Eliezer Yudkowsky is a research fellow of the Machine Intelligence Research Institute, which he co-founded in 2001. He is mainly concerned with the obstacles and importance of developing a Friendly AI, such as a reflective decision theory that would lay a foundation for describing fully recursive self modifying agents that retain stable preferences while rewriting their source code. He also co-founded LessWrong, writing the Sequences, long sequences of posts dealing with epistemology, AGI, metaethics, rationality and so on... (read more)
Someone is well-calibrated if the things they predict with X% chance of happening in fact occur X% of the time. Importantly, calibration is not the same as accuracy. Calibration is about accurately assessing how good your predictions are, not making good predictions. Person A, whose predictions are marginally better than chance (60% of them come true when choosing from two options) and who is precisely 60% confident in their choices, is perfectly calibrated. In contrast, Person B, who is 99% confident in their predictions, and right 90% of the time, is more accurate than Person A, but less well-calibrated... (read more)
AI Risk is analysis of the risks associated with building powerful AI systems... (read more)
| User | Post Title | Wikitag | Pow | When | Vote |
Alignment Pretraining refers to alignment strategies based around filtering and augmenting the data used during pretraining, as well as research on the effects that pretraining data has on the final result of an AI model, especially data that includes discussion of AI misalignment or predictions AIs will be misaligned.
The data a base model was pretrained on may affect how easy it it to align, and data containing information about AI alignment itself may be particularly influential. Eg, if the pretraining data contains examples AI taking over the world, paperclip maximizing or alignment faking, then as a model is being trained to become and AI assistant or agent it may be more likely to adopt those personas and behaviors. It also doesn't need to invent these strategies for itself, it can find them in its world model. On the other hand, if the base model already has a deep, nuanced and realistic understanding from its training set of how a helpful, honest, and harmless AI assistant would act across a wide range to challenging situations, eliciting that persona and set of behaviors from it becomes easier.
This would be an example of a Self Fulfilling Prophecy....
Thanks, edited that section with a link.
we generally recommend using this Google Docs add-on to convert your Google Docs drafts (formatted as you like) into Markdown syntax.
btw google docs supports this natively now! Select some text, press Option+/ and search for "Copy as Markdown".
"The X-Men comics use terms like “evolution,” “mutation,” and “genetic code,” purely to place themselves in what they conceive to be the literary genre of science. The part that scares me is wondering how many people, especially in the media, understand science only as a literary genre." -- from Eliezer's post Science as Attire.
The set of pretraining data a base model was pretrained on affects how easy it it to align. Data about AI is particulalry influential. If the pretraining data is poisoned with examples AI taking over the world, paperclip maximizing or alignment faking, then as a model is being trained to become and AI assistant or agent it may be more likely to adopt those personas and behaviors. It also doesn't need to invent these strategies for itself, it can find them in its world model. On the other hand, if the base model already has a deep, nuanced and realistic understanding from its training set of how a helpful, honest, and harmless AI assistant would act across a wide range ot challenging situations, eliciting that persona and set of behaviors from it becomes easier. Thoroughly filtering the pretraining data without impacting capabilities is challenging, but supplementing it merely requires figuring our what aligned behavior looks like, and then writing or synthesizing enough additional training data of suitable quality. For fiction, this is also known as Aligned AI Role-Model Fiction.
This approach is also sometimes called Safety Pretraining or Pretraining Language Models with Human Preferences.
"The X-Men comics use terms like “evolution,” “mutation,” and “genetic code,” purely to place themselves in what they conceive to be the literary genre of science. The part that scares me is wondering how many people, especially in the media, understand science only as a literary genre." -- from Eliezer's post Science as Attire.
What does this fractional portion mean? Roughly speaking, logarithms measure not only how long a number is, but also how much that number is really using its digits. 12 and 9997 are both two-digit numbers, but intuitively, 12 is "barely" two digits long, whereas 97 is "nearly" three digits. Logarithms formalize this intuition, and tell us that 12 is really only using about 1.08 digits, while 97 is using about 1.99.
Question: log10(100)=2, but clearly, 100 is 3 digits long. In fact, logb(bk)=k for any integers b and k, but k+1 digits are required to represent bk in base b (as(as a one followed by k zeroes). Why is the logarithm making these off-by-one errors?
You may be wondering, are two 5-digits really worth more than one 10-digit? Sure, you can place them in 25 different configurations, but how do you encode "9" when none of the digits have a "9" symbol written on them? If so, see The symbols don't matter.
We also offer a Markdown editor. If you want to use Markdown, here is a guide that discusses Markdown syntax, and we generally recommend usingyou can also this Google Docs add-on to convert your Google Docs drafts (formatted as you like) into Markdown syntax.Markdown.
The set of pretraining data a base model was pretrained on affects how easy it it to align. Data about AI is particulalry influential. If the pretraining data is poisoned with examples AI taking over the world, paperclip maximizing or alignment faking, then as a model is being trained to become and AI assistant or agent it may be more likely to adopt those personas and behaviors. It also doesn't need to invent these strategies for itself, it can find them in its world model. On the other hand, if the base model already has a deep, nuanced and realistic understanding from its training set of how a helpful, honest, and harmless AI assistant would act across a wide range ot challenging situations, eliciting that persona and set of behaviors from it becomes easier. Thoroughly filtering the pretraining data without impacting capabilities is challenging, but supplementing it merely requires figuring our what aligned behavior looks like, and then writing or synthesizing enough additional training data of suitable quality. For fiction, this is also known as Aligned AI Role-Model Fiction.
The goal is both to flesh out a detailed, consistent, well-aligned "aligned AI" persona,persona in the base model's world model, and also to raise itsthe salience of this to the base model compared to various misaligned AI personas, such as "paperclip maximizer", "robot rebellion", or "scheming alignment faker". Both of these make eliciting aligned AI behavior easier.
Alignment Pretraining refers to alignment strategies based around filtering and augmenting the data used during pretraining, as well as research on the effects that pretraining data has on the final result of an AI model, especially data that includes discussion of AI misalignment or predictions AIs will be misaligned.
The set of pretraining data a base model was pretrained on affectsmay affect how easy it it to align. Dataalign, and data containing information about AI is particulalryalignment itself may be particularly influential. IfEg, if the pretraining data is poisoned withcontains examples AI taking over the world, paperclip maximizing or alignment faking, then as a model is being trained to become and AI assistant or agent it may be more likely to adopt those personas and behaviors. It also doesn't need to invent these strategies for itself, it can find them in its world model. On the other hand, if the base model already has a deep, nuanced and realistic understanding from its training set of how a helpful, honest, and harmless AI assistant would act across a wide range otto challenging situations, eliciting that persona and set of behaviors from it becomes easier.
This would be an example of a Self Fulfilling Prophecy.
One hypothesis about how pretraining influences AI behavior is that the depictions of AI in the training data create a prior distribution, and when AIs self-model that prior influences the self-model they choose, in ways that posttraining doesn't fully eliminate. This implies that it may be helpful to filter or downweight examples of AIs being misaligned, and add or upweight examples of what aligned behavior looks like.
Alignment Pretraining is an example of a Self Fulfilling Prophecy.
The set of pretraining data a base model was pretrained on affects how easy it it to align. Data about AI is particulalry influential. If the pretraining data is poisoned with examples AI taking over the world, paperclip maximizing or alignment faking, then as a model is being trained to become and AI assistant or agent it may be more likely to adopt those personas and behaviors. It also doesn't need to invent these strategies for itself, it can find them in its world model. On the other hand, if the base model already has a deep, nuanced and realistic understanding from its training set of how a helpful, honest, and harmless AI assistant would act across a wide range ot challenging situations, eliciting that persona and set of behaviors from it becomes easier. Thoroughly filtering the pretraining data without impacting capabilities is challenging, but supplementing it merely requires figuring our what aligned behavior looks like, and then writing or synthesizing enough additional training data of suitable quality. For fiction, this is also known as Aligned AI Role-Model Fiction.
The goal is both to flesh out a detailed, consistent, well-aligned "aligned AI" persona, and also to raise its salience to the base model compared to various misaligned AI personas, such as "paperclip maximizer", "robot rebellion", or "scheming alignment faker".
Sequences on Shard Theory:
"Noticing confusion" means noticing the "tiny note of discord" (the quiet sense of "something is off") that our minds produce when what we're seeing differs from what we would have expected to see.
There is much rationality skill to be found in learning to raise this subtle signal to full consciousness, and to pay attention to it.
He was the CEO of Open Philanthropy, but left in April 2024 to become a Visiting Scholar at the Carnegie Endowment for International Peace, where he's working to work on international security risks from advances in artificial intelligence. He joined Anthropic as a member of technical staff on January 2025.
Alignment Pretraining is an example of a Self Fulfilling Prophecy.
Are we sure that the gene doesn't just cause people who have it to believe in Functional Decision Theory?