vhiremath4

Thank you so much for these words.

This is interesting. She would often admit that she was codependent and would forewarn that she knew she was being annoying because of that, but then there would be no actual progress/accountability on getting better. It would be this loop where she’d warn that she was being that way, but then she would go on and create an argument out of it anyway. I definitely don’t think she was a narcissist. She’s such a great human being. I think we both didn’t have the tools to deal with our competing insecurities and how they triggered each other. It was a bit of a shock to me since I had relationships with women who were much more secure in the past, and we never got into these intense of arguments. They were always resolved quickly and it never got disrespectful.

I’m very sorry to hear that things have led down this path. I hope for nothing but the best for you, her, and your kids. Do you feel like, if you could go back, you would rather end things knowing there were this large of disagreements on the fundamentals of how you would raise your kids? Or did you guys not fully broach these topics before getting married? Obviously, no pressure to answer here. These are sensitive questions to answer and put out on the internet.

Yeah great advice. We both really tried. We committed to working through things. We saw a couple’s therapist. We went through some pretty intense ups and downs on both our ends ranging from a death of a very close family member of hers to layoffs at my company and many other things. I just found that, ultimately, it still didn’t feel worth it based on how we would communicate and trigger each other even when we tried. And we would just keep going because we wanted to fight for the relationship, despite it continuing to not make a ton of progress.

I guess that’s the “not depending on books alone” part of your comment. It feels like I’ve gained some really important life experience, but I just haven’t fully processed it or been able to make sense of it. Funny enough, I’m making more sense of it with time and with comments like the ones here.

Thank you for your comment!

This is such a great distinction (teasing vs. built resentment). I think there was definitely mounting resentment from both ends (and I’m not a saint/the victim here). There were also arguments we’d have where both of us didn’t take things seriously and it was just fun/added some spice. Thank you!

Just ordered this book. Gracias!

For sure. I have friends who are in long term relationships and very happy and they bicker a lot. I have friends who never fight or raise their voices. Compatibility of sensitivity is a real thing.

Honest question - how many other people on HN are in a relationship this fiery? Are you mostly enjoying it or not enjoying it? Do you all believe this is an inevitability for most couples who have just been together for a long time?

Context: I just broke up with my girlfriend a few days ago because we would get into arguments regularly around topics that didn’t seem surpassable (values, negative feedback loop insecurity types, etc.) and would leave both of us drained. I just didn’t want that amount of anxiety in my life. I 100% understand that some people are totally cool with it though and don’t have an issue at all.

It really depends on how mature your org and stacks are. This is generally how I would do it.

1-20 people - password manager (bitwarden, 1pass, etc.) 20-30+ people - SSO

50+ people - start assigning real roles to your SSO schema

1-5 services - secrets in CircleCI and password manager is good enough.

5+ instances - use a secrets manager like Vault.

10+ instances - start using a secrets manager locally as well for dev. Start to consider using well scoped IAM policies for each of your services and team members.

15+ instances - start to think about adding additional zero trust boundaries.

Of course, this is very rough. Depending on your regulatory/compliance requirements and how much revenue you’re bringing in and from who, you might have to do this stuff sooner. In general, it should go:

1. Centralize secrets even if you can’t easily revoke people (password manager).

2. Make things easily revocable and centralized (sso).

3. Make roles and access finer grain (RBAC).

4. ^ with automation between all of these steps where it makes sense.

Something I would warn anyone of is building your own auth/secrets core tooling. This stuff is incredibly complex because of the edge cases and it’s just not worth the risk you take on by saving money unless you have a really good core business reason to roll your own. It’s also dangerous to prematurely optimize and pay the SSO tax too early. You will find that a lot of engineers appeal to emotion when it comes to risk. Something extremely helpful is going through and actually assigning a security risk score for all your systems. This might be tedious, but it brings a lot of clarity to the conversation of “what do we want to build when? What risk can we take on at any given stage?”

That’s great context. And I’m very sorry. I’m an optimist, but I’m also a realist, and, yeah, the delta between getting close to making it over and over and having that high of an expectation and then nothing coming to much fruition after so much work would make anyone discouraged.

It takes a ton of effort and courage to have hope in a situation like that. Because you will have to believe something is possible that you have no indication of being possible based on prior experience.

Or you will have to change your outlook to not care about massive financial success as much. But that’s not easy either since that requires a world view and possibly a values shift. Very very hard.

And yeah if I were stuck in that spot, I would feel cursed too and like the higher forces are fucked. It’s totally natural. I really hope for the best! I know you will be good in the end, but I’m sure that’s the last thing you really want to hear since it’s invalidating of how you’re feeling.

I am not going to try to change your outlook to be more positive, but you have to admit that it’s somewhat (tragically) amusing that we can find ourselves on a website full of people who are extremely technically capable (and highly intelligent), have access to devices that can connect to billions of others devices instantly, and can profit from said system allowing them to better their lives, yet this same community can only dwell on the negative of their situation instead of the positive.

I think, when you’re very smart, you are just many steps ahead of the average person in pattern recognition. You can literally store more information and pattern match against it, so you’re going to pattern match up the power hierarchy and think about things that have a negative impact on you that you can never directly affect without traversing up the hierarchy. If most people here (not saying you specifically) would instead just focus on “what is the biggest problem in my life that I have complete control over?”, I think they’d find that focusing on the greater powers that be has little practical consequence since they’re just not at the level where they can exert influence over those powers. There is almost certainly a path to local maxima, and from there, a greater and greater maxima.

I think it’s more appropriate to say that self-deception is a consequence of overfitting based on other primal instincts. Our #1 skill lies in pattern recognition and imitation. A lot of time, it’s faulty because it’s guided by fear/anxiety for perfectly valid reasons of survival.

It’s a bit more ridiculous because Ubuntu (and most Linux distros) do not implement their own filesystem semantics from scratch. They’re using the Linux kernel which many distros build off of. So it’s more correct to say “Linux shows its fragility” rather than “Ubuntu”. It’s all just a bit awkward, but I guess I’m being pedantic.

I wish I could address your question, but I don’t think there is “fragility” in how Linux does random writing. It’s extremely durable. MacOS (and therefore Darwin) has some pretty awesome optimizations by owning both the hardware and software, but I don’t know if random writes on Mac “succeed” more often than on Linux. It’s also a bit nonsensical since the hardware really matters when we’re talking about write durability.

It seems like small multimodal LLMs have a killer use case to be bundled with browsers for accessibility. Eventually:

* if an image doesn’t have alt text

* you need to be read the page

* you need to be described what’s happening in a video

A model built into the OS or browser seems like a no-brainer.

Ah very interesting. Thanks for making that distinction! To me, I’m wondering at what critical mass of people should you need to provide a back door to the government? Is it 2? 3? 100? 1000? Are the number of people the right indicator here? Genuinely curious of people’s thoughts. I haven’t though too deeply about this.

This might be a dumb question, but I wonder what the lifespan of this battery is. When the zinc oxidizes, the electrode eventually becomes less effective as an anode right?

So collaborative filtering?

Totally agree. And this is why companies should really profile when hiring to attract the right people with the right risk profile given the stage of the company.

Early on when we were just starting out (context as a founder), I used to think that the only types of people who would take on this kind of risk were young 20-something’s who had time and space and no significant other. Then we hired 2 early engineers in succession who were older.

One was just made for startups. He could never work for a large company, and he was ok with the risk and lower pay because it was still quite high relative to his cost of living in Europe.

The other already had kids who were older and independent. He always wanted to take a swing at a “Silicon Valley” startup, and he just cared deeply about that experience and could do so without having to worry about his kids financially relying on him.

They were 2 of the best engineers we had ever hired and stuck with us through thick and thin from a team of < 5 engineers through us scaling to 80+. When we got bigger, one left because we were once again too big for him (process, minor politics popping up, spending more and more time teaching newcomers how to own and operate a bigger codebase safely, etc.). The other left because of similar reasons but at a later scale.

I started the journey starry-eyed/overly-optimistic in both directions. Hiring people early on who obviously did not have the risk tolerance or an understanding of how much work was needed early. Holding on to people for too long who weren’t enjoying the new environment. Now I’m much more even about this - there is a right place and time based on what the individual wants, and being open and honest and kind about it is always the best path. Equity is simply a lever to compensate risk, but the appetite to take on the right amount of risk is the most important thing given the company’s scale.

There are a bunch of good answers, but I wanted to succinctly say "practically, quite a bit". Here's a good little rabbit-hole example:

> https://github.com/karpathy/nanoGPT/blob/master/model.py#L45

Karpathy's nanoGPT calling flash attention by checking if torch.nn.functional.scaled_dot_product_attention exists

> https://pytorch.org/docs/stable/generated/torch.nn.functiona...

Looking at the docs, in reality, most of the time you want this to call out to FA2 which optimizes the kernals on the device to split ops on the Softmax of the triangular matrix as well as reduce moving unnecessary batches of floating point numbers back and forth from the GPU to the CPU.

> https://arxiv.org/pdf/2307.08691

The paper for FA2 almost entirely considers itself through the hardware it's running on.

It seems like the smaller models get the largest size decrease by embedding share/weight tying between the linear head and token embeddings. Is there any research going into how to further reduce size from there?

This is kind of what I was thinking. I don’t get it. It seems like CriticGPT was maybe trained using RM/RL with PPO as well? So there’s gonna be mistakes with what CriticGPT pushes back on which may make the labeler doubt themselves?

When they say “LLM data”, does that usually include the tokenizer as well? Beginner question from someone at the end of Karpathy’s Zero to Hero course.

> Broadly speaking the only solutions are tax rises, spending cuts or more debt. But these are three very different parties, with conflicting views over borrowing and spending.

So they basically just couldn’t come up with a solution when trying to balance the budget. Definition of death by committee, which is a representation of the will of the people.

How can we do this better? This feels so important, yet it feels impossible if this is the outcome of telling our politicians they need to respect a budget.

which core exercises worked best for you?

There is currently a proliferation of robotics companies of all shapes and sizes. Lots of capital is being deployed into robotics. My best guess is give it 3-5 years before we start to see and accept that there will be many more robot friends among us both in industry and our personal lives.

Finally a title so simple to answer what I’ve been too lazy to read into the details on but was curious about.

Thank you for recording with us! And good luck with this product. There's definitely a bunch of use cases for scrappy prototyping or early mocks!

Funny story. Before pivoting my startup to Loom, we were a user testing company named Opentest. Instead of spinning up a DB and creating a dashboard for my co-founders to look at who requested certain user tests, I just dumped everything into a Google Sheet. It was so good. No downtime. Open access. Only 3 people looking/editing, so no conflict. Didn't have to deal with database upgrades or maintenance. I often think about this decision and feel like I've learned a bunch of "good engineering practices" that pale in comparison to how being truly scrappy can be a genius unlock at any level.

Agreed with the spirit of this post. OpenAI also pays very well though and has super high caliber talent (from what I can see from friends who have joined and other online anecdotes).

Honestly, you only get this kind of humility when you're working with absolute wizards on a consistent basis. That's how I read that whole analysis. Absolutely fascinating.