In such conditions, it’s not rational to work on any other problem.
I think this suffers from the fallacy of ability to forsee the future. If one were to agree with Taleb, most stuff is trial and error and positive black swans. Then paradoxically if you want improvement in AI, you need to work in so many different problems.
I think Kenneth Stanley's (a researcher in evolutionary algorithms and curiosity based algorithms) view on open endedness has to be strongly consdiered.
A recurring flaw in AI alarmism is that it treats intelligence as a property of individual minds, rather than recognizing that this capacity is distributed across our civilization and culture.
I think this has to be thought of any general AI designers. How do you incorporate diversity of intelligence required for each skill, and whether such diversity comes at a cost. ie Single agent that has to have some axioms/views/world-view in it's head to do it's task in real world. Say engineer vs product manager. Can a AI system hold such diversity of views at the same time while acting as a single will instead of a civilization of wills?
I don't buy this argument at all. Complex minds are likely to have complex motivations; that may be part of what it even means to be intelligent.
Ability to be ethical might as well be emerge from bieng a complex intelligent agent with sentience.
For all the joke of tricking AI into paradoxes in sci-fi tropes, maybe it would be easy for AI to take into Kantian deontology or atleast silver rule for it to just comprehend the word "good" and "evil".
Judea Pearl's work on Causal Inference suggests that the third ladder (in his hierarchy of statistics, intervention and counterfactuals) Counterfactuals might be important not just for "solving" tasks but also for defining what is ethical.
My hypothesis about this is a simple one about how elites fail. In general, elites fail when their relationships with each other become more important than their relationships with the world. Not just masses, the world. The inner reality of the elites absorbs all their attention: whether it is court intrigues, scholarly debates in journals, boardroom battles, product architecture arguments, rivalries among schools of economists, or media wars.
This is very important point. Those who wield the influence, whether from the entrenched or from the resistance has to realise their elite-hood.
Distributional Bellman equation?
$$\begin{array}{c} Z^{\pi}(x, a) \stackrel{D}{=} R(x, a)+\gamma Z^{\pi}\left(x^{\prime}, a^{\prime}\right) \\ \text { where } x^{\prime} \sim p(\cdot \mid x, a) \text { and } a^{\prime} \sim \pi\left(\cdot \mid x^{\prime}\right) \end{array}$$
The inversion of colors should pose no additional difficulty for a human, yet does generally impair DALL·E’s performance, suggesting its capabilities may be brittle in unexpected ways.
Is there a way to encourage models to not fit on confounders? Data augmentation, but without adding noise that might affect some other task.
How to frame this such that we don't have to have a human select those properties for the entire dataset?
Combining unrelated concepts
This implies it has a human-like compositionality on weird objects.
Could this be a pointer towards an universalist view of cognition? But how do you verify that without a fitting on alien internet image+language dataset though
success rate can depend on how the caption is phrased
More prompt engineering. At this rate we all are gonna be wizards casting spells to coax models to do the right thing.
We find that DALL·E is able to create plausible images for a great variety of sentences that explore the compositional structure of language.
Is there a way to evaluate compositionality in such datasets in a rigorous manner?
Say we use humans to say the output is correct/incorrect.
We might have to evaluate and/or/not logical operators. But the distribution in training text will be biased to those that appear in natural language (english) than anything that we generate synthetically !!
It receives both the text and the image as a single stream of data containing up to 1280 tokens, and is trained using maximum likelihood to generate all of the tokens, one after another.
Such a dumb objective works amazingly well when you use a insanely big model!!
its semantic OCR representation is quite useful. When evaluated on the SST-2 NLP dataset rendered as images, a linear classifer on CLIP’s representation matches a CBoW model with direct access to the text.
This is cool! The model must extract some representation that correlates with the word frequencies.
require trial and error “prompt engineering”
"Prompt engineer" is gaining traction.