muskmusk

Jesus.

This is starting to feel pretty **ing exponential.

I don't think anyone can consider it a success just yet. But someone just injected 2 billion dollars into the enterprise. They are probably given a lot more information than you and I are. Maybe those people are dead wrong and they lose all their money, but maybe not.

With the public information there is neither a point in positivity not negativity. It would be speculation either way. A question that could be worth pondering is what exactly the investors are shown? What would you need to see to put all your money in thinking machines?

I love it!

Finally some clear thinking on a very important topic.

This is a surprisingly good idea. The model vs model is fun, but not really that useful.

But this could be a legitimate way to design apps in general if you could tell the models what you liked and didn't like.

I guess masks are completely off now. We can see who sells out to the highest bidder and who won't sell because they care more about the mission.

"everything evolves until it becomes an operating system"

I an not sure these fairly one dimensional views are useful. Completely agree with the premises that it is the managers job to know what their people are doing. I have also worked at companies where the managers weren't doing their job. I can also completely see how an overfocus on metrics kills an otherwise good company.

I don't agree with the conclusion though: metrics are bad.

In software, in theory, you don't need metrics. You just write good code that does what it needs to do. Add tests to prove it, you are good. In reality though everyone agrees that metrics are a good thing to have. Why do we need them? Complexity and sometimes we just make mistakes. We could write software that is very close to bug free, but then we move at aerospace speed. Not always optimal.

Software metrics come with all the pitfalls of management metrics like over focus on metrics and not seeing the forest for the trees. Yet we still find the metrics useful.

Managers face many of the same problems, so they ask for metrics. I don't think that is bad. Being bad at using metrics is bad.

It's a free world. Nothing stops you from applying their findings to bigger datasets. It would be a valuable contribution.

Yes.

Heh I guess it's s matter of perspective. Your son's head is not made of silicon so in that sense it is a large neighborhood. But if you put them behind a screen and only see the output then the neighborhood looks smaller. Maybe it looks even smaller a couple of years in the future. It certainly looks smaller than it did a couple of years in the past.

Friend, the creator of this new progress is a machine learning PhD with a decade of experience in pushing machine learning forward. He knows a lot of math too. Maybe there is a chance that he too can tell the difference between a meaningless advance and an important one?

If you ask Ilya Sutskever he will say your kids head is full of neurons, so is LLMs.

LLMs comsume training data and can then be asked questions. How different is that to your son watching YouTube and then answering questions?

It's not 1:1 the same,yet, but it's in the neighborhood.

Those prompts are wild. It is deeply impressive that it works. Think about what would happen if you gave such instructions to a human. Would they be able comply? How big is the overlap between people who are creative enough to produce the kind of pictures Dall-E produces and disciplined enough to follow complex instructions so rigourously?

It also is just straight up impossible to convert those instructions to "regular" code.

That is impressive. Interesting that the prefill (i am guessing this is prompt processing) is so much slower than decoding.

Its my understanding that under normal circumstances decoding is memory bandwidth bound which prompt processing isn't due to batching. Is there some quirk in your setup?

Exactly, also 3year olds are capable of language.

How do you prove that something that lived 300.000 years ago could communicate?

The opioid addict cases are probably more when entire cities gets fired because a company moves or shuts down.

Surely there will be cases where those that are replaced are unable to find work, but I am willing to bet that

1. Those cases are somewhat rare

2. They "make sense" (e.g. close to retirement anyway)

The common case is just to get a new job. Maybe the new job is not as good, maybe it is actually better in the long run.

There are plenty of jobs to go around these days. Unemployment is very low.

If I were to give advice I would personally tell people to stop looking for work that is "safe from automation" because over a long horizon very little is safe. Figure out how to work with the technology instead. Don't need to code there are lots of other ways to contribute.

I agree :)

7-10 years ago people on this forum thought they were invincible. People who said said software work would be commoditized were ridiculed.

Rate of change is just fast and probably lots of people just thought that a career in tech was a free golden ticket for life. Now that it is threatened people are getting anxious and defensive.

The vast majority of bad things happening in the world are okay as long as they don't happen to "me". Living conditions in sub Saharan Africa are really poor, but noone cares. We are all hypocrites to some degree.

Massively if the companies are willing to let AI actually help. If the companies are just trying to delay you from unsubscribing by using lame tactics then no, but nothing fixes those companies.

Massive upside for some things, neutral for other things. That's net positive.

What if replacing people's jobs is helping people?

Milk men, computers (the people kind), horse carriage chauffeurs, people who light lamp posts (when they were oil based) we could go on here.

You are clearly coming from empathic place, but I am not sure your premise is nuanced enough. Lots of progress has "victims" but that doesn't necessarily make progress had.

Most requests are very routine and are basic exchanges of information. Think "how do I reset my password?".

Don't see why those can't be automated.

Eventually I don't see why an AI shouldn't process refunds. Its an infinitely measurable process and you always know the risk involved (the amount). So it's trivial to decide if you think the AI should handle it or send it off to a human for approval. I would not be surprised if there were already companies experimenting with stuff like this. Widespread adoption will probably take time.

Different people different experiences I guess.

I used to work in the industry and my experience had been that companies who don't care about customer service will have shitty customer service, bots or not. The opposite is true as well.

I have seen simple request bounce between up to 15 customer service reps without the end user getting what they wanted. On the contrary I think some of the best customer service is when there is no need for customer service at all.

Why?

Have you ever sat in a customer service chat with someone who wasn't very helpful and very slow sat the same time?

There were probably handling several other chats at the same time.

Why shouldn't we replace those interactions with an AI that is fast and infinitely patient and friendly?

I am willing to be convinced otherwise, but I don't see how fighting technological progress that is net positive for almost everyone is productive here.

Bingo!

> The author hits the nail on the head with his claim that imaginary problems are more fun than real ones.

Not necessarily. It's just that most developers have never worked in a setting where they got to work on problems properly.

Solving real problems for real people is very addictive. There is a reason some people like working for startups. It's because you live very close to your users and when you make them happy you know.

The second interesting fact is that if you just plow ahead and solve many real problems fast you will eventually run into problems that are both real and interesting.

After having tried that there has been no going back for me. I am allergic to imaginary problems. It feels as pointless as watching people who are famous for being famous argue on TV.

I think we are all victims of our feedback loops. (University) Education sublely teaches us that the only important problems are those that are very difficult and preferably have never been solved before. Those same problems also make for better blogposts. In the real world the incentives are mostly opposite. Problems with no known solutions (or only really difficult solutions) are generally bad. They can be worth it, but you should stay away from them until you know that are worth it. Software engineers seem to almost pride themselves on not knowing what their users want.

It takes a while to scrub all that bad learning out and replace it with something better. Unfortunately some people are stuck.

> As someone who prefers working in fast paced environments and have responded to calls like this in the past, I have found that people who say "I want someone who moves fast" often are really bad at managing and have unrealistic views of how long things take. People who write that phrase might think they have Steve Jobs reality distortion fields, but are mostly deluded. The fact that this particular posting was made by someone who actually walks the walk is not obvious - I don't judge people by what they say but what they do.

There is a very fine line between walking the walk and delusional. Amazon employee number 1 was not super fond of Bezos when he left (Shel Kaphan). Amazon employee number 2 lasted less than 2 years. As you point out it is clear after the fact that Bezos walked the walk, but he might also have been delusional depending on your point of view. Any success tends to contain a non-trivial amount of luck. Whether someone is ends up a lunatic or visionary often depends on factors that person doesn't control.

Kaphan also says he loved the first 3 years and then it gradually fell off from there. I have had a similar experience at a previous employment. Point is; startups change dramatically quite often. The same person might transition from bat shit crazy to the opposite (and back) in a rather short timeframe. You certainly can find someone who is just a through and through asshole who have learned to say the right things, but I honestly haven't met very many of them.

Most of them are just trying really hard without knowing what they are doing so they make a lot of mistakes. Managing that sort of person is a skill. Since most founders are in this category its a rather essential one. Not trying to bash the founders, the very definition of a successful startup means that most of the employees are way out of their league. When a start up doubles every year (or faster) then half the company has been there less than a year. They are just getting started and they now have double the reports they had a hear ago. Everything is on fire in both the good and the bad sense. Its just the nature of the game.

> All I'm saying is that this would have been mildly intriguing because of the ecommerce angle, which was a new problem at the time, but would require sussing out of the Bezos fellow to see if he's the real deal or deluded

Totally, you need to do your due diligence :)

> Weird that this read of it is so controversial...

Lately HN has turned into a place that is best personified as 50+ year olds looking for a stable and relaxing gig where they get paid a ton of money to play with technology. Unfortunately I see that this kind of attitude is getting quite mainstream (even among significantly younger engineers). There is nothing wrong with wanting that. I see the appeal. Its just annoying deal with. It is weird that HN which has startup roots are starting to adopt such a posture. After re-reading your post I can see that it wasn't actually the angle you came from it just read very closely to that kind of narrative. You got caught in the crossfire :P

If you play in the shark pool you might get bitten. It might also be incredibly profitable if you know how to navigate it.

If you can't tell whether the company you work for is going somewhere or just using you for cheap labor then that says as much about you as it does about the company. Employment is a two way street.

If you don't want risk, then don't choose risk. But there is nothing inherently wrong with risk.

As the comment above you pointed out, then this job wasn't for you.

If you made a job add for e.g. the job of an Olympic level swimmer most people would consider the add bonkers, not those who are actually olympic level swimmers.

Believe it or not there is in fact people who can maintain systems in 1/3 of the time most competent people think possible.

There is also people who love the thrill of going as fast as possible and who would prefer this over a more leisurely pace.

Of course the amount of people this applies to is probably less than 1%. But those people are going to really like the job.

I don't see the conflict. They see current models as mostly harmless, but what comes next is dangerous.

It sounds a little too much sci-fi for me, but I guess he knows better.

We de-linked the two because there was no practical way to autoscale the database, but autoscaling microservices is easy. Obviously you want to be very cautious about how you treat inelastic compute.

Today there are lots of practical ways to autoscale databases and colocating data and compute is generally good for performance and simplicity (your mileage may vary for simplicity).

> I don’t know why but I find this troubling.

You used the word anthropomorphize twice so I am guessing you don't like building systems whose entire premise rest on anthropomorphization. Sounds like a reasonable gut reaction to me.

I think another way to think of all of this is: LLM's are just pattern matchers and completers. What the training does is just to slowly etch a pattern into the LLM that it will then complete when it later sees it in the wild. The pattern can be anything.

If you have a pattern matcher and completer and you want it to perform the role of configurable chatbot. What kind of patterns would choose for this? My guess is that the whole system/assistant paradigm was chosen because it is extraordinarily easy to understand for humans. The LLM doesn't care what the pattern is, it will complete whatever pattern you give it.

> And part of what I find troubling is how casually people (prompters and users) are willing to go along with the ‘you are a chatbot’ fiction.

That is precisely why it was chosen :)