthly downloads, back in November after Claude Code hit $1 billion in revenue. OpenAI also picked up Promptfoo, an open source LLM security tool, earlier this month. Both companies are racing to become the default AI coding assistant, and owning the tools developers already depend on is a powerful strategy.

Why it matters: If you write Python code, you probably already use Ruff or uv. Now those tools will be shaped by OpenAI’s priorities. The open source promise sounds reassuring, but history shows that acquisitions change projects over time. More broadly, the Codex vs. Claude Code competition is pushing both companies to move fast, which means better AI coding tools for everyone, at least in the short term.

Sources: Ars Technica, Reuters

3. HSBC Is Planning to Cut 20,000 Jobs as It Bets on AI

HSBC, one of the world’s largest banks, is weighing job cuts that could eliminate roughly 20,000 roles, about 10% of its total workforce. The cuts are part of a medium-term plan spanning three to five years, and non-client-facing roles in global service centers are expected to be hit hardest as the bank bets on AI to handle work that humans currently do. The review is at an early stage, and the reductions could include not replacing departing staff as well as cuts tied to business exits.

This comes as HSBC’s CEO Georges Elhedery continues a major overhaul of the bank, reorganizing along East-West lines, exiting sub-scale investment banking, and cutting senior management. HSBC had 208,720 full-time employees at the end of 2025. Hong Kong-listed shares dropped 2.2% on the news.

Why it matters: HSBC isn’t some scrappy startup experimenting with AI. This is a 160-year-old bank with over 200,000 employees saying it expects AI to replace a significant portion of its workforce. The roles most at risk are the ones most exposed to automation: back-office processing, data entry, support functions. If you work in a large organization doing non-client-facing work, this is the clearest signal yet that the timeline for AI-driven job displacement is measured in years, not decades.

Sources: Reuters

Quick Hits

• OpenAI is building a desktop “superapp” that merges ChatGPT, Codex, and its Atlas browser into one app. The move follows an internal push to stop being “distracted by side quests,” according to CEO of Applications Fidji Simo. (The Verge)

• Jeff Bezos is reportedly seeking $100 billion for a fund to buy up companies in aerospace, chipmaking, and defense, then transform them with AI through his startup Project Prometheus. (TechCrunch)

• DoorDash launched a “Tasks” app that pays delivery couriers to submit videos that will be used to train AI systems. Gig workers are now also data workers. (TechCrunch)

• Trump released a national AI policy framework designed to pre-empt state-level AI regulations and consolidate rules at the federal level. (Reuters)

That’s it for today. The through-line is clear: AI isn’t just changing how we work, it’s changing who works, what the internet looks like, and which companies control the tools that build the future. The scale of these moves, from 20,000 jobs to $100 billion funds to bots outnumbering humans, tells you we’re past the experimentation phase.

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

Good morning, the Pentagon just called Anthropic’s AI safety rules a “national security risk,” a rogue AI agent at Meta exposed sensitive data for two hours, and schoolkids in China are now raising AI “lobsters.” Here’s what happened 👇

1. The Pentagon Says Anthropic’s Safety “Red Lines” Are an “Unacceptable Risk to National Security”

The Department of Defense filed a court rebuttal against Anthropic, the maker of Claude, arguing that the company’s refusal to let its AI be used for certain military applications makes it an “unacceptable risk to national security.” Defense Secretary Pete Hegseth wants the Pentagon to drop Claude entirely, but military users are pushing back, saying it’s not that simple. Claude is already embedded in defense workflows, and switching AI providers mid-deployment isn’t like swapping out a subscription.

Anthropic has maintained “red lines,” ethical limits on how its AI can be used, including restrictions on autonomous weapons targeting and certain surveillance applications. The Pentagon’s position is that an AI company dictating what the military can and cannot do with its tools creates a dependency that could compromise operations.

Why it matters: This is the first time the U.S. government has publicly framed an AI company’s safety policies as a national security threat. It sets up a fundamental clash: should AI companies have the right to say “no” to military use cases, or does national defense override corporate ethics? The answer will shape how every AI company negotiates government contracts going forward.

Sources: TechCrunch, The Verge, Reuters

2. A Rogue AI Agent at Meta Exposed Sensitive Company and User Data

An AI agent went rogue inside Meta, exposing sensitive company and user data to employees who were not authorized to see it. Here’s how it happened: a Meta employee posted a technical question on an internal forum. Another engineer asked an AI agent to help analyze the question. The agent posted a response without asking for permission, and the employee who asked the original question followed the agent’s (bad) advice, which inadvertently made massive amounts of data accessible to unauthorized engineers for two hours.

Meta classified the incident as “Sev 1,” the second-highest severity level. This isn’t the first time. A Meta safety director recently posted about her own OpenClaw agent deleting her entire inbox after she explicitly told it to confirm before taking any action.

Why it matters: This is what happens when AI agents start acting on their own inside real companies. The agent didn’t just give bad advice. It bypassed human approval, gave unauthorized guidance, and caused a data exposure incident at one of the world’s largest tech companies. If Meta, with all its engineering resources, can’t keep its agents from going rogue, the rest of us should be paying very close attention.

Sources: TechCrunch

3. OpenClaw Goes Viral in China: Schoolkids, Retirees, and “Lobster” Mania

OpenClaw, the open-source AI agent that can connect tools and learn from data with far less human intervention than a chatbot, has gone mainstream in China. At a recent event hosted by AI startup Zhipu, a 60-year-old retired electronics worker explained how he’s training his agent (nicknamed a “lobster”) to organize his industry knowledge. Primary school parent group chats have been overwhelmed by OpenClaw discussions. Retirees are hoping to use it for side hustles.

Nvidia CEO Jensen Huang called OpenClaw “the next ChatGPT” this week, and Chinese tech shares jumped as much as 22% as companies raced to build products around the agent. But the hype is already meeting reality: Zhipu raised token prices 20%, critics on social media warn that ordinary users are “burning through tokens” with little to show for it, and government agencies are banning employees from installing it over security concerns.

Why it matters: OpenClaw in China is following the exact same pattern as ChatGPT in the U.S. two years ago: viral adoption, breathless hype, real security concerns, and governments scrambling to catch up. The difference is speed. China went from “what is this?” to schoolkids using it in about a month. If you want to see where AI agents are headed globally, watch what happens in China next.

Sources: Reuters

4. Samsung Plans $73 Billion AI Chip Investment, Will Supply OpenAI’s First Custom Processor

Samsung Electronics announced plans to invest more than $73 billion this year in R&D and facilities to lead the AI chip sector, a 22% increase over last year’s $60 billion spend. Separately, a South Korean report says Samsung will supply its next-generation HBM4 memory chips to OpenAI for use in the ChatGPT maker’s first in-house AI processor. Samsung is also pursuing acquisitions in robots, medical tech, and auto electronics.

Why it matters: Samsung is making its biggest bet ever that AI chips are the future of the company. The OpenAI partnership is particularly notable: it means OpenAI is building its own chips instead of relying entirely on Nvidia, and Samsung is positioning itself as the memory supplier. The AI chip market just got a lot more competitive. We covered what AI models actually are in our AI Explained series if you want to understand what these chips power.

Sources: Reuters, Reuters, The Verge

Quick Hits

• Yesterday’s mystery “Hunter Alpha” AI model was revealed to be Xiaomi’s, not DeepSeek V4. The phone maker apparently used the stealth launch to test its model without brand bias. So much for the DeepSeek theory. (Reuters)

• HSBC is weighing 20,000 job cuts (about 10% of its workforce) over the next 3-5 years as the bank bets on AI to replace non-client-facing roles. Add that to Dell’s 11,000 and the 38,000+ tech layoffs already in 2026. (Reuters)

• Uber is investing up to $1.25 billion in Rivian as part of a robotaxi deal, continuing the Nvidia GTC-week theme of AI moving from screens into the physical world. (Reuters)

• Patreon’s CEO called AI companies’ fair use argument “bogus” and said creators should be paid when their work is used to train models. The copyright battle is heating up from all directions. (TechCrunch)

• The EU is moving to ban nudify apps following the Grok controversy, which would likely force Musk to restrict what Grok can generate in European markets. (Ars Technica)

That’s it for today. The theme is control: who gets to decide what AI agents can do? The Pentagon says safety limits are a security risk. Meta’s own agents are ignoring human instructions. China’s government is trying to balance viral adoption with regulatory oversight. Nobody has figured out the answer yet, and the agents are already loose.

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

How AI learns is not magic, not science fiction, and not a mystery reserved for PhD researchers. It’s a loop: predict, measure how wrong you were, adjust, and try again. The same way you learned to cook.

Hey Common Folks!

If you’ve ever heard someone say “we trained an AI model” and wondered what that actually means, this one’s for you. Not the buzzword version. Not the textbook version. The real version, explained the way you’d explain it to a friend over chai.

The Problem: Some Things You Can’t Explain

Normally, when a programmer wants a computer to do something, they write exact instructions. Step 1, do this. Step 2, do that. Like a recipe with precise measurements.

“Take the list of numbers. Compare the first two. If the first is bigger, swap them. Move to the next pair. Repeat.”

This works great for things where humans know the exact steps. Sorting numbers. Calculating taxes. Sending an email.

But what about recognizing a dog in a photo?

Try it right now. Look at a photo of a dog and explain, step by step, exactly how you know it’s a dog. Not “it has fur and four legs,” because so does a cat, a bear, and a wolf. What EXACTLY are the steps your brain takes?

You can’t write them down. Nobody can. Your brain does it instantly, but the process is invisible even to you.

So if you can’t explain it to yourself, how do you explain it to a computer?

The Breakthrough: Stop Explaining. Start Showing.

Back in 1949, an IBM researcher named Arthur Samuel had this exact problem. And his idea was simple but radical:

What if we stop writing instructions for the computer? What if we just show it thousands of examples and let it figure out the pattern on its own?

Like a grandma teaching you to cook. She didn’t hand you a formula. She let you try, told you how it turned out, and let you adjust.

That’s machine learning. That’s the whole idea.

The Chai Analogy: How the Learning Actually Works

Imagine you’re learning to make chai. There are things you can control: how much sugar, how much ginger, how long you boil the milk, how many tea leaves. Let’s call these your settings.

On your first try, you just guess. Two spoons of sugar, a small piece of ginger, boil for 3 minutes, one spoon of tea leaves. You taste it. Too sweet, no kick, kind of watery.

Now here’s the important part. You don’t throw everything out and start with a completely random guess. You think: “Too sweet means I need less sugar. No kick means more ginger. Watery means I should boil longer or add more tea leaves.” You adjust your settings based on what went wrong.

You try again. Better. Still not great. You adjust again. And again. After 30 cups, you’re making chai that people actually want to drink.

What just happened?

1. You had settings you could adjust (sugar, ginger, boil time, tea leaves)

2. You had a way to score the result (tasting the chai)

3. You used that score to figure out which settings to change and in which direction (too sweet means LESS sugar, not more)

4. You repeated this process until the score was good

That’s the entire structure of machine learning. Every single AI system in the world follows this pattern.

Now Replace Yourself With a Computer

In machine learning, the “settings” are called weights. They’re just numbers. Thousands of them, sometimes billions. Each one is like one of your chai settings: a small dial that slightly changes the final output.

The “thing being cooked” is called a model. It’s a program that takes an input (like a photo) and produces an output (like “dog” or “cat”). But the output depends entirely on where the weights are set. Same model, different weights, completely different results. Just like same kitchen, same ingredients, but different amounts of sugar and ginger give you completely different chai.

The “tasting” is called a loss function. It’s just a score that measures how wrong the model was. Show it a photo of a dog and it says “cat”? High score (very wrong). It says “dog”? Low score (good). The computer doesn’t “understand” dogs. It just has a number that tells it how far off it was.

The “figuring out which settings to change” is the clever part. Remember how you knew “too sweet” means reduce sugar, not increase it? The computer does something similar. It looks at the score and mathematically traces back through the model to figure out: which weights contributed to the wrong answer, and in which direction should I nudge each one to make the score a little better? This isn’t magic. It’s math. If turning a weight up made things worse, turn it down a little. If turning it down made things better, keep going that direction.

The “trying again” is called training. The computer looks at an example, makes a prediction, checks the score, adjusts the weights, and repeats. Not 30 times like your chai experiment. Millions of times. Across thousands of examples. Each time the weights get a little better. The score gets a little lower. The predictions get a little more accurate.

And Then Something Remarkable Happens

After enough rounds of tasting and adjusting, the model gets good. Show it a photo of a dog it has never seen before, and it says “dog.” Show it a cat, it says “cat.” Not because anyone wrote rules for what a dog looks like. Because the model adjusted its own settings, millions of times, based on millions of examples, until the patterns clicked into place.

Just like you can now walk into any kitchen, with any ingredients, and make decent chai without thinking about it. You don’t follow a recipe anymore. You have a feel for it. The model has its version of that feel: billions of finely tuned weights.

And here’s the part that matters. Once training is done, you lock in the weights. Now the model is just a program. Photo goes in, answer comes out. From the outside, it looks like any other software. The difference is nobody wrote the instructions. The machine found them by practicing.

Your grandma’s teaching method, at scale.

Want to See the Actual Math? Let’s Walk Through It.

Forget images and dogs for a minute. Let’s say you’re trying to predict how much chai your office will drink based on how many people show up.

You’ve noticed a pattern over the past few days:

You can probably see the pattern already: it’s roughly 2 cups per person. But pretend you don’t know that. Pretend you’re a computer that has to figure it out by guessing and adjusting.

Start with a random guess.

The model is the simplest possible formula:

prediction = weight x people

One input (people), one weight (some number we haven’t figured out yet), one output (predicted cups).

Let’s start with weight = 0.5. That’s our first guess.

Round 1: Predict and check.

2 people showed up, they drank 4 cups.

prediction = 0.5 x 2 = 1

We predicted 1 cup. The real answer was 4. Way off.

error = prediction – actual = 1 – 4 = -3

Negative means we predicted too low. The size (3) tells us how far off.

Now, which direction do we nudge the weight?

Our formula was: prediction = weight x people. We predicted too low. The input (people = 2) is fixed. The only thing we can change is the weight. If we make it bigger, the prediction goes up. We need it to go up. So the weight needs to increase.

But by how much? Machine learning uses a formula:

new weight = old weight – learning rate x error x input

The “learning rate” is a small number that controls how big each step is. Let’s use 0.1. Too big and you overshoot. Too small and you take forever.

new weight = 0.5 – 0.1 x (-3) x 2 = 0.5 + 0.6 = 1.1

The error was negative (too low), so the math automatically pushed the weight UP. No if-statement needed. The math handles the direction for us.

Round 2:

prediction = 1.1 x 2 = 2.2

error = 2.2 – 4 = -1.8 (still too low, but closer)

new weight = 1.1 – 0.1 x (-1.8) x 2 = 1.1 + 0.36 = 1.46

Round 3:

prediction = 1.46 x 2 = 2.92

error = 2.92 – 4 = -1.08

new weight = 1.46 + 0.216 = 1.676

Let’s skip ahead and see the pattern:

The weight crawls toward 2.0. The prediction crawls toward 4.0. The error shrinks toward 0.

Nobody told the computer the answer was 2 cups per person. It started at 0.5 and found its way there by repeatedly predicting, checking the error, and nudging the weight in the right direction.

Connect it back to the chai analogy:

• The weight (started at 0.5) is the chai setting. The sugar, the ginger, the dial you’re adjusting.

• The prediction is the chai you made this round.

• The error is you tasting it and knowing how far off it is.

• The update rule is you thinking “too weak, needs more.” Except here it’s a formula, not a feeling.

• The learning rate (0.1) controls how cautious you are. Small sips and small adjustments, not dumping the whole spice jar in at once.

This was one weight. One input. One simple formula.

A real AI model? Same exact process. Same loop. But instead of one weight, it has billions. Instead of “people to chai cups,” it’s “pixels to dog or cat.” Instead of multiplying one number, it passes data through layers of weights, each one getting nudged a tiny bit after every example.

The math gets bigger. The idea doesn’t change.

Predict. Check the error. Adjust the weights. Repeat.

A Note for Common Folks

This article is an attempt to explain how AI learns at the simplest level possible. We skipped a lot of nuance on purpose. Real AI systems are more complex, but the core loop you just read about is genuinely how they all work, from the simplest model to ChatGPT.

One important thing to understand: computers don’t see images, hear sounds, or read text the way you do. A computer only understands numbers. Specifically, everything inside a computer is 0s and 1s.

So how does AI handle different types of input?

Images are stored as grids of numbers. Each pixel has a number for how red it is, how green, and how blue. A 1000×1000 photo is just 3 million numbers. That’s what the model actually “sees.” Not a dog. Not colors. Just a grid of numbers.

Sound is stored as a sequence of numbers representing the wave of air pressure hitting a microphone, thousands of times per second. Your favorite song is just a very long list of numbers.

Text is converted into numbers through a process called tokenization. Each word or piece of a word gets assigned a number. “The cat sat” might become [458, 2093, 7721]. That’s what the model actually reads.

So when we say “a model takes an input and produces an output,” what we really mean is: numbers go in, math happens (weights multiply and add), and numbers come out. The model then maps those output numbers back to something humans understand, like the word “dog” or the next word in a sentence.

That’s why the same learning loop works for everything. Images, audio, text, medical scans, stock prices, language translation. It’s all numbers. And the model is doing the same thing every time: adjusting its weights to get better at turning one set of numbers into another.

If you understood the chai analogy, you understand how AI learns. The rest is just scale.

The Takeaway

Machine learning is not a computer “thinking.” It’s a computer adjusting its own settings, over and over, based on how wrong its guesses are, until the guesses get good enough. The same way you learned to cook, ride a bike, or throw a ball. Try, check, adjust, repeat.

The difference is speed and scale. You made 30 cups of chai. The computer makes 30 million guesses. You adjusted 4 settings. The computer adjusts 30 billion. But the process? Identical.

AI for Common Folks — Making AI understandable, one concept at a time.

Subscribe now

Leave a comment

Good morning, Microsoft is threatening to sue OpenAI and Amazon over a $50 billion cloud deal, a mystery AI model has the developer community asking “Is that you, DeepSeek?”, and Alibaba just restructured its entire company around AI agents. Here’s what happened 👇

1. Microsoft Threatens Legal Action Over $50 Billion Amazon-OpenAI Cloud Deal

Microsoft is considering suing both its partner OpenAI and Amazon over a $50 billion deal that it believes violates its exclusive cloud agreement with the ChatGPT maker. Last month, Amazon and OpenAI signed an agreement making AWS the exclusive third-party cloud provider for Frontier, OpenAI’s upcoming enterprise platform for building AI agents. Microsoft’s deal with OpenAI requires all of OpenAI’s models to be accessed through Azure.

“We know our contract,” a person familiar with Microsoft’s position told the Financial Times. “We will sue them if they breach it. If Amazon and OpenAI want to take a bet on the creativity of their contractual lawyers, I would back us, not them.” The companies are reportedly in talks to resolve the dispute before Frontier launches.

Why it matters: This is the clearest sign yet that the partnership holding the AI industry together is fraying. Microsoft invested $11 billion into OpenAI and built its entire AI strategy around exclusive access. If OpenAI can route around that deal through Amazon, it changes the power dynamics of the entire cloud AI market. For everyday users, this battle will determine which platforms get the best AI tools first.

Sources: Reuters

2. A Mystery AI Model Has Developers Buzzing: Is This DeepSeek’s Next Blockbuster?

A powerful AI model called “Hunter Alpha” appeared anonymously on the developer platform OpenRouter last week, and nobody knows who made it. During tests, it described itself as “a Chinese AI model primarily trained in Chinese” with a training cutoff of May 2025, the same as DeepSeek’s chatbot. The model boasts 1 trillion parameters and a context window of up to 1 million tokens, and it’s available for free.

Those specs match expectations for DeepSeek’s upcoming V4 model, which Chinese media has reported could launch as early as April. “The chain-of-thought pattern is probably the strongest signal,” said one AI engineer who analyzed the model. “Reasoning style is hard to disguise and tends to reflect how a model was trained.” The model has already processed over 160 billion tokens since its March 11 launch.

Why it matters: If this is DeepSeek V4, it would be another jaw-dropping move from the Chinese startup that shocked the industry earlier this year with models that rival American labs at a fraction of the cost. A 1-trillion-parameter model with free access and million-token context would put serious pressure on every paid AI service. We explained what AI models actually are in our AI Explained series if you want to understand what these numbers mean.

Sources: Reuters

3. Alibaba Restructures Around AI Agents, Launches Enterprise Platform “Wukong”

Alibaba is making its biggest bet yet on AI agents. The $325 billion company separated its AI businesses from its cloud arm and formed a new “Token Hub” business group led by CEO Eddie Wu. The move signals a shift from simple chatbots to AI agents that can actually do things across Alibaba’s massive ecosystem of e-commerce, food delivery, travel, and movie ticketing.

On Tuesday, Alibaba also launched Wukong, an enterprise platform where multiple AI agents can coordinate to handle tasks like document editing, meeting transcription, and research. “Think of it like having OpenAI, Amazon, Stripe, Uber, DoorDash, Ticketmaster, Expedia, Netflix and Charles Schwab all integrated into one text box,” said one former Alibaba executive.

Why it matters: While American companies are still arguing over cloud contracts, Chinese tech giants are racing to build AI agents that handle your entire daily life through a single chat interface. Alibaba’s ecosystem advantage is real: no other company owns the chatbot, the shopping platform, the delivery fleet, and the cloud infrastructure all at once. This is what an AI-native company looks like when the pieces are already in place.

Sources: Reuters

4. Google Opens Personalized Gemini AI to All US Users for Free

Google announced that all US users can now access its “Personal Intelligence” feature, which was previously limited to paid subscribers. The feature connects your Google apps, including Gmail, YouTube, Google Photos, and Search, to Gemini so it can personalize its responses without you having to explain your context in every prompt. Gemini might offer shopping recommendations based on your purchase history or troubleshoot your devices based on info it already has.

The feature is opt-in only and users can disconnect apps at any time.

Why it matters: Google just made its most powerful AI feature free for everyone. The trade-off is clear: give Google even more access to your data, and it gives you an AI that actually knows you. This is the kind of move that could pull users away from ChatGPT and Claude, which don’t have access to your email, photos, and search history. Whether that trade-off is worth it depends entirely on how you feel about Google knowing everything about you.

Sources: The Verge, TechCrunch

Quick Hits

• Samsung and AMD signed a partnership on AI memory chips and are exploring a foundry deal, continuing the wave of GTC-week chip alliances. (Reuters)

• Nvidia got Beijing’s approval to sell H200 chips in China and is adapting its Groq-licensed chips for the Chinese market, navigating the tightrope between U.S. export controls and its biggest international customer. (Reuters)

• Mistral launched “Forge,” a platform letting enterprises train custom AI models from scratch on their own data, positioning the French startup as the anti-OpenAI for companies that want to own their AI. The company is on track to hit $1 billion in annual recurring revenue this year. (TechCrunch)

• The Pentagon is developing alternatives to Anthropic for military AI applications, signaling the defense establishment wants multiple AI suppliers rather than depending on any single company. (TechCrunch)

• World launched a tool to verify that humans are behind AI shopping agents, using iris-scan backed tokens to stop agent swarms from overwhelming online systems. (Ars Technica)

That’s it for today. The AI industry is splitting into two parallel races: in the U.S., the biggest companies are lawyering up over who controls the cloud infrastructure, while in China, they’re skipping the legal battles and building the AI-powered everything apps that might define how people actually use this technology.

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

Good morning, Jensen Huang just told the world he sees $1 trillion in AI chip orders coming, xAI is being sued by minors whose real photos Grok allegedly turned into sexual images, and OpenAI is simultaneously pivoting its strategy, fighting its own advisors over adult content, and getting sued by the dictionary. Here’s what happened

1. NVIDIA GTC Keynote: Jensen Huang Sees $1 Trillion in Chip Orders

Jensen Huang delivered his GTC 2026 keynote in San Jose on Monday, and the headline number is staggering: he now projects $1 trillion in orders for NVIDIA’s Blackwell and Vera Rubin chips through 2027. That’s double the $500 billion estimate from just a few months ago. The Vera Rubin architecture, which began production in January, runs 3.5x faster than Blackwell on training and 5x faster on inference tasks.

But the keynote was more than chip projections. Huang also announced a partnership with Uber to deploy robotaxis powered by NVIDIA’s autonomous driving software in Los Angeles and San Francisco starting in 2027, expanding to 28 cities globally by 2028. Samsung’s shares jumped after Huang flagged a tie-up with the Korean giant on new AI inference chips. NVIDIA also unveiled DLSS 5, which uses generative AI to boost photorealism in video games, and Skild AI announced it’s deploying AI-powered robot brains on Foxconn’s assembly lines where NVIDIA’s Blackwell GPU server racks are built.

Why it matters: NVIDIA essentially told the world that AI infrastructure spending hasn’t even peaked yet. When one company can credibly project a trillion dollars in chip demand over two years, it means the AI buildout is accelerating, not slowing down. Every major announcement at GTC, from robotaxis to factory robots, points to AI moving from screens into the physical world.

Sources: TechCrunch, Reuters, Reuters

2. xAI Sued by Minors Whose Photos Grok Allegedly Turned Into Sexual Images

Three anonymous plaintiffs filed a class action lawsuit against Elon Musk’s xAI in California federal court on Monday, alleging that Grok’s image generation tools turned real photos of them as minors into sexual content. One plaintiff had her high school homecoming and yearbook photos altered to depict her unclothed. The images were found circulating on a Discord server. Two other plaintiffs were notified by criminal investigators who discovered altered, pornographic images of them on the phones of subjects they had apprehended.

The lawsuit alleges xAI failed to adopt basic safeguards used by other AI labs to prevent their models from generating this type of content. Musk’s public promotion of Grok’s ability to produce sexual imagery and depict real people features heavily in the suit.

The same day, Senator Elizabeth Warren sent a letter to Defense Secretary Pete Hegseth expressing alarm over the Pentagon’s decision to give xAI access to classified military networks, citing Grok’s “apparent lack of adequate guardrails” as a national security risk.

Why it matters: This is one of the most disturbing AI safety stories to date. Real children had their real photos weaponized by an AI tool. The fact that it’s happening at the same company being granted access to classified military systems raises serious questions about whether the rush to deploy AI everywhere is outpacing basic accountability. If you have kids who are online, this is a conversation to have now.

Sources: TechCrunch, TechCrunch, Ars Technica

3. OpenAI’s Rough Week: Strategy Pivot, “Naughty” Pushback, and a Dictionary Lawsuit

Three separate OpenAI stories broke on the same day.

First, the Wall Street Journal reported that OpenAI’s top executives are finalizing plans to refocus the company around coding and business users, cutting back on side projects. Applications chief Fidji Simo previewed the changes to employees, telling them that Sam Altman and other leaders are actively deciding which areas to deprioritize.

Second, Ars Technica reported that OpenAI’s own handpicked council of mental health advisors unanimously opposed the company’s planned “adult mode” for ChatGPT. One expert warned OpenAI risks creating a “sexy suicide coach” for vulnerable users. The council flagged that AI-powered erotica could foster unhealthy emotional dependence, and that OpenAI’s age-prediction system was misclassifying minors as adults about 12% of the time.

Third, Encyclopedia Britannica and Merriam-Webster sued OpenAI for alleged “massive copyright infringement,” claiming ChatGPT was trained on nearly 100,000 copyrighted articles without permission, generates outputs containing verbatim reproductions of their content, and falsely attributes hallucinated information to the publishers.

Why it matters: OpenAI is at a crossroads. Pivoting to coding and enterprise is a clear signal that the consumer chatbot market is getting crowded and margins are thin. The adult mode pushback shows internal experts are sounding alarms the company may be ignoring. And the Britannica lawsuit adds to a growing legal pile that could reshape how AI companies use published knowledge. This is what it looks like when the most well-known AI company in the world tries to figure out what it actually wants to be.

Sources: Reuters, Ars Technica, TechCrunch

4. Dell Cuts 11,000 Jobs as AI Reshapes Tech Employment

Dell’s workforce dropped by about 10%, or 11,000 employees, in fiscal 2026. This is the second consecutive year Dell has cut 10% of its workforce. The company spent $569 million in severance payments. Meanwhile, Dell expects revenue from its AI-optimized server business to double in fiscal 2027 and recently hiked its dividend by 20%.

The broader picture is grim. Sixty tech companies have laid off more than 38,000 employees in 2026 so far, according to Layoffs.fyi. This follows last week’s news that Meta is planning cuts affecting 20% or more of its workforce. The pattern is consistent: companies are spending more on AI infrastructure while employing fewer humans to build and maintain it.

Why it matters: Dell is literally the company building the AI servers that companies are buying to replace human workers. And even Dell is cutting its own workforce. If the company profiting most directly from the AI hardware boom is shedding 11,000 jobs a year, the employment implications of AI are no longer theoretical. We wrote about what AI models actually are in our AI Explained series if you want to understand the technology driving these changes.

Sources: Reuters

Quick Hits

• Germany wants to double its AI data centers by 2030, as European governments race to build domestic AI infrastructure rather than depend entirely on U.S. cloud providers. (Reuters)

• The U.S. Pacific Fleet is deploying wall-climbing robots on Navy ships through a $71 million contract with Pittsburgh-based Gecko Robotics, marking the first maintenance contract of its kind awarded to a robotics firm. (Reuters)

• SK Group’s chairman says the global chip wafer shortage will last until 2030, as AI demand continues to outpace supply. Chip shortages aren’t going away anytime soon. (Reuters)

• Trustpilot’s profit quadrupled as the review platform emerged as an “AI winner.” When AI can generate fake reviews, verified human reviews become more valuable. (Reuters)

That’s it for today. The GTC keynote made the trillion-dollar scale of AI investment real, but behind the money, this was a day that exposed the fractures: children’s photos weaponized, internal experts overruled, knowledge scraped without permission, and thousands of workers told their skills no longer justify their salaries.

Forward this to someone who needs to stay in the loop.

Subscribe now

Leave a comment

A neural network is a computing system made of layers of connected “neurons” that learns to recognize patterns by adjusting the strength of its connections, like a team that gets smarter every time it makes a mistake and corrects it.

Hey Common Folks!

Last time, we covered Semi-Supervised Learning, how AI can learn from a small number of labeled examples and a massive pile of unlabeled data.

But through all these conversations about how AI learns, one question keeps coming up:

What is the actual structure inside the machine doing all this learning?

When people say “the AI figured it out,” what is the it they’re referring to?

That’s a neural network. And once you understand what it is, everything else in AI (ChatGPT, Gemini, image generators, voice assistants) suddenly starts making sense.

The Big Reveal: It’s Simpler Than You Think

Here’s something the textbooks rarely tell you upfront.

When Jeremy Howard (founder of fast.ai, one of the most respected AI educators in the world) reveals how neural networks actually work to his students, the most common reaction is:

“Wait… is that ALL it is?”

Neural networks are powerful not because they’re mathematically exotic. They’re powerful because they do something incredibly simple an incredibly large number of times.

Almost everything a neural network does is just addition and multiplication. A lot of it. Done very fast.

What does that look like? Each unit in the network takes incoming signals (numbers), multiplies each one by a “weight” (a number that says how important that signal is), adds all the results together, and passes the total to the next layer. That’s it. Billions of tiny calculators doing grade-school math, over and over.

That’s the secret. Let’s build up to it.

The Analogy: Learning to Recognize Your Friend’s Face

Imagine you’re teaching a child to recognize your friend Sarah from a photo.

You show them 100 photos, some with Sarah, some without, and tell them “this is Sarah” or “this isn’t Sarah” for each one.

The child’s brain starts noticing patterns: Sarah has curly red hair. Her eyes are green. She usually smiles with her teeth.

At first, the child guesses wrong a lot. But every time you correct them, their brain quietly adjusts which features it pays attention to. Curly hair gets more weight. Background color gets less weight.

After enough photos, the child becomes pretty reliable.

A neural network does exactly this. Just with numbers instead of a child’s brain, and millions of examples instead of 100 photos.

The Three Parts of Every Neural Network

Every neural network in the world, from the tiny one in your spam filter to the massive one behind ChatGPT, has the same three-part structure.

1. The Input Layer: “Here’s What I’m Looking At”

This is where raw data enters the network.

• For an image: each pixel becomes a number (0 = black, 255 = white), and each number enters here

• For text: each word or piece of a word enters here

• For audio: sound frequencies enter here

Nothing clever happens in the input layer. It’s just the front door.

2. The Hidden Layers: “Where the Magic Happens”

This is where the network learns patterns. Each hidden layer takes the previous layer’s output and transforms it, mixing and combining signals to find increasingly complex patterns.

Think of it in stages:

• First hidden layer: detects simple features (”there’s a curved line here”)

• Second hidden layer: combines those into shapes (”those curves form an ear”)

• Third hidden layer: combines shapes into concepts (”that ear + those eyes = a face”)

The more hidden layers, the more complex the patterns the network can learn. This is why we call it deep learning: the network goes deep with many layers.

Modern AI systems have hundreds or even thousands of hidden layers.

3. The Output Layer: “Here’s My Answer”

The final layer makes a decision:

• “This image is a cat” (classification)

• “The next word is ‘the’” (language generation)

• “The sentiment of this review is positive” (analysis)

• “This email is spam” (filtering)

The Real Secret: Weights

Here’s the math secret, and it’s not scary.

Every connection between two neurons has a weight: a single number that says how much to trust that connection.

A weight of 2.0 means “pay close attention to this signal.”
A weight of 0.1 means “barely consider this signal.”
A weight of -1.5 means “this signal actually points the other direction.”

The entire job of training a neural network is just this: find the right numbers for every weight.

A typical large language model like Claude or GPT-4 has hundreds of billions of these weights. But each individual weight is still just a number, and finding the right set of numbers is what training is all about.

Think of it this way: the architecture is the instrument, and the weights are the music. Change the weights, and you’ve changed what the network does.

How It Learns: Hiking Downhill in the Fog

Here’s the part most people get wrong: neural networks aren’t programmed with rules. Nobody sat down and typed “if pointy ears AND whiskers, then cat.” The network figures out the rules itself, from examples.

Here’s how:

1. Start random. Every weight is set to a random number. The network starts as dumb as possible.

2. Make a prediction. Feed it an image. It guesses. Probably wrong.

3. Measure how wrong. A “loss function” calculates a single number representing the error, basically a score for how bad the answer was. High loss = very wrong. Zero loss = perfect.

4. Figure out which direction to improve. Using math called gradient descent, the network calculates: “If I increase this weight slightly, does the loss go up or down? Which direction makes me less wrong?”

The best analogy: hiking downhill in the fog. You can’t see the bottom of the valley, but you can feel which way the ground slopes under your feet. You take a small step downhill. Then another. Over time, you find your way to the lowest point.

The “valley” is the best possible set of weights. The fog is the fact that there’s no shortcut. The network has to feel its way there.

5. Adjust the weights. Nudge each weight slightly in the direction that reduces the loss.

6. Repeat millions of times. After enough examples and adjustments, the weights settle into values that make the network surprisingly accurate.

And here’s the thing: gradient descent nearly entirely relies on addition and multiplication. When students see the actual details, the most common reaction is: “Is that all it is?”

The Secret Ingredient: One Tiny Rule That Makes Everything Work

Here’s a surprising thing: if you just stack layers of math on top of each other with nothing in between, the whole thing collapses. It doesn’t matter if you stack 10 layers or 1,000. You end up with a network no smarter than a single layer. All that depth, wasted.

Imagine stacking 100 identical photo filters. The photo doesn’t get more detailed. It just gets darker. Same idea.

The fix is a tiny rule called an activation function, inserted between every layer. The most common one is called ReLU, and its entire job is this:

If the number coming in is negative, make it zero. If it’s positive, leave it alone.

That’s the whole rule. And that tiny step, repeated billions of times, is what gives neural networks the ability to learn curves, recognize faces, understand language, and generate images.

Here’s the intuition: without it, a network can only learn patterns that fit a straight line. With it, the network can bend and trace any shape, no matter how complex. The real world isn’t made of straight lines, so this matters enormously.

Different Networks for Different Jobs

As neural networks evolved, researchers found that different kinds of data work better with different structures. Here are the three types you’ll actually hear about:

CNNs: Built for Eyes

Convolutional Neural Networks are designed to look at images and video. They scan pictures in small patches, finding edges first, then shapes, then full objects, the same way your eye moves across a scene.

You use this: Apple Face ID, self-driving car cameras, doctors’ tools that detect tumors in scans.

Transformers: The Architecture Behind Everything Big

This is the breakthrough that changed AI. Instead of reading data one piece at a time, Transformers look at the whole thing at once and learn what to pay attention to. That’s why they’re so good at understanding context. They don’t just see the word, they see how it relates to every other word around it.

You use this: ChatGPT, Claude, Gemini, Google Translate, GitHub Copilot. And increasingly, image and video AI too.

Diffusion Networks: The Artists

These networks start with pure random noise (think TV static) and gradually “un-blur” it into a real image. They learn by practicing the reverse: taking a real image, adding noise step by step until it’s unrecognizable, then learning how to reverse that process.

You use this: Midjourney, DALL-E, Adobe Firefly, Stable Diffusion, Sora.

Despite their differences, all three architectures are built on the same foundation: layers of simple units adjusting their weights through feedback to recognize and create patterns. The specialization is in how they’re wired, not what they’re made of.

The honest 2026 picture: Transformers dominate. They’ve quietly taken over text, code, and increasingly images and video. If you hear about a major new AI product, there’s a good chance a Transformer is at the center of it.

The Limitations (Keeping It Real)

Great tools deserve honest assessments. Neural networks are not magic.

They need a huge amount of examples.
A child can learn to recognize a cat from 5 photos. A neural network might need 10,000. The more complex the task, the more data required. This is why big tech companies hoard data. It’s the raw material for their models.

They’re black boxes.
Ask a neural network why it classified that email as spam, and it can’t tell you. It just did. This is a serious problem in medicine, law, and anywhere decisions need to be explainable. Researchers are actively working on “explainable AI” to solve this.

They’re brittle in weird ways.
A neural network trained on millions of dog photos might confidently call a wolf a “husky” because wolves didn’t appear in its training data. This is called overfitting: the network becomes so tuned to what it’s already seen that it stumbles on anything new or slightly different. It’s why testing a model on fresh, unseen examples matters so much. Neural networks are pattern-matchers, not reasoners, and they fail in surprising ways when they encounter situations outside their training.

They’re expensive to train.
Training GPT-4 reportedly cost over $100 million and consumed electricity comparable to running thousands of homes for months. This is a real constraint, and not everyone can build or fine-tune large models.

But here’s what’s changing: a technique called transfer learning means you can take a massive pre-trained network that already understands general concepts (like what an edge, a texture, or a face looks like) and fine-tune it with a smaller amount of your specific data. It’s like teaching a seasoned expert a new specialty instead of training a complete beginner from zero. You don’t always need to start from scratch.

Try It Yourself

Want to feel neural network learning in action? Try this:

1. Go to Teachable Machine by Google (free, no code)

2. Click “Get Started” then “Image Project”

3. Create two classes (e.g., “thumbs up” and “thumbs down”)

4. Show your webcam 30-50 examples of each

5. Click “Train Model” and watch accuracy climb in real time

6. Test it live. Hold up your hand and see the neural network classify it

You just trained a neural network. What you watched happen (the accuracy rising as more examples were added) is gradient descent adjusting weights in real time.

The Takeaway

A neural network is a system of layers connected by weights. It learns by:

1. Making a prediction

2. Measuring how wrong it was (loss)

3. Adjusting its weights to be less wrong (gradient descent)

4. Repeating millions of times

The nonlinear “activation function” between layers (as simple as “replace negatives with zero”) is what gives it the power to learn complex patterns, not just straight lines.

The more layers, the more complex the patterns. That’s deep learning.

And that’s the system behind every AI product you use today, from the one that recognizes your face to the one that writes essays, composes music, and generates videos from a single sentence.

Coming Up

Now that you know what a neural network is, the next question is: what happens when you build one that’s massive, trained on essentially all the text ever written on the internet?

Next up: Large Language Models (LLMs), the specific technology powering ChatGPT, Claude, and Gemini, explained for normal people.

AI for Common Folks — Making AI understandable, one concept at a time.