Vachyas

The most interesting part about this to me was how they bench/compare it, like in the example with Fable:

  "Given the same prompt, Astra for Law returned two closely matching precedents; in the litigation example, Claude Fable 5.1 returned a holding that had been reversed on appeal, while in the transactional example it reported finding no such case."

It made me wonder if a good deal of law is about finding a way to work in statements with clear precedents without your opposition noticing and then later drawing upon them in court (as settled precedents, in your favor) after the opposition (perhaps implicitly) accepted it. That would clarify a lot about why some lawyers need to spend so much time pouring over and memorizing past cases (even ones that are only tangentially related); because anything they miss could be used as a potential trojan horse by the opponent.

If this is true that must mean there are a good deal of cases settled using precedent "gotchas" where both sides knew that without the "load-bearing" precedent the outcome would've definitely been the opposite. (i.e precedents almost always trump even valid arguments)

I'm not sure if "habbit" was a deliberate misspelling here, but regardless I think it makes the comment better! Lol.

Confusing enough that I see ads from major retailers mixing it up (or maybe it's on purpose?) listing "Airpods Pro 4" only to click the link and see it's the Airpods 4.

FWIW being able to effortlessly change volume with the watch dial is like 20% of my reason for wearing one.

Same, and I even always got the Pro, too. I'm still (happily) on 13 Mini.

I'm curious, what iOS are you running on it?

Lmaoo great comment

Maybe this is where all the books an LLM consumed during training goes.

$12 worth, it seems

When you put it like that, it really does, lol.

I think one hypothesis along these lines is that, if allowed, due to the limitations of human language you described, LLMs will gravitate towards "inventing" their own language (which, due to training pressures, may even resemble english from the outside, but contain deeper, "true", meaning within), but that we should do our best to prevent this even if it bottlenecks reasoning capabilities since it would cut off our ability to read its "true" thoughts and detect misalignment

See: https://openai.com/index/chain-of-thought-monitoring/

Quote below:

  Chain-of-thought (CoT) reasoning models “think” in natural language understandable by humans. Monitoring their “thinking” has allowed us to detect misbehavior such as subverting tests in coding tasks, deceiving users, or giving up when a problem is too hard.

  We believe that CoT monitoring may be one of few tools we will have to oversee superhuman models of the future.

  We have further found that directly optimizing the CoT to adhere to specific criteria (e.g. to not think about reward hacking) may boost performance in the short run; however, it does not eliminate all misbehavior and can cause a model to hide its intent. We hope future research will find ways to directly optimize CoTs without this drawback, but until then

  We recommend against applying strong optimization pressure directly to the CoTs of frontier reasoning models, leaving CoTs unrestricted for monitoring.

  We understand that leaving CoTs unrestricted may make them unfit to be shown to end-users, as they might violate some misuse policies. Still, if one wanted to show policy-compliant CoTs directly to users while avoiding putting strong supervision on them, one could use a separate model, such as a CoT summarizer or sanitizer, to accomplish that.

Agreed. Youtube recommendations are genuinely great for me. Most of the time I'll be recommended so many more good videos than I have time for, that my "watch later" playlist only keeps growing.

Compared with, say, Netflix, where even though I've been rating everything I watch on there for 5+ yrs, the recommendations still barely feel personalized (if anything, it feels like it personalizes which premade "top list" to show me, but not the titles within them...but it does personalize the cover art/thumbnail, lol).

correct, and I think games which needed action keys in addition to arrow keys (but no mouse) preferred arrow keys on the right so the left hand could handle the other moves and stuff.

I write with my left too but using right handed peripherals feels natural. I also use my phone right handed.

I also discovered the mirrored writing thing lol and it made me wonder what ambidexterity is since if ambidextrous people could presumably write non-mirrored with the other hand, aren't those completely different motions that neither hand has had any practice with?

In addition, even knowing it's not real, I feel like I can't appreciate it as much as I did (or would've) a well-made clip that I knew was CGI.

An interesting anecdote that comes to mind is playing old computer games with arrow keys, which used my right hand. I got pretty proficient with this.

Over the years, I (and I imagine many others) switched over to WASD to play newer games with mouse + keyboard, but this meant using the left hand for "arrow keys"

Now I can directly compare how proficient I am with WASD vs Arrow Keys and the result surprised me. I was way worse with arrow keys (right hand) even though back when WASD was becoming a thing I'd rebind WASD to arrow keys because it felt too weird! I would've never imagined back then that WASD could ever feel as natural as arrow keys.

Makes me wonder how much of handedness is truly innate vs learned.

Good comment, and I agree lol

I read it twice (admittedly quickly) but couldn't grasp the point even though I felt like it was there.

The link you provided is for a canvas I think rather than the convo

Good point.

So I guess while "realism" (or believability) is really good now, prompt adherence has much room for improvement.

(though put it another way, realism has always been "solved" if the model gets to output whatever it wants as long as it looks realistic, though now it looks less like a malfunction and more like an inattentive human mistake or oversight, so even when it gets it wrong it's hard to tell it's wrong without knowing what the prompt was)

That's a good example, actually.

If you asked me what I expected, since this one has "thinking", it'd be that it would've thought to do something like generate the image without Waldo first, then insert Waldo somewhere into that image as an "edit"

I really like all the compress-to-X links at the bottom and the convert from X to Y tools. Especially the Discord one with presets for different target file sizes based on Discord's subscription tiers.

I've been using server-based (online, upload required) tools for this sort of stuff, but am now going to be using this.

Pretty cool find considering I have no need for a full-fledged video editor right now, and was just checking this out for fun.

I'm honestly unsure what could be improved at this point.

Consistency? So it fails less often?

Based on the released images, (especially the one "screenshot" of the Mac desktop) I feel like the best images from this model are so visually flawless that the only way to tell they're fake is by reasoning about the content of the image itself (ex. "Apple never made a red iPhone 15, so this image is probably fake" or "Costco prices never end in .96 so this image is probably fake")

Xenon is very rare too and currently without substitute for certain medical applications, but more interestingly it produces psychoactive effects that could shed light on stuff no other substance apparently can: https://pmc.ncbi.nlm.nih.gov/articles/PMC11203236/

When you said Agent and AI, I thought there would be some way for us to resize or move elements, and have the agent figure out the right properties to change (whether it's margin, padding, top, left, and on the wrapper or whatever) ideally in a way that's cleanest WRT surrounding/existing CSS.

But I can see the more deterministic nature of the current offering being a plus too since there's no worry about the agent doing things you didn't "approve" or in the "wrong way"

The video was really good, and the UI looks fun too.

If I understand correctly, is this not as useful for frameworkless html/css/js development? Since when you make edits using browser-built-in-devtools it can and does modify the actual css files (in-memory, of course) which you can use to copy-replace with entirely (assuming no build/bundling step aswell).

If so and this allows you to use any framework and still have that kind of workflow, that's fantastic. Half the reason I don't like using frameworks is because I lose the built-in WYSIWYG editor functionality. Guess I'd still lose the usefulness of the built-in js debugger, tho :(

What you described sounds plausible (expected, even).

But

>Raw parameter counts stopped increasing almost 5 years ago

Really? 5 years ago? Until just about 3 years ago OpenAI's latest offering was only ChatGPT 3.5

Most of the models people talk about now didn't even exist 3 years ago let alone 5.

Even now, I don't know if parameter count stopped mattering or just matters less

For example, I have no idea if the new Mythos is MoE but I'm pretty sure it's more parameters.

Yea and I'm doubtful we'll see a service willing to do their own post-processing per-query while also being at the whim of Youtube's API (official or not).

Ultimately, I would like these features to come to Youtube itself since there's a lot of nice features built into it that would be hard for a third-party to replicate without permission (such as playing videos inline on hover, with captions).

Yea, webapps (even PWAs) still can't compete with native apps when it comes to responsiveness, but I still don't know why. I've yet to see even a demo PWA that passes the "native turing test" where I can't tell whether it's a native app or not.

Even native apps that were built with cross-platform frameworks feel a bit "off" sometimes.

I'd like to second that I wish a lot of those existed. For 1), 60fps is another good one.

It seems Youtube also removed "sort by upload date" if I'm not mistaken. The closest we can get now is the "uploaded today" filter but it's not the same since it still seems to prioritize popularity over recency, surfacing mostly second-hand sources or popular "reactions" to the primary-source videos (that also exist on Youtube!) I'm actually looking for.

Edit: IIRC they even used to have an "uploaded in last hour" filter, but I'm not sure. Can anyone confirm this?

I was hoping to finally see an advanced time filter so I could do something like "over 2 minutes" but it seems you've only got the same ones Youtube has (< 4 minutes, 4-20, and > 20).

If it's an opaqueness restriction with the API or something, I'd like to suggest letting us at least combine the provided ones, so I could do something like (4-20) && (> 20) to get "over 4 minutes" which doesn't exist on Youtube but seems pretty useful.

Another thing that would be useful is filter-by-channel since the search function within Youtube for searching a channel's uploads (using the search button on a channel's page) is a significantly nerfed version of their usual search function.

> And now, with the assistance of AI, I can go much farther in 10 hours and deliver a more complex project. But that means that someone else trying to replicate this execution is still going to need around 10 hours to replicate it.

The blog post does touch upon this. The key difference, I believe, would be that compute scales in a way "meat-heads" doesn't, where if the other person has 100x the capital to throw at it, they could do the same 10 hour thing in 10 minutes.

Basically, what I got from it was that innovation has never been truly scalable enough to create the "dark forest", since hiring more and more engineers saturates quickly. But if/when innovation does become scalable (or crosses some scalability threshold) via AI, that could trigger a "dark forest" scenario.