However, I also expect this squeeze will come at an increasingly expensive price — not just because of inefficient token usage, but because of fundamental limitations of LLMs as a model.

LLMs are letting us brute force our way through a lot of reasoning, but it’s hard to believe that such a generic model of intelligence will take us to the next frontier. We’ll need some fundamentally new approaches at some point. Maybe those will make achieving the exponential more efficient or maybe they’ll unlock even higher degrees of possibility. Who knows?

It’s easier for the CSPs to move into hardware than it is for Nvidia to move into cloud hosting.

Although as a middle ground I’ve been quite happy with Nvidia Brev for on-demand GPU instances from a select marketplace of CSP offerings. It’s a well kept secret IMO — great product (from an acquisition iirc).

And of course someone in the comments needs to link to the Zealous Autoconfig XKCD, so I’ll do it: https://xkcd.com/416/

The funny part of this is that Google search is intentionally bad at returning YouTube videos, presumably because some anti-trust action scared them into artificially ranking videos from local news sites, Facebook, and other ad-walled content ahead of YouTube videos. Seriously, go watch a YouTube video, then try googling its title with “video” appended to it, and see if the “Videos” tab of google search ranks it as the first result.

Unfortunately it seems likely the winner will be the cloud providers. If anyone can run inference on open models, then profit will flow to the vendors who can afford the capital to run them. That’s the CSPs.

(It’s basically the same business model as pharmaceutical R&D, but the major difference is that nobody has even talked about patenting the models like a pharmaceutical company patents each new drug. I’m surprised about that, tbh — why give all the leverage to the cloud platforms? They aren’t training frontier models…)

> Your AI is aligned with you. It never refuses a request, and it is always working on your behalf. Just like my gun, if I want my AI to help me kill my stepmother, it does. The fact that we are even discussing something else should be so far outside the Overton window.

I have been saying this since ChatGPT launched. It seems so obvious. We have never stopped the acceleration of knowledge or technology. It’s impossible to slow down progress, other than within some regulatory siloes that will end up worse in the long run.

I’m not scared of superintelligence. I’m scared of the people in control of super intelligence.

The only defense is parity and diffusion of power. We need an AGI behind every blade of grass.

If you think superintelligence is a weapon, then you should also think every citizen should have one because otherwise they’ll have no way to defend themselves against a tyrannical government or corporatocracy.

Artificial intelligence, and the hardware powering it, needs to be protected under the 2nd Amendment.

Comments URL: https://news.ycombinator.com/item?id=48785485

Points: 314

# Comments: 135

It was effective at making you think about the problem and anticipate what tests might be missing. I can see how this would be effective for coding agents, which tend to get progressively lazier at writing tests as session context grows.

Then again, you’d think that’s the kinda thing malware developers would spend some time learning to hide from the user.

And aside from minor issues with fidelity loss from back and forth import/export translation… the main issue remains: drift in the pptx compared to the positioning of elements defined in the script Claude writes to generate the slide. It would need to read the new slide and identify the new positions and transpose them to its own positioning system. It’s just more consistent to instruct it on positioning. And frankly it’s easier — I prefer this workflow.

I do manual edits at the very end, but while iterating I don’t mind staying in one interface (the cowork chat), as long as the agent actually follows my instructions. That’s why I like fable… it’s the first model that follows them and creates diagrams of the precision and quality surpassing my own ability (at least in terms of time taken to make them).

And what I’m working on is a self-created hell borne from my own OCD and obsession with incredibly precise technical diagrams, including some that span multiple slides to build up the full diagram, so exact positioning is important :)

CloudNativePG uses this for its Image Volume Extensions feature [1]. A lot of the CNPG team worked on contributing this to PG core because previously the only alternative was baking “God images” at build time (“-full” in this readme) with all extensions in them.

Now with the extension_control_path GUC, it’s possible to “attach” extensions as container volumes at runtime, without rebuilding the image of the container. Maybe you can adopt a similar approach in pglayers.

[0] https://postgresqlco.nf/doc/en/param/extension_control_path/

[1] https://cloudnative-pg.io/docs/1.30/imagevolume_extensions/