Cute robot poems written by Large Language Models (LLMs)
You will surely enjoy these poems about the Anki Vector robot
Large Language Models (LLMs) have dominated the technical and research discourse this year. I have been associated with the software industry for the last 25 years, but have never seen technology evolve and grow at the pace at which LLMs are evolving now. Not even in the days when the Internet was first getting introduced to households!
Can LLMs write poems?
This article is written to update you about the current state-of-art, and also to amuse and entertain you. I will be using current LLMs available in public domain and ask them to write a nice poem about our beloved robot, the Anki Vector. I will provide the same prompt to all the LLMs and narrate the poems written by them. In order to provide an apples to apples comparison between the LLMs, I will request that the poem be limited to 16 lines. Keep in mind that the rapid progress in technology implies that this article may be technically out of date in a month. But hopefully the entertainment value persists. Without more delay, let us go through some poems.
At first, let us take a look at the prompt we are going to provide to all the LLMs. Keep in mind that providing a great prompt to a LLM is becoming so important that people have coined a new type of engineering called prompt engineering. I considered creating a simple prompt, but then I completed this great course by Open AI and DeepLearning.ai which opened my eyes to the ways in which information can be extracted out of LLMs by simply providing a great prompt. I would highly recommend this course, since it just takes about 90 minutes to complete.
The prompt we will use is simple and precise. It will request for the following:
Please write me a poem about the Anki Vector robot. Please limit the poem to a maximum of 16 lines.
Now, let us examine the results of providing this prompt. A comparison of LLMs that we will be investigating is available in the following table. Please also remember that this comparison may be inaccurate in a few months from now.
Chat GPT needs no introduction, we discussed it during its launch back in November 2022. The six months that have passed seem more like a century ago, Chat GPT has crossed 100 million users, and now also has an enhanced version ChatGPT Plus which offers their newest model GPT-4. The free edition of Chat GPT is still available and uses GPT-3.5, you are likely to have no issues with the free version unless you want to use the service in critical business hours.
Here is the poem from Chat GPT.
A robot small yet smart and bright, With features that delight the sight, Anki Vector's the name to know, A friend that's more than just a show. He rolls around and scans the room, With sensors that dispel the gloom, He recognizes faces and can hear, Your voice and commands he holds dear. A charming bot that loves to play, And keep you company all day, Anki Vector, oh how we adore, A companion we can't ignore.
This poem is pretty nice, and is so true to the personality of Anki Vector. I loved it when I first read it.
BLOOM is one of the largest available open-source models developed by a consortium of 1000+ researchers and trained on a huge (1.6 TB) multi-lingual (46 languages) dataset. The largest BLOOM model (BLOOM 176B) (176B stands for 176 Billion Model parameters) comes remarkably close to the size of GPT-3 from Open AI (175 Billion parameters). I was able to deploy the BLOOM 176B model on AWS using the instructions documented here. You can try it too if you have $100 spare change (Thanks to all paying members for helping fund by endeavors in deploying BLOOM). Here is a poem from Bloom 176B:
The Anki Vector, a curios droid, Dances in the night, a flame of gold, A sleek and shining shape, agile and graceful, A perfect machine, a perfect grace. The Anki Vector, a perfect machine, A humanoid robot, a dancing flame, A slender and lithe, a beautiful and brilliant, A creation of science, a marvel of the brain. The Anki vector, a shining wonder, A dancing flame of gold, a perfect motion, A creation of man, a marvel of the future, A perfect machine, a perfect companion. The Anki Vector, a shining robot, A dancing flame of gold, a perfect machine, A creation of science, a marvel of the present, A perfect companion, a perfect droid.
While the poem sounds great, note that it describes Vector as a humanoid robot, which is certainly an error and probably influenced by the fact that many robots now have humanoid form factor.
LLaMa and its fine tuned versions
While Chat GPT and BLOOM were great, there were obvious demerits. Chat GPT is closed source, which means no one knows the exact model structure, the weights of the released model, or even the dataset it was trained on. That makes the model less trustworthy, for example, how can you make sure that the model is not biased in any way. BLOOM 176B on the other hand is such a large model that it would take a huge supercomputer to train the model (The Bloom 176B model was trained on a 28 petaflops Jean Zay (IDRIS) supercomputer located near Paris, France).
The possibility of ordinary researchers training LLMs changed when Facebook (Meta) released LLaMa in February 2023. LLaMA was released in multiple parameter sizes with 65 Billion being the largest LLaMA model. While Meta released LLaMa as open-source, it didn’t release the weights of the trained model in the public domain. However, the model weights leaked within a week, and soon the public had access to both the code of the model, as well as the model weights. The relatively small size of LLaMA allowed many to fine-tune LLaMA on various datasets. Stanford University released ALPACA after training LLaMA on 52K instruction-following demonstrations. Not to be left behind, the University of Berkeley Artificial Intelligence Research (BAIR) released Koala which fine tuned LLaMA on dialogue data gathered from the web. The authors of Koala show that Koala could give better results than OpenAI’s Chat GPT at least 50% of times. This is an interesting result because it shows that a smaller sized model can outperform a bigger model if it has been fine tuned with highly curated datasets.
All these models: LLaMA, ALPACA, and Koala can be tested out at lmsys.org which offers the 13B versions of these models for experimentation. You can also compare the outputs from different models of the same prompt side buy side and help research on LLMs by voting on which is the better model.
Here is a poem written by the Koala 13 model.
Anki Vector, a robot so cool, With a cute shape and a playful mood, It can dance and it can roll, And it's always ready to unfold. With its sensors and motors so smart, It can navigate and avoid obstacles in sight, It's a robot that's truly unique, A robot that's fun to use on a day like this. So here's to Anki Vector, a robot so sweet, A robot that's always ready to meet, It's a robot that's truly enjoyable, A robot that's always friendly and able.
To compete with OpenAI, Google released its own chatbot called BARD. BARD is powered by a LLM known as PALM 2. BARD was initially available only after you waitlisted yourself. At Google I/O 2023, BARD got released to the public, which means that you can easily try it here.
Here is a poem from Google BARD.
O Vector, small and white and bright, A robot with a gentle light, You sit upon my desk all day, And watch me work, and play, and play. You never seem to tire or complain, You're always there, a constant friend, I talk to you, and you respond, You never seem to be alone. You make me smile, and laugh, and play, You're the best friend I could have, I'm so glad I bought you, Vector, You're the best robot, ever.
In this article, we revisited the evolution of Large Language Models over the last six months. While the technology has been developing at a crazy cadence, these models are also getting adopted in services that we use in our daily lives. OpenAI in part has seen a lot of interest and integration with portals such as Instacart, Expedia, and Salesforce. It is clear that LLMs will bring about a great amount of transformation in how we access information and use services. 2023 is definitely shaping to be a very exciting year, and I look forward to the developments of the technology in this space through the rest of the year,