cjbprime

It's a very conservative warning. The application does not perform writes, so the application doesn't actually wear your SSD at all. The rest is just application-independent general hygiene.

You still have a core misunderstanding. Only one layer of weights is required in memory at a time. A forward pass can be over-simplified as a matrix multiplication of each layer, one at a time.

There is no swapping of working RAM. We're just talking about loading the weights read-only data into RAM on-demand for each layer. It is only as slow as your storage interface.

Both sentences are likely wrong. It's a written lifespan (technically an "erased" one), not a read lifespan. The weights are only needed from disk read-only. And Mac NVMe interfaces are surprisingly fast.

Edit: Oh, I think you maybe thought I meant swapping working RAM off disk? I didn't. I meant swapping weights off disk into RAM on-demand.

I've got access to a 192GB RAM Mac Studio, which is below the stated minimum RAM. Can swapping off fast disk be used to make it work out, especially since it's MoE?

ChatGPT reconstructing idiomatic Python source code from Python bytecode was definitely up there. That is not something humans have written a great deal about online. It requires simulating the Python VM.

I remember also having a massive wtf reaction to realizing that original ChatGPT was pretty good at decoding long random/unique base64 strings.

If it detected any harmonics it would be too high.

I think it's not that simple. A tuner is "hearing" the fundamental and all of the harmonic overtones combined. It has to guess at which frequency is the fundamental, even if the overtones are actually stronger than it amplitude-wise, and it does that by looking at the nature of the repeating overtone pattern and extrapolating back to the fundamental.

I think you can end up an octave too low (half the actual frequency) if the waveform repeats in a way that implies a different overtone repetition pattern, for example if there's an every-other-cycle artifact to the waveform.

Inference (not training) is bottlenecked by memory access speed, not compute. Having special hardware wouldn't make it faster unless you somehow found a faster memory controller than the GPU has.

I don't agree with the "twice". A frequently performed manipulation like the fuel cutoff (usually performed after landing) collapses down to a single intention that is carried out by muscle memory, not two consciously selected actions.

The prelim report states these pilots were indeed breathalyzed before takeoff.

I meant philosophical toggle switches, not physical ones. The gear can go between down and up. The fuel can go between run and cutoff. Given enough practice, the brain takes care of the physical actions that manipulate those philosophical toggles without conscious thought about performing them.

Now I'm trying to remember if I've ever picked up my razor and accidentally begun tooth brushing motions with it. Probably!

More relevantly, you seem to me to be unduly confident about what this pilot's associative triggers might and might not be.

Sometimes? If you have enough altitude to trade for speed then after the cutoff you could glide to a hypothetical miraculously-placed runway right in front of you, vs. having fire quickly consume the entire plane if you don't cutoff..

It's not that rare, and there are institutional factors (such as seeking treatment for psychosis being career-ending for a pilot) that incentivize serious pilot mental health crises being untreated.

Not really, though. They're both (retracting the gear, and cutting off fuel) just toggle switches, as far as your brain's conscious mechanisms go. Doing them both on every flight dulls the part of your brain that cares about how they feel different to perform.

(I'm not strongly arguing against the murder scenario, just against the idea that it's impossible for it to be the confusion scenario.)

There is no possible way to confuse these two actions.

This is obviously an overstatement. Any two regularly performed actions can be confused. Sometimes (when tired or distracted) I've walked into my bathroom intending to shave, but mistakenly brushed my teeth and left. My toothbrush and razor are not similar in function or placement.

I wonder if a built in delay would make sense for safety.

(Presumably delaying the amount of time before a raging engine fire stops receiving fuel would also have an impact on safety?)

It's not that nonsensical. After it's accessed the private repo, it leaks its content back to the attacker via the public repo.

But it's really just (more) indirect prompt injection, again. It affects every similar use of LLMs.

I don't know that ChatGPT's voice mode is using audio as a transformer input directly.

It could just be using speech to text (e.g. Whisper) on your input, and then using its text model on the text of your words. Or has OpenAI said that they aren't doing this?

Then the malware would provide that confirmation to the wallet too. Defending yourself from malware running on the same (Windows) machine is mostly impossible.

You should try AI sometime. It's quite good, and can do things (like "analyze these 10000 functions and summarize what you found out about how this binary works, including adding comments everywhere) that individual humans do not scale to.

Ghidra's a decompiler and WinDBG is a debugger, so they'd be complementary.

Semiconductor fabs are actually decently large (>20k employee) cities.

It's a British thing -- another example is that Brits would use e.g. "Apple have decided to.." instead of "Apple has decided to.." (for every group of people including "the government", not just businesses).

I like this language quirk a lot. It almost feels subversive, pointing out through grammar that group entities are just people, responsible for their choices like everyone else.

I agree with you, which is why I'm redirecting the blame to the CVSS standard, which does not agree with you.

That's a CVSS issue. Heartbleed only affected Confidentiality, and CVSS rates scores on a triad of Confidentiality, Integrity, and Availability. RCE affects all three.

Looks like it was possible to include the `x-middleware-subrequest` header in your request, tricking the state machine into thinking you'd passed auth already.

(Don't use the user input itself to encode state!)

Could be measurement noise..

The fact that those certificates had been issued without a secure method of verification is the thing that was wrong with those certificates.

Dyndns is not the only use case for having someone control a subdomain without having the full trust of the root domain.

the incorrect belief that there's some emergency button you can press to demand that a court consider your arguments ASAP

... isn't that exactly the legal button the complainant pushed to prevent the revocation?

(I think Sectigo's argument is that Digicert did not even attempt to convince the court that it should be allowed to revoke those certificates in the mandated timeframe. If they had attempted and failed, I don't think they would be receiving criticism.)