There is circumstantial evidence, and it’s worth noting that seemingly Lutnick (and bizarrely even once Nvidia itself) managed to kill investigations into it. When people are taking pains to look the other way it does imply rule breaking, but yes it’s still only an implication.
Red herring. The FCC doesn't have jurisdiction over the relevant export restrictions.
The Commerce Department (Secretary Lutnick) is responsible for enforcing the chip export ban / US collecting a portion of the profits according to Congressional statute.
And for the moment I haven't seen alleged violations. Exporting to DC in Asia outside of China is a legal loophole that the Trump administration preserved.
> The S11 chip uses a lightweight AI model to determine whether speech is
occurring nearby. This model doesn’t transcribe or record anything. It simply
detects whether a conversation has started. If a conversation is detected, audio
flows into a protected buffer inside the Secure Exclave on the S11 chip.
It does record audio. A microphone is constantly receiving sound waves and writing them in a digital format in a particular audio file encoding as inputs to a model. That is a recording. I don’t see how this passes legal and doesn’t expose them to class action for false advertising.
What you're describing is listening, not recording. If the bits only exist transiently on the wire and are given only as streamed input to a time series binary classifier which emits a simple true/false signal and are not written to more permanent storage, then it's hard to see how anyone could call that a recording.
A human listens. A digital system cannot listen because it must always write the analog signals into a digital representation: a recording. When you call `model(last 10 min of audio)` that is a recording. I suppose a system could "listen" if it somehow took as input the actual electrical voltages from the microphone receiver.
This might sound pedantic but it is very nefarious to me that Apple is marketing a device as "never recording" when it is actually always recording. And this is a claim going back to Siri with wake words. Again, they claim it doesn’t record and I don’t see the logic there.
> I suppose a system could "listen" if it somehow took as input the actual electrical voltages from the microphone receiver
Ah, it feels like this may be the crux of the issue, because that's pretty much what uncompressed audio is. Every value in a WAV file roughly represents the displacement of the microphone diaphragm at time t.
> A human listens. A digital system cannot listen because it must always write
There are two issues with what you're saying, depending on what you mean by that:
1) Transmission isn't storage. When you talk to me you it would be unreasonable to say that your voice is being "recorded" in the air or in my eardrum or even necessarily in my brain, even though vibrations are in some real sense physically present across the air molecules and my eardrum is being physically displaced and my brain is doing...well...some unknown ephemeral aggregation in order to produce thoughts in my head from what you're saying. So if you transmit bits over a wire, and even if you have enough wires to represent bits from many time quanta, you should likewise not call those bits recorded by the same principle of innate ephemerality.
Bits on wire in audio stream at time t ~= position of your eardrum at time t. Neither are reasonably considered storage without persistence, which we have no knowledge of, and neither are reasonably considered on their own to be representations of what is being said.
2) You don't need to accumulate data to analyze it. You can, but you don't need to, and I don't think we know that they are.
You can think of a hash function. A hash function need not record its input stream. It need only record the accumulated influence of the input on the prediction. Storing a running statistical guess about whether you're talking and recording what you're saying are not the same thing.
Fascinating, how does (did) one become a DVD programmer? I wondered as a child. What’s the tech stack like there? And were there DVD programming studios/companies, or was it more of a freelance model, or employed by the big studios in-house?
I fell into it. The company started as an MPEG-1 encoding shop making interactive CDs with Macromedia Director and Video CDs. That then naturally evolved into MPEG-2 encoding and DVDs. There were a couple of DVD programming apps, but the standard was Scenarist. I started using it on the first version released on Windows NT. We used the original 3.95GB DVD-R drive that cost $15k as test discs which made it interesting when building a full 4.7GB disc. We'd write the images to LTO4 tapes when sending to replication. This was pre-SSD, so we had interesting drive configuration. Writing disc images is full of I/O. We had a raid-5 volume where the original assets lived. When muxing, it would read from the raid-5 and write to a raid-0 volume. It would then read from the raid-0 mux data and write to another raid-0 image volume. There were plenty of places programming shiny round discs that the studios would hire out. Some were the bigger post houses, but others were boutique like the one I worked. Getting studio work was a pain as you had to get certified to hold unreleased content with a security audit. This was all Web1.0 days, so not much learning resources other than an email mailing list that wasn't so much tech support rather people talking about DVD "tech" with occasional interesting questions raised here and there. Scenarist was a drag-n-drop node based UI. It was pretty useful on learning what could connect where because it would change the cursor accordingly. The best learning resource was the DVD bible from Jim Taylor "DVD Demystified"[0]. After that, it was just watching as many DVDs as possible to see what other people were doing and taking inspiration from there. Lots of cool stuff done with the subtitle track like the robot animation over the video for MST3K.
Wow, fascinating stuff. A whole lot more than I realized but all of it makes sense. Really takes me back to the disc days. MPEG-2: haha! Thanks so much for sharing your story
Rodin has been dead for a long time, and the museum is obviously the closest thing to his representative. He specifically donated the buildings and his work to the French state on the condition that they make it a museum of his work.
Lynch also wasn’t talking about only his movies, he was talking about cinema as a whole.
Good for Japan. They are being forward thinking and protective over their existing job supply. America is a great example of what happens with too much offshoring and lax immigration policies. Especially with the effects of AI, there are no longer enough jobs to go around.
Really embarrassing to see Carmack fall so low and have such little respect for the art of programming he once advanced. Unlike swords and martial arts in combat, code is still needed. It’s not like there’s some new medium that doesn’t use code at all. It’s just a question of who is writing that code: a human or a statistical model. Guns superseded swords, cars superseded horses, but coding agents haven’t superseded code itself. A better analogy would be if, instead of guns, we ended up using advanced kung-fu fighting robots. Which would actually be sick…
> I like restrictions of another sort more. Namely, restricting the use of AI in corporate environments so that it does not destroy the economy as it advances.
Careful, you’re making too much sense. The big model providers would much rather the ball be in their court. “Deceleration” meaning Anthropic and OpenAI offer less capabilities for more money in service of “protecting humanity from extinction” rather than penny pinching. A lot of “AI research” coming out of these “labs” (increasingly enterprise) seems to be more driven by a cost benefit analysis from suits rather than genuine contributions to the field. I think the last true innovation was the idea of using doom and ending the human race as a marketing gimmick, which strangely seems to have worked in setting the narrative and captivating the sci-fi imaginations of journalists and techies alike.
It is virtual dom like.
We don't have access to a stable underlying element like we would with plain UIKit.
You can build declarative models without this virtualdomness. But SwiftUI was created during the react boom so they went with the Zeitgeist, understandably.
Now this is somewhat problematic if someone wants to implement better fine grained reactive systems. And I posit that the next paradigm is going to be in that direction given what I've been working on (furthering current reactive systems which only go halfway).
This is such a clueless take, and I see it surprisingly often. Declarative UI does not inherently mean virtual DOM nor does it "automatically" mean slower than UIKit.
The abstraction is huge advantage, especially when targeting different devices and screen sizes. We (and that includes Apple) have been learning the value of those kinds of abstraction boundaries for years, well before React.
It's also important to note that SwiftUI and UIKit are composable, so if some small part of your app needs low-level, finer grained control, then go ahead and use UIKit/Core Animation/Metal and SwiftUI for the rest.
Yes, declarative UI is just an abstraction of imperative code. :)
The rest is properties of the internal implementation.
The VDOM way obfuscates the internals more. Doesn't let you control rendering appropriately.
It is so because the diffing and reconciliation algorithm have to be able to remove any subtree of views/nodes from the UI tree.
Then you can only have islands of either paradigm within each other. UIKit and SwiftUI do indeed compose, albeit coarsely.
I can't blame them, they got influenced by react.
Don't even blame react, it was a good attempt. Basically trying to build a UI from a snapshot of a tree. Except this is too simplistic a model.
They designed it as if you could equate the number of games and the number of positions in chess. Like a markov chain.
Except playing chess has side effects. A mere snapshot does not encode those.
So what took an autonomous agentic system using a significantly more powerful internal model, totaling multi-millions of dollars of compute in training and inference, was likely to already be solved by a team of a few humans with an orders of magnitude smaller LLM budget, had OpenAI not been foaming at the mouth to jump the shark and claim “AI solves Millenium Problem.”
Also it sounds like the human research effort spanned weeks if not years from Tristan’s statement so it is extremely likely the work and prompts of these human researchers was used in the OpenAI knock-off.
reply