Hacker Newsnew | past | comments | ask | show | jobs | submit | totetsu's commentslogin

The opus models seems pretty adept and extracting structured data from ocr https://www.ocrarena.ai/battle

With llms there is a secondary training step to turn a foundational model into a chat bot. Is these something similar going on with these image generation models, that is making them all tend towards making pretty clean images and stopping them making half eaten food even if they have the capabilities?


In terms of prompt adherence, there are two issues with most image generation models, neither of which apply to Nano Banana:

1. The text encoders are primitive (e.g. CLIP) and have difficulty with nuance, such as "partially eaten", and model training can only partially overcome it. It's the same issue with the now-obsolete "half-filled" wine glass test.

2. Most models are diffusion-based, which means it denoises the entire image simultaneously. If it fails to account for the nuance in the first few passes, it can't go back and fix it.

I believe some image generation AIs were RLHFed like chat bot LLMs, but moreso to improve aesthetics rather than prompt adherence.


on a side note, I wonders if anyone submitting code to github is feeling the same way about the "duplication detection filter" type AI guardrails.


I thought these “lawful intercept” organisations had their taps inside the data centers after https tsl to the user had already been terminated. And so the infamous ssl removed here slide from prism.


> I thought these “lawful intercept” organisations had their taps inside the data centers after https tsl to the user had already been terminated.

How would that actually work?

TLS runs on the client and the server. There's no "TLS magic box" in between.


https://docs.nginx.com/nginx/admin-guide/security-controls/t...

Like let’s say you have a proxy server like Nginx on a server with a public facing ip address and then it also has access to a private subnet where your application servers are running. A visitor to your website’s browser make a secure https connection the nginx server where https would be terminated and then it would proxy traffic in plain http over your internal private subnet to the app server. And your are in a five eyes country where your intelligence services took it on themselves to follow the nsa or fbis instructions and plug a network device into those private subnets of all the big service providers inside their datacenters that is configured in something like a promiscuous way so it receives all the packets for any device on the network. Then those packets somehow end up in a big nsa datalake.. or something along those lines


That's a fancy of way of saying "not using HTTPS" which may be what average incompetent shops were doing, but isn't using HTTP everywhere which is the security standard.


But the private subnet does not leave your server.


For large sites, the private subnet is an actual network, with dozens or potentially hundreds of machines on it.

(Or was, back then. These days you can probably collapse all of that into a single medium-sized epyc or something.)


> (Or was, back then. These days you can probably collapse all of that into a single medium-sized epyc or something.)

I know where there are Sun V880s still running Oracle databases in a biggish cluster.

Their processor power, memory capacity, and storage capacity are exactly equivalent to a Raspberry Pi 4 with a biggish SD card.

We have come a long way.


> TLS runs on the client and the server. There's no "TLS magic box" in between.

If there was it'd be called Cloudflare :)


Too soon :-D


https://element.io/en/help#encryption-device-verification

> After Alice logs in on a new device, she uses her cryptographic identity to demonstrate to Bob that the new device genuinely belongs to her, rather than being added by someone else with access to her account. She can do this either by entering her recovery key (which gives the new device immediate access to her cryptographic identity ), or by carrying out an interactive verification from an existing verified device.


So is this like the Signal PIN which is required when installing on a new device? If you forget, the cryptography changes and old contacts are warned that signatures are rotated, right?


Yes, the purpose is the same but the UX is a bit different.


Quite. I have yet to manage a verification between clients.

I have had all variations of clients ignoring requests, reporting requests only for the requesting client to ignore the response. Both ends quitting declaring that the other end cancelled, asking for the other end to input a code while the other end shows no interface for doing so.

It marked the end of me using Matrix as a platform. I'd go back to the old IRC channels if there were anyone still there.


I have never failed at that. Worst case I type my recovery key and done.

I still have my encrypted messages available from 2020


People still use IRC


If by bit different you mean absolute nightmare then yes


imho it's the best out there

- no unnecessary coupling to a phone client

- no coupling to any other client - I can just put my recovery key in and be verified without having to deal with other apps.


More like the safety number / QR code.

The numerical Signal PINs are basically just for when you bootstrap your Signal identity from a telephone number.


Except Signal PIN appears to be trivial to bruteforce for Signal itself, unlike this properly secure verification.


Maybe they need a secret ‘Second Rebble’, hidden within Pebble, to take over if it collapses again.


It's open source now, so that's already taken care of.


from about 8:30 in this lecture there is a good description of the phrase "ecosystem services" https://www.abc.net.au/listen/programs/philosopherszone/envi...



I get so frustrated watching people fuss around in VSCode because they're stuck in it and they've never had the opportunity to see all the intuitive and more workable tools that a.. just part of the basic OS they are using. .. like keeping their console a tab taking up 1/4 the screen and trying to read a stack-trace ..


Zachary Stein makes the case that conferring social statuses to Artificial Intelligences is a ex-risk. https://cic.uts.edu.au/events/collective-intelligence-edu-20...


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: