Breadbrowser
@breadbrowser.bsky.social
I'm an easily jealous trans woman that likes computers and ML.
For Ai safety
Pronouns: she/her
For Ai safety
Pronouns: she/her
chutes is like super open about their service, I kind of like that. Other cloud providers should notice. chutes.ai/app/research...
November 27, 2025 at 7:53 PM
chutes is like super open about their service, I kind of like that. Other cloud providers should notice. chutes.ai/app/research...
Yeesh, I should probably comment more.
November 27, 2025 at 6:52 PM
Yeesh, I should probably comment more.
Might do it youtube style with trees and use a small Bert model for the main tree. (highly likely I will not make it)
November 27, 2025 at 6:12 PM
Might do it youtube style with trees and use a small Bert model for the main tree. (highly likely I will not make it)
There a lot more chill than Microsoft at merch, you can just buy a toaster from Microsoft. microsoftmerchandise.com/Shop/#/produ...
November 27, 2025 at 5:22 PM
There a lot more chill than Microsoft at merch, you can just buy a toaster from Microsoft. microsoftmerchandise.com/Shop/#/produ...
Lmao what microsoftmerchandise.com/Shop#/produc...
November 27, 2025 at 5:17 PM
Lmao what microsoftmerchandise.com/Shop#/produc...
honestly, it gives more of a evil vibe like this but maybe that is just me.
November 27, 2025 at 2:47 PM
honestly, it gives more of a evil vibe like this but maybe that is just me.
god i miss this site, i used to get ai news like the minute it happened on this site. i wounder why it shut down.
November 27, 2025 at 12:30 AM
god i miss this site, i used to get ai news like the minute it happened on this site. i wounder why it shut down.
i'm just saying it's heavily misleading to say it has "emergent misalignment" when there is a clear reason of the misalignment. nothing in the results is false but they clearly leave how that they trained it with positive hacking rewards and is written in a way to be misinterpreted.
November 26, 2025 at 11:07 PM
i'm just saying it's heavily misleading to say it has "emergent misalignment" when there is a clear reason of the misalignment. nothing in the results is false but they clearly leave how that they trained it with positive hacking rewards and is written in a way to be misinterpreted.
btw this paper is bs "https://www.anthropic.com/research/emergent-misalignment-reward-hacking" as "rl with hacking" has positive rewards for hacking and no hacking baseline has negative hacking rewards.
November 26, 2025 at 10:54 PM
btw this paper is bs "https://www.anthropic.com/research/emergent-misalignment-reward-hacking" as "rl with hacking" has positive rewards for hacking and no hacking baseline has negative hacking rewards.
i guess nano banana pro can just replace you with a bean. honestly i have no clue why i did this.
November 26, 2025 at 8:25 PM
i guess nano banana pro can just replace you with a bean. honestly i have no clue why i did this.
honestly i made my own with vibe coding and it works a lot better
November 26, 2025 at 7:38 PM
honestly i made my own with vibe coding and it works a lot better
I mean it deffently is, but I feel like openai should take responsibility for this stuff. Instead of hiding beside a tos.
November 26, 2025 at 6:59 PM
I mean it deffently is, but I feel like openai should take responsibility for this stuff. Instead of hiding beside a tos.
Conceptually and mathematically they are related.., but mamba predicts the next state of a continuous sequence by modeling the entire sequence as a continuous-time dynamical system, rather than processing inputs step-by-step with discrete hidden states like RNNs.
November 26, 2025 at 2:28 AM
Conceptually and mathematically they are related.., but mamba predicts the next state of a continuous sequence by modeling the entire sequence as a continuous-time dynamical system, rather than processing inputs step-by-step with discrete hidden states like RNNs.
Stuff like this huggingface.co/papers/2511....
November 26, 2025 at 12:00 AM
Stuff like this huggingface.co/papers/2511....
Is it just me or is this photo getting popular on bluesky?
November 25, 2025 at 5:02 PM
Is it just me or is this photo getting popular on bluesky?
Tried to make ship biscuits. It's just flower, salt, and water.
November 25, 2025 at 4:55 PM
Tried to make ship biscuits. It's just flower, salt, and water.
I still can't believe that Disney at one time tried to make a life size dinosaur. Source: frankmezzatesta.com/dino_demo.php
November 25, 2025 at 10:54 AM
I still can't believe that Disney at one time tried to make a life size dinosaur. Source: frankmezzatesta.com/dino_demo.php
Not to be this person I'm super confused by this paper. What are the rewards for the "rl run with hacking"? The SDF show no misalignment but when it hits RL somehow it's misaligned. The baseline is the hacking is disincentivized. So is the "rl with hacking" have positive rewards for hacking.
November 24, 2025 at 6:30 PM
Not to be this person I'm super confused by this paper. What are the rewards for the "rl run with hacking"? The SDF show no misalignment but when it hits RL somehow it's misaligned. The baseline is the hacking is disincentivized. So is the "rl with hacking" have positive rewards for hacking.
Thank goodness veo3 still has guardrails
November 23, 2025 at 4:11 PM
Thank goodness veo3 still has guardrails
i guess just doing a controlnet method will just remove it.
- inputs: input image
input image -> gemini 2.5 flash lite - > prompt
input image -> Sobel edge detection using Sobel Operator (Threshold:0, Mode: gradient) -> edge image
edge image + prompt -> seedream-4.0 -> final image
- inputs: input image
input image -> gemini 2.5 flash lite - > prompt
input image -> Sobel edge detection using Sobel Operator (Threshold:0, Mode: gradient) -> edge image
edge image + prompt -> seedream-4.0 -> final image
November 22, 2025 at 9:29 PM
i guess just doing a controlnet method will just remove it.
- inputs: input image
input image -> gemini 2.5 flash lite - > prompt
input image -> Sobel edge detection using Sobel Operator (Threshold:0, Mode: gradient) -> edge image
edge image + prompt -> seedream-4.0 -> final image
- inputs: input image
input image -> gemini 2.5 flash lite - > prompt
input image -> Sobel edge detection using Sobel Operator (Threshold:0, Mode: gradient) -> edge image
edge image + prompt -> seedream-4.0 -> final image
here is the whole chat, it didn't tell me the tool failed
November 22, 2025 at 5:39 PM
here is the whole chat, it didn't tell me the tool failed
💀 "I have the feeling that using SynthID, it will probably be negative." wtf
November 22, 2025 at 5:35 PM
💀 "I have the feeling that using SynthID, it will probably be negative." wtf