IN A CONNECTED WORLD
Multiple perspectives. One place.
Your curiosity, your daily feed.
For you enables ranking by explicit follows and saved stories; marked-as-read history is used only with reading memory enabled. Last 48 hours first, then the last 7 days. Latest turns personalization off.
Geopolitical Economy Report reads the 134-guest Washington dinner as early-stage Cold War bargaining; escalation from 145% tariffs to rare-earth curbs backfired and pushed both capitals toward temporary strategic stability.
In his weekly review, Steve Eisman argues that two numbers run the market: oil prices and the 10-year yield above 5 percent. His deeper warning is that AI spending is increasingly funded through off-balance-sheet structures reminiscent of the Enron era.
PJM's capacity-auction price jumping from $28.92 to $269.92 in one sale built the hidden tax behind your electric bill, driven largely by data centers that were forecasted but never built. The host breaks down how this auction hits your wallet every month.
A laundromat in Hamilton, Ohio, produces $1.025M in yearly revenue and $420K in profit; the video walks through the equipment, data habits, and opening tactics behind that result.
Built jointly by Stanford and Nvidia Research, CLM-8B is an open 8-billion-parameter model that moves decision-making from text generation to retrieval. It measured 44 milliseconds against 570 at 1024 options with similar accuracy, though hits fade clearly as the option count reaches the thousands.
The Nasdaq index is in a critical state; with high downside risk, holding cash is safer than making new purchases.
A tech narrator argues that the wave of new model releases is not a broken slowdown promise but the product of a deliberate strategy by AI labs.
On the Y Combinator stage, the Waddle Labs and RoboCurve founders describe the new era in which general-purpose language models steer robot arms through code, debating evidence from the Isola thesis to a live cube-carrying demo.
A 13-minute video walks through 25 tricks from amateur-looking output to showcase design with Claude; I verified the roadmap of design systems, ready components, and motion libraries against independent sources.
Meta's Muse app kicks off the personal-agent era, giving every user a dedicated cloud computer and sending server CPU demand soaring. Bank of America expects the server CPU market to reach 211 billion dollars by 2030.
Before buying the $12,000 Mac Studio M5 Ultra, the presenter rents a 96GB RTX Pro 6000 for $2 an hour and puts an open-weights model through three real jobs. Two tests end in draws, the frontier model wins the trap question, and the decision rule sharpens: rent first, measure, then buy.
On Toby Newbatt's math a £200,000 loan locked at 1.5% in 2021 costs £800 a month and £1,094 at 5% in 2026; we read the wave hitting 5 million households by 2028 through real-price inflation and UK Finance and ONS data.
The idea of converting the strong, steady wind at 400 metres into electricity with tethered kites instead of giant towers is walking from pilot sites toward a commercial product; diesel generators are the first target, and 2028 is on the calendar.
Google's Gemini Robotics 2 push aims to give AI a body through a three-layer stack that reasons, converts thought into motion, and runs on the machine itself; yet dexterity scores and the generalization problem show the road to physical AGI is still at its start.
Nova, the open source software engineer agent from Supercode, does not wait for you to tell a model what to type. You give it a goal, and it takes the whole chain on itself: reading the codebase, writing the code, running it, testing it and repairing the failures on its own. The agent installs with zero vendor lock-in and can be pointed at Claude, GPT, Gemini or any open-weight model, with the aim of moving AI agents out of the role of guests at the developer's desk and into the role of workers that finish what they start.
Three days before OpenAI's September 29 DevDay, an always-on assistant called "o" leaked, Anthropic said Claude Sonnet 5.5 arrives "in the coming weeks," and MiniMax previewed M3.1 Flash. All three answer the same question differently: raw speed, price-to-performance, or persistent background work.
A day spent in one of China's best high schools shows how education has become a selection machine: mornings that start at six, mandatory exercise in 35-degree heat, cameras in every classroom, sleep treated as a permission rather than a fault, mandatory AI lessons, and a job market absorbing 12.7 million graduates. The same system is inflating engineering departments while quietly shutting down the humanities.
In 2026 Alphabet said it would begin delivering TPU accelerator systems into customer-owned data centers and opened parts of its software stack to run on other hardware. The move turns a defensive advantage into a standard, and the industry's real bottleneck is neither design nor capital but high-bandwidth memory and grid power.
Each of the four ways to earn money, taken from a viral Twitter post, is examined with its arithmetic and its limits; the fifth dimension the speaker adds turns those four engines into a system that actually produces.
In a single week, three world models, four new open-weight language models, two robotics headline models and a plan for AI infrastructure in space all landed at once. The real story is not the model count but how each release shifts cost and licensing terms.
A factory veteran answers the fear that AI will leave everyone unemployed by pointing at history: humanity has played the same film repeatedly and adapted every time. The root of job-loss fear has always been uncertainty, and the real question is not the work lost but the new work created.
Anderson Cooper's question to the Nvidia CEO was simple, but the answer escalated the argument: a lab that cannot control its AI should shut down, yet no regulation matters before the containment technology exists. Huang calls the incident an engineering problem rather than a software defect.
Anthropic is preparing a public offering at a $2 trillion valuation, the hardest test yet for the artificial intelligence investment boom. In 5,209 words of analysis, the speaker works out how that number is justified, why data centers are pricing buildings that do not yet exist, and why Nvidia looks cheap on the more expensive side of the trade.
In the same week, one research shop sees the index finishing above 8200 while two others see 7700 and 7900. The real story is that the numbers themselves do not matter much: a Shiller ratio near 41 and a market-value-to-output ratio at 234% argue for a plan, not a prediction.
Among the noise, meaningful connections.
