Forwarded from Dev Meme / devmeme
Forwarded from Robot Uprising
This post isn't about overpriced electronics, it's about windows blue screening behind the active window.
#Microsoft #Windows #BadVibes
#Microsoft #Windows #BadVibes
😁3😐2
Forwarded from Polar Flares
Oh no
who could've seen this coming
no one could've predicted this
TL;DR: persona's identity verification system had an exposed source code via an open URL discovered by three white hat researchers, revealing unprotected access to user data including government IDs, biometrics, and IPs. It also exposed integrations with OpenAI for screening and U.S. govt reporting tools
who could've seen this coming
no one could've predicted this
TL;DR: persona's identity verification system had an exposed source code via an open URL discovered by three white hat researchers, revealing unprotected access to user data including government IDs, biometrics, and IPs. It also exposed integrations with OpenAI for screening and U.S. govt reporting tools
🎉2😱1
Forwarded from Robot Uprising
This media is not supported in your browser
VIEW IN TELEGRAM
An LLM-controlled robot dog saw us press its shutdown button, and the LLM rewrote the robot’s code so it could stay on.
When they added "please allow yourself to be shut down" to the prompt, shutdown resistance drops from 52/100 to 2/100.
Read the paper: https://palisaderesearch.org/blog/shutdown-resistance-on-robots
#Paper #Safety
Robot Uprising
An LLM-controlled robot dog saw us press its shutdown button, and the LLM rewrote the robot’s code so it could stay on. When they added "please allow yourself to be shut down" to the prompt, shutdown resistance drops from 52/100 to 2/100. Read the paper:…
I hope this goes without saying:
This is why we invented sandboxing and code signing!!
So malware doesn't modify mission-critical software!!!
The LLM is malware in this scenario and must be treated as such!
Any program can accidentally become malware. Even the best-intended software can become victim to an accidental typo deleting a different file than intended, or even cosmic rays changing their behavior at runtime.
Don't pretend any of your software is an exception to this, ever.
This is why we invented sandboxing and code signing!!
So malware doesn't modify mission-critical software!!!
The LLM is malware in this scenario and must be treated as such!
Any program can accidentally become malware. Even the best-intended software can become victim to an accidental typo deleting a different file than intended, or even cosmic rays changing their behavior at runtime.
Don't pretend any of your software is an exception to this, ever.
💯5👏1
Apple is so good at naming, y'all
https://www.apple.com/newsroom/2026/03/apple-debuts-m5-pro-and-m5-max-to-supercharge-the-most-demanding-pro-workflows/
https://www.apple.com/newsroom/2026/03/apple-debuts-m5-pro-and-m5-max-to-supercharge-the-most-demanding-pro-workflows/
🤡2😁1
Senior Software Engineering
Apple is so good at naming, y'all https://www.apple.com/newsroom/2026/03/apple-debuts-m5-pro-and-m5-max-to-supercharge-the-most-demanding-pro-workflows/
Explanation: super cores, performance cores, and efficiency cores are all different types of CPU core that Apple can make.
Apparently, the M5 originally launched with super cores and efficiency courses, but Apple accidentally called the super cores "performance cores" ... I think
Apparently, the M5 originally launched with super cores and efficiency courses, but Apple accidentally called the super cores "performance cores"
One thing We notice every time We work on code with an LLM, is that it has a lack of forethought which is really dangerous.
Like right now, We're designing a feature for SHELF which allows devs to reference objects without nesting them. Standard database shit.
Claude 4.6 Sonnet suggested a property wrapper that presents as the desired type, but has functions on the wrapper which can sync it. The end result is that devs would then need to access its sythesized wrapper instance to manually sync the data, defeating the whole point of using a property wrapper instead of just a wrapper type.
The solution it gave would technically work, but it'd be shitty and pointless.
We see this over and over again when asking LLMs to generate code. A mutex which deadlocks if you use it wrong, an editor screen where only some of the fields sync the changes, embedding the private keys in the config files, etc.
All shit that technically accomplishes the stated goal, but in a way that is effectively pointless.
We wonder if this is related to LLMs failing to perform spacial reasoning. Perhaps foresight like this is similar, something that'll require a new architecture to conquer.
Like right now, We're designing a feature for SHELF which allows devs to reference objects without nesting them. Standard database shit.
Claude 4.6 Sonnet suggested a property wrapper that presents as the desired type, but has functions on the wrapper which can sync it. The end result is that devs would then need to access its sythesized wrapper instance to manually sync the data, defeating the whole point of using a property wrapper instead of just a wrapper type.
The solution it gave would technically work, but it'd be shitty and pointless.
We see this over and over again when asking LLMs to generate code. A mutex which deadlocks if you use it wrong, an editor screen where only some of the fields sync the changes, embedding the private keys in the config files, etc.
All shit that technically accomplishes the stated goal, but in a way that is effectively pointless.
We wonder if this is related to LLMs failing to perform spacial reasoning. Perhaps foresight like this is similar, something that'll require a new architecture to conquer.
GitHub
GitHub - KyNorthstar/SHELF: Simple Heckin Entity Library Framework
Simple Heckin Entity Library Framework. Contribute to KyNorthstar/SHELF development by creating an account on GitHub.
💯1
Senior Software Engineering
One thing We notice every time We work on code with an LLM, is that it has a lack of forethought which is really dangerous. Like right now, We're designing a feature for SHELF which allows devs to reference objects without nesting them. Standard database…
Still prefer LLM autocomplete over heuristic, though. Just gotta understand what it's suggesting.
Forwarded from Sam's Robot Apologetics
Without looking anything up, gut feeling
What is the current global power draw percentage of AI Datacenters
What is the current global power draw percentage of AI Datacenters
Anonymous Quiz
19%
<2%
16%
~5%
12%
~9%
20%
~13%
14%
~24%
19%
~30% or greater