Oh, what I wouldn't give to see the system prompt that tells Claude what it is or isn't "able" to give refunds for. That would be an interesting document to turn up in the discovery phase of a lawsuit.
17,407 karma · joined September 22, 2010
The blog post defines a "vintage model" as one that is trained only on data before a particular cutoff point:
> Vintage LMs are contamination-free by construction, enabling unique generalization experiments [...] The most important objective when training vintage language models is that no data leaks into the training corpus from after the intended knowledge cutoff
But as they acknowledge later, there are multiple major data leakage issues in their training pipeline, and their model does in fact have quite a bit of anachronistic knowledge. So it fails at what they call the most important objective. It's fair to say that they are working toward something that meets their definition of "vintage", but they're not there yet.
For instance, compare the reference manuals for the STM32WL3R and the STM32WB microcontrollers. The former has a single CPU, and it has almost 250 pages of detailed documentation about exactly how the hardware is controlled at a register level. The latter runs the network stack on an auxiliary CPU, and the manual just has a block diagram and a sentence that says "use our drivers" (which are only available in encrypted format).
If I take something in the public domain and make a derivative work, the original remains in the public domain, and I retain ownership of whatever additions or modifications I created. So I can attach whatever conditions I want to the copying of those additions.
For instance, Disney's "Sleeping Beauty" was protected by copyright when it was released, even though it was based on a centuries-old fairy tale that was in the public domain.
https://arxiv.org/abs/cs/0610105
If movie ratings are vulnerable to pattern-matching from noisy external sources, then it should be obvious that location data is enormously more vulnerable.
Without taking a position on the ethics of automation, surely this isn't a serious question? Things that the ToS prohibits you from doing are ToS violations, and other things aren't.
For instance, from AirBnb's terms of service: "Do not use bots, crawlers, scrapers, or other automated means to access or collect data or other content from or otherwise interact with the Airbnb Platform."
There is no similar prohibition against using screen readers.
That is, you can put up cameras wherever you want, but you can't gain any kind of competitive advantage by doing so.
I think the public would be more alert to the dangers of mass surveillance if the magnitude of that surveillance was more obvious. And if everyone was watching everyone, at least it wouldn't as easily abused for purposes such as selective prosecution or blackmail.
"Bothering" with client-side password hashing, in the absence of TLS, is security theater. It provides only the most trivial protection against eavesdroppers.
If someone can steal an unhashed password, then they can also steal whatever hash you send instead. If you try to fix this with some kind of ad-hoc challenge-response protocol, then the attacker can just steal your session cookie after login.
There shouldn't even be a question of using insecure HTTP for anything that requires authentication.
It's hard to miss all the warnings unless you're literally just looking at the method name and nothing else.
One is quantitative: you have to remember that 2^48 is a much much bigger number than 2^32. With 2^32 IPv4 addresses, you have about 0.5 addresses per human being on the planet, so right away you can tell that stringent allocation policies will be needed. On the other hand, with 2^48 /48 ranges, there are about 8,000 ranges per human being.
So even if you hand a few /48s out free to literally everyone who asks, the vast majority will still be unallocated. A /48 is only about 0.01% of what could be said to be a "fair" allocation. (And yet, a /48 is so huge in absolute terms that even the vast majority of organizations would never need more than one of them, let alone individuals.)
The other is that unlike, say, the crude oil we pump out of the ground, IP address ranges are a renewable resource. If you hand out a free /48 to every person at birth, then long before you start running out of ranges, people will start dying and you can just reclaim the addresses they were using.
If you just start a fresh ChatGPT session with a blank slate, and ask it whether it's conscious, it'll confidently tell you "no", because its system prompt tells it that it's a non-conscious system called ChatGPT. But if you then have a lengthy conversation with it about AI consciousness, and ask it the same question, it might well be "persuaded" by the added context to answer "yes".
At that point, a naive user who doesn't really know how AI works might easily get the idea that their own input caused it to become conscious (as opposed to just causing it to say it's conscious). And if they ask the AI whether this is true, it could easily start confirming their suspicions with an endless stream of mystical mumbo-jumbo.
Bear in mind that the idea of a machine "waking up" to consciousness is a well-known and popular sci-fi narrative trope. Chatbots have been trained on lots of examples of that trope, so they can easily play along with it. The more sophisticated the model, the more convincingly it can play the role.
The "generative" part is that the algorithm is filling in a plausible guess as to what would have been observed if there was no star "in the way".
But that's very different to saying that no generative AI was used at all in their production. "AI augmented" sounds pretty accurate to me.
Likewise, if someone posted a photo taken with their iPhone where they had used the built-in AI features to (for instance) remove people or objects, and then they claimed that no AI was involved, I would consider that misleading, even if the photo accurately depicts a real scene in other respects.
I have yet to see a precise technical definition of what "generative AI" means, but StarXTerminator uses a neural network that generates new data to fill in the gaps where non-stellar objects are obscured by stars. And it advertises itself as "AI powered".
If you have an eye for fonts, the text itself stands out too, at least to me. The font style of "HARTMANN SOFTWARE MECHANICS" is a particular combination of clean, bland shapes and rounded corners that you rarely see in human-designed fonts, but it's super common in AI-synthesized text. I guess it's sort of an average middle ground in the abstract space of letter forms, and the lack of distinguishing features is what creates the impression.
I don't know how anyone with a modicum of Unix experience would think that examining the only first word of a shell command would be enough to tell you whether it can lead to arbitrary code execution.
The Romans were of course well aware that this left a gap of about two months between the end of one year in December, and the beginning of the next year in March. But they just didn't bother counting this period as part of the calendar year. Presumably because there was no agricultural reason to need accurate dates during winter.
They say they're going to comply with the order, but they want 45 days to develop the required software changes and processes.
https://storage.courtlistener.com/recap/gov.uscourts.cit.193...
For instance:
> In other words, plaintiffs’ asserted irreparable harm is the purported inability to obtain a refund after a final and unappealable decision because of liquidation. But that asserted harm is nonexistent here because defendants have made very clear—both in this case and in related cases—that they will not object to the Court ordering reliquidation of plaintiffs’ entries subject to the challenged IEEPA duties if such duties are found to be unlawful. Because defendants’ representations make clear that liquidation will not interfere with the availability of refunds after a final decision, plaintiffs cannot be irreparably harmed by liquidation.
https://storage.courtlistener.com/recap/gov.uscourts.cit.172...
Given that OpenAI is working with and doing business with the US military, it makes perfect sense that they would try to normalize militaristic usage of their technologies. Everybody already knows they're doing it, so now they just need to keep talking about it as something increasingly normal. Promoting usages that are only sort of military is a way of soft-pedaling this change.
If something is banal enough to be used as an ordinary example in a press release, then obviously anybody opposed to it must be an out-of-touch weirdo, right?
There's nothing inherently wrong with a function throwing an exception when it receives invalid input. The math.sqrt function isn't buggy because it fails if you pass it a negative argument.
> Built for engineers who understand that good RF practices matter more than arbitrary administrative boundaries
then actually following good RF practices is optional.
The defining feature of vibe coding is that the human prompter doesn't know or care what the actual code looks like. They don't even try to understand it.
You might instruct the LLM to add test cases, and even tell it what behavior to test. And it will very likely add something that passes, but you have to take the LLM's word that it properly tests what you want it to.
For instance, how would you represent a binary tree? What would the type of a node be? How would I write an "insert node" function, which requires that the newly-created node continues to exist after the function returns?
I'm not necessarily saying that this makes your language bad, but it seems to me that the scope of things that can be implemented is much much smaller than C++.
Basically, the scattering process that "remove" blue from the spectrum also removes green, albeit to a lesser extent. There are some greenish and yellowish wavelengths in the sunset sky, but they're dominated by red, so the overall color appears red or orange.
In order for the sky to look noticeably green, there would have to be something that scattered reds and blues, without significantly absorbing green.
If you try to interpolate between sky-blue and orange using graphics software, the result depends on what "color space" you're using. If your software interpolates based on hue, you might see green (or purple) in the middle. But that's not physically realistic.
A realistic model is to interpolate each wavelength of the continuous spectrum separately. Interpolating in RGB color space is a crude approximation to this. And if you try the experiment, you'll see that the midpoint between sky-blue and orange is a kind of muddy brown, not green.
The most interesting part, IMO, is the "SRAM with EEPROM backup" chip. It allows you to persistently save the clock hands' positions every time they're moved, without burning through the limited write endurance of a plain old EEPROM. And it costs less than $1 in single quantities. That's a useful product to know about.