> For network operators, it raises a different question: could the infrastructure that already generates subscriber revenue also generate a second, sensing-based revenue stream?
That's kind of strange. What are you going to sense that the cellular network doesn't already know? Air traffic? Car traffic? Drones?
The most useful application for this is probably drone detection.
We might see this rolled out in China, which is setting up traffic control for the "low altitude economy".[1] This includes drones, now, and flying cars, soon. There's also interest in tracking balloons, traditional hot-air flying lanterns, and other small sky traffic obstructions. As usual, Shenzhen leads in this area.
Historically, that happens now and then. Examples include color television cameras and fluorescent lamp ballasts.
RCA developed the first good color TV cameras. They were monsters. Originally, electronic TV (as opposed to mechanical scanning) used Zworykin's iconoscope. This needed huge amounts of light. It didn't have a photo-multiplier stage to boost sensitivity, but it did accumulate charge from light over the whole frame time. There was an alternative approach, Farnsworth's image dissector, which also needed huge amounts of light. It had a photo-multiplier stage but only the light from the dot currently being scanned by the scanning beam mattered.
So RCA threw money at the problem. They developed the image orthicon camera tube. It combined the advantages of both approaches, integration over a full frame and a photo-multiplier stage. It was insanely expensive, costing thousands of dollars in the 1950s. It needed about twenty different voltages and multiple independent power supplies. It worked reasonably well.
Then came color. That took three image orthicons (RGB) inside a box the size of a small refrigerator, along with optics to split the incoming image into three colors. Plus
a rack of analog electronics per camera, connected through a large cable. Plus an alignment problem from hell, with a huge number of adjustment knobs.[1]
Scaling worked! From the 1950s to the 1970s, color TV broadcasting was powered by those monsters.
Now, independent of this, other people were working on other approaches. Vidicons, silicon detectors inside camera tubes, and finally fully solid state image sensors. Eventually, in
1983, after thirty years of image orthicon monsters, RCA produced a usable solid state camera. Today, a cell phone camera costs about US$10 in bulk, and it's far better than the monsters.
It took thirty years of low-priority development to get past the giant image orthicon cameras.
Electronic fluorescent lamp ballasts had a different problem. That was a suppressed invention.
The magnetic ballast industry was dominated by Magnetek. Someone invented a solid state replacement, which was smaller and cheaper, and licensed their patent to Magnetek for royalties. Magnetek just sat on it, not wanting to gut their core market. The inventor hired Townsend, Townsend and Crew in Palo Alto and sued. (Read this in the law firm's brag book in their office.) Eventually, the inventor got a hundred million or so, solid state ballasts went into production, and Magnetek seems to be gone. (The company called Magnetek today, which makes crane controls, is an offshoot of a Litton unit, and seems to be unrelated.)
There are also parallels from the steam locomotive industry, the Baldwin Locomotive Works being the best known. Baldwin actually built the first Diesel locomotive. They made a few half-hearted tries at Diesels over several decades, then gave up and went back to steam. Baldwin ceased all locomotive production in 1956. The corporation lingered on for a while, not doing much.
(So it must be those Orthicon cameras which cause the distinctive weirdnesses of early TV images like the reflections from highly-polished surfaces (such as the bells of brass instruments and saxophones) which gleam black? https://www.youtube.com/watch?v=2uTQQKU6ge0&t=111s )
Nice, it got a press mention, with a quote from Crew, one of the law firm's partners.
Suppressed inventions are very rare, but that's one of them.
One of the few others is Airadar, from 1973.[1] Not "AI", "air radar". This
was a phased-array radar for small aircraft. Way ahead of its time. So far
ahead that the USAF was upset about it being available commercial before it
was in use by the military.
Suppressed inventions aren't "very rare" per se. You simply don't hear about those that aren't being rescinded while the inventor is still alive.
According to official data from the U.S. Patent and Trademark Office (USPTO), there were 6,653 active secrecy orders in effect at the end of fiscal year 2026. Roughly 100 to 350 new secrecy orders are added every year. Many remain locked away for decades or more.
If the U.S. government puts a secrecy order on a patent application against the inventor's will, they are legally forbidden from talking about it. The inventor cannot talk about how it works, publish research on it, sell it, or market it. They cannot even discuss the exact classified details with a regular lawyer unless that attorney has the proper security clearance.
For as long as the big companies are committed to focusing on scaling LLMs it does seem reasonable that they don't want their overall efforts led by people like LeCun or Hassabis who don't particularly want to scale LLMs. But it does feel as if there's something a bit extra about many LLM maxis' active hostility to deviationists.
I think Hassabis’ position is far more nuanced than LeCun’s. From what I can see, LeCun believes LLMs are a dead end for AGI while Hassabis position seems to be that some new technique is needed combined with LLMs.
JEPA's method isn't dead, non-generative self-supervised learning that predicts abstract representations (embeddings) of missing information. It is understanding latent structures rather than creating content.
(2016), from the New York Fed, back during the Obama administration. Reading this, my first thought was amazement that this came out under Trump. It didn't.
This paper is mostly about risk associated with debt. That's the Fed's job, after all. The current AI bubble is more on the equity side, although, when it pops, we'll see who's been leveraging their equity positions with debt. Historically, up until the 1980s. corporate debt was supposed to be low-risk. If a company went bust, that usually took down the top executives, who were unlikely to get good jobs ever again.
As a result, the bond market was boring. Not much trading. Buy and hold.
Then came "high yield bonds" - debt instruments with high returns but a significant
risk. The concept was that a broad portfolio of high yield bonds became, statistically, a low-risk loan. Risk management and the Black-Sholes model became a major part of Wall Street activity. That's the background of the paper. The audience to which that paper is addressed knows that, so they only hint at it.
This tends to break down when there's some across the board change. The 2008 US housing crisis is the big modern example. Housing prices went down nationally. The assumption of many Wall Street risk management models was that the risk of independent debt instruments was largely uncorrelated. Sometimes, it's not. The Big Short (2015) is the movie version of this, and it explains all this in an entertaining way. (That's quite an achievement for Hollywood - making an entertaining movie about mortgage risk correlation. It's not about bad people. It's about bad risk models.)
That's the underlying basis of the culture problem. Once investors start accepting that some debt instruments will routinely go bust, perverse incentives appear. Then you get Milken, Boesky, LTCM, and Enron. (Not Madoff; Madoff was a straight Ponzi scheme.) The paper tries to tie this to financial culture.
That's not wrong. But it's not the basis of "postmodern capitalism".
All this is to some extent a result of tax policy. Interest received is taxed much more favorably than stock dividends. That encourages debt financing, equity to debt conversion, high-risk lending, and leveraged buyouts. ("Private equity" is mostly leveraged buyouts. Usually the private buyer doesn't have the funds for an all-cash deal.) The line between debt and equity has been blurred. Both are ways to pay for money, but with different tax treatment. There are instruments such as convertible bonds which combine aspects of debt and equity. This effect is so strong that the number of publicly traded companies in the US is down, while the number of tradable financial instruments is way up.
So, the paper is incomplete in that it fails to identify how tax policy increased Wall Street's appetite for risk.
> Once investors start accepting that some debt instruments will routinely go bust
Debt always carries a default risk. This is a fact and has since the beginning of time. I don't see how anything fundamentally changes in the 80s.
You are wrong about the tax treatment of dividends vs interest received. In fact, dividends have the superior tax treatment to investors because of the qualified dividend rule. What you are looking for is the tax shield effect at the level of the company. Interest is deducted from taxable income, so if you have a company with no leverage earning $5 million a year in pre-tax profit, you pay $2 million in tax, with $3 million left to distribute. If the company takes on debt with $2.5 million in interest, now taxable income is $2.5 million with $1.5 left over after tax. This means now you have income distributable to investors (debt and equity) of $4 million vs the $3 million unlevered.
> Take a sentence as volatile as "I'm going to kill you." Standalone, I’m calling 9-1-1.
Had to do that once, pre-emoji. I have a domain in .com which is also a domain in .co.uk. That's owned by one of the famous British "public" (i.e. private boarding) schools. I used to have a catchall email address for that domain, and once in a while I'd get student emails, which I'd return with a note about why the right TLD matters.
One day, I got a student-to-student email "I am going to kill you tonight."
No indication of humor.
This was shortly after the Columbine massacre. I couldn't ignore it.
I'm in California, and the message was from the UK, where it was late at night. I had some trouble getting the right phone number, but after a while I was on the phone with the very British headmistress of the school, who'd clearly been awakened for this.
Once I read her the message, she was thoroughly annoyed, but not at me.
She recognized the student's name.
It was made clear that someone was going to have a Very Bad Day. "He's only 12", she said.
If that message had come from a US school, there probably would have been a SWAT callout.
Does anyone still use emoji? Except for ones on forum replies.
I sometimes get an emoji as a one-glyph reply to texts. That's about it.
They were far more popular a decade ago.
Yes. When there is a post awaiting moderator review on one of my sites everyone can see the toilet paper emoji next to the board listing. I also have animated CSS that makes an emoji of a spy pop up behind threads in a couple themes. During Halloween I have an emoji bat emoji fly across the screen and a spider emoji drop down assuming the client does not have animations disabled on their desktop as I respect motion preferences. [1] (click set to use the spooky theme)
It may be silly to some these days and that's fine with me. I enjoy it and I have an odd sense of humor.
Marketing copy and job offers are full of them, and AI has made that much more frequent. I assume you include all kind of pictures punctuating sentences, such as "someone at a computer" when you mention working, etc., not just yellow faces.
It's going to take a year or two more. Will the humanoid robots being built be useful? Will Tesla's robotaxi ever work without a huge number of remote monitors? Will the free models start to displace the expensive ones? Will someone figure out a way to do all this at lower cost? We don't know yet. But there are too many huge companies with rather inflated market caps. Some of that has to be air.
Really? I find local models such as Qwen 3.8 and Muse Glimmer to be pretty good. Not Opus or Sol, but.... Close to Sonnet maybe? So if I had to only use the local models from now on, it would not be a huge loss. Granted, I can't work on parallel on multiple tasks locally. But it's still 48 GB of RAM vs 750 GB or what is required for GLM5.3
You’d happily take a TB of RAM if offered though. The fact you can’t get it is evidence that hardware manufacturers have a huge amount of life left in them and the current correction is a buying opportunity.
> There’s no version of the immediate future where this isn’t the case.
Pity you qualified your prediction with "immediate"; a simimlar sentiment was around just before the dot-com crash: "There's no version of the future where we won't be able to use all that extra network infra."
Turned out, in the immediate future, just post-crash, we actually didn't need them. They only got used a decade later.
> So we have enough RAM kicking around in the world today that will suffice for the next 5-10 years?
No, but we may have enough GPU (with their RAM) sitting around in datacentres that our AI needs will be taken care off for the foreseeable future.
I mean, RAM is not really fungible - these valuations are not tied to RTX5090s, they are tied to the server GPUs.
The investment was not used to buy RAM sticks for desktop and server m/boards. ANthropic did not use their VC money to buy RTX5090s. The VC money went to power generation, datacentres, and server GPUs with 768GB of RAM.
So, yeah, if the bubble pops, we're well taken care of in the immediate future by the stuff the VC money already bought.
And they’re gonna let you access that super cheap at fire sale prices yeah? Because of the oversupply. Right.
And what about local AI and agentic stuff? We can all stick with 32Gb of ram at home and piddly lil GPUs?
In my opinion this turmoil is hedge funds and market makers flushing out liquidity, same as they always do. Paying Bernie Sanders to tweet “no more datacentres” is pretty cheap to do.
> And they’re gonna let you access that super cheap at fire sale prices yeah?
That's literally the definition of a bubble-pop. The assets get sold for pennies on the dollar (sometimes even fractions of a penny on the dollar).
> And what about local AI and agentic stuff? We can all stick with 32Gb of ram at home and piddly lil GPUs?
Won't matter if there's demand for that, that specific market can't use the stuff already purchased by VC money. TBH, I won't even care about self-hosting if the bubble pops, because all those things the VC money purchased will get bought up for basically scrap prices, then hosted for cost of hosting + tiny markup.
There's no principle to repay, so prices are going to be opex + markup, not opex + training costs + repayment on loan + commitment penalties + markup.
> Extreme fear = always a buying opportunity.
Which brings me back to your assertion: that all these things already purchased by VC money will still be needed. We found that to be untrue during the dotcom crash: Hardware makers (those that survived) stock took 17 years to get back to where they were pre-crash, the 0.01% software stock that survived the crash took a decade to recover.
The only way this is a buying opportunity is if there is no crash.
These assets aren’t going to get sold for pennies on the dollar. Too many people would be interested in them and they’ll get bid back up to 90 cents on the dollar.
We’re currently in a situation similar to the dot com crash but one that has enough differences for one to not apply the logic of that crash to it. The main difference being hardware companies are making huge amounts of money and cannot keep up with production due to excess of demand compared to supply. Their only approach in your crash scenario would that they would drop prices, but that will lead to yet more demand and prices rising.
I’d happily heat my house on a rack full of GPUs and memory and use all the AI capacity that came with it. But there’s no way hardware prices are going to also crash to allow me to do that.
> The main difference being hardware companies are making huge amounts of money and cannot keep up with production due to excess of demand compared to supply.
You're describing exactly what happened during the dot-com crash and calling it different. Sun, Cisco, etc literally could not keep up with demand pre-crash, and yet they still tanked (Sun tanked enough to go under, Cisco took another 17 years to recover).
I mean, you're calling it a main difference, but describing exactly the same thing.
You’re describing different hardware with drastically different users and use cases calling it exactly the same.
I don’t need a rack of Cisco routers on my desk. I’d happily take a Sun workstation loaded with GPUs though.
The only real question that will come out of this is whether we buy the hardware directly or rent the hardware from Anthropic/OpenAI etc. no doubt we’re seeing the winners and losers get sorted out right now, and the strategy of doing the growth behind the scenes and then doing IPO (instead of IPO on no revenue as in the dotcom era) is proving to be yet another stumbling block. Valuations can come down as much as they need to but the AI genie has been freed now, regardless of how the market is choosing to price it.
Nasdaq will be higher in 12 months than it is today. Democrats or no democrats, no politician is going to vote to crush the primary driver of revenue growth in the S&P. Pooh Pooh it all you like.
Motor winding is an amusing problem. Everybody making motors has a powered winder. It's handling the wire ends that can be either manual or automatic. If you look at videos of winders, automating the wire end handling is most of the machinery.
Here's an overview of winding automation:
Full auto, robotic transfer.[1]
Full auto, manual transfer [2].
Semi auto. [3].
Manual, powered. [4]
Manual, unpowered motor overhaul.[5] (Country girl in China who fixes motors and engines.)
Those winding machines are all from the same company, NIDE. You can buy as much automation as you want, depending on your labor costs. Option [4] is cost-effective if you can have one machine tender per machine.
The opportunity for the west is that an engineer with todays resources can work on building [1] for $5-10k and get into the DC motor business from your shed. Im working on this myself right now, I want to try and start an autonomous manufacturing company here in Australia, maybe I'll start with motors. I predict once the war is over, Ukraine will become a robotics & manufacturing powerhouse because they've already got this model working
1. The terrible quality of the locally manufactured items. I am in Sydney and will absolutely not buy anything marked Made In Australia. I have been burnt too many times.
2. The nearly 100% mortality rate among the companies.
Rotorlab (rotorlab.com.au) are or were doing this in Australia.
They had a couple of posts about their machinery: a winding machine, CNC mill, etc.
I believe they were bought out pretty quickly by a US company. Can't recall who though.
Are you winding something in an incredibly specialized kV / wattage / torque or something? DC motors are incredibly inexpensive, how could a hobbyist sell them?
Also really difficult to do, with huge tradeoffs depending on whether the construction is 2-core, 4-core, interpoled 4-core, coreless, slot-wound, salient poles, and many more configs and winding types. In particular cheap salient-pole motors can be wound directly with automated machinery while better slot-wound motors need to have the coils formed first, then inserted into the rotor slots, which is a manual process and complicated further by the fact that the coils are skewed to minimize torque ripple, no-one's really managed to automate that process yet.
So if you saw a factory making a mixture of salient-pole motors for toys and slot-wound motors for serious use you'd see exactly what TFA is describing, salient-pole automated, slot-wound with human operators.
If you want to see what real highly-automated Chinese manufacturing is like, look at one of Bunny Huang's videos on electronics assembly. Then for contrast look at the videos of how it's done in Pakistan.
As to prediction, there were some accurate early hits. A Logic Named Joe nailed the search engine with a chatbot front end, and pointed out the problems of systems smart enough to tell people in detail how to do bad stuff.
Much early SF thinking was that AIs would be logical. That's what I was taught, going through Stanford back when the logicians were in charge. They were trying to hammer the real world into predicate calculus. That turned out to be totally wrong.
Almost everybody in SF underestimated the difficulty of robot unstructured manipulation.
So did people in AI. They still do. Most SF assumed that robots would be able to do
most eye/hand tasks before they could talk. No. Still unsolved. LLMs don't seem to help all that much. Written language doesn't have much info on grasping, and watching without doing is only marginally helpful.
Asimov's robot novels were all about alignment, and why alignment is hard. That was prophetic.
Banks' Culture novels are great, but they're for a future that has a lot of new physics.
Now what? What we seem to be headed for in the near term is the fast automation of most white collar work, but not blue collar work. SF foresaw that future a few times. One example is the Rosinante Trilogy from 1981, where there's a boss of an orbital space colony, an AI as middle management, and union labor doing the real work. That's a hard SF view. There's How to Live Safely in a Science Fictional Universe, from 2010, where the protagonist's supervisor is "Microsoft Middle Manager 2.0 with passive-aggressive set to low." It's a minor part of the comedic plot there. And, of course, there's Manna, from 2003, which is looking all too likely.
I have enough ad and tracking blocking that the ads I get are totally misdirected. I get ads for drugs for diseases I've never heard of. I get ads for equipment for processing and bottling olive oil in bulk, despite being nowhere near that industry. I get ads from China for very large hydraulic presses, despite not having any great need for metal-forming equipment. And World Dairy Expo wants me to attend, although I don't have any cows.
That's kind of strange. What are you going to sense that the cellular network doesn't already know? Air traffic? Car traffic? Drones?
The most useful application for this is probably drone detection. We might see this rolled out in China, which is setting up traffic control for the "low altitude economy".[1] This includes drones, now, and flying cars, soon. There's also interest in tracking balloons, traditional hot-air flying lanterns, and other small sky traffic obstructions. As usual, Shenzhen leads in this area.
[1] https://eastasiaforum.org/2026/10/16/chinas-low-altitude-eco...
reply