trifurcate

> There is no good reason to react emotionally to professional criticism.

And yet when it happens (not a rare event), it can be quite unproductive for both parties. I would hope that you deal with other worst-practices in your domain with greater grace than a total refusal to deal with and prevent pathological outcomes.

Additionally, if you truly believe what you stated, then so long as I avoid sacrificing significant efficiency/productivity to keep my words from bothering people, then I do not see why it should bother you either. If I had to self-judge, I really don't think that the added conversational padding pans out to more than 15 minutes per day. That's worth spending to make sure that the people you work with are engaged with you and the task at hand in an accommodating way, rather than fearing being rebuked at each turn.

I always have doubts about others' work and suggestions on how they could do it better. To avoid constantly dominating others, I either have to shut my mouth or try to be more accommodating in how I deliver my feedback. I do a mix of both, but if you tell me to be straight with you, then I'll be straight with you.

Especially in large corporate organizations where there are vast differences in how people communicate and take input, and you talk to too many people in a given day to really get close with any of them, it requires a handshake to know that I can always indicate my roughly unfiltered technical opinion to you and that you'll take it in stride, and if you feel strongly about it, you'll have the confidence to shoot my idea down without me having to give you an opening by posing my suggestion as a question.

It's unfortunate for the potential purchaser of the business.

"Unfortunately, my friend ate his entire birthday cake." It's not unfortunate that my friend ate the cake (though it probably actually is, that's not why it's used there). It's unfortunate that I don't get to have any of his cake.

You're overestimating how long it takes to brute force SHA256 all 10 digit phone numbers from scratch.

That's not a minimum, just what the author had on hand. I can get 65B and 70B running just fine with exllama on 2 3090s.

You're really surprised that computers can outperform humans at pattern recognition?

From some Tor exit nodes (not all!), YouTube does present a modal to make you agree to the ToS before you are allowed to interact with the page. Probably some localities actually legally require you to do this, and so Alphabet complies.

The risk (mostly) isn't fake email screenshots, the risk is that I create postings for (or even commit to paying for fake reviews for) my competitor, and tarnish their reputation. It's hard to establish provenance of such an arrangement.

Even if the full text is fully deduplicated, there is just so much more content about Harry Potter on the internet. And not just retellings of it, but discussion of it, mentions of it, bits of information that convey context about the Harry Potter story, each instance of which will help further strengthen and detail the concept of Harry Potter during training.

No, I have not stopped beating my wife.

> Do you maintain a freely-available repository of all of your knowledge and experience, in case someone else wants to consult it one day?

No, but I try to maintain some of it, and I see the value in maintaining as much of it as possible.

> If they came out and said "We are pausing hiring..." the markets would rightly be like "Man they are screwed..."

This hasn't been the case with the ongoing tech layoffs though? Investors for the most part congratulate these companies for diligently cutting costs in preparation of a recession.

The slowdown here is going to be more about the increased context length than throttling.

You really don't think Google is using a language model in its spam filter?

I don't understand, GPT-4 answers this very easily:

> To calculate the total cost, first find the cost for each item:

> Two baguettes at 1 euro each: 2 * 1 = 2 euros

> Five chocolatines at 1.40 euros each: 5 * 1.40 = 7 euros

> Five croissants at 1.20 euros each: 5 * 1.20 = 6 euros

> Five raisin breads at 1.60 euros each: 5 * 1.60 = 8 euros

> Now, add up the cost of each item:

> 2 euros (baguettes) + 7 euros (chocolatines) + 6 euros (croissants) + 8 euros (raisin breads) = 23 euros

> You should take 23 euros with you to purchase all the items.

Are you sure you're using GPT-4 and not 3.5? GPT-4 is incomparably more competent compared to GPT-3.5 on logical tasks like this (trust me, I've had it solve much more complicated questions than this), and you aren't using GPT-4 on chat.openai.com unless you're paying for it and deliberately picking it when creating a new chat.

Edit: Here's an example of a more complicated question that GPT-4 answered correctly on the first try: https://i.imgur.com/JMC7jsw.png

Funnily enough, this was also a problem that a friend posed to me while trying to challenge the reasoning ability of GPT-4. As you can see (cross-reference it if you like), it nailed the answer.

Regarding [2], dang, I am working on exactly this! I mean, it's not that novel of a technique once you start controlling the sampling process directly, but you beat me to the punch.

This technique generalizes to pretty much any grammar one can specify. I weakly hypothesize that by making it impossible for the LM to output syntactically invalid text, the model's task performance improves not just because all of its outputs are valid, but also because part of the model's "processing power" gets "rerouted" from trying to understand and follow the grammar it's writing, to applying improved reasoning overall.

> Hence, your daughter being or becoming an incredible artist would have no meaning, except perhaps for herself enjoying the process of creating art.

There are lots of points and arguments to be made in this general area, but I have to ask, is this really so bad? I mean, what is the point of our lives and everything we do, other than to generally spend the rest of our time doing things we enjoy for their own sake?

If we're comparing "your daughter is an incredible artist, and here's a job for her designing product packaging for a multinational conglomerate" to "your daughter is an incredible artist, and the multinational conglomerate is using a diffusion model to design their packaging", I think it's really hard to say that the former is better than the latter. Of course, it all depends on the economic model, but the line I am quoting is within that assumption you made of the economic model being able to support this. In that case, I am for the latter wholeheartedly.

Economic incentives are great to get people "hustling", but they are rarely aligned with the human values you wish to protect, and mostly by chance if they are. Your daughter's artistry is better "spent" on art for art's (and personal enjoyment's) sake than on drawing clip art for an obscure HR form somewhere, IMO.

It's not really fair to ask someone to leverage themselves to display their conviction.

> Which means those exchanges are ... completely pointless.

I wonder what entices people to execute hundreds of thousands of trades totaling USD equivalent billions in daily trading volume on these platforms then. Would you say that everyone using these platforms is basically out to scam others or get rich quick?

Decentralized exchanges are... exactly not that. They only trade on-chain assets for on-chain assets. Fiat ramps are not part of what makes a decentralized exchange.

Also, conversely, the kernel doesn't do that much. Most of the Linux kernel source consists of device drivers, which compile to modules rather than get bundled into vmlinuz. Many of these modules are also rarely built if ever. The kernel itself really is a pretty small fraction of the complete software bundle that makes a Linux system functional.

All you need is 2 3090s.

In my experience, the smaller models are almost completely worthless as-is. 65B is the only decent one (I'd say just behind gpt-3.5-turbo, and obviously it's not instruction tuned but I mean the coherency of the core language model), and understandably people aren't really paying attention or devoting much resources to the largest one. 30B shows promise for specific tasks with fine tuning, but 7B and 13B are just toys.

Is this some kind of copypasta? Too many tropes all at once. "GTP" on top of all this is too on the nose.

Many, many people are working on this right now. Having had a shot at it, the details become very important very quickly and it feels to me that getting it right will require a ton of iteration at this stage. Such that it feels easier to wait for GPT-N+1 rather than work on improving its harness. Regardless, again there is a ton of effort being poured into this very specific thing across the world.

> Also, if all it takes to mess with altimeters is a rogue 5G cell, the right approach is to harden the altimeters, instead of hoping no 5G cell ever will be misconfigured or pushed wrong configuation parameters by a enemy country or ransomware hacker group.

Wow, that is a lot of misconception about the threat model and real issue at hand packed into a single sentence.

First of all, great, let's say you're right. How do you harden the existing altimeters out there, equipped on thousands of aircraft? These are sensitive, tightly controlled and life-critical instruments. Radar altimeters have been in use for decades, almost a century now, and so the FAA is for the most part talking about already installed altimeters which were designed far before 5G was ever a thing. Likely most of their designs will not readily accommodate hardening against new kinds of RF interference. If you managed to patch all the different kinds out there and get fixes pushed out, that'd be a great effort in labor, logistics, and certification. Of course, one could simply replace the old radar altimeters with new ones, at a much greater cost too.

Separately from the difficulty of actual hardening, there's the question to be asked of, why ask old technology to adapt to new one when we can adapt the new technology to suit the old? Especially because, again, these are safety critical instruments that you don't wanna mess with, whereas 5G towers are much better suited to new experiments or fixes, or just being designed correctly and safely from the get go when the technology hasn't even launched yet.

And then, why do you think this is a security issue, as in an attacker aiming to disrupt air travel? You must be aware that most all aviation radio comms are in the clear and extremely easy to jam and spoof, including critical instrument systems like ILS. That is not the threat model! The FAA is worried about interference from ordinary people trying to use their cellphones, rather than a specialized attacker targeting radar altimeters. Of course, because these are two very different situations!

The bond is impaired in that sense. The concept of present value isn't made up just for fun, it's because the value of money depends upon the time at which it is available. $10 in 10 years is obviously worth less than $10 right now, which is not only captured by present value calculations but it's also plainly and intuitively visible if you make the chain of associations of high interest rates -> higher price levels -> lower monetary value. So yes, if your bond sells for $87, you can be reasonably sure that $87 is the value of all of its payments back to you. It doesn't matter that the nominal payouts sum to $100, because they are denominated in future dollars which are worth less than present dollars! You need to cover the shortfall when you move those payments from the future to now!

None of the siblings are right. The models themselves are idempotent: given the same context you will get the same activations. However the output distribution is sampled in a pseudorandom way by these chat tools. You can seed all the prngs in the system to always have reproducible output using sampling, or even go beyond that and just work with the raw probability distribution by hand.

Which law caused the banks to fail? SVB failed because they took the wrong risks. I don't understand what your question means.

You need to realize that essentially bond pricing reflects the present value of all cash flows you expect from the bond. For long term bonds especially, this makes it a particularly risky book, because you are very sensitive to interest rate changes.

If the interest rate goes up while you are holding your low interest bonds, that means that your future cash flow from the bond (the repayment) is literally worth less than what it was. Your bond payments have a lower real value due to the higher interest rate of the surrounding environment and the increasing price levels, despite being the same nominal amount. That's why the bond's price plummets in the market, which is why this is a solvency crisis: because the assets really are not good for the liabilities at present value, which is the only kind of valuation that makes sense here.