Denzel

> As far as I can tell, chatgpt performs the action of finding and simplifying the information needed to take the next step forward [...] But that's all it does at the moment.

ChatGPT's output is certainly limited by its input. If you ask questions that can be answered by finding and simplifying information, then that's the answer you'll get. However, if you treat ChatGPT like a functioning software engineer, describe the problem space, explain the requirements, and define the boundaries, then ChatGPT will do much more. From that point on, you can ask it to explain why it chose a certain data structure, or to adjust an implementation approach, among other things.

With 15 prompts in 2 hours, I was able to pair program a novel 200-line program in Go that used AST parsing, command execution, pattern matching, file parsing, and custom data structures for specific data processing. All to solve a problem one of my teams experienced in the industry. ChatGPT wrote 99% of the code; I tweaked 2 or 3 lines because it had problems correcting an error in a regex (most likely due to the fact that ChatGPT sees tokens not characters). The code compiled, ran, and produced the exact result I wanted.

Researching, prototyping, and writing this code by hand would have taken me a few days probably. That's why the project sat on my TODO list for months

The value prop on this is unbelievable. I've been trying to get an affordable whole home heat pump system for over a year now.

You've explained the problem space, the solution, and why your background is relevant, extremely well. Kudos. I put a preorder in.

Minor nit: the grey on white on your preorder page is practically unreadable for my eyes.

> I am sorry for being so critical

You should apologize for being disrespectful. There's no reason to attack the founder personally with words like: pathetic, desperate, or insulting their critical thinking by saying "you haven't thought this through at all." You can write an equally critical comment without the vitrol.

> If you don't work at Google, convincing your engineering organization into learning Bazel is almost always a non-starter. Who uses Bazel in the wild? Xooglers primarily.

My response shows these hypotheses to be false. You’re correct, no one’s arguing whether Bazel is useful or worthwhile, not even I.

> Large companies […] who can support Basel within an engineering organization definitely exist.

Ok great, we agree. You’ve revised your point from earlier.

> There’s probably a few Xooglers […]

A baseless hypothesis that’s confirmed false by all the references shared with you. Feel free to plug all the lead engineers’ names into LinkedIn to verify that the majority are not ex-Google.

> Bazel-lite may be enough.

Sure, it may be. That’s why I’m offering the constructive feedback to go talk to the thousands of active Bazel users to see whether and what they want from a Bazel-lite, if that’s the positioning Moonrepo wants to choose.

Usually, there’s a strong rationale and need behind an organization’s adoption of Bazel. Some of that rationale is captured in the references shared earlier. These organizations need Bazel; they’re generally happier with Bazel than their previous systems. These baseline needs represent table stakes for any build system that wants to compete with Bazel.

If Moonrepo wants to compete where Bazel is weak, then I’m suggesting that they need to sharpen their communication such that an engineer well-versed in Bazel has an “aha!” moment within 2 minutes. The proverbial elevator pitch.

Once again, I’m not dissuading Moonrepo from pursuing their vision. I’m offering constructive feedback on their messaging/positioning — the type of feedback I’d like to hear as a founder.

Off the top of my head, some companies that successfully use Bazel: SpaceX [1], Datadog [2], Databricks [3], Stripe [4], Uber [5], etc. That's a good 10k+ engineers using Bazel successfully in the wild. There's more, of course.

[1]: https://www.youtube.com/watch?v=t_3bckhV_YI

[2]: https://www.youtube.com/watch?v=H67uuwVO1tc

[3]: https://www.databricks.com/blog/2019/02/27/speedy-scala-buil...

[4]: https://stripe.com/blog/fast-secure-builds-choose-two

[5]: https://www.uber.com/en-SE/blog/go-monorepo-bazel/

I'm not trying to be overly critical here, in fact, I want to share the type of empathetic feedback I'd hope to receive as a founder.

Have you actually talked to Blaze/Bazel users to understand what frustrations (if any) they have with their current build system? Have these users asked for a Bazel-lite? If so, and you still want to position yourself as Bazel-lite, then you should include some of their direct feedback and write your messaging accordingly.

As a daily Blaze/Bazel user, I don't have a desire for a Bazel-lite. I've worked at a midsize company that used Bazel, and I'm working at the company that created Blaze/Bazel.

Disclaimer: Opinions expressed are my own; not representative of my employer.

Not to mention using an actual language aids readability, extensibility, and static analysis, unlike a data exchange format. Starlark is a benefit, not a con, so YAML feels like a major step backwards.

I'm generally happy with Blaze/Bazel, so I'm not necessarily in the target market for Moonrepo, I guess.

EDIT: This isn't really competing with Blaze/Bazel either when I look at the execution model. It goes back to imperatively defined tasks instead of declaratively defined dependencies, which feels more spirtually aligned with Make than Blaze/Bazel.

Yes, this is why some people are afraid of Artificial General Intelligence (AGI). We can't even control or predict simple, by comparison, LLMs and yet we have the hubris to believe we'll learn how to control or predict AGIs.

I would normally just upvote, but I want to personally thank you for following up with a motivating example.

Do you have any actual hard data you can link to?

When I test your hypothesis by searching for "2019 honda cbr1000rr wiring diagram" [1] the first results are images of the wiring diagrams and then links to the exact non-SEO optimized forums you're talking about with a post that contains a PDF file available to download for the exact motorcycle I'm interested in.

[1]: https://snipboard.io/nkQY6U.jpg

> You are never going to get to 100% comprehension. 80/90% depending on the situation.

Key insight, people should set the realistic goal of being at most as good as their native language.

I can confidently say, in my native language, there are many times throughout daily life where I have way less than 80/90% comprehension. The other key insight though is that I can remove that ambiguity with follow up questions and elaboration if I want.

And that's what people should strive for when learning a second language. Once you get to the point where you can learn in and with the language, then you're golden. That requires far far less effort than most people estimate.

All these conversations have one glaring omission. As it stands right now, ChatGPT is a net negative on the ecosystem it exists in. What does that mean?

ChatGPT extracts value from the ecosystem by hoovering up primary sources to provide answers, but what value does ChatGPT give back to these primary sources? What incentivizes content creators to continue producing for ChatGPT?

Right now, nothing.

ChatGPT (or its descendants) must solve this problem to be economically viable.

> almost always out of date

Perfection need not be the enemy of good enough. Usually, whether the documentation is out of date or not, _some_ documentation is better than none. I've rarely come across an instance where documentation was actively harmful -- not to say those cases don't exist on the margins.

> And I do not even know what 'code labs' even means.

Code labs are self-guided lessons, like a tutorial, about the systems and technologies that your company is built on. They're called code labs because producing code/artifacts as part of your learning is one of the goals. For example, your company may have a custom enterprise installation process that's built upon Ansible, Terraform, AWS tooling, and a little proprietary work sprinkled on top. Some code labs may take you step-by-step through how to add a new service; configure, migrate, and wire up databases; etc. such that you'd have working code at the end of the code lab and new knowledge around the company's best practices.

Similarly you can do this for the standard language at your company. Most companies accumulate some amount of proprietary helpers, libraries, or frameworks in their standard language. A code lab can walk you through how to write code in the way your company expects using the helpers, libraries, and frameworks that are common across your teams.

> My experience is fairly typical SV stuff, so I am wondering where these companies that still do all these things are.

I have over a decade of similar experience across startups and large companies in NYC, Seattle, and SV. In the last 5 companies I've been at, all of them have had some combination of the above.

> Where else will people get habituated […]?

Standards documentation, code labs, pair programming, instructor-led group trainings. I agree that code review is a necessary piece of the puzzle, but there are other places for engineers to acclimate too.

I’m sorry, I’m not a marketer, I’m a SWE working on a team within Google Search. Google Search is composed of many teams, and my team is not responsible for following up on any type of public user feedback because our users are internal.

Hey Michael, I appreciate the effort you put into describing a few examples:

1. If you could link to specific examples and queries that’d be super helpful for someone like me that’s not active on the forums you’re describing.

2. Algolia is a fuzzy matching search engine. Searching for memeplex [1] returns matches like “megaplex”, “memepher”, “meeples”, etc. Unchecking typo tolerance in the settings returns < 100 results in line with Google’s results.

3. Again, if you could link to specific examples and queries that’d be helpful.

[1]: https://hn.algolia.com/?dateRange=all&page=19&prefix=true&qu...

You should try to assume good faith when engaging with a person who’s just like yourself.

First, we can agree that Google Search is attempting to solve an astronomically hard problem. Like mind boggling hard. Indexing the entire web and serving quality results to unstructured queries from billions of users every day in under one second is no small feat.

Second, Google is not monolithic. We employ more people than most cities have citizens. Furthermore, many more people than our current staff have come and gone over our 20+ year existence. It’s better to think of Google as an organic entity than a rigid command-and-control hierarchy. Are you able to think of a city in the world that does everything perfectly? I certainly can’t, and yet, there are cities that are better and those that are worse for any set of criteria that one may care about. As it is with large companies like Google.

Third, while there’s an objective element to search result quality, there’s also a significant amount of subjectivity. Your idea of quality results may differ from another person’s idea.

Checkout Paul Haahr’s talk on “Improving Search over the Years” [1]. He summarizes our work the best when he says things that look easy on the outside can take a lot of work to implement.

As it was with our state-of-the-art automatic synonym system that works on any written language in our corpus. (More details in his presentation.) This system is a transparent workhorse from the user’s perspective.

Here’s a simple example you can compare between Google, Bing, and DuckDuckGo: “united flight formations”.

Two of those search engines will show a bunch of things about United Airlines as top results because that’s what you would expect to get when you’re only focused on matching terms. Only one of those search engines understands the meaning behind the query and returns everything to do with formation flying as the top results.

If you use our products and you mostly enjoy our products, it’s in your best interest to give feedback when you feel we’re not serving your needs. You’ll find that most of products, Search included, have open feedback channels that we do review and act upon.

[1]: https://youtu.be/DeW-9fhvkLM

I’m not authorized to disclose data that’s not public knowledge.

What I can say is that we have a feedback process in place for Google Search that we use to improve our product. You can send feedback and check the box to allow our teams to contact you if you’re interested in a follow up. Of course, given our scale, we’re not able to follow up on every bit of feedback but that doesn’t mean we don’t review or act upon that feedback in some way.

Thank you for the kind words. Long time HN member here like you (going on 11 years) that recently started working on Search as a SWE.

Yes, that meme is very common. I hope I can contribute positively to these discussions by offering an outlet for feedback, and humanizing our organization. Google’s Search organization is large, so it’s certainly not monolithic, but we’re staffed with a bunch of normal, hardworking, genuine human beings like most companies, that care about the impact we’re having.

I’m happy you’ve found some value in our discussion. :)

Would you mind providing details like the search query and link to the page you expect to be found?

To test your hypothesis, I did a basic search for exact matches on "we do not synchronize on the update of the broker node" and Google returned 2 search results in 240ms:

- https://github.com/a0x8o/kafka/blob/master/core/src/main/sca...

- https://jar-download.com/artifacts/org.apache.kafka/kafka_2....

Which contain exactly the source code from GitHub that I was looking for. You'll notice that the first result is actually a0x80's fork of apache/kafka. Google states that some entries very similar to the 2 already displayed were omitted, and I'm able to remove that filter. With that filter removed, I can see the same document indexed from apache/kafka on GitHub.

There's nothing I can do or promise directly, but I can assure you that Google takes the quality of our search results very seriously. If you believe we're not delivering quality results, I strongly encourage you to click that "Send Feedback" link at the bottom of your results so that our teams can act upon your feedback.

Disclosure: I work on Search at Google.

Disclaimer: The words, views, and opinions expressed in this post are my own. They are not representative nor do they represent my employer in any capacity.

Thanks for the clarification!

Thanks for the response. I was responding to the comment more so than advocating for adjusting your log retention. :) Looking forward to part 2.

Are you able to reconcile some of the numbers and calculations in the article for me? (Understanding that you don't want to reveal any confidential info.) I see:

- 31 PB data + 10 PB application logs = 41 PB logs (uncompressed json) costs 7-figures (say ~$5M)

- 41 PB logs * 5% ORC compression = ~ 2 PB logs (compressed ORC) costs low 6-figures (say ~$300k)

I don't know what timeframe that cost is measured over. But that brings us to $300k / 2 PB = $0.15 / GB which is far above S3's quoted costs so I must be missing something.

Like markets... executives can stay irrational longer than you can remain sane, sometimes.

Yes, shipping computations instead of data is a reasonable design goal. Your proposed system only works when the predicate is independent across all logs though, correct? If you have to correlate or join your logs to anything, then this model becomes more complex. Not to mention, you're adding an additional performance tax to your prod machines which could be more costly than shipping logs to a centralized store. (A team should profile and make a tradeoff decision appropriate to their context.)

Additionally, what happens when we want to correlate these logs with tens of other systems?

I guess I don't agree that distributed log analysis simplifies the problem any more than centralized log analysis does. If the primary concern is cost, then you can save equivalent amounts of money with a different lifecycle policy for centralized logs.

EDIT: Btw, don't get me wrong, you are asking the right questions that HubSpot's performance team should be asking. The first phase of a cost savings program should observe benefits against cost, or stated another way, requirements vs cost. You're asking the right question, i.e., uhm, how do we actually use this data after we log it? I find it striking that this cost analysis didn't say anything about the end-user's use cases or benefits. Sure, we can optimize a system and save 40% the cost, but what if no one is using the system? Then we could save 100% the cost.

Software Engineering at Google has lots of valuable advice and best practices to accomplish this - https://abseil.io/resources/swe-book/html/toc.html

Can confirm. Migrated from LastPass -> 1Password last month.

Does Hypothesis require a cloud account that syncs your annotations? If so, these products are a non-starter for any corporate environment, which is unfortunate.

Obsidian gets this right. Most software these days doesn't which is a shame.

Thanks for taking the time to put together a well-reasoned argument. I had adopted some of your assumptions in a follow-up response to my other reply. [1] While I don't disagree with the perspective you've provided, in fact some points make a lot of sense, I think it misses the forest for the trees.

My point-by-point response should be read respectfully, since you've taken the time to do the same, I do not mean to sound argumentative. :)

> It is reasonable to assume that running a search operation, especially one that does not require any ad sales personel, would require much less people and infrastructure.

This is a reasonable assumption for running a steady-state business, but I challenge the worldview. Google does not exist in a vacuum without competitors, and contrary to what most outsiders believe, we're constantly iterating, innovating, and improving on Search alone to provide a better product and compete with our competitors. Google can't rest on its laurels.

A lot of laymen take Google Search's progress as inevitable, but I can assure you it's not. Consider the example of YouTube Snippets in Google Search. That feature was created within the last 5 years. The average person has used and found value in that new feature. (Both anecdotally and quantitatively.) That wasn't an easy feature to ideate, create, develop, or deliver. It took a lot of effort by a lot of smart people.

That's just one new feature. Google Search has been delivering several new features consistently.

Therefore I disagree with this assumption. If Google Search chose to run in steady-state, it'd soon find itself dethroned.

> That means ~$45B in cost needed to break even, or 5 times less than your starting point.

No, the financials don't work this way. [2] GOOG's Cost of Revenue is 50% Revenue. Cost of Revenue is your infrastructure, your financial floor, you can't go below this cost. Employees are accounted for under Operating Expense, specifically, your Sales under SG&A ($35B) and engineers under R&D ($35B).

So, if you'd like to banish all salespersons, you'd only save $35B. (Of which Ad Sales is only a part because Google sells many other things.)

> It is also reasonable to assume that the most of those which do not have access to internet, belong to the <$10/day income group.

This is not a reasonable assumption. Many people surviving on < $10/day have a low bandwidth cellular connection that they utilize for their family. Hence why Google innovated here with offline maps and landmark map directions for families that can only spare a little bit of bandwidth to calculate their route, and then make their way there without online point-by-point directions.

> quality of search has detoriated a lot in the last decade as documented by many discussions here

HN is a unrepresentative sample of the world population with respect to wealth, income, knowledge, interests, etc. I wouldn't consider HN as documentation for this. In fact, once again, the real-world data disagrees with HN's characterization.

> YouTube Premium exists (unlike Google Premium), giving an opportunity for people who don't want ads to pay, opportunity that 25 million people took, paying a $12/month subscription. So if anything, this just proves the point of viability of this as a business model.

That proves a 5-10% take-rate from YouTube's MAU. :) Which changes the break-even math on delivering a service. With a 5-10% take-rate on "Google Premium", your price must be 10-20x higher.

[1]: https://news.ycombinator.com/item?id=34301742

[2]: https://finance.yahoo.com/quote/GOOG/financials?p=GOOG

I’d call it an assumption, not so much a flaw. :) We can make other assumptions.

To break even, Google would at least need to cover their expenses. Google had $200B in expenses for 2022, with a net profit of $50B. So, that won’t really change the math.

Especially when we consider the fact that the take-rate would be much less than 100%. Maybe 5-10% is a fair take-rate assumption? (Seems fair since YouTube has roughly 50M paid subscribers on 1B MAUs, 5%, from the public data I’m seeing.)

At 5-10% take-rate, the service would cost 10-20x more to break even ($50-$100/month on average) which would be a nonstarter for the global middle class and lower.

One could make the argument then for Google to lower its costs in an effort to lower the consumer’s price, but then we must realize this runs opposite of innovation. Investment is necessary for innovation, and profits are necessary for investment. Without profits, there’s no more innovation.

Reality doesn’t lend support toward your arguments. I assume you’re from a first-world country, based upon your comment.

- Google’s 2022 revenue was $250B. (We’ll assume that’s all ads.)

- 8B people in the world; we’ll assume only 4B people have internet access to the web.

How much would Google have to charge to break even?

$250B/yr / 4B people = $62.5/yr/person = $5/month/person

What does the world income distribution look like? [1]

60% of the world population survives on < $10/day ($300/month) for their household (multiple people).

We expect more than half the world’s population to pay more than 2% of their income for a single service? To put that in a first-world perspective, that’s $100/month for someone earning $60k; $200/month for someone earning $120k; etc. Does the average American spend that amount of money per month on a single web service?

Well of course, you may say, let’s charge more for those who earn more so we can ease the pain on those who earn less! That sounds great in theory, but once again, how many people do you know dropping $200/month on a web service even if they can afford it?

Back to two of your points:

1. “Imagine how different the incentives would be” - Yes, imagine, only the wealthy would have access to state-of-the-art search for the web and other services, further increasing the disparity between the two groups. Consider, maybe, that ads represent one of the greatest wealth transfers in our history. One perspective to consider is that the rich (advertisers) are subsidizing the poor (information access) via ads. That sounds like a net positive to me?

2. “Keep their memberships active with a small monthly fee” — YouTube premium exists. What do you think the uptake is on that? YouTube has provided immeasurable benefit to people across the world in the form of knowledge, resources, training, etc. Yet people would sooner reach for an ad-blocker than pay the “small monthly fee” even in nations as rich as the U.S.

What causes you to believe that people will pay even more than that for Google’s other services?

Disclosure: I work for Google. The opinions and data represented in this post are my own, and not representative of my employer.

[1]: https://www.pewresearch.org/fact-tank/2021/07/21/are-you-in-...