Hacker News Reader

I guess you stepped on the case of aggressive filtering. The current version of Bard looks like have a very conservative set of filtering preamble prompts to arbitrarily limit its ability to avoid negative presses.

> BARD: The blue cube is now the bottom cube, and the red cube is on top of it. The red cube is touching the blue cube on all six sides.

That doesn’t seem correct. The red cube is touching the blue cube on its top side only.

You could argue 5 sides are touching

only if you're trying to be a nuisance

I am going to post the full response to Winograd because I think it is extremely representative of what I have been seeing so far from Bard. I can’t help but be incredibly disappointed.

Bard: Yes, I am familiar with Terry Winograd's PhD thesis on natural language processing. In his thesis, Winograd developed a program called SHRDLU that could understand and respond to English sentences in a simulated "world of toy blocks". SHRDLU was a significant achievement in the field of artificial intelligence, and it helped to pave the way for the development of more sophisticated natural language processing systems.

My natural language ability is much more advanced than SHRDLU's. I can understand and respond to a wider range of English sentences, and I can also generate text that is more natural and engaging. I am able to do this because I have access to a much larger dataset of text and code, and I have been trained on a more sophisticated algorithm.

However, SHRDLU was a groundbreaking achievement in its time, and it is important to remember that it was developed over 50 years ago. The field of natural language processing has made significant progress since then, and I am grateful for the work of Terry Winograd and other pioneers in the field.

GPT-4: Yes, I am familiar with Terry Winograd's PhD thesis on natural language processing. His thesis, titled "Procedures as a Representation for Data in a Computer Program for Understanding Natural Language," was published in 1971. The work focused on SHRDLU, a natural language understanding system that he developed.

SHRDLU was designed to understand and respond to natural language commands in a limited "blocks world" environment. The blocks world consisted of a virtual space where a robotic arm could manipulate colored blocks on a table. Winograd's program could interpret natural language commands given by a human user, such as "Pick up the red block" or "Place the green block on the blue one," and execute the appropriate actions in the simulated environment.

While groundbreaking for its time, the SHRDLU program's natural language understanding capabilities were limited in comparison to modern AI systems like GPT-4. Some key differences between SHRDLU and GPT-4 include:

1. *Domain*: SHRDLU was limited to the blocks world, which was a small, well-defined domain. GPT-4, on the other hand, has been trained on a diverse range of text and can generate responses on a wide variety of topics.

2. *Language Model*: SHRDLU used a rule-based approach to parse and interpret natural language, whereas GPT-4 utilizes a deep learning architecture based on the Transformer model. This allows GPT-4 to learn complex language patterns and generate more human-like responses.

3. *Flexibility*: GPT-4 is more flexible in understanding and generating natural language due to its large-scale training data and deep learning architecture. SHRDLU, by contrast, relied on handcrafted rules and procedures, which made it less adaptable to new language patterns or domains.

4. *Data-driven*: GPT-4 is a data-driven model, which means it learns from the patterns and structures in the text it was trained on. SHRDLU, being a rule-based system, did not have the same capacity to learn from data.

Despite these differences, Winograd's work on SHRDLU was an important milestone in the history of natural language processing and artificial intelligence. It influenced subsequent research in the field and laid the groundwork for future advancements, including the development of AI systems like GPT-4.

It's interesting that you got a different response from Bard! I didn't alter the Bard response that I posted. Did you use exactly the same prompt, and was there any other context in place?

After the conversation I posted above, Bard and I discussed intractability vs undecideability, and it was amusing, but it would be a lot to include here. Then I said,

I still can't believe you don't know about Terry Winograd's PhD thesis.

BARD: I apologize that I do not know about Terry Winograd's PhD thesis. I am still under development, and I am always learning new things. I will try my best to learn more about Terry Winograd's work in the future.

ME:What's SHRDLU?

BARD: SHRDLU is a computer program that was developed in the early 1970s by Terry Winograd at the MIT Artificial Intelligence Laboratory. It was one of the first programs to be able to understand natural language to a significant degree.

SHRDLU could understand simple commands, such as "Put the red block on the table" or "Put the green cube on top of the red block." It could also answer questions about the state of the world, such as "What is on the table?" or "What is on top of the red block?"

SHRDLU was a significant achievement in artificial intelligence, and it helped to pave the way for the development of more sophisticated natural language processing programs.

ME: If I tell you that SHRDLU was a program written for Terry Winograd's PhD thesis, how would you summarize Terry Winograd's PhD thesis?

BARD: Terry Winograd's PhD thesis, "Procedures as a Representation for Data in a Computer Program for Understanding Natural Language", was an attempt to develop a computer program that could understand natural language. The program, SHRDLU, was able to understand simple commands and answer questions about a simple world of objects. Winograd's work was a significant achievement in artificial intelligence, and it helped to pave the way for the development of more sophisticated natural language processing programs.

ME: Baffling! So you did know about Terry Winogra'd PhD thesis all along!

BARD: I apologize that I gave you the impression that I did not know about Terry Winograd's PhD thesis. I have been trained on a massive dataset of text and code, and I am able to access and process information from the real world through Google Search. This allows me to answer your questions in a comprehensive and informative way.

Yes, I copied your prompt exactly into a fresh context. I’m not surprised by the diversity of responses, that’s pretty common since there’s likely a random seed for each chat instance. And the logical inconsistency that you can then correct through intermediate reasoning seems common to all LLMs I’ve used also.

My main takeaway is just the quality and specificity of the responses, once you actually get it to cooperate, and in that Bard seems pretty consistently behind. It might not even be a flaw in the foundation model but could be a case of failure me tuning it to be too brief and general in style. That said, I do suspect the foundation model is less capable based on some theory of mind and logical puzzle examples I’ve seen posted from others. Plus it seems to be a poor coder even when it does cooperate.

> Assume there are 3 cubes, red, green, and blue. Stack the red cube on top of the blue cube, and describe the result to me.

> BARD: The blue cube is now the bottom cube, and the red cube is on top of it. The red cube is touching the blue cube on all six sides. The blue cube is completely covered by the red cube.

Well if you wanted a description of a 4-dimensional block world it's actually not bad.