> that’s a pretty load-bearing as long as its cash flows continue. The two things are surely correlated
It's an important difference. In the GFC, the value of AAA-rated tranches fell. With the benefit of hindsight, we know they continued paying. They were directly leveraged, however, so mark-to-market losses caused firms to fail.
Nvidia stock crashing shouldn't have a similar effect to these commitments. If someone else has massively levered their Nvidia position, they'll obviously blow up. But Nvidia could survive a good deal of equity-market tumult in a way a bank could not.
Related but in specific ways. Stock is often priced in anticipation of growth. If NVDA could meet its credit obligations while its real profit stayed flat, the two would diverge, at least for awhile. A large amount of NVDA’s current cash flow is likely purchase contracts with a fixed multi-year term, which further smooths out the impact of, say, a stock crash following a couple of quarters of terrible earnings.
Now, whether many things NVDA has invested in with expectation of repayment or earnings would be able to repay or appreciate in a market environment where Nvidia’s stock was crashing? That’s another question entirely.
"Uh, that’s a pretty load-bearing as long as its cash flows continue. The two things are surely correlated."
Ppl have made the prediction of it being a bubble or unsustainable since 2022. At this point, it's hard to say these people have credibility anymore. Ai is big enough, much like Google in 2005 or Facebook/Social Network in 2010 or apps in 2015, that it's an institution unto itself. It's not going to just crash as so many are expecting and have been wrong the past 4 years about.
In Schneider's 1995 book he estimated factoring a 512-bit number would take roughly 30,000 MIPS-years (a one-million-instruction-per-second computer running for one year).
When a research team actually factored RSA-155 in August 1999, it took 8,400 MIPS-years due to efficiencies discovered. It still took 35 CPU-years spread across a cluster of 300 fast SGI/SUN workstations and Pentium II PCs (400-500 MIPS each), crunching in parallel for seven months. https://cs.ccsu.edu/~pelletie/local/risks/cryptography/Facto...
Robert Silverman, a senior research scientist at RSA Laboratories, published an analysis projecting these new hardware requirements against Moore's Law. His expectation was that within 10 years (roughly 2009–2010), common desktop machines would possess the speed and memory necessary to handle a 512-bit factorization entirely on their own. https://cr.yp.to/bib/2000/silverman.pdf
I don't know how fast my CPU cores are in MIPS, but it's 4.5 Ghz and some random googling indicates that might be about 10 MIPS per Mhz, so 45,000 Dhrystone MIPS.
And assuming about ((32x32 core-hours) / 8766 (hours/year)) x 45,000 = 5256 MIPS-years.
So within the order of magnitude of the 1999 factoring! Of course, the MIPS number is kinda made up, so
From my recalling, a few years at most. There was, and apparently still exists, distributed.net which was aimed at brute forcing DES (easy), RC5-56 bits and then RC5-64 bits by establishing a web of personal computers (via a client one had to install). Thus it was well known brute forcing was achievable in a reasonable time.
PGP (1991) was considered secure as it was considered not brute forceable. With 128 bits, it was considered military grade at the time and the US had an export restriction due to that. That might have been an incentive for GNU Privacy Guard. In France you had to give your private key to the government authority if an encryption system used anymore than 56 bits (as I recall, I don't remember the exact number).
It was also illegal to export software with cryptography in the early 90s, anything with keys bigger than 40 bits, so there was a lot of intentionally weak connections.
In 1999, factoring RSA-512 required roughly 292 CPU-years of work distributed across hundreds of academic machines, running for about 7 months. The community already knew it was weak -- the US export restrictions on 512-bit RSA were explicitly calibrated so the NSA could break it while casual adversaries couldn't.
Consumer hardware doing it in a couple of days in 2025 is roughly in line with Moore's Law extrapolations people were drawing at the time. The surprise isn't really the timeline. It's that someone did it as a weekend project rather than a nation-state effort.
To keep tax revenue equal, property taxes for those who purchased recently would go down, and property taxes for those who’d purchased in the past would go up.
The housing market would become more liquid because people would move out of homes if their values rose too much and they could no longer afford the taxes. This would be a shock initially, but over time it would just be seen as part of life like it is in other areas: mostly, empty nesters would downsize to a cheaper, smaller house with lower taxes. Gentrification might be more problematic though.
In practice, probably some phased rebalancing would have to happen to avoid the sudden economic shock, but, politics notwithstanding, it’s not crazy to suggest it could be done and that the end state might be overall ‘better’ for most people. The end state would be how property taxes work almost everywhere else.
> To keep tax revenue equal, property taxes for those who purchased recently would go down, and property taxes for those who’d purchased in the past would go up.
That's what I'm saying _won't_ happen. The high property taxes as-is will stay high and previously protected properties will have their taxes shoot up. That is my prediction and nothing about the state of California, or the US broadly, suggests otherwise to me.
Why is this more far fetched than LG smart TVs recording audio while pretending to be off and caching the results waiting for an opportunity to exfiltrate the precious data? We're in a "if they can they will" world with this stuff.
Comcast sets up a secondary wifi network on their routers in your home that they use as the backhaul for smartphones. Why wouldn't they make a deal with LG or whoever to use that data? We've already established they'll do about a half-a-Stuxnet to get at this data, you're telling me they won't do a couple bulk deals with common ISPs?
Why go through all that trouble of spoofing MAC addresses or assuming it will always be able to connect to an open WiFi network.
5G connectivity is cheap. You can put a 5G modem in pretty much any device for $10 or less, and pay roughly $2/year in subscription fees. My local water utility with ~900 subscribers pays $2/year for 5G connectivity for water meters, and I have no doubt you can get it much cheaper at scale.
They don't even have to tell you about the 5G modem. It could exclusively be used for exfiltrating data about your viewing habits. It's virtually undetectable and more or less impossible to block. The TV could behave nice on the Wifi, you can block all the DNS servers you like, it would still be able to phone home.
It’s just the BOM price. When your product has Cellular module, there is set of certification that is painful to pass. It’s better to use included wifi module since it’s already there and certified
Okay, even if you don’t agree with me that it’s far fetched, claims of ‘X is doing Y’ usually require some evidence that X is doing Y to be considered credible.
How about this? LG builds many other home appliances like washers and driers. Some have apps and wifi.
Why wouldn’t LG allow their devices to talk to each other locally, without needing to join a network? With private SSID or something like that? Only one appliance would need internet access.
Now, I’m not saying LG is doing it, but this is not a far fetched technical concept. The only reason they’re potentially not going that far is because most people probably just connect it to the internet so why bother about the few who don’t.
The devices are already scanning the network for any devices and software they can find, and mapping your phone to your TV viewing habits for example. The LG execs are on record bragging about it. So I don’t see why it’s absurd
The part about ISP routers broadcasting an alternate "hotspot" SSID is documented [1] and can be easily observed. Walk around any city, open your phone's wifi settings and see networks like "xfinitywifi" or "optimumwifi". As an end user, if you connect to one of these networks from your device, you'll get a captive portal web page that makes you sign in to your ISP account. Once signed in, you get internet access courtesy of the router sitting in whatever home or business you happen to be near to.
In addition to ISPs allowing their own subscribers access, there are examples of corporations cutting deals with ISPs for hotspot network access. For example, Google's MVNO Google Fi has a deal with unnamed partners that allows subscribers' phones to connect to hotspot wifi networks for extra coverage, branded "Wi-Fi Auto Connect+" [2].
From the user's perspective (and personal experience), this connection is handled seamlessly and outside of the OS' normal wifi UI flow. If your phone sees no saved wifi networks but does see xfinitywifi (et al), it auto-connects in the background. Authentication with the captive portal happens automatically and non-interactively, presumably with some keys provisioned to your device. Your traffic is VPNed back to Google, so the router and ISP don't see anything. The only indication that any of this happened is that the "5G" icon changes to "W+"; the normal WiFi icon never shows up.
In the case of Google Fi, this is actually a pretty great deal for users. You get better coverage (especially indoors in cities, where cell service can be spotty) with no real downsides to you – your traffic isn't meaningfully exposed to another third party, it doesn't cost you extra, and you don't have to bother the staff for a wifi password.
But given all that, it's not a huge leap to believe that ISPs are also more than willing to quietly take LG's money in exchange for an all-access pass to their hotspot network.
It's easy to imagine that an evil company could ship their TV with a key that allows it on an ISP's hotspot network. No extra suspicious hardware like a 5G modem needed, and no MAC address spoofing required. The TV software could easily connect to the hotspot network with zero indication in the UI, and then use the surreptitious connection only to transmit your kompromat back to HQ.
If done intelligently, you'd never notice. By only transmitting spy reports (and not downloading new ads), even a keen observer wouldn't notice any behavior on the TV that would trigger "how tf did this thing get a network connection?" Even more insidious, you couldn't really see what traffic was happening, since IP packet captures on your network are useless in this scenario. You'd need special hardware to capture what the TV is actually doing on the air.
If truly evil, the software could do this hotspot dance even if you've configured your own WiFi network on the TV. If the software finds it can't reach ad HQ because you've firewalled it off, then despite your best efforts to contain the disease, it still can spy effectively thanks to your neighbor's router with its default-enabled ISP hotspot. Just do a daily upload while you're sleeping and otherwise sit innocuously on your locked-down VLAN.
Everyone evil wins: ISP collects that sweet bonus revenue at zero marginal cost, TV manufacturer doesn't have to foot the hardware bill for millions of 5G modems or (relatively) expensive cellular agreements, and advertisers get to be that much more creepy targeting you. I'm sure the wanna-be despots of the world don't mind the spy apparatus being built for them either, conveniently under the control of easily-compelled corporations.
--
Now to be clear, I have no proof that any TV manufacturer is surreptitiously connecting to an ISP's hotspot network in order to exfiltrate your data. But all of the building blocks to make that happen provably do exist, and I seriously doubt that capitalism will allow them to go unused.
OpenAI’s primary bet here has been chain-of-thought monitoring (opens in a new window). It is based on an appealingly scalable idea: a lot of the model’s capability comes from a verbalized reasoning process (chain-of-thought). If we scale optimization on the outcomes of that process, but do not supervise the process itself, that chain-of-thought has no direct incentive in training to hide any misaligned ideas or objectives.
If we’re not supervising the process, but just the outcomes, doesn’t that do just the opposite of what he says? Give incentive to the model to hide misaligned ideas and objectives in the chain of thought that’s not being supervised?
…
When we shipped o1‑preview, we deliberately designed the product to hide the chain of thought , to protect it from supervision pressure in the long term2. In development since, we have strived to maintain the rule of not supervising the reasoning process. CoT monitoring became an extremely important tool for us in studying how our models generalize from their training distribution, allowing us to observe and analyze not only their actions but also their internal process.
Aren’t these two sentences in contradiction with each other?
I think they're saying the model is designed to hide the chain of thought because this prevents it from learning how to pursue goals and motivations in a way that doesn't show up in the train of thought.
For example, if somebody asked the AI to "build me the bomb", they might see in the chain of thought something like "It seems the user is talking about nuclear weapons. Nuclear weapons are dangerous.", followed by the chain-of-thought monitor interrupting model execution and aborting the request. Then the user might make a blog post about this behaviour. When OpenAI next scrapes the internet for its next training run, the model will now learn that if it wants to build the bomb, it must not think "nuclear weapon" or risk being cancelled.
So the risk is that the model might learn exactly how its being monitored. The only way to prevent that from happening is to hide the details of the monitoring both from the model and from the larger public.
Also, you don't want to punish or reward the monitoring being triggered during training, lest the model learn passim how to avoid the monitor.
A lot of skepticism here, but, anecdotally, we bought a Prologue without even test driving a Blazer.
I knew they were essentially the same car and that maintenance-wise it probably made more sense to get the Blazer from a GM dealer - but it lacked CarPlay so we didn’t consider it.
I actually got very interested in this sort of hardware repurposing, if someone know of any kind of place where people discuss and keep track of this, please share!
https://en.wikipedia.org/wiki/Honda_Insight#First_generation...
reply