707 karma · joined June 9, 2025
Previously, I would need to do the trade-off calculation. How urgently does this need to ship, and do we have time to rework this? What are the deal breakers that need to be addressed, versus what things are best practice/ideal for maintainability? How did their last code review go and do they need a small win right now?
There's no more "nit" comments tagged as nits: just things to fix. It's de-personalized in the sense that we can both at least pretend/have plausible deniability and blame the model for being dumb, as opposed to the person making mistakes. I flat out told someone that a PR was not solving the right problem earlier, and neither of us thought it was a big deal. I could give the technical guidance and suggest a path forward to "help Claude understand better".
Then someone makes a shitty comment. Is that correct?
And even when they are: they sure seem to bet against Moore's Law or just the general tendency for things to get better/efficient over time.
It's frankly remarkable how capable the models have become that we can run locally now on a decent laptop.
The same thing happened with image generation. I've had arguments with people that image generators are killing the environment, but I can do it in 20-30 seconds on my GPU. No one bats an eyelash when I play 20-30 minutes or even hours of a video game on my GPU, but the images are burning down the planet.
It's slightly maddening.
I'd argue this isn't true today, but that the loop for incorporation is long (ie, the next training or finetuning run).
>A multimodal model knows about images of pipes and facts about pipes, but doesn't know pipes; it doesn't have literally first-hand experience with them.
Wouldn't this mean that any human who hasn't seen a pipe in person or interacted with it, similarly doesn't "know" a pipe? Most of us haven't interacted with the vast majority of "things" in the world, yet we're still able to build a model and abstractions for them such that we can reason about them, right?
I suspect this may have depended on the specific framework. I quite literally could not get Claude (in Cursor) to give me a basic Micronaut setup in a fresh workspace with essentially a "hello, world" API. I would guess that if you're using something like Python and FastAPI, it might have been an easier task or better represented in the data.
The difference that I observed in the Opus 4.5 era is that Claude could take a service framework it has never seen before (proprietary corp) and figure it out.
I find this somewhat puzzling. I thought things were moving quickly, but at this time last year I couldn't even get Claude (using Cursor) to spin me up a service skeleton that would compile, let alone do anything meaningful.
I know it feels like a long time somehow, but it was only between November and February that things started to actually somewhat work without significant hand holding. Even now, it seems like we're still figuring out how to fully leverage the current models and tooling, even in organizations that have largely gotten on board.
It'd be weird if they were doing this, since it would mean the context window size was a lie and that the API would presumably reject requests whose expanded form went over the 1m limit. For someone using pxpipe with an effective context compression of 90% in some instances, it'd hit the limit at barely 100k.
There are many dogs that have been trained to press buttons corresponding to words, in the extreme case tens/hundreds of buttons/words, and they can even construct rudimentary sentences. It doesn't seem insane to me that we could perhaps do a very rudimentary version of this for dogs, given a large enough training set.
Honestly, I think it's less so (for some of us) that we think they're "more than matrix multiplication gadgets good at compression", so much as thinking that perhaps what our brains are doing is not so dissimilar.
A materialist view of the world could support the idea that intelligence itself may just be a series of predictions from a big compressed multi-modal dataset. That's not to say that LLMs are doing it in a way that is even close to how our brains are doing it, but we also don't understand how different it may be, and how much utility we can get out of them even with the current architecture.
There's a reason Yudkowsky only thinks AI can be stopped by literal missile strikes against data centers.
I think people aren't paying enough attention this part right here: I think the architecture of atproto creates a different set of incentives. People running communities are more likely to face community pressure to ban users, or de-federate with other communities they don't like, and your open/decentralized network fractures apart, creating little silos.
People running infrastructure connecting them to a global network/resource are less likely to feel those pressures. No one asks an email provider to ban another webmail host. The architecture colors the perception.
The biggest problem Bluesky has had is that by virtue of being the only effective relay at launch, people viewed it more like the former and less like the latter. But I think if we start to see more places like Blacksky and Eurosky get mindshare, we can shift it back.
The job of writing is to be read. The job of code is not to be read: it merely helps the humans who write the code. If the code is no longer read by humans, then it no longer needs to be well composed.
We want human writers because humans have a better grasp on the ideas and stories that can best move us, and are just better at crafting writing that engages with us. Machines will likely struggle with that for some time, though they can certainly augment and help human writers improve their prose.
I get the sadness. I was the one decades ago comparing beautiful code to poetry, and I think that's still true. But also, this type of beauty is craftsmanship that only the craftsman can actually see and appreciate. The reality is that the code just needs to work, and AI can (largely) make it work.
Also: the above is assuming that the models continue to improve to human performance, which is of course an open question. I'm not taking a stance here as to whether that will come to pass, merely arguing that it is fundamentally different from writing.
I don't think that was a recommendation for people to do, but to demonstrate that it doesn't require much in the way of resources. That VM could probably host hundreds of people, since relays scale and de-dupe shared content.
If someone said that email or a personal website was inexpensive and could be hosted on a cheap VM, I don't think you'd make a comment like the above dismissing it as impractical: the idea is that if it's that cheap, then there are people that can host cheap/free (maybe ad supported) relays, similar to how cheap/free webmail exists today.
But I'm sure for at least some folks, this is true, given recent valuations.
Is it, though? This scrappy startup was able to take a large(-ish) open weights model and adapt it. Why can't the frontier labs do the same cost effectively?
>If they want to get acquired, then they should show that they know what they're doing.
I'm sure they would do so under an appropriate NDA as part of negotiations. I'm not sure why you think a full public disclosure is necessary.
My guess is that they're angling for an acquisition.
There's a dark future timeline where snark becomes the new "don't use any em-dashes" that people inject even more of to intentionally look human, since empathy and patience read as..robotic.
Words apparently don't mean anything anymore.
I'm suggesting that while LLMs are bounded by physical reality, that you actually don't know what that bound is. Just a few years ago we would have thought it a fantasy to have a conversational model run on a phone.
Even if you could compute it now, that would still be tied to current architectures. With appropriate incentives, we'll continue developing hardware to make these models more efficient to execute. It's very likely that you'll be able to run a Fable caliber coding model on your phone in the next five years.
>This matters because unlike cars LLMs are only doing stuff we can already do using our brains, just several orders of magnitudes less efficiently. Cars can at least take us distances we would never be able to using our muscles.
But that's not largely true of cars. The majority of trips are five miles or less and could easily be replaced with a bicycle. While I might personally use a bicycle, the majority choose a car to save a bit of time and effort.
So, please continue to enjoy your car, and I will continue to enjoy ready access to an LLM for a variety of other tasks. My inference energy costs are almost certainly less than your vehicle usage. ;)
This is a silly comparison. There is a certain quantity of energy stored in oil, so we know what peak efficiency looks like. We don't actually know what amount of energy is required to solve certain problems. We quite literally have models with quite a bit of capability that can run locally on a phone today, right alongside Stockfish, for example.
And this is to say nothing of work happening now on new hardware approaches, such as Normal Computing's work on thermodynamic matrix math: https://www.normalcomputing.com/blog/a-first-demonstration-o...
That said, this feels like a strange tangent: I'm not sure it's that important that the models be as energy efficient as a human brain. We don't avoid cars because they're less energy efficient than our legs. ;)