Bjorkbat

Still impressive that it can beat top models with fine-tuning, but now I’m mostly impressed by the fact that the 70b model was so good to begin with.

To some extent I'm glad we live in a time where websites are lean and efficient for the reasons you mentioned, but man, I miss the days when the internet was actually worth exploring.

The book is an interesting read. The death of Flash feels like the start of a dark age for web design when you compare what came before with what came now. Granted, this may be less due to the death of Flash and more due to marketers and UX professionals trying to simplify the journey from visit to purchase.

If you take this at face-value, they probably aren't profitable (to the tune of a $5 billion loss). I don't like idle speculation, but having actually read this their analysis seems pretty sound. Their numbers on training costs and hosting actually came from an anonymous insider. Impossible to say how reliable this anonymous insider is, but planting false, exaggerated numbers to outwit competitors seems a bit too much like a hare-brained scheme, hence why I opt to go with the numbers. If anything, companies tend to downplay costs when lying.

The one thing that caught me as a little bit crazy was spending $1.5 billion on employees. They have ~1500 employees, so that's $1 million per head. That seems insane, but then again, AI researchers are a valuable commodity these days.

Even if the cost of employees was $0, you'd still be looking at estimated losses in the billions.

Like, ideally, it shouldn't really take that much code to implement a thing. I like to think of programming as writing a bunch of levers, starting with simple levers for simple jobs, incrementally ratcheting up to larger levers lifting the smaller levels. Before too long, it'll feel as though you've written a lever capable of lifting the world...or at least one that makes an otherwise wickedly difficult project reasonably manageable.

If you say that LLMs make you more productive because it allowed you to finish a project that would otherwise take forever to write, then I'm skeptical that an LLM is the best solution. I mean, it's a solution at least, but I can't help but wonder if there's a better solution.

If the problem is that you lack the understanding to take on such a project, then perhaps what we really need are better tools for understanding. I myself have found that LLMs are great for gaining a quick understanding of languages that otherwise have sparse information for beginners, but I have to wonder if perhaps there's a better way.

If, on the other hand, the problem is that writing that much code would take forever, then I have to wonder if the real solution is that we need a better way to turn programming languages into patterns (levers) and turn said patterns into larger patterns (larger levers)

A partial solution works, but only partially well, and occasionally has consequences one has to reckon with

Now, granted, your risk of dying from heart disease is actually greater than that of cancer, so it's entirely possible that the low cancer rate is because everyone is dying from heart disease instead. It would make a lot of sense since New Mexico is one of the poorest states in the country.

Thing is though, the heart disease mortality rate is also low. (https://www.cdc.gov/nchs/pressroom/sosmap/heart_disease_mort...)

So I don't know, I'm kind of inclined to that our low mortality rates might be due to the sun instead. On top of being located further south than most states, we also get more UV exposure due to our higher elevations, and most days are relatively cloud free. Other theory of mine is a more Latin-influenced diet and exercise, but again, we're a poor state, poverty tends to be negatively correlated with exercise and diet.

It's a trend followed by most of the Mountain West, which is roughly as arid and sun-rich as New Mexico. I just find New Mexico so interesting because it definitely dispels any notion that people are healthier because they're richer or have access to better health care.

A while back someone on Twitter seemed to confirm that Claude-3.5 was aware of the Github issues inside the dataset by mentioning them, but I couldn't find the original post.

30% performance on the full SWE-bench benchmark is quite the leap, but just how "real" of an achievement is this? Anecdotal reports mention that GPT-4o is marginally better than GPT-4 turbo at best, and yet agents leveraging the LLM did perform better.

What would happen if SWE-bench was updated, top to bottom, with completely new Github issues? Would all these agents just completely shit the bed?

I mean, going by their own model evals on various benchmarks (https://llama.meta.com/), Llama 405b scores anywhere from a few points to almost 10 points more than than Llama 70b even though the former has ~5.5x more params. As far as scale in concerned, the relationship isn't even linear.

Which in most cases makes sense, you obviously can't get a 200% on these benchmarks, so if the smaller model is already at ~95% or whatever then there isn't much room for improvement. There is, however, the GPQA benchmark. Whereas Llama 70b scores ~47%, Llama 405b only scores ~51%. That's not a huge improvement despite the significant difference in size.

Most likely, we're going to see improvements in small model performance by way of better data. Otherwise though, I fail to see how we're supposed to get significantly better model performance by way of scale when the relationship between model size and benchmark scores is nowhere near linear. I really wish someone who's team "scale is all you need" could help me see what I'm missing.

And of course we might find some breakthrough that enables actual reasoning in models or whatever, but I find that purely speculative at this point, anything but inevitable.

https://www.indiehackers.com/post/holy-heck-this-is-hard-8eb...

Granted, this is Indie Hackers, so take this with a grain of salt, but I think the general idea still applies. Getting to $10k a month, let alone making $250k a year (so ~$20k/month) is really hard, probably about as hard as making $250k a year as an employee, maybe even harder.

It's as if the easiest people to fool are the researchers themselves

It gets complicated when you have to build websites and apps to actually solve some kind of business / consumer need, and I would argue it isn't even the programming aspect that makes it complicated. Rather, it's the fact that the thing you're building is built atop a layer of dependencies that you don't really know anything about, who in turn have their own layer of dependencies. To use these dependencies effectively, you need to know about some of the concepts embodied within them, such as a render lifecycle if we're talking about some kind of frontend js framework, knowledge that isn't precisely "programming knowledge".

And I as I mentioned before, this is why I think I'm not too worried about LLMs getting good at programming.

Good to know I suppose.

I had a brief teaching stint out there on Shiprock. One of the first peculiar things I noticed was a clinic that I first thought offered some kind of radiation therapy, like an oncology clinic or what have you. Upon closer inspection I realized it was a clinic for those affected by radiation-related illnesses.

Meanwhile, given all the oil and gas activity, it certainly seems like that has given them far fewer problems.

Something which I'm kind of convinced about though is that I think after this we're going to see an end of an era in tech in that I don't foresee another "next big thing". Feels like we've kind of played all of them out, they've all more-or-less matured. Like, maybe there will be another hype cycle in AR/VR, but unless Apple really outdoes itself with the next Vision Pro launch I don't foresee it being all that buzzy. Hell, even if Apple launches a headset that's both affordable and ridiculously good, I still don't foresee AR/VR being all that buzzy.

And if AI really does live up to the hype, well, there still won't be a "next big thing", just for an entirely different reason.

I'm not personally upset about it. There's still a wealth of good ideas out there to chase that don't have the scaling potential to attract VC interest but nonetheless could potentially make a decent chunk of people rich and employ a lot of people. Perhaps society will give more of its collective attention to the problems that aren't so easily solved with a mobile app.

Getting both kids and adults (especially kids) to figure out how to program is easy if you understand that most of the concepts are better taught visually through p5js or what have you. Once they leave that sandbox, however, and have to contend with what has to go into developing a production React app, it's a different animal.

Programming is easy. Putting a bunch of black-boxes together in order to build some app or whatever is much, much harder and more complex (and, arguably, I think that calling it programming is kind of deceptive. You're technically doing programming, but you really feel like you are? I can't say I do.)

Incidentally this is perhaps why I'm calmer than others when it comes to AI getting better and better at programming. All these researchers and companies have done is given me another black box to manage. They mean to assault my castle by first repairing its walls.

God I love Paradox Interactive. They aren't even trying to radicalize gamers against rent-seeking capitalism. They're just putting out games with complex simulations under the hood and inadvertently discovering that landlords ruin the game.

Kind of funny that someone's syllabus from a class taught in 2015 has done more to broaden my mind concerning speculative computing futures and "what-ifs" than anything else I've come across.

I think the only point of disagreement we might have is that I don't necessarily think that the speadsheet, even an improved one, is the best graphical model/structure for articulating complex processes, but it seems to be the best we've discovered thus far.

It's made me realize that it's usually the rule that intent at least feels easier to convey through code rather than through natural language. This is especially true if I'm trying to perform some complex mathematical task that's geometric in nature. Put another way, it's probably easier to understand how poisson disc sampling works if you saw the code for it, vs if you read the actual paper, especially if the code had comments providing additional context for different blocks of code. Doesn't help that the paper uses complex mathematical terminology, whereas the code, at worst, might have very terse variable names

And yet, I think a better example might be found in markup languages like HTML. Trying to convey the layout of a webpage using ONLY natural language seems really hard compared to using a markup language.

A simple explanation is that the devil is in the details when it comes to implementation. Edge cases and granular behavior are hard to capture except in granular snippets of code. I'm not convinced that LLMs necessarily solve this problem. Even if you are working with a theoretically quite competent LLM, there are still going to be instances where describing what you want is actually challenging, to the point where it would be easier if you just did it yourself. You could argue that this doesn't matter in the case of simple software, but I think we underestimate what people really expect out of something that's "simple", and I think we underestimate people's tendency to grow bored of old toys, especially if they don't work as expected.

If anything, my belief is that LLMs by themselves aren't necessarily going to make this a reality. You need an accessible and empowering UX/UI to go along with it. Otherwise, asking an LLM to build software for you probably won't be much of a fun experience aside from those who are AI enthusiasts foremost.

Side note, I have painful feelings about so many UX researchers I used to admire jumping on board the AI hype train so uncritically. I kind of get it, their job is to speculate on new possibilities a tech offers without getting too hung up on external complications. Still, I feel disillusioned. It seems that prior to all of this these same people were questioning our implicit assumptions with how we interact with computers in really interesting ways (in the same vein as Bret Victor). Now, their takes are starting to converge with that of the usual anonymous midwit AI enthusiast on Twitter who pivoted from crypto.

Put more bluntly, the idea that LLMs will usher in a golden age of people making simple software is kind of a boring speculative future, a speculative future shared and talked about by the most uninspired and boring people on Twitter.

I think we underestimate the difficulty of our jobs, or perhaps we underestimate the chance that something better will take its place. Perhaps a bit of both.

The other thing that helps me sleep at night is that I really don’t believe that AGI is as near or inevitable as people make it out to be. I don’t necessarily think models will simply get better with more data and I don’t think it’s reasonable to expect that we’ll just somehow figure out intelligence. When I think about AI’s bitter lesson, that nothing has really worked better than scale and data, I take that as a sign that we really don’t understand intelligence. I don’t take it as a sign that “scale is all you need”. Maybe to a certain point, but for AGI? I don’t know man.

Of course, suppose I’m wrong and we create AI that’s smarter than your typical software developer and costs the equivalent of pennies an hour. In which case, well, game over, but I don’t think I’ll feel like much of a loser. Sure, there goes my job, but the economy also probably gets wrecked in the process for a variety of factors. Massive unemployment aside, there’s also the fact that paying for software is kind of stupid when you can just pay a competent AI to make it for you or use some open source alternative made by people pooling AI resources together. This is kind of a big deal considering that tech companies basically carry the stock market. I haven’t even touched on the damage done to the rest of the economy by the existence of intelligence-on-demand for cheap. Either the government considers drastic measures and a radical rethinking of economics, or they ride out the worst economic crisis in history and try to figure out which holy cow gets cut from the government budget now that tax revenue is in free fall. Despite how little faith I have in the government, I trust their self-preservation instincts will kick in and they’ll resort to taking certain socialist economic policies more seriously.

I like my little green squares

Don't get me wrong, I kind of fear what laws and legislation a tech voting bloc would pass, but I'm also optimistic that we could push forward some genuinely progressive legislation as well. Maybe we could pass laws that encourage more housing development. Perhaps some bureaucratic reforms similar to the ones that occurred in Estonia, the kind that result in major cost reductions without any loss in quality of service. Universal healthcare? Those slightly strange policy ideas we like to talk about that are clearly progressive, even "socialist", but are also palatable to libertarians?

Mixed feelings about lobby for open source AI, but nonetheless, glad it happened. Hope to see more of it.

It's unfortunate that on the day of his death all I can really think about is how troubled the man was, between dealing with alcoholism, cheating on his past wives / gfs, the sexual assault allegations, and resigning from his company because of just how negative his public image had become. He became something of a household name because of a misleading documentary that was arguably more about the dangers of alcoholism rather than fast food.

Strange processing the death of someone with such a mixed legacy.

Sometimes the AI misinterprets what you want. Sometimes this is due to lack of sophistication on the model's part, but I think the bigger issue at hand are the shortcomings of natural language. Even when working with an intelligent human being, sometimes you have to communicate your thoughts/ideas through pictographic means.

It's something I've thought a lot about lately. No-code/low-code is about as old as GUIs themselves, yet it's never become the de-facto way of creating software. However, for everything else we do we use a GUI. No one creates illustrations on the computer programmatically (unless you're into generative art like me). You probably aren't doing your taxes in a terminal. You get the idea.

This is a segue into my other critique of prompting as a UI. A lot of UX people think it is the ultimate UI, and natural language the premier way of communicating jobs to be done. I disagree. I think symbols and diagrams communicate more with less. Concerning language, I think that for whatever reason programming languages are the most natural way (that we've currently found) to communicate intent in a programmatic context, markup languages are the most natural way to communicate intent in a layout context, and that's the reason why despite 50+ years of GUI innovation programmers still use text-editors. The old guard has all aged out of the profession, the new guard still uses text-editors.

So that all being said, I think the most incredible AI demo I've seen so far is the tldraw stuff done by Steve Ruiz, namely Make Real, and that's because you're communicating with the AI in a pictographic / symbolic context, and because it's a whiteboard, there's no limit to the symbols at your disposal to communicate your intent. The limit is AI's ability to interpret what those symbols mean.

For example, AI website builders. It's impressive as a tech demo that you can prompt an AI to make a website for you, but otherwise I don't think this is going to be disruptive to the web design industry since it's already pretty easy to make a website using a drag-and-drop builder like Squarespace or Webflow. Granted, it can get complex when you add certain special features on these platforms, but same goes for an AI-generated website. AI doesn't eliminate the complexity, it just hides it. At best it'll get you maybe 80% of where you want to go and you can get to the remaining 20% by hand. At worst, the complexity is so thoroughly hidden that addressing it isn't an option. It simply does not allow you to manually edit the outputs, or makes it unreasonably difficult to do so.

In that same vein, an AI game maker is pretty impressive as a tech demo, but I feel like it isn't a huge improvement over a code-light game maker. You're not eliminating the complexity of making a game, just hiding it.

Nonetheless, it's a pretty interesting tech demo.

But if it were related, then that would presumably be because people within the company (or at least two rather noteworthy people) no longer believe that OpenAI is acting in the best interests of humanity.

Which isn't too shocking really given that a decent chunk of us feel the same way, but then again, we're just nobodies making dumb comments on Hacker News. It's a little different when someone like Ilya really doesn't want to be at OpenAI.

https://github.blog/2022-09-07-research-quantifying-github-c...

That's where the 55% number is coming from. It's coming from this experiment, and only this experiment.

So yeah, if you're wondering why you aren't somehow 50% more productive when using Github Copilot, it's probably because you're not implementing a simple task that's been done to death in countless tutorials.