mochomocha

I've also lived in San Francisco for many years, in different neighborhoods. Now with my family including 2 young kids. The usual media narrative around SF is also very foreign to me and doesn't match my personal experience. Living in SoMa has very little in common with living anywhere west of Twin Peaks.

Therefore the need to rely on data to find more objectivity when debating public policies. Yes data can be twisted and most of the time is poorly reported and interpreted by people, but it's better than relying on personal anecdotes.

Since it's tax season... Did you ever benefit from a tax credit on your return? Should we consider you state-affiliated?

Less than 1% of their funding is from the government.

Tesla received billions in government subsidies... Is it state-affiliated?

Yann LeCun pioneered CNNs for this actually. OCR of zipcodes for USPS. Early 90's.

I understand. In the typical corporate world this is solved with "hire interns to work on all the cool things we don't have bandwidth to explore, while being simultaneously jealous of them". Or I'm sure a lot of people such as myself would happily volunteer to help out.

Do you have a simple AB testing system in place to test hypotheses like this? If you're the only sherif in town running experiments, it doesn't take much work to build a simple one, probably around ~100 LOC ("do things that don't scale").

? Considering Elon as its own category is a bias.

No, the "weights" of your model are trained from the input data. What is usually AB tested are hyperparameters of the model, or different "flavors" of (model+input data).

> But, we do have a bit of code that measures metrics on his account, so can we find the bit of code that increases the engagement on his account?

There doesn't need to be. When they run AB tests, it's possible that they'd pick the winning cell if it makes the Elon metrics look better.

Even if the algorithm doesn't do anything explicit about boosting him, it can be tweaked through AB testing to favor him.

... Metrics tracked in AB test. So even if it's not explicitly encoded in the algo (or implicitly through some of the features plugged in), they'll pick the winning cell as long as it doesn't hurt Elon's metrics (I'm just parroting the comment you quoted).

It doesn't have to be in the algorithm for the systems to be tweaked to please Elon vanity metrics.

[I've been running lots of ML AB tests over the years, some in organizations of similar size & complexity as Twitter]

While this looks like a step in the right direction to provide user-space per-process control of the latency VS throughput tradeoffs (lower runqueue latency at the expense of shorter timeslices), I'd rather see the SCHED_EXT patch [1] merged, so there's a lower barrier to entry when it comes to implementing alternative scheduling policies (eBPF vs in-kernel). It's also much faster to iterate user-space if you don't have to upgrade kernels every time, especially in large fleet deployment setups where AB testing is available.

On a more controversial note, I trust more Meta & Google engineers proposing alternative schedulers that have been properly AB tested on millions of servers running a wide range of heterogeneous software VS small-scale synthetic benchmarks even when run by experts/maintainers.

[1]: https://lwn.net/ml/linux-kernel/20221130082313.3241517-1-tj@...

> legally mandated

> real free market environment

I'm hoping you're just kidding.

And it's not only the names that are stupid. 50% of the time on my TV when I'm trying to open the Youtube app, I end up clicking on the Youtube TV app, because the logos look exactly the same.

That's been my experience. However when fallback to CPU happens, it sometimes end up making a specific graph execution slower. But that's explicitly mentioned by the warning and pretty much expected.

Ha yes, the good old "privatize profits, socialize losses" American capitalism.

I think we indeed hit "peak open-source" for AI and there unfortunately won't be as much sharing in the coming years. When the economy is down, people and companies think more in "zero-sum-game". I hope to be proven wrong.

> But from the way it's built and trained, it's not like there's even any kind of factual confidence level you could threshold, or anything. The concept of factuality doesn't exist in the model at all.

I'm not super familiar with ChatGPT internals, but there are plenty of ways to tack on uncertainty estimates to predictions of typical "large scale ML models" without touching Bayesian stuff (which only work for small scale academics problems). You can do simple parametric posteriors estimation or if all you have is infinite compute and don't even want to bother with anything "mathy", bootstrapping is the "scalable / easy" solution.

... To the surprise of no one. That's what regulatory capture + corrupted democracy (ie "lobbying") looks like.

Pretty click-baity title. This article is a good "ELI5" summary for non-ML people of some of the techniques & infra described in the arxiv paper, but it has nothing to do with TikTok "secret sauce".

> the fuzzy nature of the recommender problem means I am never going to get into the 90's

Right. If your AUC is in the 90's you usually either have a bug, or a predictive task that isn't very challenging.

I do not disagree with you, but that's exactly what the author of the blog post emphasizes: focusing on the "risk -> money" is an example of missing the forest for the trees. It's not the company raison d'être (unless you're talking about an insurance), you're just summarizing how a company operates under a very specific lens, the same way you could describe a company as something converting paycheks into human labor. A company makes products satisfying customers: that is its core purpose. Focusing on the process instead of the product is exactly why mega-corporations fail, which was the main point of the blog post.

> Mandating every manager do a meaningless, symbolic one hour stint in customer support isn’t going to change anything.

I am of the opposite sentiment. I think it would do wonders. It's very easy in a mega-corp to completely lose track of what the product is, who the customers are. And my intuition is that most VPs at mega-corps think about products in the abstract and focus more on politics & empire building/strategy than actually caring about the customers - and it definitely shows for companies like Google.

Just last week I tried to spawn a GPU instance on GCP, and despite being very familiar with the space (I work on cloud infra), after 20 minutes of fighting the overly complex UI and ridiculously broken flows, I still hadn't my instance up. In the next 3 minutes, I signed up for Lambda Labs, entered a CC, clicked a single button and ssh'ed into my VM.

That's what any start-up focuses on: the product. And that's why they always win over bloated mega-corps focused on internal processes and infighting, whose employees self-select for being complacent about mediocre products as long as the monopoly keeps printing fat paychecks.

> turning risk into money is what a company, any company, does.

I must say that's one of the most unorthodox description of what a company does I've seen.

How is a bread factory or a plumber "turning risk into money"? It sounds like you're describing an insurance company or a bank, not, you know a company like Google who produces free software & services to billions of people in exchange for monetizing their attention. Focusing on risk mitigation in a company is explicitly optimizing for protecting existing income stream at the expense of innovation, customers, and new income streams. It's not what "any company does": it's what uninspired companies who have lost their ways focus on.

It sounds like what a consultant used to "financialize" everything would say to solve Google's problems and slowly turn it into an intangible conglomerate like GE who is just the shadow of its previous self. But maybe this time has arrived for Google?

"Native advertising 2.0"

I have a hard time being on board with the logic. Neither k8s nor mesos guarantee optimal placements (which as you pointed out is unrealistic given the NP-hardness of the problem), in fact they are explicitly trading placement quality for lower scheduling latency - and are mostly designed to schedule really fast.

Mainstream schedulers make no distinction between service & batch jobs out of the box - you need to explicitly go out of your way to implement differential preferences when it comes to scheduling latency for example.

I am surprised cost isn't a concern for you guys, I actually had assumed you went all in on bin-packing+firecracker oversubscription to maximize your margin.

I have my own fair share of complaints about k8s, but I can't say the author articulates clearly what is wrong with k8s scheduler exactly.

IMO it's one of the better part of k8s. The core scheduler is pretty well written and extensible through scheduling plugins, to implement whatever policies you heart desires (which we extensively make use of at Netflix).

The main issue I have with it is the lack of built-in observability, which makes it non-trivial to A/B test scheduling policies in large scale deployment setups because you want to be able to log the various subscores of your plugins. But it's so extensible through NodeAffinity and PodAffinity plugins that you can even delegate part (or all!) of the scheduling decisions outside of it if you want.

Besides observability, one issue we've had to overcome with k8s scheduling is the inheritance of the Borg design decisions around pod shape immutability, which makes implementing things like oversubscription less easy in a "native" way.

I have a somewhat biased opinion since I work in ML close to people building recommender systems at Netflix, but I think Tidal is a great showcase of the pros and cons of algorithmically curated content VS human curated. Tidal playlists are mostly human-curated and excellent. However it does not scale: it does not cover the long tail of wide ranges and nuances of musical preferences, nor does it cover such preferences over a large range of time. Their recommendation engine is orders of magnitude worse than Spotify's one. On the other hand, I also find that Spotify sometimes pushes algorithmic curation too far: one example of that is when they used to force an algorithmically-driven order of songs when playing an album instead of a linear order (until Adele complained loudly). I think there's a balance here: Spotify could benefit from some human-curated playlists that don't only cover mainstream genres, but Tidal should definitely invest into their recommendation engine which is really subpar, especially so because a much smaller user-base makes this problem harder to solve well.

> Like, if this person cant say "Look at me, I am UNUSUALLY INTELLIGENT!" then who can?!

No one, that's my point. If academia has not taught to the author that his intelligence isn't unusual, the workforce of his new employer certainly will. Listing github stars and twitter followers in the second paragraph as an achievement to me transpires lack of maturity and a need for external validation. On the bright side, being good at self-promotion when entering a mega-corporation will make sure he has a great career in front of him.

The tone of the post is pretty off putting. It reads as a "look how smart I am!" article - the author doesn't even pretend to be modest.

Regarding denormal/subnormal numbers mentioned as "weird": the main issue with them is that their hardware implementation is awfully slow, to the point of being unusable for most computation cases with even moderate FLOPs

Because enforcing the law is the basis of a functioning society? If citizens don't like the laws, it's in their power to change that (assuming of course a functioning democracy where elected officials are not just vessels for corporate & wealthy people's will).

Promoting the non-enforcement of the law is both immoral and erodes the foundations of a civilized society.