Business

A new piece argues that AI productivity gains are real but narrow: a small number of people are getting dramatically faster, while many others see little difference or even more overhead. This matters because if your team just rolled out AI tools and declared victory, the scorecard is probably more uneven than leadership thinks.

OpenAI slashed prices on GPT-5.6, with one version dropping 80 percent. If you or your company are building anything on top of these models, the math on what you can afford to build just changed significantly.

Andrew Ng, one of the most respected names in AI education, just launched LearnVector, a company focused on building one-to-one AI learning experiences. If this works even halfway as well as the promise suggests, the private tutor that used to cost a fortune becomes something a lot more people can actually access.

A detailed investigation found a live marketplace where people pool stolen or shared API keys and resell access to big AI models at a discount. If you or your company pays for AI services, this is the ecosystem that undercuts you on price while also making the services less reliable and less safe for everyone.

Iran's Revolutionary Guard claimed to have destroyed an Amazon data center in Bahrain. If confirmed, it's a reminder that the physical infrastructure behind cloud AI, the actual buildings full of servers, is a geopolitical target now, not just a tech company problem.

A blog post documenting restaurants that used AI to redesign their menus and made them genuinely worse went viral, pulling over 300 upvotes on Hacker News. It is a good reminder that 'AI can do this' and 'AI should do this' are two very different sentences, and real customers are already noticing the difference.

OpenAI has opened up an advertising platform, meaning the free version of ChatGPT may soon surface sponsored answers alongside your questions. If you have ever wondered why a free product this expensive to run exists, here is part of the answer, and it is worth thinking about how that changes what the tool is optimizing for.

A new benchmark tests frontier models on actual business reasoning, the kind of open-ended case analysis done in strategy, finance, and operations, rather than trivia or math. How models score here will matter more to most working professionals than any coding leaderboard.

Consultant Nik Suresh writes that the companies he works with are letting AI hype short-circuit normal decision-making, skipping research, rushing timelines, and dismissing caution as being 'anti-AI.' The concrete consequence is that real business decisions are getting worse, not better, because the tool became the goal.

Perforce is selling a training course for $500 that is narrated entirely by AI, and people on Hacker News are not thrilled about it. The issue is not just the price. It is that paying a lot for professional training and getting a synthetic voice without being told upfront feels like a quiet bait-and-switch.

Anthropic quietly pushed back the sunset date on one of its Claude plans, apparently responding to competitive pressure from a newer model. Upgrade schedules are starting to feel like phone carrier contracts, and it is worth keeping an eye on what you are actually locked into.

The circular financing story is one to keep an eye on not because collapse is imminent, but because the AI buildout is increasingly self-referential in ways that are hard to audit from the outside. The next time you hear that a company raised billions for AI infrastructure, it is worth asking where exactly that money started.

Nvidia sells GPUs to companies like CoreWeave, CoreWeave uses those GPUs as collateral to borrow money, then uses that money to buy more Nvidia GPUs. It is a loop, and when one part wobbles, the whole thing can wobble with it. If you use any cloud AI service, the financial health of that loop is part of what keeps the lights on.

The three GPT-5.6 sizes, Luna, Terra, and Sol, are priced very differently from each other, which means the model a company picks for their app will often come down to budget, not what is smartest. The next time an AI product feels a little dumber than you expected, cost is a reasonable guess.

A well-argued post is making the rounds about how cheap, capable models from China are squeezing the profit margins of every AI company. If the cost of intelligence keeps falling this fast, the business models that everyone built in 2023 are going to need a serious rethink, and that includes the price you pay for the tools you use at work.

Think of it like a company car that costs more in gas and maintenance than the driver earns. A new analysis argues that by 2029, AI spending at many companies will exceed what they would have paid a human to do the same work, which means the ROI math that made AI a no-brainer is going to get a lot harder to defend.

Josh Comeau, a well-regarded web developer who sells online courses, says his latest launch sold about one-third of what a normal launch used to, and his existing courses are down too. He suspects AI tools are replacing the need people used to have for certain tutorials, which is a concrete signal for anyone who teaches, writes, or sells expertise online.

Meta is not just building AI for its own apps anymore. It wants to sell you the spare computing power underneath them, which means it is quietly becoming a cloud infrastructure company competing directly with Amazon and Google, and that changes who has leverage over the AI tools your company might buy next year.

Anthropic also launched Claude Science, a dedicated product aimed at research workflows, on the same day as Sonnet 5. Whether this becomes a genuine tool for scientists or mostly a branding move is worth watching, but it signals that AI labs are now competing on specialized audiences, not just raw capability.

Moondream, a small AI company, published a sharp argument that GPU prices are inflated by speculation and fear-of-missing-out procurement, not by actual demand. If they are right, the cost of building and running AI products could drop significantly over the next year or two, which changes the math for a lot of businesses.

The argument here is that frontier model companies make most of their money in a short window right after a release, which creates pressure to ship fast and charge high. That cycle affects what gets built, what gets tested, and what lands on your screen.

The US government is now scrutinizing OpenAI directly, and the broader AI economy report gives a rare honest look at where the money is actually going. These two things together will shape which AI tools survive long enough for you to keep using them.

Apple is reportedly skipping a whole chip generation to go straight to one designed around AI workloads, which suggests they are betting that running models locally on your device is the next big battleground, not just cloud AI you ping from an app.

OpenAI now has its own chip, called Jalapeño, built with Broadcom. Right now you pay for every word an AI generates partly because renting chips from Nvidia is expensive. A company that owns its own silicon controls those costs, and eventually, those savings or that leverage.

TLDR AI flags orchestration models, a wave of departures from DeepMind, and new thinking on how AI agents loop through tasks. The DeepMind exits are worth a second look because talent moving out of a lab often signals either frustration with direction or the quiet start of something new.

The Anthropic and SK Telecom partnership is drawing scrutiny, and the details behind it matter for anyone thinking about how AI deals between big labs and big telecoms actually get structured. Who controls the model, who profits from it, and who is accountable when things go wrong are questions that do not have clean answers yet.

Engineering leader Charity Majors argued that 2025 flipped the economics of writing code: generating it got cheap and fast, so the hard part is now reading, judging, and maintaining what the AI produces. If your job involves software in any way, the skill that just got more valuable is not typing, it is knowing whether the output is any good.

Tim Ferriss and others are asking whether AI has already replaced the self-help book, and the Hacker News crowd has a lot of feelings about it. Worth reading less for the answer and more for what it tells you about how people are actually using these tools day to day.

Arvind Narayanan and Sayash Kapoor make a careful, evidence-based case for why software engineers have not been replaced by AI and are unlikely to be fully replaced anytime soon. It is a useful anchor the next time someone tells you a whole profession is about to vanish by next quarter.

The McSweeney's joke about Jenny's crematorium is funny because it is accurate: a lot of AI investment is one company paying a second company, which then pays the first company back for compute. Understanding that loop helps you read any headline about billion-dollar AI deals with the right amount of skepticism.

The Techdirt piece making the rounds argues that replacing employees with AI is a symptom of poor management, not smart strategy, because the hard part of most jobs is judgment and relationships, not the keystrokes. It is a useful counterweight to the replacement narrative that tends to flatten every nuance out of the conversation.

AWS Bedrock will start retaining your conversations with Anthropic's most capable models for 30 days and sharing that data with Anthropic directly. If you are using Bedrock for anything sensitive, that is worth a conversation with your legal or compliance team before the change takes effect.

Apple announced that its new AI architecture leans heavily on Google Gemini under the hood, which means two longtime rivals are now deeply intertwined in the device sitting in your pocket. If you care about whose hands your data passes through, this is worth a second look.

A writer followed up on their earlier post about LLMs slowly eroding their freelance career, responding to reader pushback. It's a small post with a real conversation attached, and it's one of the more grounded accounts of what job-level AI friction actually feels like from inside it.

Google employees are apparently sharing internal memes about their own AI products being bad, and a reporter caught it. When the people building the thing are the ones making jokes about it, that tells you something honest about where the technology actually stands right now.

The S&P 500 has a simple rule: you need to be profitable for four straight quarters to join the club. OpenAI and Anthropic do not qualify, which means the companies building the tools you use every day are still spending far more than they earn. That is worth keeping in mind the next time pricing changes or a free tier disappears.

Apple apparently doubled production of the MacBook Neo because demand is so high. The reason this matters for AI is simple: more powerful personal chips mean more people can run smaller AI models locally, on their own hardware, without sending data anywhere.

32GB of DDR5 RAM now costs $375, up sharply, because AI data centers are hoovering up memory chips faster than factories can make them. If you were planning to build or upgrade a PC this year, you are now subsidizing someone's GPU cluster. That is not a metaphor.

Uber started capping how much its engineers can use AI coding tools like Claude Code, because the bills got uncomfortably large. This is the part nobody puts in the press release: AI tools are genuinely useful, and genuinely expensive, and at some point the spreadsheet wins. It is a healthy reminder that 'transformative' and 'affordable' are two different questions.