In a prior post I wrote about my efforts to use GPT to explain jokes. GPT 3 didn’t do a great job, and 3.5 still relied on word association with varying results. I tried again with GPT 4 and it pretty much got them all. The difference between 3.5 and 4 is subtle for most questions that 3.5 gets, but pretty drastic for questions that 3.5 fails at.
I tried out GPT-4 on memes and it worked very well. The OCR is very good and it can read words very well even with low resolution and obnoxious meme fonts.
But here’s one from a Balkan meme account that I didn’t understand:
I prompted GPT 4 “Explain this meme for me” for which it answered:
The meme is a humorous representation of the historical events surrounding the Treaty of Sèvres and the subsequent Nationalist movement in Turkey led by Mustafa Kemal Atatürk.
The figure with the British flag represents the Allied powers, mainly Britain, expressing anger and frustration. The text indicates that the Allied powers wanted the Ottoman Empire (represented by the flag of Turkey and the character with the fez) to accept the conditions of the Treaty of Sèvres, which would have greatly reduced the size and influence of the Ottoman Empire.
The map in the middle shows the territorial losses the Ottoman Empire would have faced had the treaty been fully implemented. It displays a significantly reduced Turkish territory.
The figure with the hat represents Mustafa Kemal Atatürk, the founder of modern Turkey. His expression and the word "I REJECT" reflect his refusal to accept the Treaty of Sèvres. Atatürk led the Nationalist movement against both the occupying Allied forces and the Ottoman Sultan's government to establish a new, sovereign Turkish state.
The map on the bottom right shows the modern borders of Turkey, indicating the success of the Nationalist movement in establishing the Republic of Turkey and rejecting the terms of the Treaty of Sèvres.
In summary, the meme humorously depicts the tension between the Allied powers and the Nationalist movement in Turkey over the fate of the Ottoman Empire after World War I.
I can’t share the actual conversation because sharing conversations with images is not yet supported.
The depth and background behind the explanation surprised me. It even correctly identified the territorial losses map on the top-middle part of the meme and the modern version. And it even identified the man in the hat as Mustafa Kemal Atatürk, who was a stern looking mustached man and founding father of modern Turkey.
It shows that GPT is gluing various concepts together very well in a comprehensive way. First you have to describe the image and transcribe any text. Then you have to use those components to arrive at a context.
For the meme above, it says something about treaty being rejected (text), it has something to do with Turkey and UK (flags from image), it has something to do with territory (map from image) and it has a few people involved (from image). From that you can find out it probably has to do with Treaty of Sèvres. Next step is to find the important people involved in the treaty and someone that rejected. And then tie it all together through a meme lens (seething Wojak vs defiant Chad).
Or here’s how GPT explained the reasoning:
Sure! The image consists of several distinct elements, and making sense of it requires knowledge of the historical context and the ability to associate the visual cues with specific events or people. Here's a breakdown of the steps and associations:
British Flag and Crying Figure: This figure is a visual representation of frustration or distress. The presence of the British flag suggests this figure represents the sentiment of Britain or the Allied powers.
Association: The Allied powers, especially Britain, played a major role in drafting the Treaty of Sèvres post-WWI, which sought to dismantle the Ottoman Empire.
Text and Turkish Flag: The text "NOOO! YOU MUST ACCEPT THE CONDITIONS OF THE TREATY LIKE EVERYONE ELSE EVEN YOUR SULTAN HAS ACCEPTED IT!" along with the Turkish flag and the character with the fez suggests an interaction between the Allied powers and the Ottoman Empire regarding a treaty.
Association: The Treaty of Sèvres was imposed on the Ottoman Empire by the Allied powers. The reference to the Sultan indicates that while the ruling Sultan might have been in favor of or pressured into accepting the treaty, others in the empire were not.
Central Map: This map showcases territorial divisions, with parts of the Ottoman Empire being allocated to various other nations.
Association: The Treaty of Sèvres aimed to partition the Ottoman Empire and allocate various territories to the Allied powers, drastically reducing the empire's landholdings.
Figure with the Hat: This figure represents a specific individual, given the distinct look and attire, saying "I REJECT."
Association: This figure is representative of Mustafa Kemal Atatürk, who led the Turkish Nationalist movement against the conditions of the Treaty of Sèvres and the decisions of the Sultanate. His statement of rejection symbolizes the refusal of the Nationalists to accept the treaty.
Map on Bottom Right: This map displays the modern borders of Turkey.
Association: After a series of military and political maneuvers, Atatürk and the Nationalists succeeded in redefining the borders of the new Turkish Republic, which are reflected in this map. This was solidified with the Treaty of Lausanne in 1923, which superseded the Treaty of Sèvres.
Overall Association: By connecting all these visual and textual cues, one can infer that the meme is a humorous portrayal of the tension and conflict between the Allied powers and the Turkish Nationalists regarding the fate of the Ottoman Empire and the establishment of the Republic of Turkey post-WWI.
Tying everything together is very impressive as every layer of interpretation is a risk of misalignment, much like a small change in the telephone game can cause wildly different results at the end of the line.
What ever happened to the Turing Test?
People often move goal-posts when we talk about AI. Intelligence is often defined by “the thing computers can’t do”. For instance, the Turing Test was a proposed measure of intelligent behavior. The argument was that when a computer’s output is indistinguishable from that of a human, we should carefully consider whether its conscious or not.
When was the last time you read anything about the Turing test? People stopped talking about it because GPT 3.5 obviously passed and even did so without pretending to be a 13 year old Ukranian boy. And it was obvious that GPT was not conscious primarily because it consisted of a set of numbers and metadata that can be used to perform mathematical computations on numbers that result in more numbers that can be translated into words that have some meaning. Are numbers represented in digital form conscious? If we print them out, would the book be conscious as well? It’s all silly.
You can argue that a brain is a biological object that could be represented in digital form no different from an LLM. But a brain in a jar isn’t conscious. Nor is a brain in a jar that gets some electrical stimuli that causes some parts to light up. Consciousness is just he secularist attempt to define soul without religion. But being unable to define something doesn’t necessarily mean you’re unable to identify it. As Supreme Court Justice Potter Stewart wrote when presiding over an obscenity case:
I shall not today attempt further to define the kinds of material I understand to be embraced within that shorthand description ["hard-core pornography"], and perhaps I could never succeed in intelligibly doing so. But I know it when I see it, and the motion picture involved in this case is not that.
You see the same thing with the AI doomer religion. As we get more and more advanced models with no sign of desire to kill humans, the doomers will move the goal posts and argue that the next leap of intelligence will surely imbue some magical desires and abilities to these numbers that will lead to extinction.
But this stuff moves fast and its worth looking back on how far the state of the art has come in just the last few years. There’s still a lot of room for improvement in terms of planning. But there are already signs of effective planning and multi-step reasoning.