4,562 karma · joined August 10, 2025
It's why they have plans to elongate the stack further. If they have this much TWR to go around at liftoff, they can afford to cash in that extra thrust to pack more fuel.
Haha! Nope!
LLMs have consequentialist thinking sometimes. And sometimes they don't. Sometimes they follow the prompts, and sometimes they don't. There's NO single magic prompt that fixes all the weird behavior of modern LLMs, and it's baffling that anyone who has ever interacted with an LLM would expect there to be one.
Believe it or not, deciding that you weren't wronged and not suing isn't a crime. It happens all the time. What people do with each other is up to them.
Did Bob allow his friend Jack to borrow his truck? No. Does Bob want to sue Jack for taking his truck anyway, and driving it into a ditch? No. Does Jack owe Bob big time for the mess he caused? Yes, but not in any formal legally binding way.
This works for corporations too. When two corporations find themselves at odds, threat of legal action is often used by one company against another as a leverage to resolve things behind closed doors instead. In a more amicable fashion - with no legal expenses of a protracted court battle and no loss of reputation on either side.
It never fucking worked that way and maybe never will.
Prompts don't define model behavior. Prompts steer model behavior. Instruction-following over long horizons is NOT a guarantee in LLMs. Instructions doing what you want them to is NOT a guarantee in LLMs.
Saying "don't exploit the box please pretty please" might actually cause an LLM to exploit the box more often, for bizarre "don't think of a pink elephant" reasons. 3% rate of exploiting the box (no prompt) -> 11% rate of exploiting the box (with prompt). Because fuck you, that's why. Increased salience -> increased incidence. Welcome to AI tech - good luck and have fun.
Frankly, I expect weirdness like this to be even worse in internal unreleased models that had their behavior fried with who knows what experimental training techniques.
AI can't be an actual powerful, dangerous technology! Thus, any indication that an AI may attempt concerning things or may possess dangerous capabilities must be secretly a marketing effort!
Especially if an AI has actually succeeded at pulling off a concerning thing out in the wild. Can't have that happen in real life! Nuh-uh! Must be staged!
"It's a marketing stunt" is just denial trying to look like it's being clever.
US rolled in, convinced Iran's AA to stop doing its thing, and proceeded to bomb targets with impunity. And I mean "impunity" when I say it. When you see B-52s take off with JDAMs under their wings, you know someone's AA is not picking up the phone.
If a B-52 with JDAMs is somehow survivable, an F-16 is not even a question.
The bulk of US losses in Iran in the air were: MQ-9 Reapers, with 10-20 airframes lost reportedly. That's an aging low/slow turboprop platform that's highly vulnerable even to MANPADS.
No F-16s or B-52s were lost to Iran in the air. Zero. One F-15 was shot down by Iran. One F-35 was damaged - and that specific airframe has crawled back to the nearest airbase under its own power and landed gracefully. Clearly, the jets involved were having a good time overall.
(Except over Kuwait. No airspace is more hostile than an incompetent friendly airspace, it seems.)
Now: a (manned) F-16 flight hour is 20-30k, and a single F-16 can truck 2 to 4 GBUs or JDAMs. An unmanned F-16 can shave a few thousands off that. Trucking cheap bombs with an expensive aircraft is quite a lucrative gig if you can reliably get that aircraft back. And the possibility of using those aircraft in SEAD/DEAD missions that are a bit too spicy for meatbag pilots, but can keep airspace open for other platforms? There's considerable value in that too.
Going "full missile" makes sense at extreme standoff distances, or in extremely low survivability scenarios. That, as Iran shows, isn't at all a given when US levels of SEAD/DEAD are in play. Missiles and OWA drones had a place there, but so did high end SEAD jets and jet-powered missile carriers.
MQ-9 might have underperformed, but that seems to be as much of an argument for more cheap missile/OWA hybrids as it is for cheaper strike/loiter drones, and more expensive, more survivable drones. An "F-16D" and even "F-35D" unmanned conversions might make economic sense, given the prices of MQ-9, F-16, F-35, and what we've seen happen in Iran.
Variance in numerical instability across hardware is basically a given, because all the different accelerators implement their "fast paths" in slightly different ways. Modern AI just takes it on the chin.
Anthropic, for example, already has to support both Nvidia CUDA and Google/Broadcom TPUs - AMD aside. And I expect those two to have more architectural and software differences than AMD ROCm and Nvidia CUDA accelerators would.
And yes. If you're wondering how they solved the problem of variance between those two: they didn't! Claude instances that run in the TPU land produce slightly different outputs than those in Nvidia racks! Just not different enough for, you know. Anyone to give a shit. No statistically significant effect on model performance.
Anthropic has Mythos. That thing's low level code "AI slop" is better than the "meatbag slop" most software developers write, and it can keep cracking at a given problem with persistence.
OpenAI has GPT-5.6, and also that rabid dog of an AI model that was last seen out in the wild tearing HuggingFace open.
Modern LLMs are very, very capable - not just of writing raw code, but also of persistent, methodical problem solving. Which is what you want to tackle things like "port from an exotic system A to an exotic system B and smoke test the port". Persistently hunting for testable optimizations is a good fit too.
Now, the dominant compute-hungry workload is AI, where precision takes second place to the independent parameter count. To the point that the capacity of BF16, which were originally designed as a radical optimization for AI workloads, is sometimes considered wasteful now.
AI workloads have some truly peculiar and counterintuitive properties - the kind of things you might expect to see in biology instead of conventional computing. Intrinsic error tolerance, for one. It did necessitate some rethinking and reprioritization, and I'm not quite sure if we converged to the general shape of an "optimal" AI accelerator as of yet.
Second, US already has a lot of F-16s, and not a lot of dedicated high end drones. US has more F-16s than it has trained F-16 pilots.
Things like this enable turning "manned" airframes into "optionally manned" airframes. Which enables using those F-16s for cheaper for routine "JDAM truck" runs, and also for the kinds of missions that would be too risky for a human pilot. No pilot rescue operations for a downed Nvidia GPU rack.
You could think of an AI F-16 as of a "higher end variant of Reaper", and you wouldn't be wrong exactly.
Third, the best thing? This AI approach transfers. If you can have an AI fly F-16 through reasonably complex missions, you can also adapt it to an F-35, or a dedicated high end drone platform.
By all accounts, US is much better at "Scud hunt" now than it was back during the last attempt at it.
One-way attack drones? That's a new threat, and one US is far less equipped to tackle.
But I don't think that "OWA threat" was where US strategic planning has failed. From the outside, it seems like US plan was to make a single decisive strike, and negotiate with cowed Iran from there. Iran has failed to be cowed by a single decisive strike, and there was no workable Plan B for forcing it to submit if it didn't. That's the key strategic failing. Persistent OWA threat is downstream from that.
A fleet of AI-controlled F-16s? That contributes to the latter.
In all situations where autopilot was disengaged shortly before a crash, the autopilot is implicated in the crash. Both by Tesla's own standards and by NHTSA standards.
And yet, people who really should know better repeat this as gospel.
Iran had a solid AA system - and the level of impunity with which US operated with over Iran should give every country that wants to rely on its AA a pause. US SEAD/DEAD capabilities are fucking scary.
Of all the things that went wrong with Iran: USAF's capability to get in, punch through AA and rain fires from the sky sure wasn't one.
The contact data is real, but you'd have to at least look like a decent sized company to get anywhere with it.
It's not CMOS. It doesn't even have a digital interface. It has basically nothing in common with the kind of image sensors you'd find in a flagship smartphone, like Samsung ISOCELL. It's a very specialized device made for a very short list of uses, none of which are in consumer electronics.
Most sensors you'll see there are similar. Not all of them are as hideously exotic as the one you linked, but very few of them are the SKUs you'd find in a smartphone. The closest thing to a smartphone camera sensor would probably be STMicroelectronics VD56G3, which is similar to VD56G0 that Apple uses in iPhone FaceID. And even that is a specialized structured light NIR piece.
Sorry, but no. I trust that "the Old Way is just Better in an invisible, unproven way" about as much as I trust any other "appeal to tradition".
I like Neal Stephenson's writing, but I wouldn't take this advice seriously.
Being an "AI inference middleman" is pretty lucrative, and AI companies have only started to crack down on it.
As a ready-made example: the "AI models are unprofitable, API inference is subsidized, OpenAI is naked" line he keeps stressing. That doesn't seem to be likely even on the basic napkin math estimates. And is also repeatedly refuted by both AI company officials and various AI industry insiders - of varying trustworthiness, and with different incentives.
Having "the numbers" isn't the same as having objectivity, and again: I'd trust Altman to have a semblance of objectivity before I would trust Zitron.
And no, I wouldn't actually trust Sam "King of the Cannibals" Altman. But I would still expect him to start with far, far more accurate data, and distort it a lot less to suit his narrative than Zitron would. That's how low the bar is.
The man made saying "the AI industry is failing" his entire personality. He'll keep saying that regardless of what the AI industry is doing, forever. You can write "tech companies often overpromise and underdeliver" on a post-it note, stick it to a whiteboard, and that will provide about as much perspective as everything Zitron has ever written, or will ever write.
You'd buy instead an "industrial/automation/automotive" sensor - one that has a fraction of the raw resolution, but maybe spots some other perks. Like a global shutter with no rolling shutter distortions, a "no RGGB Bayer" option that lets you set up your own filters and pick what wavelengths you care about, a package that's amenable to low volume manufacturing, longevity guarantees, availability in quantities of tens instead of tens of thousands, or actual documentation that you can get without 3-6 months of salesman and lawyer negotiations.
Or you'd buy a ready-made "scientific instrument" camera from someone else. Which probably has another "industrial" sensor - maybe worse, maybe better than one you could get directly, depending on how up to date the catalog is and how good of a working relationship does the instrument company have with the sensor vendors. And a price tag in 4-5 digits range. Then you would integrate that thing as a subsystem into whatever science hardware you wanted to make.
But if you truly want the "flagship smartphone" mix of small sensor size, low cost and high resolution? There are no good options! Clearly, the tech just hasn't advanced far enough for that!
A part of it is just how involved the manufacturing of those camera modules is. The sensor chips ships as bare silicon dies. They're precisely placed, glued and wire bonded to the substrate PCB, then adorned with a focus/OIS voice coil frame and a lens assembly. Absolutely nothing about this process is hobbyist friendly.
The other part is that corporations are incredibly stupid about how "valuable" their precious proprietary data is, and would rather jump into a volcano than give a sensor datasheet to someone who doesn't look like they have at least 20 lawyers employed and can take a MOQ of 100000 units. And how would one use a sensor without a datasheet, or a bring-up register sequence, or anything at all?
Vendor buying agreements and NDAs often prevent hobbyist-grade sensor boards from even existing. Especially for cutting edge high performance sensors, like the ones found in flagship smartphones.
Modern LLMs kick ass, but they aren't magic.
Fully vibe coding a driver for things more complex than, perhaps, a well documented CMOS sensor (that's a small part of what's covered in the article) is still a no-no.
But an LLM does wonders at emitting boilerplate, "vibe checking" your implementations, suggesting how to implement certain things, helping debug some of the things, etc. You can get the LLM to do a lot, but you still need a lot of understanding, and a lot of applied handholding.
People who never worked with corporate software written by underqualified, underpaid and overworked developers often have some incredibly inflated code quality expectations. An average open source project has code that's ten times as neat and a hundred times as battle tested as what's common in tooling inside corporate perimeters.
As a rule of thumb for this kind of corporate code: assume the software was written by a drunk developer at 3am, and you wouldn't be too far off.
All the more reason to mock the braindead "it's all marketing". There's no magic in a year 2026 agentic AI being able to traverse poorly secured corporate networks.
FM radio support is a part of a Bluetooth chip, an audio codec chip, or even a standalone IC.
In modern high end SoCs, even the 5G modem is largely evicted from the SoC itself. Only the digital processing happens in the SoC - while ADC/DAC and anything analog lives in its own IC.
By now, I'm pretty confident that some people would keep screeching "it's just a marketing stunt, AI capabilities and AI risks aren't real, they're just doing this to prop up their stocks" even if they find a Cyberdyne Systems T-800 armed with a shotgun breaking down their front door.
"It's a marketing stunt" is just denial trying to look like it's being clever.