twelfthnight

Wrote this elsewhere, but this printer prints a full volume in 2 hours. In many scenarios it's 10x faster than FDM with better resolution and compatibility with advanced materials, like food grade / skin contact, ultra strength and rubber-like consistency.

You actually would probably want the wash/cure and VOC extractor upgrades so it's more like 1200$ before it's useful, but if you have the clean room space and ability to handle the toxic resin, functionally resin printing is typically better than fdm for single color prints.

My first printer was resin without a dedicated wash/cure set up and it was difficult enough that I sold the printer and got an FDM. Currently I'm using an Snapmaker U1 and it's really awesome and much easier. It also prints in multicolor seamlessly.

That said, if you do spend the extra for this along with the wash/cure and the voc extractor, the demos are really impressive. It seems like this printer is blazing fast, for single color prints it will print basically anything at better resolution than an FDM in <2 hours, whereas FDM could take anywhere from 4-12+ for a similar part with bigger layer lines.

Why not use a text embedder for the unstructured data and concatenate with the structured data?

For example you could freeze most of the layers of the embedder but let the final ones learn. Then you wouldn’t need to do either feature or prompt engineering?

I wonder if most people like/don’t care about recommendations and don’t want to spend time giving companies more feedback.

I work in streaming recommendations and basically no one voluntarily gives extra feedback, and most initiatives that aren't machine learning based on true behavior fail on A/B tests.

I’ve been in so many meetings where the outcome is to plan another meeting, and include even more folks. Whichever team brings the most folks steers the decision in their favor, and thus manages hire more unnecessary employees for political will (which then increases the need for even more meetings).

The way out is creating a singular vision (eg leadership) and assigning teams goals they can work independently on towards that vision. It is to remove dependences between teams (and thus the need for them to communicate as much), not to increase communication or Jira tickets or Gantt charts or RACI matrices.

Seems equally valid to come out of this with the takeaway that code quality _does_ matter, because poor coding practices are what led to the leak.

Sure, the weights are where the real value lives, but if the quality is so lax they leak their whole codebase, maybe they are just lucky they didn’t leak customer data or the model weights? If that did happen, the entire business might evaporate overnight.

I agree AI could probably do a decent job on Kaggle problems. Of course, almost no DS job is building models with well-defined objectives and perfect data. The DS and MLE folks I work with mostly spend their time reframing ill-posed product requests into ML systems that can be maintained and improved with feedback loops.

A _huge_ part of a DS is saying "No" to bad ideas posed by non-experts. The issue with LLMs is all they ever say is "Yes" and "Wow, that's such a great idea!"

Classic 2024 company. Started out with a few genuinely cool products and a quirky, fun attitude. Got popular, decided to exploit that good will and now everything they do seems greedy and inauthentic (regardless of whether any specific product is or not). Maybe this is a great product? But after the OP Field pricing shenanigans and OP-Z quality issues... hard to give them a second chance.

That sounds right to me. I don't work in physics, but have many colleagues who have left it for that reason.

~The said, I think Dune still does seem useful. Hard to say whether the cost overruns are a lack of good management or if the experiment is just a quagmire, though. Guess I lean towards the scientists here that it's the former~

EDIT:

Doing some research it seems the Hyper K experiment might just be better, and funding might be better directed there than Fermilab.

I don't know all the details, but I think they are starting to do novel work with the accelerator to produce neutrinos in an experiment called Dune. So there actually is important stuff being done that isn't just superceded by Cern.

The larger scale models aren't necessarily built on the smaller models, they are independent and generally less specific. They are probably even inputs to the more specific models. So, like rising sea levels and increased temps (on average) is basically guaranteed, the _exact_ change for a particular place is more uncertain.

Neil deGrasse Tyson in his Cosmos show described it like a person walking a dog. We don't know where the dog will be exactly, but we can see the person with the leash is moving in one direction, and the dog isn't going to ever get too far from him (because of the leash).

Huh. I actually can understand devs not wanting to need permission to install libraries/versions, but with a pull-through cache there's no restrictions save for security vulnerabilities.

I think it actually winds up speeding up ci/cd docker builds, too.

Because of containers, my company now can roll out deployments using well defined CI/CD scripts, where we can control installations to force usage of pass-through caches (GCP artifact registry). So it actually has that data you're talking about, but instead of living in one person's head it's stored in a database and accessable to everyone via an API.

Agreed that the argument "overprotection leads to lack of free speech in academia" is tenuous.

That said, I do wonder if we all _are_ being protected from opposing views these days. Like, we come across the opposing views but usually in a filtered / characature form on Social Media/Fox News/MSNBC. It's actually kinda hard to find stuff without spin in my experience.

EDIT

My hypothesis is that's it's just cheaper to create speculative / opinion based journalism rather than real investigation. Since the former gets enough clicks, there's not a strong financial reason to create good journalism.

The article mentions Clarence Thomas has been courted (even bribed?) by the very people paying the lawyers trying to overturn Chevron. I think it's naive to believe the judges don't have policy preferences that are strongly reflected in their rulings... If that wasn't the case the GOP wouldn't have blocked nominations from Obama to get their preferred judges in.

If you had an efficient market, wouldn't a competitor charge just a little less than others who are using surge pricing? Isn't the whole efficient market thing about how eventually prices should reach the cost of production, which would not be affected by surge demand?

Honest question, what different decisions could you make with a 48-52 interval than a completely uninformed interval of 0-100?

True, I suppose my disagreement is that I believe it doesn't go far enough to explain how big of a deal it is, and how there _aren't_ ways to deal with it without substantial, subjective intervention from the forecasters.

I've worked on weighting code for online polls, they literally rely on dozens of hand picked decisions to stay "reasonable". These decisions aren't factored into the error bars, making them appear smaller.

And as far as the fundamental style predictions, how can you use a single GDP number when Fox tells its viewers one number and MSNBC tells its viewers another?

This article does describe a faithful statistical effort, but to me it doesn't emphasize the risk of a "black swan" event enough.

This is great statistics, but it avoids the problem that moving to online polling has made it very difficult to get representative populations, so the data itself is biased in ways that cannot be counteracted by methods (only by assumptions, priors, etc). Which makes this misleading because it gives the forecast air of confidence that is unjustified.

I've worked as a data scientist for political campaigns (but no longer). The elephant in the room is that there are so many forces that make this election different from historical ones, from polls moving online away from phones (making them much much less reliable) to Trump's conviction to a completely unprecedented/fractured media landscape. Even if these forecasts were accurate, their usefulness for anyone other than people spending ad money is absolutely zero. Now that these forecast have so so much bigger error bars, I think public forecasts are actually harmful and just a scummy money grab from the media that posts them.

EDIT:

Why are they harmful? At best they make lots of people angry about something that hasn't happened. At worst they skew the election by making people not vote, or afterward serve to justify election fraud complaints ("my candidate wasn't supposed to lose based on forecasts, so fraud occured").

I think we all need to take a deep breath, accept we have little understanding what will happen, _vote_, and hope for a sane outcome.

Unless I'm missing something, I don't think this blog really defines "reason". So, like, this is a completely pointless question.

Mostly agree, although I think "good" engineering culture is so incredibly rare that it feels strange to call it "good" and everything else "bad". It's more like "regular" engineering culture vs "exceptional" culture.

In my experience I think the larger any team gets the more likely the team is to hire "regular" engineers and the quality regresses to the mean. So it's only a small set of specialized teams with unusual luck and control over hiring that can become "exceptional".

This resonates with me, although I would go a step further to say the makeup/experience of the developers is a product of the competency of leadership. Most leaders wind up in their positions out of pure luck and chicanery. I can't tell you how many lay offs I've been through where the best people are let go and the worst are kept on and unintentionally sabotage the entire engineering team.

I for one would love to have a boss that cares about those things.

So true. Communication is insanely difficult. Hiring competent, motivated engineers is insanely difficult. Working around short-term thinking executives is insanely difficult. Iterating until you stumble on a success product is insanely difficult.

How often does anyone talk about these problems in meetings? It's wild, I feel like all I ever do is argue over mostly irrelevant technical minutia.

Totally agree. And I honestly don't think Starbucks removing seating has much if any impact on reducing "third spaces" since almost everyone buying Starbucks was getting it to go anyway. This article reads more like someone being annoyed about Starbucks changing than any real evidence that third spaces are going away.

If anything, if more people are now using coffeeshops as remote offices we are likely to see new shops catered to this opening up.

I'm an American and I don't get it either. There are tons of local coffee shops _everywhere_ in America, most with better coffee, happier baristas and unique atmosphere. Have so many Americans/Canadians never... like... been to a coffee shop that's not Starbucks?

The author's library is being renovated, not disappearing. One large corporation (Starbucks) is enshittifying its experience by removing seating. I don't think the premise "third spaces are disappearing" is a valid conclusion at all.

I don't know if third spaces are indeed going away or not, but the argument here doesn't really address that.

By the way, if you don't like it don't buy Starbucks. Don't use Amazon. Don't use Twitter .(If you have the option). Not spending money/attention is more effective than complaining at changing companies' behavior.

Even in production my guess is most teams would be better off just rolling their own embedding model (huggingface) + caching (redis/rocksdb) + FAISS (nearest neighbor) and be good to go. I suppose there is some expertise needed, but working with a vector database vendor has major drawbacks too.

I think they are talking about bag-of-words. If you apply a dimensionality reduction technique like SVD or even random projection on bag-of-words, you can effectively create a basic embedding. Check out latent semantic indexing / latent semantic analysis.