PheonixPharts

Essentially a funnel report where you want to know the total percent of the population that has reached a given path, but you only know the output probabilities of each step in the funnel (node). This is a fairly common situation.

As a simple example: you know after signup 20% of customers purchase, 80% don't, but what you want to trivially add in is the fact that of the users in a marketing campaign, 10% of them signed up, which means for the marketing funnel 2% purchase. Now consider that you have 20 or more such events in your funnel and you want to combine them all with out doing all the math by hand. Likewise you want to be able to add a newly discovered step in the funnel at will.

Using a topological sort you can take an arbitrary collection of nodes where each node only knows that probability the next nodes are, sort them and then compute the conditional probabilities for any user reaching a specific node fairly trivially, given the assumption that your funnel does represent a DAG.

If you don't perform the topological sort then you can't know you have calculated all the conditional probabilities for the upstream nodes, which makes the computation much more complicated. Topological sort is very useful any time you have an implied DAG and you don't want to have to worry about manually connecting the nodes in that DAG.

> a quote of himself

Curious, did you seriously not recognize that this is a famous quote from the final scene in Blader Runner?

I read that more as quirky call out to a famous film, not that he was claiming this was his own view of the world.

> Asking me to solve an extremely esoteric problem that has zero relevance to my day-to-day

I'm always surprised how useless something is when I don't know it, and suddenly once I do know it, I solve lots of problems with it!

I've heard programmers grumble about how useless calculus is, before I learned calc I used to grumble about that too. After I learned it there were countless problems I unlocked solutions for by applying the thinking I learned in calculus.

I've heard programmers say that you'll never need to implement your own sort for mundane tasks, but, it turns out that after really grokking topological sort I used it countless times for fairly mundane problems like creating plots.

I've heard programmers say that learning the lambda calculus is a waste of time, and nobody uses functional programming. Yet it was people that understood these things that transformed Javascript from a useless browser oddity into one of the most widely used languages. It was seeing that Javascript was essentially a Scheme that unlocked it's true potential.

Over my career it's remarkable how many "esoteric problems" have lead to me solving hard tasks or even shipping entirely new products. If you're only focused on what is required of your day job today you're only going to be at best a mediocre engineer.

> It really sucks right now

For me, it's been the opposite: the last 2 years have been the best time I've had working in tech since the early 2010s.

Around 2019 I was seriously considering leaving the field (if it didn't pay so much) as the entire industry had turned into a bunch of leet code grinding, TC chasing, mediocre drones. It was incredibly hard to find people working on actual problems let alone challenging/interesting ones. Nobody I worked with for years cared one bit about programming or computer science. Nobody learned anything for fun, nobody hacked on personal projects during the weekend, and if they were interest in their field it was only so they could add a few more bullet points to their resume.

But the last two years I've worked with several teams doing really cool work, found teams that are entirely made up of scrappy, smart people. Starting building projects using a range of new tricks and techniques (mostly around AI).

Right now there are so many small teams working on hard problems getting funding. So many interesting, talented and down right weird programmers are being sought after again. People who like to create things and solve problems are the ones getting work again (my experience was these people were just labeled as trouble makers before).

I'm probably getting, inflation adjusted, paid the least that I have in a long time, but finally work is enjoyable again. I get to hack on things with other people who are obsessed with hacking on things.

This can be solved by just running "restart and re-run all" as a matter of habit whenever you feel like you are in a good state. If you're looking at someone else's notebook, this should also be your first step.

If you're using Python, Jupyter still offers a major advantage for sharing prototyping/research work in that it it is the only real way to provide literate programming. There are cases where code inside of a notebook should be pulled into a library, but pure code will always be inferior when most of your work is expository and exploratory.

Literate Haskell, and R Markdown are superior forms of literate programming, but Jupyter is not terrible if you just take some basic precautions before sharing.

Notebooks work great as... notebooks. Which is primarily how they're used in my experience. When the bulk of your work is prototyping and and exploring an idea, they're still an excellent choice.

Many people misunderstand what this case was about.

> Homeless need shelter and help. It is inhumane that we let them rot on the streets.

The issue being argued was not that homeless should be able to camp where ever and whenever they want. This issue was that if you don't provide shelter then you cannot kick them out of public spaces.

If that was the law, then it would mean you need to build shelters if you don't want homeless in the streets.

What this new verdict means is:

- You can forcibly remove homeless people who have nowhere else to go

- Thus homelessness can be effectively illegal

- The only realistic solutions now are:

    - Put the homeless in prisons

    - Move them to cities and towns that don't have the resources to remove them.

> Rather than educating founders on more fundamental topics such as how to get from 0-1, how to hire, fostering positive culture etc.

My experience is that none of these things actually matter anymore. I wish they did, but my experience has been that getting any of these things right in today's environment is in the best case inconsequential to success and in the worse an actual determent.

I've worked at horrifically toxic startup cultures, but it never seemed to hurt them getting funding from outside teams that didn't care in the slightest how healthy the org was or even the product was even remotely good for customers.

I've worked at companies that literally did not know how they were every going to make a profit IPO, then continue to fail to figure this out and changed their strategy to shrug-your-shoulder-and-wait-for-the-return-of-zirp, with no observable consequences on their stock price or investor support for years. Maybe this will eventually catch up to them, but so far they've been quite successful burning a dollar to make $0.70.

I've worked at companies that hired hundreds of completely useless data scientist, filled management with toxic management consultants, driven out all the serious talent they had... with again, no real consequences.

And all of the companies I've known that checked all of your bullet points? Well they've mostly stayed small, saw revenue decline, and eventually start falling apart. All of the best companies I've worked for ended up having to get rid of all the things that made them good in order to appease investors and keep growing.

I wish that we lived in a world where your advice was correct, but I haven't seen any evidence that that is the world we actually inhabit.

> no revenue and no profit.

Revenue has gone up, but fewer publicly traded companies are making a profit than ever before [0].

I could never understand why so many people just talk about revenue. Revenue without profits is meaningless. There's the old logic of "get enough revenue and then figure out profits and you're highly profitable", but it's very clear that switching the "profit switch" is not so easy in practice.

Investors are still basically waiting for the fed to drop rates, which means that people have abandoned rationally thinking about businesses and are just holding until the free money starts pouring in again.

I honestly don't think the AI bubble is anything like the dotcom bubble. There's something much stranger happening here since the entire market is basically hallucinating and AI is just one manifestation of that.

0. https://finimize.com/content/beware-the-rise-of-unprofitable...

> compensation packages that truly gifted AI researchers can make now

I guess it depends on your definition of "truly gifted" but, working in this space, I've found that there is very little correlation between comp and quality of AI research. There's absolutely some brilliant people working for big names and making serious money, there's also plenty of really talented people working for smaller startups doing incredible work but getting paid less, academics making very little, and even the occasional "hobbyist" making nothing and churning out great work while hiding behind an anime girl avatar.

OpenAI clearly has some talented people, but there's also a bunch of the typical "TC optimization" crowd in there these days. The fact that so many were willing to resign with sama if necessary appears largely because they were more concerned with losing their nice compensation packages than any of their obsession with doing top tier research.

> things like chain-of-thought/reasoning perform way worse in structured responses

That is fairly well establish to be not true.

> Why do I want to read an intro CS book? I've read those before.

"We shall not cease from exploration, and the end of all our exploring will be to arrive where we started and know the place for the first time." - TS Eliot.

You aren't a curious person who studies things for their own sake and finds wonder in exploring ideas. That's fine, this book is clearly not for you so why concern yourself with it? Most of my friends who are obsessed with their field count many intro books as their favorites and frequent re-reads.

> condescending promises of enlightenment.

It sounds like you're more upset that other people enjoy and therefore recommend this book. You're the one asking for proof that it's worth your time. It's clearly not. People who are curious and like to explore ideas in computing recommend this book frequently to other like minded people.

If you don't like sushi, why question somebody's recommendation for their favorite omakase place?

Have you never been to a book store?

You don't have to read every word of a book to understand if it's interesting to you. I purchased a bunch of technical books the other day that I had never heard of based on opening them up, reading a bit of the intro and flipping through the examples.

Relatively few of my favorite books have come through recommendations compared to those that I have come across through serendipitous discovery.

For anyone who has no time to browse books, then most of the best books in existence would be of little interest to that person.

"Don't cast your pearls before swine".

When I was younger I used to passionately defend those things I've seen as beautiful, but after years experience talking with people passionate about their fields and learning and those who never will be: If you lack the innate curiosity to explore those things others have declared marvelous, then this book will offer you no value.

Every time I crack this book open I get excited and I've read it multiple time and one most of the exercises. I can think of few other books that really expose the beauty and simultaneously strong engineering foundations of software.

You have "tons of experience programming" and sound like you've already decided you know what needs to be known (otherwise why even ask rather than just read it free online), I doubt this will offer you anything you haven't already seen before.

I don't think it's that fake. There were people, researchers, that were obsessed with ELIZA back in the day. I personally have a friend who is constantly thinking about fine-tuning a model on his ex's texts to recreate chatting with her. I suspect many others are less extreme but still seek out a form of companionship they have full control over.

In the visual space, if we're honest, most of the work being done on Stable Diffusion is with the aim of creating "intimate" art. Might not be "companionship" in the family sense, but absolutely a good chunk of the Civit.ai world is, more-or-less, concerned with the mass creation of "waifus".

I don't think that we're going to have many startups succeed in this space, because I suspect the main driver is to have conversations/generate images that people don't entirely want there to be a record off/watched by a private company. But I actually suspect that the number of people using AI for companionship of some form or another is much higher than it appears.

Boston at number 6?

I find this list hard to take too seriously. Despite the city's incredible university culture it's always been fundamentally anti-startup from it's very nature.

Boston is a city were credentials matter far more than ideas, and even then there's a strong, strong culture of "you aren't special, nobody is". People are skeptical of new ideas, compensation for tech has always been abysmal, and people tend to be surprisingly risk adverse.

I've lived there a few times, and there are parts of the city I love, but nobody seriously interested in anything disruptive is going to be hanging in that area too long after graduation.

I use VScode every day on linux and have never had an issue with it.

I want to second this regarding desktop Linux.

Each major PC build I've had I've tried to go Linux desktop and ultimately failed because I just got tired of switching to windows for gaming along with periodic cases of essential devices/apps just not working right.

I tried a final time a few years back. I even made sure I had a separate drive just for the Windows partition I felt was inevitable... turns out it never happened.

Gaming on desktop linux is phenomenal now, and I haven't run into any issues with some device or essential app not working. There is no personal task that I ever have to use an old macbook for and it never made sense to add that Windows partition. Occasionally I have to get more fiddly with driver settings to get things to work just right, but not enough for it to be a major annoyance.

Overall if you're tired of your personal OS spying on you increasingly more often and slowly turning into an advertising machine, the linux desktop is ready and quite enjoyable!

I just use a single monitor/mouse/keyboard/mic/camera and then have a usb switch they all plug into so I can easily swap between work and personal. It make switching between workstations just a few seconds and with a bit of wire management can be quite clean.

> hard not to regret the choice

If you can't handle "regret" in these cases, then you probably shouldn't be in a position where you're deriving the vast majority of your income/weatlh from investments (which is fundamentally what a CEO does).

It's astounding how many ICs can't wrap their heads around the concept that holding onto your RSUs make absolutely no financial sense. With rare exceptions, this doesn't make sense for anyone. And yet, fear for "regret" keeps people holding.

But it's not shocking that even in tech many ICs are not good at reasoning financially. But if you want to be a co-founder, and hold a lot of your wealth in investments it's essentially that you learn to reason, plan and accept outcomes accordingly. Otherwise you're more-or-less a professional gambler.

> Now, is it 10k examples? No, but I think it was on the order of hundreds, if not thousands.

I have kids so I'm presuming I'm allowed to have an opinion here.

This is ignoring the fact that babies are not just learning labels, they're learning the whole of language, motion planning, sensory processing, etc.

Once they have the basics down concept acquisition time shrinks rapidly and kids can easily learn their new favorite animal in as little as a single example.

Compare this to LLMs which can one-shot certain tasks, but only if they have essentially already memorized enough information to know about that task. It gives the illusion that these models are learning like children do, when in reality they are not even entirely capable of learning novel concepts.

Beyond just learning a new animal, humans are able to learn entirely new systems of reasoning in surprisingly few examples (though it does take quite a bit of time to process them). How many homework questions did your entire calc 1 class have? I'm guessing less than 100 and (hopefully) you successfully learned differential calculus.

Coming up with and quickly adopting new terms to sound "hip" is one of the most important skills for AI practitioners. We've had "agent-based" concepts in CS for decades, but if you're "in" you'll of course refer to "agentic" workflows and the like.

It make sense to come up with terms to describe common patterns: Chain-of-Thought, RAG etc. are good examples of this. But the passion some members of this community have for being intentionally confusing is tiresome.

> but no-one in the comments here is even mentioning it

The post has an entire section discussing this.

The problem with text-to-sql is that, as the post elaborates, writing SQL is not the problem. It's understanding the context and the data:

> On the other hand, a technical person would notice that the question doesn't make sense, and they would ask for more context. They would ask for details about the business person's hypothesis and the problem at hand. Then, they would explain what type of data is available, and work with the business person to formulate a precise and useful question.

Text-to-sql in practice is a solution that nobody was asking for, despite the insane number of SV startups shipping GPT-text-to-sql wrappers as products.

There certainly is like places where LLMs can help (post touches on this briefly), and that is in semantically exploring databases/tables/etc and contexts around data, but this is a very different project and would require a lot of curation from data teams to make it happen.

> I think that posting at least one commit per day

Notice that the article has been edited.

> Edit: A better title would've been "commit every day that you work". I don't mean you should work on weekends or not take time off, and whatever you work on doesn't need to "ship to prod".

I'm pretty sure parent comment was replying to the original version which, at quick skim, gives the impression that the article was about pushing to prod everyday.

I've worked at startups to do push to pod every day at it's an absolute frenetic nightmare.

Posting a commit everyday is so uncontroversial that I doubt the headline would have made it to the FP of HN. I've also never heard the term "shipping" in software refer to anything other than getting something into prod.

I would agree except the gap between what's realistically cool about LLMs (make tons of traditional NLP tasks easier/better) and what people are hyping them up to be (potentially world ending sentient beings) is so tremendously large it essentially guarantees an AI winter despite the fact LLMs represent a major advancement in practical SotA.

It also does concern me how basically nobody is building a real product around the current state of "AI", but are rather hinging their success on what they believe it will be in the near future.

Is this written from a different time line were TypeScript doesn't exist?

I have a long history of doing FP in a bunch of the traditional FP languages (Scheme/Haskell/etc). Ended up working at a young startup with equally young employees awhile back and was sort of surprised to see that the long heralded fantasy of cranky FP enthusiasts has, for better or worse, come true in Type Script.

I saw a generation of new programmers truly doing type-driven development and using a range of functional programming techniques without them even realizing that this was a big thing. Funnily enough they were terrified of anything resembling object oriented programming. If the word "class" appeared in the code base (their was some Python), they would quietly walk away.

What I saw was also incredible because these, largely junior, programmers were using types for exactly the thing old-Haskell people had hoped they would. Things would ship to prod lightning fast as all engineers had to do was make sure the types all lined up and the compiler was happy and they would send the PR off!

But then I saw the downside: these type-happy programmers almost never tested anything. I'm not talking about formal unit tests or integration tests, some of those existed. I mean these were programmers that had entirely lost (or maybe never had) the ability to play around with the code that they just wrote and make sure it worked. I kid you not, "print debugging" was viewed as some advanced technique to do in emergency situations. When bugs were introduced to prod I would ask "did this work when you ran it locally?" only to be met with quizzical stares. If it pasted the type check it was good to go.

It also had the negative consequence of inadvertently discouraging abstraction in favor of just adding more complexity to your existing types. Because refactoring code takes time and the entire point of the type check doing all your thinking is to ship fast.

All that said, I'd much rather work with that code base than one made by a similarly proficient team in say Ruby.

There are tons of skilled (and less skilled) functional programmers out there. The real problem is if you're looking for both highly competent programmers who also happened to be hung up on a particular niche language. I'm sure there are plenty of excellent "Gooby" engineers out there, but they likely choose jobs based on other factors than the language being used.

As mentioned repeatedly, HN doesn't block self-promotion.

But reddit something else entirely. It so clear that all of the major subs are constantly being manipulated by firms working PR for large companies, but the second someone posts "I made this!" people get up-in-arms.

Of course who really knows how much of reddit is even real people anymore.

To be clear, as an ex-smoker, I do believe we have agency in the cases where patterns are disruptive. Smoking tobacco got in the way of a range of activities, and I had to put in a serious effort to curb this behavior. Certainly drinkers who find their drinking interferes with other things are able to change their habits. Though even this is probably more environmental than not. I haven't smoked in 20+ years but I also no longer know any smokers. I'm not sure I would be a non-smoker today if smoking rates were closer to what they were in the 1950s. Similarly I have known people with problematic drinking behavior and their ability to stop has always been strongly correlated with having good reasons to stop.

However, for the smaller things that "aren't good for you" in a less immediate sense, I don't think we have as much control over our behaviors as we'd like to believe.

Another example is obesity. Many people still chalk this up to a "moral issue" where people are making "poor choices", but that doesn't seem like a good explanation for why we live in an obesity epidemic. I personally don't think people in 2024 are dramatically less "moral" than they were in 1990.

My personal pandemic realization was that I'm far more of a node in a network of cells in a vast social organism that is humanity than I am an individual actor.

In my lifetime virtually everything that was once good for you (wine, high grain diet, non-fat foods) is now a bad and vice versa (eggs, high fat - low carb foods).

And articles like this one have been appearing in some variation my entire life. My take away, long ago, is that nutrition is fundamentally complex and poorly understood topic and any extreme opinions are likely to be inverted.

On the topic of alcohol, one things that has really become clear to me, is how directly tied to my environment drinking is. I've always liked to have a beer with dinner, but whether or not that was my only drink or one of many has much less to do with my personal decisions and much more to do with my environment, and I've noticed the same goes for most people.

Many of us became pretty serious drinkers during the pandemic. As it eased up I never made the decision to drink less, I just naturally drank less.

Point being is that no only am I skeptical of the claims of what I should and should not consume, I'm skeptical of entirely how much agency I have to change what I should consume baring case where the impact is immediate.

People keep saying it because that's literally how LLMs work. They run Montecarlo sampling over a very impressive latent linguistic space. These models are not fundamentally different than the Markov chains of yore except that these latent representations are incredibly powerful.

We haven't even started to approach the largest problem which is moving beyond what is essentially a greedy token level search of this linguistic space. That is, we can't really pick an output that maximized the likelihood of the entire sequence, rather we're simply maximizing the likelihood of each part of the sequence.

LLMs are not reasoning machines. They are basically semantic compression machines with a build in search feature.

>_because that's the only way it can compute anything_

I'm fairly certain we'll soon realize that what's happening here is that the markov chain being run over latent space needs a certain amount of "warmup" before it starts sampling from the optimal region. HMC samplers for Bayesian methods have this same property.

The terms "reasoning", "computing" or "thinking" for this stage should be considered metaphors rather than explanations for what's happening, which is really waiting for a random walk to start sampling from the typical-set.