Felz

About 30 seconds for a smaller response on an EC2 instance. Mind you this is without a GPU because I couldn't figure out how to set one up.

Oh wow, I was literally just making a toy Discord bot for GPT-2. Guess I'll update it with the bigger model.

EDIT: Done! It takes about 3x longer than it did before to generate a response, so if you try it be very patient. Also, I made this in three hours so I wouldn't be surprised if it goes up in flames at some point.

https://github.com/ScottPeterJohnson/gpt2-discord

Assuming that were true, wouldn't it be a massive existential risk to Google? The finely tuned ML model that classified everything correctly before might catch fire and explode if the thing generating its data changes.

In practice I assume that data has highly marginal value for them so the small percent of people who bother to use this won't matter much.

This article didn't mention whether or not their servers are attempting delivery using TLS/StartTLS, which is a good thing to check. I think Google penalizes email delivered insecurely.

Which does make it rather curious that the music's kept playing, so far. Is "common knowledge" that it's probably a complete fraud (like this filing should give) required to pop the tether bubble? Or will even this not be enough?

S3QL does this with sqlite. An actual external DB would give you multiple users at once, though.

I've been using a combination of S3/Postgres to get filelike semantics myself, but not with a POSIX interface.

The biggest problem for general use is that S3 is really for immutable file storage, and not so much things that'll change a lot. That can be worked around, obviously.

> Smartphones are equally as portable, too—but it’s hard to argue that they’re even half as lovely.

Actually, why not? You could find an endearing custom printed case, and to an extent personalize the software in a lovely and unique way.

Is the problem that they're not rare/irreproducible enough to be a collector's item? If so, will we eventually see collectors obsess over out of production phones?

Memorizing APIs is... not how I'd approach productivity, but if it works for you, great.

I'm doing a lot of Kotlin right now, where simple, useful extension functions are an autocomplete away. I don't need to memorize how to get a slice of an array; I know it's going to be named something like .slice(), and I can jump right to the documentation from my IDE.

When I run into something not in the standard library, I write a little extension function of my own. And then I remember it the next time I run into that problem, and get to reuse it, just like it were part of the language.

A bit of a wild idea:

Deduce how important a game is to people by introducing small amounts of artificial delay when playing it. This could be input delay, or maybe the game freezes up for a second every minute, or it takes a bit longer to launch or something.

Measure the dropoff rate. People should be really willing to keep playing great games even when inconvenient, and drop marginally-fun grindy games faster. With a large enough install base, you shouldn't need to degrade anyone's experience very much to get enough data.

> I don't think most people are putting down Go as an outright terrible language.

Beware of evaporative cooling. All the people who can't put up with a lack of generics will wander off, and you'll only be left with people who are okay with the status quo. That doesn't mean you don't have a problem.

Lucene's pretty great. The segments file model even works pretty well with S3.

I'm skeptical of any model that throws up its hands and says "this is so for a number of reasons I guess". Multifactor trends are unlikely [1], bad at explanation and prediction, and easy to fit to anything. It's better to admit you don't know than say it's a combination of N things.

Remember, most things are hard to explain until you know the truth. That doesn't mean the truth is super complicated.

[1] https://slatestarcodex.com/2015/02/14/how-likely-are-multifa...

Yea. The entire argument appears to be a single anecdote about some guy's dog acting really upset after biting his hand.

Quoting the article:

> For the longest time, science has depicted animals as stimulus-response machines while declaring their inner lives barren. This has helped us sustain our customary “anthropodenial”: the denial that we are animals. We like to see ourselves as special, but whatever the difference between humans and animals may be, it is unlikely to be found in the emotional domain.

Isn't this just a strawman in service to an applause light?

Is it concerning that there are short, sudden drops in prediction in the middle of a block otherwise solidly classified as rabbit/duck? I don't know much ML, does anyone know why it'd be so discontinuous?

I actually used to bike that road all the time 3-4 years ago. The bridge section before it could be disturbingly dark, but the point where she was crossing was decently lit, as I recall. It was also somewhere I would never consider crossing, mostly because there was a crosswalk just up ahead.

Still, if I absolutely had to it was definitely an intersection to watch carefully for cars when crossing, but one with fairly long sightlines for that.

I noticed Google Voice was broken in Firefox yesterday. (Shocking, I know.)

I generally recommend .com. You can still get a pretty decent domain name if you try a bunch of options and put some thought into it. (It has to be 5 letters or above; I recently confirmed that every 4 letter or fewer domain name is taken.)

.com has legitimacy, it's not going to have hiccups (some country code domain names are pretty iffy, like .io had issues a while back), and contracts with ICANN ensure it's not going to extort a huge price out of you.

It should support that through Sieve (filters, basically). If you want to redirect a whole domain, that's a routing level function.

Sent externally means to a different account or provider. So basically, sending email to yourself or other users using your account is much cheaper. (There's no spam worries there.)

Yea, that was my reaction too some months ago when I was looking to move away from gsuite, which was beginning to get ridiculously slow and expensive. (Seriously, what is gmail _doing_?)

We have this weird dichotomy in services pricing that's either free or pretty expensive. If you're a free customer, you're not a customer. You're a potential customer in need of sales. And everyone is looking for the features they can put behind a pricing gate.

Let me know if you run into any issues!

Well, thanks. Feel free to let me know if you come across any issues.

I'll definitely try not to become full of bullshit :). I hope it's not as inevitable as it seems. (It probably helps that there's no VC funding involved, and I can stay small.)

Yes, which is part of why I recommend getting your own domain. (Although you gotta pay for that too.) Is that a big issue for you?

Don't SPF/SSL also provide proof that email comes from where it says it does? As far as I can tell, DKIM's advantage is to allow for direct message forwarding (not remailing), which is a fair use case, but pretty specific. In return it introduces a fair chunk of complexity to do right, with signing and message key rotation and keeping track of who you gave keys to.

SpamAssassin assigns small positive scores for valid SPF/DKIM/whatnot headers (and larger negative for lacking either), but it's not really an effective spam deterrant. Spammers can set up their own domains that pass all the checks (although I've heard they're having good times just sending from Gmail).

I didn't take it as rude! I appreciate your input. And I agree that the billing will seem more complicated until and unless I can win enough trust that users realize the costs don't matter, because they'll come out ahead.

You're right that sending/receiving don't add to overhead much. The receipt is actually way overinflated right now, and should be down to a more reasonable number tomorrow (about $0.03/1000). I might just waive it completely if it complicates things too much.

Sending is much harder, because I need to deter spammers. I priced it significantly above Mailchimp's price for that reason. But the cost will definitely go down as my ability to detect and ban spammers improves, and we'll likely end up pretty much with a scheme like you're proposing.

All good ideas. I think it'll also help a lot if in a year or so I'm able to say "99% of our customers only spend $X a year".

I actually just checked the storage price, and I notice that at some point I added a $2/10,000 charge to the equation at some point, which is almost all of the charge price. I'm going to remove that tomorrow, since it was probably just paranoia and not actual cost plus margin.

Good to hear! I'll be sleeping soon, but feel free to email my support address or just comment directly if you have any issues.

Thanks for the suggestions! I'll look at the privacy policy again and see how I can make it more clear that we're not abusing users' trust.

I did look at mailbox.org. It's about double for their lowest monthly plan, which is reasonable.

Where I hope Purelymail will shine is that it scales a lot more naturally (once you hit mailbox.org's 2 GB cap, you upgrade to a plan that's 2.5x more expensive), it enables use cases that'd get you frowny faces from other providers (because we charge appropriately), and there's no surcharge for things that don't actually cost more.

Want a dozen usernames? Sure, whatever. Need a hundred users? Sure. Store 2 TB of mail for some reason? We've got you covered.

The downside, as you point out, is that the pricing gets a bit scary and I need to work on making it feel safer.

Hah, you're right- that's terrible phrasing.

1. We require SPF, but don't have DKIM/DMARC yet. DKIM as far as I can tell doesn't actually... do... anything, but I need to look closer.

2. Same way anyone does it, I'd assume. Shared IPs, rate limits on sending, banning users who send spam emails. I'll likely need to hone exact approaches more.

3. Not that I know of? The mailserver I'm using might handle that.

4. Yea, fairly generic Spamassassin setup that I'm tuning.

5. I think signatures work through Roundcube, and maybe clientside encryption too.

6. I don't have SLAs yet (it's a beta!). My architecture allows for continuous deployment with no planned downtime, though. Delivery from gmail -> my servers and back seems to take about a minute as far as I know, but I can't answer the delivery time metrics without more data. You should get a bounce email on final delivery failure, which should take a maximum of about 208 minutes.

7. They're stored encrypted and compressed in S3, with an encryption key based off of a derivative of your password. Specifically: The encryption is AES/GCM with a per-message key encrypted by a libsodium crypto box, whose private key can be retrieved with a derivative of the user's password. The bucket also has AES-256 encryption in place.

Good questions! I'm going to work on documentation tomorrow. And yea, I realize that sending emails is going to suck.

The thing about actually using Fastmail's 25 GB for $50 is that if you go even slightly over, you're going to need a $90 plan. Most users aren't going to use anything near that amount, especially since our compression is pretty good. (I'm not sure if Fastmail's 25 GB count is compressed or not.)

You're right that I need to add more value. My planned direction is more along the lines of utility than security/privacy, which I think Protonmail covers pretty well. There's a lot of interesting value-add to be done in email.

The pricing is more that we offer honest pay-as-you-go pricing, with caps to make sure you don't actually get a ruinous charge, in direct contrast to "billing something reasonable for a whole year". If you'd rather do the former, then yeah! You're really well covered in the email space already. But I think people are getting subscription fatigue, where every service imaginable wants its $5/month cut of the pie.