> driving a car, industrial control, vibe coding are bottlenecked by hallucinations

No they're not, this isn't 2022 anymore. Let me reiterate, they're fairly trivial to spot and largely accounted for in serious autonomous systems. The issue is monkey paw property of the LLM - they're maliciously compliant by seeking shortcuts too greedily while skipping over subtler ontologies that aren't verbalized in corpus enough so as to differentiate into a concept ("carwash is typically somewhere you might want to drive to" wasn't solidified -> hilarious stupidity ensues). This sort of mistake isn't hallucination, but just plain stupidity of someone who's never quite understood what carwash is for, but only heard em mentioned few times in a book.

Yann LeCun does point at this, though I think he's too optimistic that DNNs could fill those gaps by having online learning experience in a closed loop. The way GPTs work is embarassing parallel SGD that you simply can't do when learning one speck of experience at a time. Would need to invent entire new optimization algorithm that scales just as SGD if not better, yet works sequentially.

I mean dumb humans are confidently wrong (hallucinating) all the time and especially humans in managerial positions he's positing this should replace.
Although LLMs hallucinations are driven by biases alien to humans.

This seems like a quick cash grab aimed at AI antis of Ed Zitron ilk, who famously insist that "AI is real only if it is never wrong and never hallucinates", but they never provide any insight why the AI needs to be always 100% correct.

Meanwhile, the real problem is that LLMs simply descend into madness and go off the rails it's not hallucination per se, but simply wandering off meaningful goals. Mere hallucinated errors are comparably simple problem and possible to correct for with checks and balances on authority, compared to more fundamental inability of LLM "intellect" not turning into a mad king (whether always correct or not, being mad is trouble).

The operational costs are neglible, and run at most in hundreds of millions. Google made ~$10B on Play Store in 2025 (20% of estimated $50B revenue). 99% of what you pay for is monopoly pricing, fundamental cost is two magnitudes less than that. Moreover, the developer "support" is abysmal, as Google has near zero motivation to deal with small time developers in monopoly market.

> Cell phones are not required to offer apps at all. Don't like smart phone terms? Don't use a smart phone.

That's a relevant argument for as long Google doesn't exploit network effects to lock you in. Once it does, it's more akin to cable company charging you 100x more above board for the privilege of using their wires. You're free to not use the service duopoly of Comcast or terrible ISP #2, aren't you?

Same thing with low-resource languages. The more problem is the more obscure domain/language, the less actual training corpus, the more nonsense output from the robot. Eg ask it to write reasonably complex win32 and posix command tool for same problem, with the only differences there being the semantic differences between the systems. It will overall hallucinate more nonsense for win32.

The robot is very good at what it knows - especially at the center of distribution (not too much extrapolation). And it knows what's common. The more you touch edge of the domain, the accuracy nosedives.

> More importantly, I wonder if the positioning and orientation of a person between the Wi-Fi antenna and the sensor matter.

One way to conceptualize it is consider the wifi to be ordinary light, and person's profile their silhouette they cast as a shadow. You can tell someone is fat or not just by diffused reflections in the next room. Resolution of this can be substantially improved by number of observers (all in different rooms), but from various angle (= light bulbs, and measuring SNR they "see") with least squares matching (think something akin to triangulation, but pixel-voxel volumetric solving). The model most likely learns it internally, but one doesn't necessarily need to use brute force by stacking moar layers on this.

These commands are executed from HCI pipe, not radio end. There are no exploit scenarios, as the HCI host has full control in any case. It's about as much of a "backdoor" like a wifi radio with hacked (or similarly permissive) firmware to inject raw frames. If anything it makes the chip moderately more useful now.

Reminder Qualcomm was aggressively pushing for RISC-V extension since 2022 that would allow for a trivial frontend swap on Nuvia. If this went through last year, we'd be quite possibly seeing "RISC-V Elite", not Oryon right now.

For better or worse SiFive isn't going along with it. First its a bad look to let QC dominate the consortium on the market, plus they're probably right on that ARM (uarch-wise) compatibility approach would just add bloat just to save Qualcomm a boatload of money so that they don't have to do expensive redesign of Nuvia's wide issue schedulers.

Some earlier context:

https://www.forbes.com/sites/r...
https://old.chipsandcheese.com...

It's his panicked reaction to Zuck firing the shots about Meta's commitment to dominance through open-source with Facebook entrenched as BDFL.

ICANN are central-control-by-committee bureaucrats of his ilk who'd be quite keen to hop on this (and turn it into a corrupt shitshow, just like domain registries) if given a chance. He has an eye to recognize and groom potential allies well.

The whole business model of clickbait is the audience sharing it all like "look how bad this is" and feeling all smug. A lot of stock photo uses the same form of bait. AI seems perfectly suited for this sort of thing.

To block malware domains (Google SafeBrowsing, Firefox Phishing Protection). The reasoning of bureaucrats is along the lines "well, the feature is already there, might as well use it for whatever we consider malicious".

The fundamental market in the foreseeable future is dark markets and trade with sanctioned countries. This is already starting to show - certain untraceable coin popular in this sector outperforming by 20% everything else during the 2022 crash.

Market cap of shadow economies could only ever account for a very tiny fraction of world trade though.

Why blame the machine, and not the executive? Same old, guns don't kill people, people kill people.

I didn't RTFA either, but I think it is about romance scammers, not really the usual flaky Tinder matches.