tzhenghao

Duplicate of https://news.ycombinator.com/item?id=38998914

Good point. By the way, I checked out both projects above and starred them. Pretty cool!

This. I've seen a couple open core startups that are less than a year old with thousands of stars on their repositories, so I decided to have a look at some of these profiles who starred the project. Most of them have weird usernames that resemble spam accounts, almost all of them cannot be pointed to some other profile on a different platform (Twitter, LinkedIn, HN etc.).

Another giveaway is the ratio of stars to watchers / forks. I remember one project with thousands of stars but only 10 users "watching" it. They went on to raise a sizable seed round too.

I think they're just playing a larger distribution game to get people locked into Mojo then eventually pay for their Modular engine hosting services, just about the same as most open core, VC backed startups.

Yup. I've personally used both YAML and TOML for configurations, much more the latter recently and can see pros and cons for both.

> How well suited are their syntactic choices to the community they're targeting?

Also, "best" practices. One could reduce the pain of the other, but by no means is the right solution to a deeper problem at hand. For example, if one has very deep and complex nesting for configs, TOML "may be a lot nicer" compared to YAML, but that doesn't mean using TOML will make all the config parsing problems go away. It just mask away code smell. Maybe time to check if they're overcomplicating configurations in general.

I wear them every day, but take them off if I'm going to high crime cities. The thing is I've been collecting watches for over a decade. My parents are into it, and so is my younger brother. I think it's very rare to share a common interest that won't bore anyone at the dinner table. Now onto the horological aspects - they're like the iPhone / app store back in the day. Keeping track of leap years and all in a 36mm package (think Patek 3940s). Various complications to address various "limitations" of mechanical time telling like the remontoire, co-axial escapements or solid block case constructions for better waterproofness. It's not too different than some of us on HN who fall in love with old Apple IIs, NES or Sega Genesis :)

> On mechanical watches, they are smoother than Quartz since the escapement releases power multiple times / sec. But still ever so slightly jumpy since power is still released in discrete increments.

Yup, and the higher the beat rate, the "smoother" it looks. Grand Seiko Hi-Beats and Zenith El Primeros come to mind. There's a good Hodinkee article describing the tradeoffs of different beat rates [1].

[1] - https://www.hodinkee.com/articles/watchs-frequency-hz-vph-me...

Malaysian here. Yeah some of the historical sites in Penang or Melaka come to mind. I think the Peranakan [1] culture is one of the most obvious examples in using colors and is widespread to the region, not just in Singapore.

[1] - https://en.wikipedia.org/wiki/Peranakans

VCs aren't always the best capital allocators - a lot has succumbed to the money management + fees disease. They push you to raise more, force you to hire and burn when you really shouldn't / haven't figured out product market fit yet.

> Whatever magical market forces that might change how funding works, they don’t seem to be at play.

I think money is scarcer these days, and founders who are constantly being burned by VCs will think twice about riding the big VC train in the coming years. As a founder myself who work crazy hours getting the business going, it's painful / bad optics to see full time VCs doing weekend Vegas trips, browse art galleries on weekday afternoons and fine dining every couple of days, knowing I've sold a chunk of my / my team's hard earned equity on that bs. Of course there are great VCs, and some businesses needing to be VC backed, but oh boy, are there really bad apples out there.

Hey there, OP and author here!

> alluding to with respect to the ONNX being an Azure on ramp

Yeah sorry, should've worded it better in the original article. ONNX is one of the industry's first attempt at standardization of ops across various deep learning frameworks. Here's a way to think about it: Like Google, Microsoft has a cloud business. Unlike Google, it doesn't have a deep learning framework (Tensorflow) and is a significant threat if the cloud players continue to vertically integrate their ML stack. So ONNX makes sense, ie. you, the ML engineer should convert to ONNX so now at least you can run it on multi cloud providers!

Tangential to this, there's also a rise in heterogenous environments, and picking the "best" is non-trivial. A lot of serving / inference hosting companies capitalize on this, and ONNXRuntime is sort of Microsoft (Azure)'s answer: "give me an ONNX or PyTorch (which we will subsequently convert to ONNX), and I'll show you how easy it is to deploy it to Azure". The key nuance here is how much user tooling friction there is to deploying models, and it's in the best interest of these cloud providers to sometimes use their open source initiatives as on-ramps to their larger cash generating products / services.

Same. When I was at my first SWE job after college, I’d get up at 6am and try to beat Chicago traffic to be at my desk at 7am. That’s when I’d pull up HN or my list of technical topics I’d like to learn, then go for it until my teammates arrive. Rather shamelessly, I even had Leetcode open and went through exercises weeks before I left the job. Good reflex training having to force close it right around 8:30am though :)

These days my parents based in Asia use the two iPads I gifted them almost exclusively. They watch YouTube videos, Netflix, Facetime the kids and browse the web on them. The only time my dad boots up my old college MacBook is when he has to do important document editing - never more than once every month. To my surprise, he's even using the iPadOS Files app extensively. We're just not the target market.

> Plus, I love the disconnectedness of a sketchbook. I can sit and think and plan in a sketchbook, and it doesn’t distract me from what I’m doing.

This is why I still keep pens and physical notebooks on my desk. It's so much faster to flip it open and start writing than opening a 99th browser tab and navigating to an empty page on Notion, all while more desktop notifications start popping up on screen. Another side benefit is the act of writing / drawing out stuff actually helps me remember things better. Sure, my handwriting is atrocious but I rarely go back and read them anyway. ¯\_(ツ)_/¯

> motivate everyone (Intel, AMD, ARM, Google, etc.) to try and tackle the problem by making new chips

Yes, there has been repeated efforts to chip at Nvidia's market share, but there's also a graveyard full of AI accelerator companies that fail to find product market fit due to lack of software toolchain support - and that applies even for older Nvidia GPUs and their compatible toolchains, let alone other players like AMD. This isn't a hit on Nvidia, I'm just saying things move so quickly in the space that even the only-game-in-town is trying to catch up.

Nvidia is also leading by being one or two hardware cycles ahead of their competition. I'm pretty confident AI workloads in enterprise is their next major focus [1]. I think this more than anything else will accelerate AI adoption in enterprise if well executed.

To your point, I think the industry needs to focus more on the toolchains that sit right between the deep learning frameworks (PyTorch, Tensorflow etc.) and hardware vendors (Nvidia, AMD, Intel, ARM, Google TPU etc.) Deep learning compilers will dictate if we allow all AI workloads run on just Nvidia or several other chips.

[1] - https://www.nvidia.com/en-us/data-center/solutions/confident...

They’re not, but there’s non-zero probability most R&D below a deep learning framework gets fully vertically integrated (or even aggregated in the short term) to just the big cloud players.

Open source also serves as a "defense strategy" for larger companies whose core business does not directly / indirectly depend on commercializing its open source project(s). They in turn serve more like hedges.

For example, Meta with PyTorch in context of the machine learning tooling space. PyTorch is probably strategically important for them as they do not have a cloud business. The other large competing deep learning framework is Tensorflow, designed and open sourced by Google but also owns a cloud business (GCP). They also have in-house AI accelerators like TPUs. Microsoft didn't really have a deep learning framework, but they have ONNX, and is now using ONNXRuntime as an on-ramp to their cloud business (Azure).

If Meta wants to continue doing ML, and all the other cloud players jacked up their prices, that's gonna hurt Meta's operating costs. So it's still "cheap" to have hedged against such a scenario. All these on top of upsides like developer marketing etc.

> In many cases, however, that’s not what happens, and the CTO was basically the smartest person with a CS degree standing close to the CEO when the company was formed. The end result is that a lot of startups wind up giving a huge chunk of equity to someone who is essentially doing a job a semi-decent engineer could have done.

I think fundamentally (early stage) CEOs are judged the same way as well right? First time founders, especially the CEO, having to figure out sales and GTM etc. - very rarely do they get it right the first time. I certainly didn't. That's the risk of running a startup and investing in a startup. As the company scales, both CEO and CTO have to grow into their "new" roles every 6-18 months anyway. So, it's weird to single out the CTO, or really any founding member at all for "cap table dead weight".

Agree! I've also noticed that the # of restaurants vs. price distribution for eating out in SF falls off a cliff especially <$25 per person range, whereas there are many more options in NYC and LA. In other words, prepare to pay up, stand in line for hours or be on a reservation waitlist for the most part if you want "good" food in SF.

Loved his "7 Habits For Effective Text Editing 2.0" (2007) talk and humor [1]

"How many of you are mostly using Emacs?"

a bunch of raised hands

"Okay, we'll try to convert some people today!"

RIP, Bram Moolenaar

[1] - https://youtu.be/p6K4iIMlouI

Their github repo README has the following: > Performant ...

Do we have benchmarks for this? More generally speaking, I don't see more technical differentiation besides it being a Python DSL. Nor do they provide numbers to back it up.

I also went through this phase about a year ago, then just gave up and went with VSCode for its superior out of the box experience. That said, I still fire up vim to read files in a server or make one line config changes, but everything else....VSCode.

Related: Imagine having to remember the full scp / rsync command to sync source file(s) between your local machine and server, but it's just a drag and drop on VSCode.

I've just come around to accept that we've been blessed with great hardware over the years, and it's a waste if those extra CPU / GPU cycles aren't translated to significant dev productivity improvements. I'd take slower (~ ms) editor paint times over saving hours messing with vim configurations.

Yeah, I was just chatting with another principal eng friend recently where maybe a whole new generation of programmers may be several degrees away from topics like operating systems, compilers, computer security and networking.

Once in a while, (lower level) software needs to be rewritten as underlying invariants go out of fashion. Think filesystems due to hardware access patterns etc. - to unlock magnitudes of performance improvements. The scary part is maybe we'll never have enough programmers to know how to do "the hard stuff" anymore. I hope I'm proven wrong though!

"Show me the incentive, I'll show you the outcome." - Charlie Munger

Aside from the many software application examples on this thread, have you also considered how they also pushed innovation in the semiconductor industry that in turn contributed to non-obvious advancements in non-Big Tech sectors, from particle physics to drug discoveries?

Good point. "AI accelerators" is a thing now that competes directly with "GPUs used for AI as its sole purpose".

Eh, time will tell. Nvidia is killing it because they have "semi decent" software toolchains like CUDA when just about hardware player botched software. That said, a lot of interesting development on the XLA and PyTorch 2.0 sides that lower straight down to LLVM, bypassing Nvidia's CUDA moat today.

I remember working with several ML teams on specifically inference performance (latency, memory usage etc.), and it's not a surprise to see some object detection performance variance depending on a scene.

Sometimes, even capping the model architecture to bound us from exceeding performance thresholds is non-trivial in itself, but convincing "some" researchers why p99, inference latency for example, is more important than the p50 case in a safety critical system...that's surprisingly several magnitudes harder.

Amazing speech. I love that he touched on "the CUDA bet". How their early adopters in the sciences (particle physics, fluid dynamics etc.) eventually made it to the larger AI boom. Many of my former teammates came from Lawrence Livermore where they became CUDA experts while doing research work. This skill is now very high in demand for foundational ML work.

Interesting! Thanks for sharing, and definitely quite surprised how big that bandwidth premium is [1]

[1] - https://blog.cloudflare.com/bandwidth-costs-around-the-world

This. Modular and OctoML are building on top of MLIR and TVM respectively.

> It's pretty hard to really beat NVIDIA for developer support though, they've invested a lot of work into the CUDA ecosystem over the years and it shows.

Yup, strong CUDA community and dev support. That said, more ergonomic domain specific languages like Mojo might finally give CUDA some competition though - it's still a very high bar for sure.