jsolson

Generally I'm a huge Stephenson fan (I own all of his major works in first editions), but I hated Seveneves. The third part is the only bit I can claim to have really enjoyed.

That said, I loved Anathem, so perhaps something of your theory tracks :)

Funny story, if you're using Rook Ceph and say "I'm going to try to update these parameters to see if it will let me (and trigger a re-encoding)" it absolutely will let you change them, but it does not trigger anything to re-encode.

It just uses --force and leaves you with a corrupt filesystem.

I suppose that's only really funny in a "you had to be there, and not be me" sense.

You're sorta describing TPUs, especially early ones: https://cloud.google.com/blog/products/ai-machine-learning/a...

That first generation also included the ability to apply a handful of fixed activation functions, but really that's about it. The array is bigger than 32x32, also.

GCE fares a little better than this:

- VMs with SSDs can (in general -- there are exceptions for things like GPUs and exceptionally large instances) live migrate with contents preserved.

- GCE supports a timeboxed "restart in place" feature where the VM stays in limbo ("REPAIRING") for some amount of time waiting for the host to return to service: https://cloud.google.com/compute/docs/instances/host-mainten.... This mostly only applies to transient failures like power-loss beyond battery/generator sustaining thresholds, software crashes, etc.

- There is a related feature, also controlled by the `--discard-local-ssd=` flag, which allows preservation of local SSD data on a customer initiated VM stop.

I should've aimed for more clarity in my original comment -- the first link is to locally attached storage. The second is network attached storage (what the GP was likely referring to, but not what is described in the article).

Persistent Disk is not backed by single devices (even for a single NVMe attachment), but by multiple redundant copies spread across power and network failure domains. Those volumes will survive the failure of the VM to which they are attached as well as the failure of any individual volume or host.

In most cases, they're physically plugged into a PCIe CEM slot in the host.

There is no network in the way, you are either misinformed or thinking of a different product.

Same for Google (the GP is incorrect about how GCE Local SSD works).

This is also true for GCE Local SSD: https://cloud.google.com/local-ssd

The GP is incorrect.

This is untrue of Local SSD (https://cloud.google.com/local-ssd) in Google Cloud. Local SSDs are PCIe peripherals, not network attached.

There are also multiple Persistent Disk (https://cloud.google.com/persistent-disk) offerings that are backed by SSDs over the network.

(I'm an engineer on GCE. I work directly on the physical hardware that backs our virtualization platform.)

The memory doesn't have ECC, the controller does.

You lose data space for syndrome bits. This is not, in practice, different from fixing data space and adding chips for syndrome bits, but it reflects differently in specs.

I mean, they weren't exactly shy about it: https://go.dev/blog/generate

Yeah, zero favorites, joined in Feb 2007.

The second one has a link labeled "free up or get more storage."

I suspect it links here: https://support.google.com/a/answer/12005619?sjid=1774262812...

There is a difference between "my lawyers advised me that it was probably ok" and "Microsoft's legal team spent 100,000 billable hours pouring over case law to demonstrate that it was not, in fact, ok."

> I don't think OpenAI board has any responsibility to care for Microsoft's stock price.

They control an entity that accepted $10B from Microsoft. Someone signed that term sheet.

I agree that the act of voting itself is too squishy/personal, but the things that led up to it and their handling afterwards?

My curiosity stems from whether the board was involved in signing the contract for Microsoft's investment in the for-profit entity, and where the state might set the bar for fraud or similar crimes. How was the vote organized? Did any of them put anything in writing suggesting they did not intend to honor all of the terms of the agreement? Did the manner in which they conducted this business rise the level of being criminally negligent in their fiduciary duty?

I feel like there are a lot of exciting possibilities for criminality here that have little to do with the vote itself.

... and also +1 to your whole last paragraph.

They are a small board, and Microsoft has a very large number of lawyers.

I do not believe it is possible for them to have thought this through. I believe they'll have read the governing documents, and even had some good lawyers read them, but no governance structure is totally unambiguous.

Something I'm immensely curious about is whether they even considered that their opposition might look for ways to make them _criminally_ liable.

I worked for Amazon for a bit over two years, followed by Google for a bit over the last decade.

It never would have occurred to me to visit an Amazon office, so I didn't know about the ticket thing. I rarely worked with anyone remote.

When I joined, Google encouraged travel (less so lately). In the time I've been here I've visited over a dozen offices around the world. Some of them on vacation, because I knew there would be something unique about the local office, but also because with the exception of one office (Copenhagen), I'd worked closely enough with someone there to drop by their desk, say hello, shake hands, and get the best restaurant and bar recommendations I've ever found (and some great unexpected dinners with colleagues!)

A perspective from an American:

- Yes, typically we can refinance whenever we like, _but_ it extends the mortgage for a 30 year term, along with additional direct immediate costs (plus human inertia). Unless interest rates were alarmingly high for your last go-round (ehem), you're directly incentivized and indirectly likely to not do so.

- I own properties in Canada (yay Commonwealth!). The notion of a 30-year fixed does not exist. One can get a 25-year amortization, but typically only with a 5-10 year guarantee for a fixed rate.

- As an American, Canadians are insane for buying into this system. Our system is so much more favorable to anyone with good enough credit to be approved for a loan it's literal comedy. Also our standards for approving someone for a loan seem to be lower (that said, I had no credit history in Canada when I started this adventure, so perhaps residents get a better deal).

- As a property investor, I'm happy to control for the cash I sink into my investments in interest versus the returns I get from rental revenue. Combining that with exchange rates and US interest rates versus Canadian, I <3 Canada.

- Fully variable interest mortgages are for suckers (and in that regard, I do have some regrets).

(bias: I <3 Canada regardless -- I'd live in Whistler, BC if circumstances allowed)

I have one of those, although I bought it for the Zynq FPGA SoC.

I tried to find a use for the Epiphany cores, but the speeds and feeds were too unbalanced for anything I was looking at at the time.

The way that particular story played out suggests that I wasn't alone.

TAP is a protocol, though, not a framework?

Yes, but sometimes that's what's called for. I think I may have found my answer to where LLMs might be useful for short communications: softening something that might be interpreted as "curt" or even "rude" while not really changing the message.

I find they're useful as a tool for achieving a particular structure or tone that isn't "my voice," including populating boilerplate from bullet points. Sometimes I'll go back and revisit everything they've produced, but they let me put something tolerable in place early. This lets me focus on the meaty parts of the text sooner than I'd otherwise find myself capable of.

I personally don't find them useful for quick / short / informal communication like email, or at least not yet.

Note that GKE Sandbox allows GKE users to sandbox the workloads running on GKE nodes. The GKE nodes themselves are still GCE VMs.

Following more or less directly from this:

- If you walk backwards from where the non-deterministic Turing machine halted, the list of taken state machine transitions is polynomial in length (naturally, as the machine stopped in polynomial time).

- Walking that polynomial length list of actually-taken edges forward through the Turing machine constitutes a polynomial time verification of the solution.

This is the essence of the equivalence between being able to solve problems in NP in polynomial time on a (hypothetical) non-deterministic Turing machine and being able to deterministically verify a solution to those problems in polynomial time.

Some of this is not current.

Constant-rate TSCs are ~15-20 years old. Synchronized TSCs are at least ten.

Also RDTSC produces a 64-bit result, so it does not overflow at a rate of several Hz.

They break things apart and preserve interface protocols, but leaving the die isn't free. AMD's approach has all of the upsides and downsides that come with microservices, and taking effective advantage of it requires a similar level of planning.

That makes sense, as you're fucked by any false positive.

If it's a graph of potential stacks, though, wouldn't you eventually find the one that unwinds, if one exists?

Instead, Google created Google Offers. It then shuttered it a year or so later, refunding every transaction that had gone through it.

I was working for Amazon's bet (AmazonLocal -- https://local.amazon.com/) at the time. It lasted a bit longer, but was also eventually shut down.

Mount Evans is NIC silicon -- great chip, but it doesn't have the lane count / topology to be an effective switch chip as far as I'm aware.

With clients in the cloud (CitC) the natural/default is to have a complete view of Piper. No more narrowed clients/explicitly tracked subtrees.