Artificial intelligence (AI) has seen remarkable advances in recent years, from self-driving cars to defeating humans at games like chess and Go. However, leading AI expert Melanie Mitchell argues that achieving truly human-like "general" AI is much harder than many experts predict. Here are the key takeaways from her paper for business leaders:

• Narrow AI versus general AI: Success in specialized applications like chess or image recognition does not necessarily translate to broader capabilities. We cannot assume today's AI systems are on a steady path to human-level intelligence.
• "Easy" things are hard: Basic human skills like perceiving the world and carrying on a conversation have proven very difficult to replicate in machines. Conversely, AI can excel at things that are quite hard for humans.
• Wishful vocabulary: Terms like "learn," "understand," and "think" are often applied to today's AI, but these systems do not have the same underlying capabilities as humans. Using human vocabulary can misleadingly imply advanced intelligence.
• Intelligence is embodied: Human intelligence relies heavily on our experiences and interactions using our entire bodies, not just abstract reasoning. Attempts to achieve human-level intelligence must consider embracing human-like bodies and environments.
• Common sense is key: To operate successfully in the real world, AI needs the vast background knowledge humans accumulate about how the world works. We still do not understand how to enable machines to acquire this "common sense."

Key implications:

• Avoid overconfidence about timelines for achieving human-level AI based on hype and narrow successes. True general intelligence likely remains far off.
• Focus investment on applications of existing AI capabilities, not attempts to replicate human thinking. Manage expectations of near-term outcomes.
• Monitor advances in embodied AI and research on common sense reasoning as indicators of progress toward general AI.
• Ensure AI systems have transparent workings, clear objectives, and human oversight. The orthogonality thesis that intelligence can be coupled to any goals does not hold for human-like general intelligence.

The path to human-level AI is long with much still unknown. By avoiding unfounded assumptions and acknowledging the challenges ahead, business leaders can make wise strategic decisions about how to apply AI technology today and anticipate what may come tomorrow.

Sources:

A new study from New York University and the Allen Institute for AI has shown that large language models called transformers can be expressed in a simple logic formalism. This finding challenges the perception that transformers are inscrutable black boxes and suggests avenues for interpreting how they work.

Transformers are a type of neural network behind major AI achievements like chatbots and language translation. They are trained on massive datasets to generate human-like text. Despite their impressive capabilities, how transformers arrive at their outputs has remained poorly understood.

The researchers proved transformers can be translated into symbolic logic sentences that replicate their function. Specifically, they showed transformers fit within a logic called first-order logic with majority quantifiers. This logic allows logical sentences with familiar constructs like "AND", "OR", and "IF-THEN", as well as majority quantifiers that check if a condition holds for over half of the elements.

While real-world transformers are complex neural networks, this study theoretically shows their reasoning can be captured by simple logical expressions. For instance, the logic could recognize patterns like "three As followed by three Bs", which transformers are known to identify.

The findings disprove the notion that transformers are inscrutable black boxes. Instead, they suggest transformers implement a form of reasoning not radically different from familiar logical formalisms. The possibility of expressing transformers in interpretable logic could enable explaining how they arrive at outputs, like detecting biases.

For business leaders deploying AI, this research opens possibilities for making transformers more transparent and accountable. It provides a path toward debugging models to avoid failures or bias. The ability to translate transformers into logical sentences could allow systematically checking if undesirable reasoning patterns occur.

Overall, this theoretical advance challenges prevailing views of transformers as hopelessly opaque. It demonstrates their thinking can be characterized in understandable logic, unlocking new ways for technologists to interpret these increasingly critical AI models. The research brings transformers closer to human-level reasoning by showing their outputs are not ineffable, but rather can be explained through logic.

Sources:

Artificial intelligence has made great strides in recent years, especially in natural language processing. Systems like ChatGPT and Claude can now hold impressively human-like conversations. However, a major limitation of these AI language models is that they operate like a black box - their internal workings are complex and opaque.

Researchers at Stanford have proposed a new AI architecture called Backpack that aims to fix this problem. Backpack models have an internal structure that is more interpretable and controllable compared to existing models like BERT and GPT-3.

Here's an analogy to understand how Backpack works:

Think of words as Lego blocks. Each block can be connected to other blocks in many ways to build something. Existing AI models are like throwing all the Lego pieces together in a pile - there are endless ways to combine them, but you can't understand or control the resulting structure.

A Backpack model is more like having clearly labeled Lego pieces in different bags. For each word, there are "sense vectors" that represent its different meanings and uses. When the model sees a word in a sentence, it decides which sense vectors to pull out of the bag to understand and predict that usage.

This structure offers two key benefits:

1. Interpretability: We can inspect the different sense vectors for a word and understand what aspects of meaning they represent. This is like looking inside the bags to see the different kinds of Lego pieces.
2. Control: We can directly edit the sense vectors to change the model's behavior. For example, reducing a gender-biased sense vector for the word "nurse" can reduce sexist outputs. This is like removing certain Lego pieces from a bag to change what can be built with it.

In initial tests, Backpack models matched the performance of existing models like GPT-2 while offering far more transparency. Researchers were able to do things like swap associations (so "MacBook" predicts "HP" instead of "Apple") and reduce gender bias in occupations.

The inventors stress that Backpack is still early stage research. The approach needs to be scaled up and tested across different languages and applications. But it represents an exciting step towards AI systems that are not black boxes. Instead of blindly trusting model outputs, users can interpret why it behaves in certain ways and directly edit its knowledge.

As AI becomes more powerful and ubiquitous in products and services, retaining human agency is crucial. Approaches like Backpack could make future AI not only smarter but easier to understand and actively improve. Business leaders should track developments in interpretable AI closely, as it is an important competitive differentiator down the line.

Source:

arxiv

The researchers focused on political leanings across two dimensions - social values (liberal vs conservative) and economic values (left vs right). First, they evaluated the inherent biases of 14 major language models by analyzing their responses to statements from a standard political compass test. The models occupied a range of positions across the political spectrum, with BERT variants being more conservative and GPT models more libertarian.

Next, they examined if further pretraining these models on partisan news and social media corpora leads them to shift their political stances. The results show that left-leaning data induces liberal shifts, while right-leaning data causes conservative shifts. However, the overall shifts were small, suggesting inherent biases persist.

Finally, they tested if these biased models perform differently on social-impact tasks like hate speech and misinformation detection. While overall performance was similar, models exhibited double standards - left-leaning ones were more sensitive to offensive speech targeting minorities but overlooked attacks on dominant groups. Right-leaning models showed the opposite.

For business leaders, these findings highlight risks of unintended bias and unfairness creeping into AI systems built on large language models. While some bias is inevitable given the training data, being aware of its extent and impact can help inform ethical AI practices. Companies should proactively probe for biases, use diverse evaluation datasets, and leverage different perspectives through ensemble approaches.

Tracking bias from data to models to decisions is vital for ensuring AI transparency and accountability.

Source:

Arxiv

Recent advances in AI image generation have led to impressive results, with systems like Stable Diffusion able to create highly realistic images from simple text prompts. But while these models can generate remarkably detailed pictures, they remain something of a black box. We don't have much insight into how they actually convert the text into pixel representations.

New research from computer scientists is helping peel back the curtain on these AI systems. In a paper titled "What the DAAM: Interpreting Stable Diffusion Using Cross Attention", researchers propose a method called DAAM (Diffusion Attentive Attribution Maps) to analyze how words in a prompt influence different parts of the generated image.

DAAM creates heat maps showing which pixels are most related to each word in the text prompt. For example, for a prompt like "a blue bird flying", the word "blue" would highlight the blue parts of the bird, "bird" would highlight the full bird, and "flying" would highlight the motion blurred wings and body.

By aggregating attention scores between text and image patches across the AI model's layers, DAAM produces interpretable maps linking words to visual features. The researchers validated DAAM by testing how well it can perform noun segmentation, a common computer vision benchmark where the goal is to identify the regions in an image that correspond to noun objects. DAAM achieved competitive scores, despite having no explicit training.

Experiments with DAAM revealed new insights about these generative AI systems:

• The relationships between words in a prompt translate to visual relationships in the image. For example, verbs like "flying" encapsulate their subjects like "bird".
• Using similar words like "giraffe" and "zebra" leads to worse image generation, likely because their features get entangled. The DAAM maps heavily overlap for such co-hyponyms. Essentially, co-hyponyms are words that share the same superclass or category but are not synonyms. They have a sort of "sibling" status as distinct members of their parent group. Giraffe and zebra are co-hyponyms. They both belong to the hypernym "wild animals".
• Descriptive adjectives like "blue" attend too broadly across the whole image, suggesting objects are entangled with their surroundings. Changing the adjective modifies the entire scene.

For business leaders, research like DAAM is important because it improves the explainability of AI systems. As generative models become more ubiquitous, understanding how they operate will help identify limitations and better assess risks. Models that entangle features more could potentially suffer from bias or produce unrealistic outputs.

DAAM also demonstrates how attention mechanisms in AI models can be repurposed for interpretation, without retraining the models from scratch. This allows transparent analysis without compromising performance.

Overall, DAAM represents an impactful step toward explainable AI in generative models. Demystifying these systems will be key as businesses increasingly look to utilize powerful generative AI capabilities in their products and processes. Interpretability helps ensure these technologies are trustworthy and dependable.

Sources:

arxiv