bschne

(how) do you (try to) correct for this with interviews?

Here’s Time — https://time.com/6898967/ai-extinction-national-security-ris...

I agree that’s a bit of a weird move, especially considering I got it in an email within minutes of requesting it (and I’m in no way prominent, influential, government-associated, or in any other category that seems like it might matter if they want to limit dissemination)

Twitter thread by the main author -- https://twitter.com/harris_edouard/status/176718361480072012...

not a trick question, I'm just genuinely ignorant about the topic

> You can handle conditionals much better

Because you can have a core set up for each branch and just pass to that vs. “context-switching” your core to execute the branch that ends up being taken?

What makes the resulting topology great for matrix math, vs. non-matrix math workloads? Naively if you know you‘re „only“ going to multiply matrices, what do you need the flexibility and fast context-switching for? Is the end-game here that you can lay out the workload s.t. you have a series of closely colocated cores carrying out the operations of some linalg expression one after the other and the memory for intermediate results right in between, or something like that?

> I like the "Network On Chip" reversed toruses

What about them do you like as a design decision? (genuinely curious, as again, I don't understand it)

but isn't that aspect common among this, CPUs, GPUs, ...? And it feels like the whole NOC thing would add quite some overhead to moving things around.

Are you saying proximity here more than offsets this vs. e.g. each core having its own cache as I think they do in a "normal" CPU? And if so, is this more true of ML inference workloads than other workloads, for some reason?

Been following this for a while b/c Jim Keller, but every time I look at the arch [1; as linked by other commenter] as someone who doesn't know the first thing about CPU/ASIC design it just looks sort of... "wacky"? Does anyone who understands this have a good explainer for the rationale behind having a grid of cores with memory and IFs interspersed between and then something akin to a network interconnecting them with that topology? What is it about the target workloads here that makes this a good approach?

1. https://docs.tenstorrent.com/tenstorrent/v/tt-buda/hardware

Do they? I'm not aware of cases (but generally not up to date on topics like this, genuinely curious)

Yes, but you can't easily use your non-smartphone glucose monitor as a communications device to manage a bunch of drug dealers which can't be confiscated because your life depends on it. That changes once it's your smartphone.

This reminded me of this guy I was friends with starting in kindergarten, really smart kid, everyone kind of thought he would go on to study physics or something. Then sometime around 14 or so he started juggling and got so obsessed he would constantly practice during every break and even on the train to/from school. I think he's working as a circus artist now? Always kind of fascinated how you could get so absorbed in a topic, especially given my own propensity for rapidly cycling interests.

> "(...) best seen in the popularity of mass lay-offs – a strategy that has been resoundly proven to be counter-productive, costly, even disastrous, along multiple dimensions, over multiple decades of study."

This is the first time I read that big layoffs are "disastrous" for companies vs. some other approach in these situations (other maybe than the rather obvious fact that companies who are conducting mass layoffs are often in dire straits to begin with). I'm assuming "for companies" is implied here, because the way I read the sentence, the author is saying doing mass layoffs is misguided from the point of view of company management, not someone altruistically trying to maximize welfare in society overall or something like that. Does anyone have a good source for this?

If you run a trial like this hoping to get an idea of what effect the thing you're doing has, you have two things to worry about (using the terms common at least in econ here). The first is internal validity: Does what you're doing give you a good estimate (i.e., not biased) of the effect you're trying to estimate. The second is external validity: Does your result tell you enough about what would happen if you did the thing you did, but outside your trial context.

If you do a trial of four-day workweeks, you're probably always going to have issues with the latter for any feasible setup. As you allude to, the companies that self-select into participating in such a trial are probably not a good representation of the wider economy, even if you randomize among them.

Now if you randomize rollout of four-day weeks at least among those willing to participate, and you didn't lose about 50% of participants between the start of your trial and the follow-up survey, you would at least get an unbiased estimate of what the policy did for those companies. But alas, they don't even do that part.

I have some gripes with this group and their publications:

- They don't randomize pilots, so any effect you see is likely to be confounded

- In this publication, it appears there was ~50% attrition between the initial pilot and the follow-up study — again, a huge source of potential confounding

- As another example, in another publication, they showed a plot of GDP per capita against average working hours, and insinuated from the negative relationship that less working hours somehow made workers (causally) more productive, without even hinting at the obvious alternative explanation that people work less as they get more productive because they don't have to work as much to maintain standard of living.

I like that things like this are being tried, but I wish the research conducted on it were more intellectually honest and less obviously geared towards pursuing an agenda. The level of analysis here is more like a company marketing whitepaper than anything bordering on scientific.

Don't get me wrong, personally I think companies where it's feasible should just define minimum presence where the business needs it and leave it to employees where and when to do the work within those constraints depending on preferences and their situation. I don't have an axe to grind against working less (and in my circles it seems like many people are making this decision by reducing their workload to e.g. 80%, albeit at a corresponding salary cut). But the whole thing just seems a big disingenuous.

seems hard to pin down exactly, but in this Computerphile interview the gist seems to be "if you get to the point where you want a distinct code for upper- and lowercase characters, digits, and a few punctuation symbols, you land a little north of six bits, an odd number of bits would be annoying to implement in hardware, so let's go for eight."

https://www.youtube.com/watch?v=vuScajG_FuI&t=184s

This reminds me of when, as a teen, I got into tinkering with various security tools using the BackTrack live linux distro (I think it was the precursor to Kali?), and on my then laptop had to run about ten commands I had painstakingly cobbled together from various google searches in order to get the GPU drivers running and a decent screen resolution — on every reboot

Oh god, for some reason my brain just jumped to "International Parsing of Badly Documented and Absolutely Cursed Legacy File Formats Olympiad"

Rituals work!

Interesting, for me that sounds like exactly the montage thing, just without the set endpoint necessarily. But if you focus on your goal number too much, you may get disheartened by the slow progress there, or the brief setbacks along the way, whereas if you learn to enjoy the underlying process, you're just doing your thing and then before you know it your weight is sort of where you want it.

I think this is great! It may not work if you're trying to build a five-minute-a-day daily habit, it may not work for everyone, but there's def something to it! There's also a pattern of some pretty productive people making playlists of a few upbeat but simple songs they'd never listen to otherwise, and putting that on repeat when they go heads-down on something, see e.g. one of my favorite examples: https://ryanholiday.net/the-guilty-crazy-secret-that-helps-m...

Seconded! As an add-on to just navigating things, my native language is German, and when I first learned to read Cyrillic I was extremely surprised to realize how many words in Russian are similar to German. It's not like you get reading comprehension "for free", but I understand more than I would have expected given the totally different alphabet.

Huh, TIL this is in Switzerland

https://www.pv-magazine.com/2021/07/02/exploring-the-depths-...

> The 10M context ability wipes out most RAG stack complexity immediately.

1. People mention accuracy issues with longer contexts 2. People mention processing time issues with longer contexts 3. Something people haven't mentioned in this thread is cost -- even thought prompt tokens are usually cheaper than generated tokens, and Gemini seems to be cheaper than GPT-4, putting a whole knowledge base or 80-page document in the context is going to make every time you run that prompt quite expensive

+1 for Karpathy, the series is really good

I always wonder how much of this sort of thing is disdain for the specific tool vs. disdain for what working in a mid-to-large org. is like and the overhead it usually brings with it.

> The list wasn't the important part, but evaluating my own preferences & deciding what I truly cared about was important.

I get the temptation, I got very curios too and my first impulse would be asking the same, but maybe it might be better to just sit down and make one's own list instead.

the point of 3b1b's video wasn't "how can I figure out what words are still an option", but "what is the optimal word to guess at each step to minimize overall attempts needed to solve". Picking a "good" word will give you more information about the actual solution than picking a bad word, so you will get there faster.

ICYMI, 3blue1brown (Grant Sanderson) has a video applying information theory to solving Wordle, which is conceptually similar to mastermind -- https://www.youtube.com/watch?v=v68zYyaEmEA