HNHacker News
TopNewBestAskShowJobs

gjulianm

4,409 karma · joined November 21, 2011

Mathematician and computer engineer.

http://twitter.com/gjulianm

[ my public key: https://keybase.io/gjulianm; my proof: https://keybase.io/gjulianm/sigs/nfNutn3AksS22AzUBMHIkR3wqz46pUi-dOCF6GN2SbY ]

submissionscomments
gjulianm··on Mathematics is hard for mathematicians to understand too
I wonder why so many people are under the impression that the notation is what is keeping them away and if only the notation was easier then the underlying concepts would be clear. For example, if you don't know what the pullback of a differential form is, it doesn't matter if I write it in clear text or if I write the common notation φ^* ω.

> It's a domain reserved for a few high priests inducted into the craft and completely inaccessible to everyone else.

It's a domain reserved for people who want to learn it, and there's ton of resources to learn it. Expecting to understand it without learning it does not make any sense.

gjulianm··on Mathematics is hard for mathematicians to understand too
I'm not sure that symbols are the thing actually keeping you away. Clear text functions might not be as clear, as it will be harder to scan and it will still contain names that you might not be familiar with. Those "weird symbols" are not there because people liked to make weird symbols. No one likes them, it's just that it makes things easier to understand.
gjulianm··on Mathematics is hard for mathematicians to understand too
> I suspect they do so as a means of gatekeeping

I'm surprised at how could you get at this conclusion. Formalisms, esoteric language and syntax are hard for everyone. Why would people invest in them if their only usefulness was gatekeeping? Specially when it's the same people who will publish their articles in the open for everyone to read.

A more reasonable interpretation is that those fields use those things you don't like because they're actually useful to them and to their main audience, and that if you want to actually understand those concepts they talk about, that syntax will end up being useful to you too. And that a lack of syntax would not make things easier to understand, just less precise.

gjulianm··on Mathematics is hard for mathematicians to understand too
3blue1brown, while they create great content, they do not go as deep into the mathematics, they avoid some of the harder to understand complexities and abstractions. Don't take me wrong, it's not a criticism of their content, it's just a different thing than what you'd study in a mathematics class.

Also, an additional thing is that videos are great are making people think they understand something when they actually don't.

gjulianm··on Ask HN: What is interviewing like now with everyone using AI?
The interview works as intended because the main priority is to avoid hiring people who will be a negative for the company. Discarding a small number of good candidates is an acceptable tradeoff.
gjulianm··on Anthropic: "Applicants should not use AI assistants"
> So what are we afraid of? That people are going to copy paste from AI outputs and we won't notice the difference with someone that really knows their stuff inside out? I don't think that's realistic.

Candidates could also have an AI listening to the questions and giving them answers. There are other ways that they could be in the process without copy/pasting blindly.

> To me it's always been about how someone reasons, how someone communicates, people understanding the foundations (data structure theory, how things scale, etc).

Exactly, that's why I feel like saying "AI is not allowed" makes it all more clear. As interviewers we want to see these abilities you have, and if candidates use an AI it's harder to know what's them and what's the AI. It's not that we don't think AI is an useful tool, it's that it reduces the amount of signal we get in an interview; and in any case there's the assumption than the better someone performs the better they could use AI.

gjulianm··on Diffusion for World Modeling
Which capabilities are being shown off here? The ability to take an already existing world-model and take lots of compute to have a worse, less correct model?
gjulianm··on Diffusion for World Modeling
Generalisable how? The model completely hallucinates invalid input, it's not even high quality and required CSGO to work. What's the output you expect from this and what alternatives are there?
gjulianm··on Diffusion for World Modeling
What exactly would be the benefit of that? We already have Counter Strike working far more smooth than this, without wasting tons of compute.
gjulianm··on Run CUDA, unmodified, on AMD GPUs
OpenCL was released in 2009. AMD has had plenty of time to push and drive that standard. But OpenCL had a worse experience than CUDA, and AMD wasn't up to the task in terms of hardware, so it made no real sense to go for OpenCL.
gjulianm··on Leaving Substack
With volume-based fees you only have one type of user. Costs and benefits are predictable, and it's clear who you're providing the service to (the writers) so the company's goals are better aligned with those of their clients: attract more people to the platform. It's the writer the one tasked with deciding how to monetize those subscribers.

On the other hand, taking revenue off paid subscriptions means they have two types of users, free and paying. Even inside paying customers they'll have classes, as not all subscriptions cost the same. This means that the company has an incentive to implement dark patterns that convert users from free to paid, or from lower tiers to higher tiers, even when that is not in the best interest of the writer.

gjulianm··on Stable Diffusion 3
It definitely is easier without AI. Before, if you saw a photo you could be fairly confident that most of it was real (yes, photo manipulation exists but you can't really create a photo out of nothing). Videos, far more trustworthy (and yes, I know that there's some amazing 3D renders out there but they're not really accessible). With these technologies and the rate at which they're improving, I feel like that's going out of the window. Not to mention that the more content that is generated, the easier it is that something slips by despite being fake.
gjulianm··on Stable Diffusion 3
I feel like this analogy is not very appropriate. The main problem with AI generated images and videos is that, with every improvement, it becomes more and more difficult to distinguish what's real and what's not. That's not something that happened with literacy or printing press or computers.

Think about it: the saturation of content on the Internet has become so bad that people are having a hard time knowing what's true or not, to the point that we're having again outbreaks of preventable diseases such as measles because people can't identify what's real scientific information and what's not. Imagine what will happen when anyone can create an image of whatever they want that looks just like any other picture, or worse, video. We are not at all equipped to deal with that. We are risking a lot just for the ability to spend massive amounts of compute power on generating images. It's not curing cancer, not solving world hunger, not making space travel free, no: it's generating images.

gjulianm··on A connection between after-hours work and decreased productivity
To be honest, I don't know if I would have said that. I think that employees in "creative" positions (e.g., you're not serving customers at a bar and need to keep the shop open and working) are paid based on the output they produce. Whether that output is produced in 2, 6 or 8 hours is not really their problem. Looking at it this way, they would have happily paid you 100% of your salary for the same output you're now giving them with a 20% discount.
gjulianm··on Ask HN: 20% of LinkedIn's recent layoffs were managers
q
gjulianm··on Daily energy expenditure through the human life course
One time I tried to gain weight just to check if I was able to, as I have been unable to gain any significant weight in a variety of situation (sport, not sport; young, not that young, better/worse diet). I ate a calorie surplus diet (I don't remember the exact numbers, it was a few years ago), which meant eating more than usual but not too much either. This was measuring calories, weighting everything, all the drill; meanwhile I wasn't doing exercise other than walking/biking to places. Again, I don't recall the numbers but I gained maybe 10% of what I was supposed to gain. I've also seen how eating the exact same as other people (sometimes more) results in the other person gaining weight and me losing it, despite the differences in exercise being fairly irrelevant.
gjulianm··on Finding a Tech Job Is Still a Nightmare
> I wonder what the biggest hindrances are for software development to become this way.

Context and complexity. Jumping into a codebase and suggesting changes that add value to the company usually require having context of what the company is doing, who are the users, what's the product used for, etc. Not to mention how a lot of codebases are complex and understanding where and how to do changes is not something you can just do straight away.

gjulianm··on September was the most anomalously hot month ever
Not really, I don't distinguish between the different ways to solve it. I wish that was the discussion, but I feel like any solution that has any costs of any kind is just ignored because a lot of people don't feel like this is serious enough to do anything serious about it, doesn't matter whether it's reducing consumption, massive switch to nuclear, or whatever.
gjulianm··on September was the most anomalously hot month ever
> There are such obvious first steps we could take without a complete understanding of cause if we really cared.

That's the issue, a lot of people just don't care or refuse to believe it. Questions like the one in the parent comment are starting to seriously annoy me because they're demanding an unachievable level of evidence before doing anything, a level that's never asked for in any other kind of intervention (for example, interest rate hikes).

gjulianm··on Days of Awe: The clinical trial drug that might save my husband's life
> But when it comes to the very things that keep us alive, we are not so eager to play with computer code. That, I think, has less to do with the complexity of the subject and more with a certain moral disposition...which is the thing I find perplexing.

Honestly it's the first time I hear about moral stigma having anything to do with research. Again, I think you underestimate how much harder it is to simulate biological processes than LLMs. Even from the paper you've linked, you'll see the massive amount of simplifications they had to do: they're using a minimal cell, not all metabolites are included, multimeric proteins are left out/replaced, spatial distributions are simplified, the simulation timescale is below 10μs, they do not simulate reactive processes, they do not talk about how much time did they need to perform the simulation... Don't get me wrong, it's a massive achievement. But the amount of computation that needs to be done just to simulate a single cell is absolutely massive, let alone simulating multiple cells in a system. Considering how much it would cost it's no wonder other avenues are explored first.

gjulianm··on Days of Awe: The clinical trial drug that might save my husband's life
I don't think it's as easy as "simulating cellular processes". I feel like sometimes HN tends to severely underestimate the difficulty of problems in other fields. Just off the top of my head I can see several hard problems in this situation

- How precise do you want the simulation? Do you want physical processes fully simulated? You'd need that to get an accurate simulation of molecular interactions, but full physical simulation is computationally expensive even for single, simple molecules.

- Even simplificated simulations are hard. Simulating just the shape of DNA strands is computationally very expensive (usually done with Monte Carlo simulations).

- How many cells do you simulate? The body has ~37 trillion cells. Even if it only took one processor cycle to simulate a cell, you'd need 9250 4GHz processors. That should give an idea of how hard it would be to simulate just an organ.

- How do you take into account interactions? There's a lot of difficulty in understanding how drugs affect the whole human system. Lots of trials show promising results against in-vitro cells and then fail spectacularly in animal models. There's a lot influencing how drugs work and single cellular processes are just a small part of it. The body is incredibly complex.

- How do you validate the models? It's not like we can go into a cell and see where the molecules are. We don't have enough visibility into actual cellular processes to build such complex models to a sufficient degree of accuracy.

Compared to even a single human cell, a monstruous language model is a trivial thing.

gjulianm··on μMon: Stupid simple monitoring (2022)
I honestly don't get the criticisms of the Prometheus + Grafana stack.

> A full-blown time-series database (with gigabytes of rolling on-disk data).

Prometheus has a setting that allows you to limit the space used by the database. I'm not sure however how one can do monitoring without a time-series database.

> Several Go binaries dozens of megabytes each, also consuming runtime resources.

Compared to most monitoring tools I've tested, the Prometheus exporters are usually fairly lightweight in relation to the amount of metrics they generate. Also, "several dozens of megabytes" doesn't seem like too much when we're usually talking about disk spaces in the gigabytes...

> Lengthy configuration files and lengthy argument lists to said binaries.

Configuration files, yes if you want to change all the defaults. Argument lists, not really. In reality, a Docker deployment of Grafana + Prometheus is 20 lines in a docker-compose.yml file. Configuration files come with defaults if you install it to the system.

By the way, I'm not sure that configuring a FastCGI server will be easier than configuring a Docker compose file...

> Systems continuously talking to each other over the network (even when nobody is looking at any dashboard), periodically pulling metrics from nodes into Prometheus, which in turn runs all sorts of consolidation routines on that data. A constant source of noise in otherwise idling systems.

Not necessarily. Systems talk to each other over the network if you configure them to do so. You can always install a Prometheus + Grafana on every node if you don't want to do central monitoring and you'll have no network noise.

> A mind-boggingly complex web front-end (Grafana) with its own database, tons of JavaScript running in my browser, and role-based access control over multiple users.

Grafana, complex? I think dragging and dropping panels with query builders that don't even require you to know the query language are far better than defining graphs in shell scripts.

> A bespoke query language to pull metrics into dashboards, and lots of specialized knowledge in how to build useful dashboards. It is all meant to be intuitive, but man, is it complicated!

Again, this is not a problem of the stack. Building useful dashboards is complicated no matter what tool you use.

> maintenance: ongoing upgrades & migrations

Not really. Both Prometheus and Grafana are usually very stable and you don't need to upgrade if you don't want to. I have a monitoring stack built with it in my homelab and I haven't updated it in two years, and it still works. Of course I don't have the new shiny features, but it works.

To me, it seems that the author is conflating the complexity of the tool with the complexity of monitoring itself. Yes, monitoring is hard. Knowing which metrics to show, which to pull, how to retain them, it's hard. Knowing how to present those metrics to users is also hard. But this tool doesn't solve that. In the end, I don't know how useful it is to make a custom tool that collects very limited metrics based on other ancient, limited, buggy tools (SNMP, RRD, FastCGI...) that is missing even basic UX features like being able to zoom or pan on charts.

gjulianm··on Linux network performance parameters
This is great, not just the parameters themselves but all the steps that a packet follows from the point it enters the NIC until it gets to userspace.

Just one thing to add regarding network performance: if you're working in a system with multiple CPUs (which is usually the case in big servers), check NUMA allocation. Sometimes the network card will be in one CPU while the application is executing on a different one, and that can affect performance too.

gjulianm··on Goodreads is terrible for books – why can’t we all quit it?
I use it (sparingly) for tracking which books I've read/want to read, getting recommendations and seeing what people say about certain books as (partial) information to check if they're worth the time
gjulianm··on Ask HN: Is it possible to get a conversation with a CEO from here?
> preferably one that doesn't feel they got into their position out of mostly luck.

This part stood out to me. I think it's important to see that almost all successful CEOs (or founders, as it seems closer to what you wanted) have gotten into their position out of mostly luck. I don't mean this in a disparaging way and I am definitely not implying that they haven't worked their assess off to get there. But the reality is that there are a lot of hard working people, with sellable ideas, with experience, knowledge, resources... and most end up failing. And that's because there's a lot of luck in having the right idea at the right time (e.g., Jeff Bezos wouldn't be in the position he's in had he been born 20 years later or 20 years sooner) and finding the right people (customers, employees, relationships) at the right times too.

I think assuming that there's a massive amount of luck at play will help you a lot if you want to get those projects off the ground. First, from a personal, mental perspective: it's not necessarily that you're failing or doing things wrong or are not capable, but that it's a hard process and a lottery in many ways. I think that helps relieve some pressure and mental stress.

The second is that, knowing how much of it is luck, you can be more critical about the "advice" you receive. Most founders and CEOs will tell you the things that they think worked for them, but they might not be reproducible or, as i_dont_know says, might not even apply to a different problem-space. And the character traits that they have might be just a byproduct of being a founder/CEO and not what helped them get there.

gjulianm··on Reduction of sulfur emissions from ships may be causing rising sea temperatures
> while putting a substantial portion of GDP into building nuclear reactors and removing carbon from the atmosphere.

The first thing takes years and we still don’t have feasible technology for the second. Not to mention that we would still have a lot of emissions from transport, manufacture, and meat consumption.

A technological silver bullet would be cool but the reality is that we don’t have one and there’s nothing in the near future. But reducing our consumption is something we can start doing now.

gjulianm··on What's up, Python? The GIL removed, a new compiler, optparse deprecated
Of course a lot of libraries are not thread safe. However, that's not at all rare, lots of libraries for other programming languages aren't thread safe either. My point is that those libraries won't start magically crashing when running in no-gil mode unless the dev using them starts using threads in Python. Yes, it's hard to know which libraries are thread-safe and which ones aren't, and just like any other language you should default to "not thread safe" unless the developer explicitly says otherwise or you inspect the code.
gjulianm··on Ask HN: How do you monitor your systemd services?
Short answer: Prometheus + Grafana + Alertmanager. prometheus_node_exporter has an option to export SystemD service status and you can alert on failed services, and you can use Alertmanager to configure multiple types of alarms, including repeats so you don't forget.

Long answer: Whenever I've started to add alerting and monitoring to a system, I end up wanting to add more things each time, so I find it valuable to start from the beginning with an extensible system. For me, Prometheus has been the best option: easy to configure, lightweight, doesn't even need to run in the host, and can monitor multiple systems. You just have to configure which exporters you want it to pull data from. In this case, prometheus_node_exporter has a massive amount of stats about a system (including SystemD), and there are default alarms and dashboards out there that will help you create basic monitoring in a minute.

You can choose to use Grafana for visualization, and then either the integrated Grafana alerting or use the Prometheus alerting + Prometheus Alertmanager. I think in the latest versions Grafana Alerting includes basically an embedded AlertManager so it should have the same features.

Regarding the type of alert itself, I send myself mails for the persistence/reminders + Telegram messages for the instant notifications. I find it the best option tbh.

gjulianm··on What's up, Python? The GIL removed, a new compiler, optparse deprecated
Woops, yeah, you're right. I thought the default executor was a threadpool, my mistake. However, in that case, I assume that the default executor will not change to multithreaded when no-gil comes.
gjulianm··on What's up, Python? The GIL removed, a new compiler, optparse deprecated
async/await code already runs in threads, so that's not really a change.
← PreviousPage 3 of 34Next →