Rafael Pardinas
@muchomuchacho.bsky.social
AI (NLP, RL) Researcher at ServiceNow Research (London, UK)
Ok, but you're linking (advertising) their device rather than the free android app?
December 19, 2024 at 8:20 AM
Ok, but you're linking (advertising) their device rather than the free android app?
One of the main features is their ability to use tools to complete a task. If any of those tools allow them to reason, plan or interact with the environment you have all those definitions covered.
December 19, 2024 at 8:08 AM
One of the main features is their ability to use tools to complete a task. If any of those tools allow them to reason, plan or interact with the environment you have all those definitions covered.
device = torch.device(torch.cuda.is_available() and "cuda" or torch.backends.mps.is_available() and "mps" or "cpu")
December 8, 2024 at 5:05 PM
device = torch.device(torch.cuda.is_available() and "cuda" or torch.backends.mps.is_available() and "mps" or "cpu")
You could probably expand your analogy much further than Canada. Most of the developed world would have you covered.
December 8, 2024 at 4:29 PM
You could probably expand your analogy much further than Canada. Most of the developed world would have you covered.
Yes, and I agree. The Markov assumption may not add much/any value beyond what the LLM already provides in its context handling process
November 29, 2024 at 6:58 PM
Yes, and I agree. The Markov assumption may not add much/any value beyond what the LLM already provides in its context handling process
hmmm, I guess an LLM does comply with:
* Their internal state is effectively unknowable
* Each response depends on the current context
* Long-term state transitions are less important than immediate response quality
Am I missing anything?
* Their internal state is effectively unknowable
* Each response depends on the current context
* Long-term state transitions are less important than immediate response quality
Am I missing anything?
November 29, 2024 at 6:41 PM
hmmm, I guess an LLM does comply with:
* Their internal state is effectively unknowable
* Each response depends on the current context
* Long-term state transitions are less important than immediate response quality
Am I missing anything?
* Their internal state is effectively unknowable
* Each response depends on the current context
* Long-term state transitions are less important than immediate response quality
Am I missing anything?
they probably meant "massive drama"
November 28, 2024 at 6:42 PM
they probably meant "massive drama"
Excellent points, I completely agree.
I hate the way articles about BSky keep framing choice to move here in terms of left/right and echo chambers. Fundamentally, the bigger issues for me are wanting to see links/current news, wanting to see posts that aren't rage-baiting/content farming, and wanting actually diverse conversations.
November 27, 2024 at 10:40 PM
Excellent points, I completely agree.
I had a WireGuard service running on my edge box in 2018. It used to report around 200 intrusion attempts per day. As long as you have a public IP address, you're a target. But to be fair, it needed very little maintenance, the uptime was something like 3 years before I shut it down.
November 26, 2024 at 9:53 AM
I had a WireGuard service running on my edge box in 2018. It used to report around 200 intrusion attempts per day. As long as you have a public IP address, you're a target. But to be fair, it needed very little maintenance, the uptime was something like 3 years before I shut it down.
are you also expecting to use the rig controlling the Hadron Collider to do your online shopping?
November 26, 2024 at 9:23 AM
are you also expecting to use the rig controlling the Hadron Collider to do your online shopping?
Reposted by Rafael Pardinas
This is actually a really cool interview, wish I could read it properly (here's a google translate link: mp-weixin-qq-com.translate.goog/s/r9zZaEgqAa...)
November 25, 2024 at 5:57 PM
This is actually a really cool interview, wish I could read it properly (here's a google translate link: mp-weixin-qq-com.translate.goog/s/r9zZaEgqAa...)
I let's keep going, I guess: lmsys.org/blog/2024-02...
Fast JSON Decoding for Local LLMs with Compressed Finite State Machine | LMSYS Org
<p>Constraining an LLM to consistently generate valid JSON or YAML that adheres to a specific schema is a critical feature for many applications.
In this blo...
lmsys.org
November 25, 2024 at 5:07 PM
I let's keep going, I guess: lmsys.org/blog/2024-02...
R1-lite is a maths-beast.
November 23, 2024 at 1:23 PM
R1-lite is a maths-beast.
Reposted by Rafael Pardinas
About to send this to my newsletter list tomorrow morning.
Did I miss any important points?
justinjackson.ca/twitter-blue...
Did I miss any important points?
justinjackson.ca/twitter-blue...
Leaving Twitter for Bluesky
After 15 years and 50,000 tweets on Twitter/X, I'm moving to Bluesky, which (in the last month) has become 1,000x more fun than X.
justinjackson.ca
November 23, 2024 at 7:15 AM
About to send this to my newsletter list tomorrow morning.
Did I miss any important points?
justinjackson.ca/twitter-blue...
Did I miss any important points?
justinjackson.ca/twitter-blue...
Did you keep both available for download as a social experiment?
November 23, 2024 at 11:18 AM
Did you keep both available for download as a social experiment?
AI inference is like compression for thought. Faster paths from input to insight mean AI can explore more possibilities.
November 23, 2024 at 11:09 AM
AI inference is like compression for thought. Faster paths from input to insight mean AI can explore more possibilities.
Thanks for sharing! This is very inspiring. Refusals play a critical role in any safety work.
November 22, 2024 at 1:51 PM
Thanks for sharing! This is very inspiring. Refusals play a critical role in any safety work.