You can’t predict a coin flip because it is random. However, we have an accurate understanding of the random process producing coin flips and therefore, we can make accurate predictions about large quantities of flips.
Weather may or may not be random. It could be entirely deterministic for all we know. However, we lack the ability to fully model all the factors that contribute to weather and therefore our predictions are inaccurate.
Now let’s consider long term climate predications. Do you think these predictions are more like coin flips, where we have an extremely accurate model of the process, or more like weather, where unknown unknowns have outsized impact on accuracy?
That’s not to say climate change isn’t real, but your analogy doesn’t make sense.
All responses are so focused on exact predictions. We have high certainty that 50% of flips will be tails over long enough timespan. We don't know what any single flip will be. Climate science works the same way. But climate is not a coin, let's say it's a multisided die and it appears the sides are changing sizes as we compare data year over year.
Our models of weather are so accurate that literally trillions of dollars per year bank on them in the agriculture sector, the shipping sector, and everywhere else. Similarly, our models of climate change have been refined and refined, and now are essentially irrefutable.
Our “models of climate change” have regularly been falsified at this point. It is absolutely unknown how much “climate change” is attributable to humans right now.
Nor is it actually known what the net favorability of mild warming might be…including the possibility of mitigating the next Ice Age!
Predictive models are not the same as historic data analysis and trend fitting.
Flipping coins: no predictive models, very definitive statistics
Weather: +/- 2 week predictive models, 100 years of measurements getting more definitive each year where trend are headed
Hypothetically. None of what you said is testable, and there is no evidence that models today are any better than before except at matching historic data.
Compare with another topic like, say, evolution. Here outcomes are testable and verifiable because we can observe the theory at work by watching micro-ecosystems, or small animals with fast reproductive cycles.
Meteorology is short term accurate based on a linear regression of data points from historical data. Deviation like "warming" or "cooling" are relative descriptions of how closely aligned one theory is to the line, and how far back the specific model goes along with the number and quality of relevant factors you want to look at.
No matter which model you go with, you're proving the accuracy of a math function at matching historical data, and then hoping that it will match the future. And as we know, none of them match the future very accurately, which tells us there's something wrong with the theory.
This is only slightly better than day trading in the stock market. And much like the stock market, everyone thinks they know better than everyone else but statistically, most fund managers and professional stock callers underperform the market. They earn by selling you on the idea that they have the next model that finally DOES make accurate predictions. They tell you that they know that because this new model matches the historical data more accurately. No shit. Because there's more data now in a growing set of data. So the most recently calculated linear regression is the most accurate.
But we don't know how it works. That's the key here. More data, doesn't mean the theory is better. More accuracy in making predictions about the future, on the other hand, is a strong indicator, and maybe the only indicator, that something is worth believing in. That is to say, it's more likely to be true.
Making overzealous claims about how much we know is not science, it's ignorance. Let's help interest people in science by being cautious about what we claim to know for sure. At least don't claim to know the next 200 years, until we can at least make accurate predictions beyond the next few days.
I majored in biochem with a lot of extra classes I took for fun on environmental chemistry. You?
> more like weather, where unknown unknowns have outsized impact on accuracy
"Unknown unknowns" aren't the reason weather forecasts are inaccurate.
Weather is path-dependent. Small changes to starting conditions or minor differences between modeled and actual conditions shortly after the simulation begins lead to large differences by the end of the simulation. Errors propagate and magnify.
Isn’t a better question: why would they migrate off COBOL? Their business is working. What’s the impetus to change? It’s not like they need to use COBOL for every new project.
I disagree. The purpose of writing is to convey ideas. If written language had just been invented, I’m sure you’d be saying “IMO any important stories you expect others to know should be communicated orally. It’s kind of disrespectful to convey stories as if they were hearing you speak.”
Your conflating the medium with the source of the message in this analogy.
Writing and oration are two different media, the question of which is preferable in which context is completely unrelated to LLM authorship.
If "the purpose of writing is to convey ideas" (which I largely agree with), which ideas are being added between whatever you prompt the model to convey and what the model conveys? Are you proposing that an LLM can extract some meaning from your initial prompt that a human being couldn't?
> Are you proposing that an LLM can extract some meaning from your initial prompt that a human being couldn't?
No, I’m asserting that an LLM can help formulate ideas in a coherent, understandable way. You can give it a brain dump with a rough outline for an argument and it fill in the details. If the argument isn’t to your liking, you can try again. But the end result is basically equivalent to the human-written equivalent.
It’s fine to disagree with my analogy. But I find it a little ironic your dismissal of the analogy is a non sequitur. The invalidness of the analogy doesn’t directly follow, logically, from the fact that the written word and the spoken word are both mediums.
Your friends don’t produce much content yet people had a need for frequent entertainment. Also, people realized that posting things to social media meant that it was there forever. This led to a bifurcation: friends / family updates are mostly relegated to temporary formats like stories while “feed” content is professional produced.
I don't know, I mean for most SaaS products this is true. But for something like Salesforce, the feature set is incredibly broad. The coding is not hard, so much as it is just an enormous volume of code.
It was never the only hard part, but it definitely was a hard part (at least in most cases; obviously there are some monopolies with relatively simple software - mostly where there are network effects like WhatsApp).
But give me the source code for something competitive with Solidworks, Jasper Gold, FL Studio, After Effects, etc. and I'm sure as hell making a business out of it!
Furthermore while good software may not guarantee business success, it is pretty much a requirement. I have seen many projects fail because the software turned out to be the hard part.
Yeah, but its still usually cheaper to pay for software than build and support it. I think that will be true for a long time going forward, its just that you can't plan on extracting a ransom for your SAAS.
We're not going to see the end of software; we're going to see the end of margins.
I don't know for sure, but I suspect that other industries have experienced this. Would love to know which. Photography comes to mind, but I'm sure there are more meaningful examples.
But a query optimizer only matters once you have an established business with large customers.
You seem to be implying Salesforce’s business is successful because they have their own query optimizer. But the causality is reversed. Salesforce has their own query optimizer because they’ve built a successful business.
My point is that a lot of people think it'd be really easy to build the next Salesforce until they actually try to compete with Salesforce in the market. Like it or not, if you want to build a Salesforce competitor (or try to get your company to build its own) you're going to be compared to actual Salesforce, not the version of Salesforce that existed when the market was new.
Many of the built-in types in Objective C all have names beginning with “NS” like “NSString”. The NS stands for NeXTSTEP. I always found it insane that so many years later, every iPhone on Earth was running software written in a language released in the 80s. It’s definitely a weird language, but really quite pleasant once you get used to it, especially compared to other languages from the same time period. It’s truly remarkable they made something with such staying power.
>It’s truly remarkable they made something with such staying power
What has had the staying power is the API because that API is for an operating system that has had that staying power. As you hint, the macOS of today is simply the evolution of NeXTSTEP (released in 1989). And iOS is just a light version of it.
But 1989 is not all that remarkable. The Linux API (POSIX) was introduced in 1988 but started in 1984 and based on an API that emerged in the 70s. And the Windows API goes back to 1985. Apple is the newest API of the three.
As far as languages go, the Ladybird team is abandoning Swift to stick with C++ which was released back in 1979. And of course C++ is just an evolution of C which goes back to 1972 and which almost all of Linux is still written in.
And what is Ladybird even? It is an HTML interpretter. HTML was introduced in 1993. Guess what operating system HTML and the first web browser was created on. That is right...NeXTSTEP.
In some ways ObjC’s and the NEXTSTEP API’s staying power is more impressive because they survived the failure of their relatively small patron organization. POSIX and C++ were developed at and supported by tech titans - the 1970s and 1980s equivalents of FAANG. Meanwhile back at the turn of the century we had all witnessed the demise of NeXT and many of us were anticipating the demise of Apple, and there was no particularly strong reason to believe that a union of the two would fare any better, let alone grow to become one of the A’s in FAANG.
I actually suspect that ObjC and the NeXT APIs played a big part in that success. I know they’ve fallen out of favor now, and for reasons I have to assume are good. But back in the early 2000s, the difference in how quickly I could develop a good GUI for OS X compared to what I was used to on Windows and GNOME was life changing. It attracted a bunch of developers to the platform, not just me, which spurred an accumulation of applications with noticeably better UX that, in turn, helped fuel Apple’s consumer sentiment revival.
Good take. Even back in the 1990s, OpenStep was thought to be the best way to develop a Windows app. But NeXT charged per-seat licenses, so it didn't get much use outside of Wall Street or other places where Jobs would personally show up. And of course something like iPhone is easier when they already had a UI framework and an IDE and etc.
Assuming you mean C (C++ is an 80s child), that’s trivially true because devices with an ObjC SDK are a strict subset of devices that are running on C.
Yes, that is why I don't find it "insane" like the grandparent does, like yeah, devices run old languages because those languages work well for their intended purpose.
You should feel that C’s longevity is insane. How many languages have come and gone in the meantime? C is truly an impressive language that profoundly moved humanity forward. If that’s not insane (used colloquially) to you, then what is?
When you frame it that way, it’s really not that different. The issue isn’t the access control system itself, more so that it’s really asking too much of people who don’t have the skills or understanding to manage it. Teams of trained professionals get it wrong when the scope is limited to a single application or suite of applications, and you think grandma is going to properly manage access control over her entire digital footprint?
Well, its maintained by humans to start with, peer reviewed by humans. They fuck up from time to time in extremely limited scope, depending on how much given company is willing to invest into getting quality work, but nothing like this. Humans are clearly not the weak link to be automated away, in contrary.
I work in one of the special legal jurisdictions, such fubar would normally mean banning such product from company for good. Its micro$oft so unfortunately not possible yet, but oh boy are they digging their grave with such public incompetence, with horrible handling of the situation on top of that. For many companies, this is top priority right behind assuring enough cash flow, not some marginal regulatory topic. Dumb greedy amateurs.
It’s worth asking: what do Wall Street traders know about building software companies? Almost nothing. Anyone who has attempted to start a startup knows that the software is always the easy part. Building the business is hard. The notion that we’re going to undo 100K+ years of specialization just so that companies can run mediocre, buggy versions of SaaS tools just to save a few bucks is crazy to me.
SaaS stocks are currently the buying opportunity of a lifetime.
Also, has nobody thought about operations? I pay SaaS companies so I don't need to think about operations. It is insane to me to think that people think a restaurant company is going to suddenly want to get into managing the operations of its vibe coded accounting app. Absolutely not.
Yes, I’m very aligned with “minimal dependencies, live off the land, roll your own” to avoid lock-in and vendor bullshit. Shit I’d rather rent baremetal from Hetzner than VPS from EC2/GCP/etc if I have enough workload to justify it.
But for my startup I still use a ton of SaaS services for things that I could probably do just fine myself. (Clerk/StackAuth, Supabase/PlanetScale, Cloudflare STUN/TURN, Clickhouse, Vercel, Calendly, Google Workspace, ngrok, Tailscale).
Spiritually, I hate using these. Any one of these would be dead-simple to replace. But my time is genuinely better spent on my startup’s particular value-add. Maybe I’ll replace these some day when we can hire someone to manage internal replacement services - some of which are as easy as “a postgres database” or “wireguard on some VPS instances”. But it’s just not worth my time right now when I’m focused on building revenue.
Even if they all cost $300/mo in total, and we’re bootstrapped, it’s a lot easier to cut back on UberEats or shiny nerdy toys than it is to replace all of these SaaS offerings. I recognize there’s a lot of ”I don’t know what I don’t know” and I’m liable to subtly misconfigure something in a potentially disastrous way.
There's enough unhappiness with commercial SaaS EHRs that I expect as few health system CIOs will decide to operate their own vibe coded replacements. This won't work out well for most of them but I think it's going to happen.
We spent the last 20 years building a whole set of specialised little tools on the web to do specialised little tasks for businesses.
Now an LLM can do most of that, easier, and effectively for free. The first pass on a new problem is not googling to see if there's a SaaS for that, it's prompting an LLM to see if it can do it, or if it can build a tool that can do it.
Case in point: in my job we have to data-enter invoices. I have dealt with or in this industry for 30-ish years. I worked on various projects trying to get computers to read invoices, to various degrees of success. It's a hard problem; there's no standard format, or layout. Every company does its invoices differently. Some are Excel files, some are PDFs, some are Word docs, etc.
This entire problem vanished this year. You get an LLM to read the invoice. It does this more accurately than humans do. Job done.
There are entire SaaS businesses that read invoices that are now obsolete and have no moat.
I think that’s different. You have a problem: invoice management. LLMs have made that cheaper and you should expect disruption. But you’re not building your own invoice scanner. You’re using another, cheaper product on the market.
However, the hypothesis in the SaaS market is that LLMs have made software have zero value and therefore the SaaS companies will be less profitable. That’s like if wood was suddenly free, expecting home builders to go out of business. If anything, home builders are going to do better, because they can apply their expertise while deploying capital elsewhere. We should expect software companies to be more profitable, not less.
Of course, there are exceptions. Sometimes AI replaces the product itself, e.g. image generation models vs. contractors on fiverr.
There has been a standard X12 EDI format for invoices for decades. It's kind of a hassle to work with but it can at least be reliably parsed. A lot of huge businesses like Walmart use it successfully, and even require their suppliers to submit all invoices that way.
I don't object to using LLMs to parse PDFs but over the long run it's going to be less efficient and reliable than other options.
Yes, there has been a standard format for invoices for decades, but it was only ever used if both companies were using a ERM system (and as you say, large enough purchasers could force their suppliers to). We have to deal with small business who don't use the standard format, which is the vast majority of them.
Please go ahead and try parsing non-standard invoices without an LLM. I spent 20+ years on and off dealing with this problem. It's not as simple as it looks. And then LLMs came along and made it simple.
Anecdotally, my partner does experience design for fintech systems akin to Bloomberg for a particular niche. They do the standard lifecycle of user research, design iteration in figma, handoffs to devs and all that. The tech department at this company has also been building OpenAI integrations for more than 2 years and are neck deep in LLM technology doing exactly what this thread describes. My partner is still doing exactly the same work she's been doing the entire time and getting the same level of adoption for all the bespoke UI development while the chat interface is just kinda there. I'm sure it's getting usage for some tasks, but it's supplemental.
You look at what Claude’s doing to make sure it doesn’t go off the rails? Personally, I either move on to another ask in parallel or just read my phone. Trying to catch things by manually looking at its output doesn’t seem like a recipe for success.
I often keep an eye on my running agents, and occasionally feel the need to correct them or to give them a bit more info because I see them sometimes diverge into areas I don't want them to go. Because they might spend much time and energy on something I already know is not gonna work.
yeah I watch it but not like staring at every line - more like checking when something feels off or when it's been chugging for a bit. if it starts pulling in files I didn't expect or the diff looks weirdly large I'll stop and check what's happening. the problem isn't that you can't catch it after the fact, it's that by then it's already burned 5 minutes reading your entire utils folder or whatever, and you gotta context switch to figure out what it misunderstood. easier to spot the divergence early
I don't think this is necessarily a fair comparison. In your sample of David Greene, he's being _interviewed_, which is different than hosting a radio show or podcast. For instance, turn on the nightly news and listen to the very bizarre intonation used by the newscasters. This is something they do for the broadcast, it's not how they normally talk.
Weather may or may not be random. It could be entirely deterministic for all we know. However, we lack the ability to fully model all the factors that contribute to weather and therefore our predictions are inaccurate.
Now let’s consider long term climate predications. Do you think these predictions are more like coin flips, where we have an extremely accurate model of the process, or more like weather, where unknown unknowns have outsized impact on accuracy?
That’s not to say climate change isn’t real, but your analogy doesn’t make sense.
reply