I have had fairly bad luck specifying the JSONSchema for my structured outputs with Gemini. It seems like describing the schema with natural language descriptions works much better, though I do admit to needing that retry hack at times. Do you have any tips on getting the most out of a schema definition?
Constrained generation makes models somewhat less intelligent. Although it shouldn't be an issue in thinking mode, since it can prepare an unconstrained response and then fix it up.
Not true and citation needed. Whatever you cite there are competing papers claiming that structured and constrained generation does zero harm to output diversity/creativity (within a schema).
That is clearly not possible. Imagine if you asked a model yes/no questions with a schema that didn't contain "yes".
In general you can break any model by using a sampler that chooses bad enough tokens sometimes. I don't think it's well studied how well different models respond to this.
I mean that's too reductionist if you're being exact and not a worry if you're not.
Even asking for JSON (without constrained sampling) sometimes degrades output, but also even the name and order of keys can affect performance or even act as structured thinking.
At the end of the day current models have enough problems with generalization that they should establish a baseline and move from there.
To add to the confusion, you can also just use Gemini via API (without Vertex AI). It shows up as a separate item in billing.
In the (latest, of three different) Go SDK, you can use either Vertex AI or Gemini. But not all features exist in either. Gemini can use uploaded files as attachments, and Vertex AI can use RAG stores, for example. Gemini uses API key based authentication, while Vertex AI uses the traditional credentials. All in the same SDK.
I suspected as much. It would have been too obvious of an attack vector for something so sensitive. Then obviously my argument falls apart, since it’s no longer saves any config.
That said, you can still benefit from pub keys by having good infra and key rotations to prevent some attacks like message replay after months. Putting such a requirement on customers is pretty doomed because of the workload, processes and infra required.
1. Haruhi is based on light novels, so has to actually perform to get a release. Japanese market is upside down, the anime often goes to free to air to support a manga release where the real money is made (I have no idea how this works economically this is just how its explained to me) as there isn't any more manga or light novels to release, the likelihood of another season is low. It was sort of always a passion project.
The studio being firebombed probably does not factor much into it. Kyoani and Kadokawa have beef, but Kadokawa can easily contract it to another studio to do. They just don't want to because of 1.
Also don't forget to watch Disappearance after the 2 seasons.
It is more likely that a given anime is used as advertising for a light novel/manga series than it is for the entire thing to be faithfully adapted. Just off my the top of my head, shows I've watched that are never, ever going to get animated conclusions: Grimoire of Zero, 86, Kemono Jihen, Drifters, High School of the Dead. Might seem short, but I'm lucky in that my tastes tend towards big-name shounen stuff and anime originals that either lurch towards a conclusion over a decade or were written from the beginning to be one or two cours, respectively. For less-well-known series, seinen, shoujo, etc... Well, I hope you enjoy S1.
Tanigawa released a new volume a few years ago, like 10 years after the last one. But still, it probably doesn't have the popularity it once had so a new animation sounds unlikely.
The most recent LN is volume 13, which came out in November 2024; it's on my shelf but I haven't got round to reading it yet. (The big gap between volumes was between 11 and 12: frankly I'd assumed that the author intended the end of volume 11 as the ending of the series. Volume 12 was in 2020 and was about half stuff the author had already published in other places.)
I don't know how anime series economics works these days -- AIUI traditionally the live late night TV broadcast was effectively an advert to get the hardcore fans to buy the extremely expensive Japanese market DVD/bluray sets, which were what brought in the money. But I expect streaming has changed things a lot.
They still do Blu-ray releases with suspiciously obvious art fixes. My intuition is that the model/merchandise market is also a lot bigger than it used to be in the early 2000s (thanks in part to Haruhi). 3D/higher quality printing and hig-fidelity digital crafting probably played a big role there.
I was disappointed that it was a bit bitter, yes, but not in the category of “won’t do it twice”. Such that I even tried multiple cartridges. I’m retirement age, though, so maybe my taste buds are shot.
I don't think the "be nice to everyone" is the thing people are annoyed with, rather it's the "you will be canceled if you step out of line even once" that comes along with it.
what's an example here of unfair cancellation that you are thinking about? I feel that most of those who have been "cancelled" have been so because of fairly egregious stuff like sexual assault, clearly racist speech, etc.
The irony being of course, that the current wave of performative anti-woke articles from PG, Zuck et al is based mostly on fear that they'll be cancelled by the incoming administration if they appear to be too kindly disposed to minorities...
I mean, it'd be a remarkable coincidence if firms queuing up to renounce their DEI initiatives and argue that the President's new bestie was doing a great job of promoting free speech were doing so without regard to the imminent inauguration date of a President who's never exactly disguised his loathing of such things and his desire to make life miserable for people he sees as "woke" (or his fondness for recalcitrant critics for that matter). Even PG himself is busy observing (on Twitter) the trend of removing pronouns from profiles. In fairness to PG, unlike Zuckerberg he's not actually rolling back anything he's been doing for most of the past couple of decades, but it'd be a helluva surprise if he'd just forgotten who was in charge when he got suspended from Twitter in his haste to lavish praise on the president's right hand man for doing such a great job of removing censorship from Twitter
The incoming administration has never been unkind to minorities. If they have been unkind to criminals, yes, but minorities, never. If you conflate that with minorities, then that kind of demonstrates who you protect: Not the minorities.
It’s all words anyway. Wokes hate white people with passion, and that’s why they are not getting half the USA against them.
And yet we're not observing a trend of CEOs rushing to talk about crime in their haste to accommodate the new regime, we're seeing them rush to update their moderation guidelines to stress that "gays are freaks" and "immigrants are grubby, filthy pieces of shit" must be understood as valid political commentary and not abuse, delete pronouns from their profiles and shut down their diversity programmes.
Can you get a working game generated? All I ever see is at least "Script error." with no clear way to debug. Having bolt try to fix it seems to just regenerate bits of the app but has failed to actually fix any problems so far.
After some initial pains with colima, I tend to agree. Mostly, just needing to specify some VZ args[0] so I could run x86_64 docker images on my M-series.
Is there something in these desktop UIs that colima is completely missing?
"some initial pains" = Colima VM running out of resources running kind, so I had to raise the CPU and RAM, and then raise the fd's in the VM itself to get it to work. but now it works!
Could this be the answer I needed to run an SQL Server image that refused to run on my M3 MBP? I was about to, sadly, try Docker Desktop, because of that.
Yep, that whole example after the float cast is bit silly. After the cast, the value can never be an empty string, or null. Not to mention, comparing a floating point value to an arbitrary, literal value (of zero in this case) is potentially problematic [1].
