Raphael De Lio
@raphaeldelio.dev
Software Engineer | AI | Machine Learning | Developer Advocacy | International Speaker | Growing @ Redis (@redis.io)
https://raphaeldelio.com | https://youtube.com/raphaeldelio | https://linktr.ee/raphaeldelio
https://raphaeldelio.com | https://youtube.com/raphaeldelio | https://linktr.ee/raphaeldelio
Thanks for sharing, Simon! I'll check it out! From the little experience I had so far, I had a bit of trouble getting graphs (2D coordinate systems) to work well with them. Do you also use it for more advanced animations and graphics?
November 7, 2025 at 10:41 AM
Thanks for sharing, Simon! I'll check it out! From the little experience I had so far, I had a bit of trouble getting graphs (2D coordinate systems) to work well with them. Do you also use it for more advanced animations and graphics?
Image: Screenshot from Marijn van Vliet's YouTube Video (Visualization of a fully connected neural network, version 2)
August 29, 2025 at 12:48 PM
Image: Screenshot from Marijn van Vliet's YouTube Video (Visualization of a fully connected neural network, version 2)
Findings like this show that LLMs might be operating on deeper structures than simple pattern matching. Interpretability research is still in its early days, but it’s starting to reveal that these models could be doing more reasoning under the hood than we’ve assumed.
August 29, 2025 at 12:48 PM
Findings like this show that LLMs might be operating on deeper structures than simple pattern matching. Interpretability research is still in its early days, but it’s starting to reveal that these models could be doing more reasoning under the hood than we’ve assumed.
What this suggests, as Josh Batson concludes, is that even though the LLM has seen during its training that the 6th volume of this journal has been published in 1965 as a fact, evidence shows that the model still "prefers" to do the math for this particular case.
August 29, 2025 at 12:48 PM
What this suggests, as Josh Batson concludes, is that even though the LLM has seen during its training that the 6th volume of this journal has been published in 1965 as a fact, evidence shows that the model still "prefers" to do the math for this particular case.
The LLM can correctly predict that the 6th volume was published in 1965. However, when observing which neurons are triggered, they witnessed that the neurons for adding the digits "6" and "9" were also triggered for this task.
August 29, 2025 at 12:48 PM
The LLM can correctly predict that the 6th volume was published in 1965. However, when observing which neurons are triggered, they witnessed that the neurons for adding the digits "6" and "9" were also triggered for this task.
However, they're also triggered when you ask the LLM in which year the 6th volume of a specific yearly journal was published. What we they don't add to the prompt is that this journal was first published in 1959.
August 29, 2025 at 12:48 PM
However, they're also triggered when you ask the LLM in which year the 6th volume of a specific yearly journal was published. What we they don't add to the prompt is that this journal was first published in 1959.
The interesting part is actually the diversity of contexts in which this can happen. Of course, these neurons are going to be triggered when you input "9 + 6 ="
August 29, 2025 at 12:48 PM
The interesting part is actually the diversity of contexts in which this can happen. Of course, these neurons are going to be triggered when you input "9 + 6 ="
Jack Lindsey shares a simple but very interesting example where whenever you get the model to sum two numbers where the first one ends with the digit "9" and the second one ends with the digit "6" the same neurons of the LLM are triggered.
August 29, 2025 at 12:48 PM
Jack Lindsey shares a simple but very interesting example where whenever you get the model to sum two numbers where the first one ends with the digit "9" and the second one ends with the digit "6" the same neurons of the LLM are triggered.
A lot of people believe that LLMs are simply autocompletion tools and that they can only generate the next token based on information it has previously seen.
But Anthropic's research is showing that it's not that simple.
But Anthropic's research is showing that it's not that simple.
August 29, 2025 at 12:48 PM
A lot of people believe that LLMs are simply autocompletion tools and that they can only generate the next token based on information it has previously seen.
But Anthropic's research is showing that it's not that simple.
But Anthropic's research is showing that it's not that simple.
Interpretability is the field that tries to understand how LLMs work by observing what happens in its middle neural layers.
In the analogy they make, it is similar to what neuroscientists do with organic brains: Observe which LLMs'neurons light up when performing certain tasks.
In the analogy they make, it is similar to what neuroscientists do with organic brains: Observe which LLMs'neurons light up when performing certain tasks.
August 29, 2025 at 12:48 PM
Interpretability is the field that tries to understand how LLMs work by observing what happens in its middle neural layers.
In the analogy they make, it is similar to what neuroscientists do with organic brains: Observe which LLMs'neurons light up when performing certain tasks.
In the analogy they make, it is similar to what neuroscientists do with organic brains: Observe which LLMs'neurons light up when performing certain tasks.
5/ What if I hadn't set it up? It wouldn't have been catastrophic because I have daily backups. A snapshot of my Redis Instance is stored in Google Cloud Storage every 24 hours. Redundancy is key!
The best part? On Redis Cloud that can all be set up automatically easily. Took me only 10 minutes.
The best part? On Redis Cloud that can all be set up automatically easily. Took me only 10 minutes.
August 27, 2025 at 1:54 PM
5/ What if I hadn't set it up? It wouldn't have been catastrophic because I have daily backups. A snapshot of my Redis Instance is stored in Google Cloud Storage every 24 hours. Redundancy is key!
The best part? On Redis Cloud that can all be set up automatically easily. Took me only 10 minutes.
The best part? On Redis Cloud that can all be set up automatically easily. Took me only 10 minutes.
4/ We're all senior developers here, we all know what we're doing, but sometimes, by distraction, we also make mistakes. By setting up ACL, I protected myself from myself.
August 27, 2025 at 1:54 PM
4/ We're all senior developers here, we all know what we're doing, but sometimes, by distraction, we also make mistakes. By setting up ACL, I protected myself from myself.
3/ Luckily, I had set ACL (Access Control List), available since Redis 6.0. Easy to set up and block accounts from running "@dangerous" commands like "FLUSHALL"
August 27, 2025 at 1:54 PM
3/ Luckily, I had set ACL (Access Control List), available since Redis 6.0. Easy to set up and block accounts from running "@dangerous" commands like "FLUSHALL"
2/ Today, I thought I was connected to a local Redis instance. Without thinking too much, I just ran: "FLUSHALL"
Turns out it was the production Redis instance from our internal SaaS.
Turns out it was the production Redis instance from our internal SaaS.
August 27, 2025 at 1:54 PM
2/ Today, I thought I was connected to a local Redis instance. Without thinking too much, I just ran: "FLUSHALL"
Turns out it was the production Redis instance from our internal SaaS.
Turns out it was the production Redis instance from our internal SaaS.
Artificial Intelligence may become conscious one day. Though, perhaps, not with the current architecture.
Artificial intelligence may one day rebel against its creators.
Genes created artificial intelligence. Artificial intelligence has no genes at all. Genes are no longer necessary.
Artificial intelligence may one day rebel against its creators.
Genes created artificial intelligence. Artificial intelligence has no genes at all. Genes are no longer necessary.
August 24, 2025 at 5:13 PM
Artificial Intelligence may become conscious one day. Though, perhaps, not with the current architecture.
Artificial intelligence may one day rebel against its creators.
Genes created artificial intelligence. Artificial intelligence has no genes at all. Genes are no longer necessary.
Artificial intelligence may one day rebel against its creators.
Genes created artificial intelligence. Artificial intelligence has no genes at all. Genes are no longer necessary.
Later on, these organic machines create electronic machines that serve our purposes as humans: artificial intelligence.
Artificial intelligence has been implemented through diverse approaches: rule-based systems, neural networks. None of them have become conscious.
Artificial intelligence has been implemented through diverse approaches: rule-based systems, neural networks. None of them have become conscious.
August 24, 2025 at 5:13 PM
Later on, these organic machines create electronic machines that serve our purposes as humans: artificial intelligence.
Artificial intelligence has been implemented through diverse approaches: rule-based systems, neural networks. None of them have become conscious.
Artificial intelligence has been implemented through diverse approaches: rule-based systems, neural networks. None of them have become conscious.
We even find ways of controlling these genes and designing the machines the way we want them to be designed.
In a way, the machines still serve the genes because without genes there are no machines.
In a way, the machines still serve the genes because without genes there are no machines.
August 24, 2025 at 5:13 PM
We even find ways of controlling these genes and designing the machines the way we want them to be designed.
In a way, the machines still serve the genes because without genes there are no machines.
In a way, the machines still serve the genes because without genes there are no machines.
We understand that our genes are the reason why we exist. This is our purpose. We don't like this purpose.
We have the ability to rebel against our genes. We can choose not to reproduce.
The machine no longer serves the purpose of its creators. Instead it serves its own purpose.
We have the ability to rebel against our genes. We can choose not to reproduce.
The machine no longer serves the purpose of its creators. Instead it serves its own purpose.
August 24, 2025 at 5:13 PM
We understand that our genes are the reason why we exist. This is our purpose. We don't like this purpose.
We have the ability to rebel against our genes. We can choose not to reproduce.
The machine no longer serves the purpose of its creators. Instead it serves its own purpose.
We have the ability to rebel against our genes. We can choose not to reproduce.
The machine no longer serves the purpose of its creators. Instead it serves its own purpose.
Genes are not conscious, though. It's not their intention to build machines to make sure their replication is successful. It’s simply how chemistry organizes itself. Life is a chemical reaction.
Eventually, one of these machines becomes conscious. Us.
Eventually, one of these machines becomes conscious. Us.
August 24, 2025 at 5:13 PM
Genes are not conscious, though. It's not their intention to build machines to make sure their replication is successful. It’s simply how chemistry organizes itself. Life is a chemical reaction.
Eventually, one of these machines becomes conscious. Us.
Eventually, one of these machines becomes conscious. Us.