hintymad

You'd be surprised how VPs love to insert PMs. A famous game platform company, for instance, had two PMs for their storage team, one PM for their data team, one PM for the compute team, one PM for their dev tools team, if I remember correctly (the numbers could be larger, but won't be smaller)

It may appear so, but if you you talk to the Anthropic employees and management, especially those researchers, you'll see that they truly believe that the AGI is coming, if not has been here already.

> Google, a search engine which does not have human emotions, assumed that I had been spurned by a man in my life named Dario and decided what I wanted was an empathetic digital friend.

I'm not sure why this is a surprise. Google most likely uses an efficient yet limited language model to give us search answers. When a search query does not have strong or specific enough context, the model will sample and find what most of the training material will cover. That is, the most likely average behavior covered in the training data. In this case, "hes never coming over dario" is such a generic sentence, so unless Google happens to have indexed what the author mentioned and ranked it high, the answer would appear natural to the AI. Or using the AI jargon, the answer appears to be properly aligned.

> so they’re using LLMs to build native applications for every platform they support.

People say LLMs will let us build native apps instead of Electron, but is that missing the point? I thought people turned to Electron for a reason: building a fast, rich, collaborative Markdown editor on macOS is notoriously hard. Neither SwiftUI nor AppKit makes handling those rich interactions easy. AI can do wonders, but can it really overcome the inherent limitations of macOS's native UI frameworks?

I thought the difference is that clearly written rules or SLAs are acceptable and can even be negotiated before signing a contract, but having a person, in this case Amodei himself, manually approve DOJ's usage case by case with his own moral judgement after a contract is signed is not okay.

> Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.

I'm quite puzzled about why Anthropic is so hellbent on blocking other coding agents. It's not like Claude Code has any secret sauce, right? And doesn't Anthropic make monkey off API usage, and their magic is on the model side anyway?

Finally Apple has a screen time that allows different configurations for different time segments. What took them so long? /j

> Right now, overall compute is a moat.

Very true. Or further, access to capital is the moat. It is the very reason that we don't have a real open-source community that trains frontier models - individuals simply can't afford the training infrastructure, nor sufficient high-quality training data.

> I guess what they really want is to limit sale of AI models to compliant vendors and then raise the bar to compliance just high enough so they can pass it but smaller labs can’t.

If it's true, it makes them evil. Especially Dario, who speaks about moral high ground and fate of humanity all day, yet it's not that different from a cult leader does.

> To him, the true AI doomer scenario is for all the immense power of frontier AI to wind up in the hands of a single powerful, proprietary provider.

Isn't this exactly what Dario wanted? He thought he knew what's best for the humanity...

Maybe the real threat is that our industry has been stalled for years, or to put it more politely, that it has been mature for a while. Back before 2010 or so, people actually read books like The Art of Computer Programming, the Dragon Book on compilers, and the Lions’ Book on Unix, or followed sites like lambda-the-ultimate.org. Then sites like High Scalability became popular. Not that everyone studied them cover to cover, but plenty of engineers considered them essential reading. Yet long before LLMs came along, that kind of depth had clearly become niche. The creator of High Scalability even put the site up for sale. Whenever someone posts an article titled "Top X Data Structures for Y," every single structure mentioned was invented decades ago. If most of what we do now is just slice and dice abstractions created and refined by previous generations, we are really just relying on our once-unique ability to transfer knowledge - something AI is rapidly replacing.

This isn't unique to software engineering. In Renaissance Italy, mathematicians like Tartaglia and Fior hoarded cubic formula shortcuts like proprietary algorithms and challenged rivals to public math duels. Today, we solve cubic equations without a second thought. Special functions used to be a staple college course for physics and engineering majors. Are they still? The US military used to employ thousands of people just to calculate PDEs by hand. Do we need anyone doing that today?

Our only hope is that our society moves fast enough to create new demands and problem domains that genuinely require new systems and algorithms. Look at AI: it’s evolving rapidly, driving massive demand, and forcing the development of new systems and architectures. As a result, the lucky few[1] working at that frontier are having rewarding careers. Without frontiers like that, the rest of us risk becoming irrelevant.

[1] One unfortunate factor is that building AI now requires lots of capital for accessing GPUs, which means individuals in the open source community have a hard time working on it.

> LLMs are able to write almost perfectly correct code.

This is kinda vague. Correct at what scale? I wonder if there's a measurement on the correctness per scale, and hopefully the scale is not just CLOC.

Curious why it is so hard to find an owner to the issue. Gates' experience started with the web UI, then I'd expect that the team who owns the microsoft.com or owns the experience. To them there are only two types of bottlenecks: their web pages (usability, JavaScript performance, ways to get backend data, and etc), and their immediate dependencies. So, they drive improvement on both types of the bottlenecks, and the owners of the immediate dependencies recursively handle their own. For instance, the web team will identify that calling the catalog API has a P99 latency of 5 seconds and the network is fine, then they ask the catalog API to improve the API latency. If the catalog team does not do the obvious, the team's manager gets punished.

Of course, I'm being naive here, as I've seen too many companies fail to achieve such basic ownership. So, curious what I have missed. Of if the ownership is not a clean DAG, well, it goes back to Gates, as he was responsible for both the org charts and the company culture.

All luxuries indeed. Are they enough, though? Particularly, don't many books talk about how important it is to have productive hobbies or output-oriented hobbies? The underlying thesis is that one would quickly get bored and start seeking the meaning of life if he does not output something consistently. I was wondering if being able to find such hobby and being able to afford it is also a luxury.

There's an interesting dynamic, too. Even if an engineer reads the output of the AI and understands the root cause of the problems and how to diagnose the incident, somehow it's hard for them to internalize the learning and apply it next time to a new incident. As a result, the engineer loses touch with the system anyway.

It looks like our brains somehow have to experience the failures during a diagnosis and in gemerak perform this kind of pathfinding by themselves to truly understand the system. I don't know if this has to do with how our brains actually learn.

I think using open-source AI is no longer about API cost but about company survival.

Take Anthropic for an example. Anthropic has successfully destroyed customer trust, at least for me. DHH in a recent interview mentioned that Claude refused to translate an article about immigration. Not summarize. Not editorialize. Translate! I think this reveals an unacceptable level of paternalism: Anthropic fundamentally believes that it possesses a moral authority superior to the people actually paying for the API. If such basic and mechanical translation is already too sensitive to touch, the goalposts have moved from safety into outright censorship. What prevents them from quietly deciding tomorrow that your proprietary business logic, financial data, or legal documents cross their invisible moral line?

Let alone how Anthropic treats Cursor and Figma - not that they are wrong as companies are free to compete legally, but nonetheless it shows that companies can't outsource their intelligence to a potential competitor.

I'm quite curious why few people are interested in what Uber has done. 70% of the code gets auto merged is a pretty impressive number. Companies have achieved way more than that? Uber's AI infra turns out to be not so useful? Or something else?

In some interviews, OAI mentioned that they didn't think that GPT-3.5 would be a success. They thought it would be a cool toy and they decided to launch it to see how users react. That means that they didn't think GPT-3.5 was intelligent enough. But somehow once GPT-3.5 became a huge hit, people conveniently ignored the anecdote, and started to believe that AGI had been eminent.

> they can afford the penalties and continue doing it.

I thought they could've bought just a single copy of each book and use the content to train their models. In that case, it falls into the fair use doctrine and they wouldn't need to pay the fine. And that will be way less expensive than the $1.5B price tag.

What puzzles me is this: research means that we are exploring something that has not done before, yet using Claude to generate code means slicing and dicing what has been done many times before. So, I'm not sure how to make sense of both at the same time: Anthropic is pushing the boundary of AI, yet all the knowledge and engineering in form of code can be generated from the previous work?

> The coding part of my career is over. LLMs are capable of doing everything I've ever been paid to write.

Remember we used to spend enormous amount of time in school and in our spare time studying computer science? Algorithms, operating systems, compilers, and etc. All kinds of insights. All kinds of fun. All kinds of hard engineering. Yet, how much time do we really need to spend in our day-to-day work implementing or using the algorithms and etc that we have learned?

Engineers have done amazing work of abstracting away the hard algorithms and data structures. In the meantime, there has been little progress or few new fields in the past 10 years or so in business that ask for implementation of new algorithms. In contrast, getting LLM to work is a new field, so it requires tons of new implementations: KV caches, speculative decoding, all kinds of variants of attention like FlashAttention, all kinds of parallel processing techniques, RL pipelines, post-training pipelines, and etc. It's just that the field is so concentrated that only luck few get to work on them.

So, maybe it's not LLM per se that removes the need of writing code. It is the maturity of the software engineering that has done so. It's just that LLM fills the last gap: making knowledge transfer so much faster and cheaper - if all that's left for most of us is slicing and dicing of what has been already been implemented, then LLM can reliably take over.

Boris: "I don't prompt Claude anymore. I have loops prompting Claude and figuring what to do".

Boris: "I haven’t written a line of code by hand in, I think, eight months now… Claude Code, 100% written by Claude Code".

Boris: "There’s no manually written code anywhere at the company… All of the SQL is written by models. Everything is just built by the models... Claude instances communicate with each other (e.g., over Slack) in autonomous loops"

This does not sound like they review the code either. So, either the frontier labs like Anthropic have figured out something that very few companies could replicate, or they are being incredibly deceptive. I don't know which is true.

> I don't know why anyone should care about understanding the results if the AI is better at math than us

This is a big if, right? AI can still generate subtle or even silly mistakes that any normal human, let alone a mathematician, wouldn't make. Besides, math is more than just getting a conclusion but to understand and to generalize new ways of solving problems. After all, mathematicians are a curious bunch. To quote Hilbert's epitaph: We must know. We shall know.

I see a disconnection here. Every once in a while we see a article criticizing management consultancy, yet the business is still booming and very powerful and sophisticated corporations still hire them. So, I'd assume that the consulting companies do offer some value, in fact billions of dollars of value. What are those values, then?

> Learning new CS concepts

A key reason is that so many CS concepts are already well packaged in very nice libraries or frameworks. And unfortunately many people do not have to use very advanced CS concepts. Look at the so-called top 10 algorithms or data structures for <your favorite area>, how many of them are invented recently, and how many of them do not have an amazing library of production quality? Note this does not mean that the CS field is not advancing. It's just that we have yet to find another hyper-growth area that demands novel algorithm, except probably the AI infra - except that only very few people get to work on that area.

> The problem with US healthcare isn't who pays for it, its how damn expensive it is

Exactly! There's so much paperwork that a doctor sometimes need two assistants just for the paper work. There's so much cost for independent practice that increasingly more doctors end up joining big hospitals. Charges with and without insurance have a huge difference. Just to name a few.

It sounds like Go didn't follow the suggestion of pretty much every CS books on concurrency: favor containers over concurrency primitives.

> while Go forces channels for `select` whether they model your problem nicely or not, and they're very difficult (often impossible) to wrap without changing semantics.

I understand that Go's concurrency model is based on CSP and fork-joins and the primitives like locks, but they are not mutually exclusive with concurrency containers, right? It's okay if the Go team's core philosophy is that channels are the universal abstraction, but I don't get why the community didn't produce 3rd-party containers as robust as JCTools.

Java has so many excellent concurrency containers, plus robust 3rd-party containers like JCTools. It puzzles me why Go communities do not offer such containers.

They do, and they help their users all the time

And that should be the engineers' job instead of outsourcing it to PMs who do not even use any of the infra services -- I'm not insulting the PMs, of course, but to state a fact. Infra is used to serve the internal engineering teams, and the PMs don't code, so they don't have a need to use the infra.