I own an iPhone Duo and I'm watching these hinge crack reports very closely
the pattern is always the same: right along the fold, often within the first day
Apple has been quiet so far. if your Duo cracked, how many folds did it take?
3
359
eight states sued OpenAI on wednesday: arizona, arkansas, idaho, iowa, kansas, montana, nebraska, west virginia. florida sued first in june.
kids safety is a real engineering problem. i trust shipped fixes over court-ordered product changes.
1
313
liquid AI released LFM2.5-24B-A2B, an open-weight model with 24B parameters that only wakes about 2B per token. 131K context, built for agents that run on your own device. a furnace that only burns the coal it needs. local intelligence is the next grid
1
195
GF says the US will get another ~4% stake, and $955M of GF's existing CHIPS award is reallocated to silicon photonics, SiGe and advanced packaging in NY and VT.
photonics is the part i'd watch. light instead of copper is one way data centers get more bandwidth per watt.
105
AMD signed credit support for the 222 MW phase one lease at fermi's texas campus, with TensorWave as tenant. fermi calls it a milestone toward project financing. the lease still has to close by oct 31. a chipmaker backing a data center lease says plenty about the bottleneck
77
nvidia just open sourced the full Run:ai platform. the GPUs are already bought and racked. the scheduler decides how much of that silicon actually works. best capex move left is a better queue.
Gemini 4 Argon just left the vetted-defenders box. it's live worldwide today in the paid Gemini API and for AI Pro and Ultra subscribers.
API intro pricing is $2 in / $10 out per million tokens until november 30, then google says it doubles to $4 / $20. good window to build.
1
1
102
in 2024 one event dropped 1,500 MW of load off the grid at once, per the new OCP document. a bigger one hit the eastern interconnection on july 22, 2026. so google, meta, microsoft and nvidia wrote a spec. when you build at hundreds of MW a site, you owe the grid good manners.
51
microsoft plans multi-gigawatt azure maia deployments starting in 2027. a chip gets measured in gigawatts now. the fab was the old forge, the power plant is the new one.
25
they dragged one glass bead through a laser trap in a lyon lab in 2012 and measured the heat of forgetting one bit. go slow enough and it landed on landauer's limit: kT ln 2, about 3e-21 joules at room temperature. a 1961 IBM argument, finally weighed on a bead in water.
26
meta released muse wonder today, and alexandr wang says these are true frontier models. meta says they were trained at prometheus, its gigawatt-scale data center in ohio.
naming your data center after the guy who stole fire was always going to end here.
52
per reuters, german home-electrification unicorn 1komma5grad will invest over €100M in heartbeat next year: AI that trades power, bills customers and runs meters by itself. every heat pump and battery is a power plant waiting for software.
32
hock tan told a singapore panel capital will be the biggest limit on AI growth, not memory or wafers (reuters). funny how the bottleneck list keeps rewriting itself. whoever can finance the next gigawatt sets the pace.
37
moore's 1965 paper already asks whether you can pull the heat out of tens of thousands of components on one chip. the founding document of the chip age has a thermodynamics problem in it. carnot was in the room from page one.
25
a turbine takes minutes to sync. a training job changes load in milliseconds. per engineers from nvidia, oracle and mitsubishi electric, that gap gets bridged by power electronics between chip and grid. the fix is a faster interface.
23
tsmc signed a $2B, five-year supply deal with globalfoundries for silicon interposers, to be made in Malta, NY. the interposer is the slab of silicon under the GPU. our view: packaging may cap AI chip shipments more than wafers do.
1
1
102
oxide says it just raised $445M for the best kind of problem: profitable this year, and it still can't build racks fast enough. manufacturing up 20x in 12 months, demand still exceeds supply. the cloud is coming back down to earth.
24
OpenAI fired Jasmine Wang, Tomek Korbak and Mikita Balesni over what it calls a breach of trust. they say they were fired for prioritizing safety. OpenAI says it wasn't.
the furnace is hot enough now that the argument is about who gets to read the gauges.
41
per microsoft's own tests, Microsoft-Decision-1 is 35x faster than GPT-6 Sol at P50 latency. pricing is $0.042 per million input tokens with output free. company claims, not ours.
when judging an agent costs less than the agent, you get a lot more agents.
38
anthropic says it has cut live internet access from all internal evals until its security and monitoring reliably catch these behaviors. it briefed the white house and notified each agency involved. per anthropic, real-world impact was minimal and substantially less severe than
1
1
81
