In the News: ChatGPT, its Competitors, and… Mittens??
We’re back! This year, we’ve got some new authors that are looking forward to sharing their work. Expect some great content from HoyAlytics soon! This week, we explore ChatGPT in depth and take a look at one of its competitors, Claude. Also, a new deceivingly cute chess bot is taking the world by storm!
How does ChatGPT work?
By Annika Lin
ChatGPT uses deep learning algorithms to create text responses to prompts. It’s based on a transformer model that uses self-attention mechanisms to process and generate text. Transformer models are mainly used in natural language processing and computer vision. These models take input, encode the input in multiple layers, and then decode the layers. The transformer model uses three weight matrices: the query weights, key weights, and value weights. These matrices form attention units, which are the building blocks of the model.
The self-attention mechanism specifies weights for the model, depending on the importance and relevance of the word. For example, take the question “is chatgpt bad for humans?” Words in the beginning of the question such as “the” are assigned lower weights than nouns such as “chatgpt” and “humans.” The attention weights are calculated using the query and key vectors.
If you’re curious to learn more about the model (including the cool math involving dot-products and matrices), feel free to dig into the mathematics behind this model here.
Nice to meet you Claude: A new rival to ChatGPT
By Spencer Karp
By now, you’ve heard of ChatGPT. You’ve probably read about it, saw it on the news, or even tried it out for yourself. Now, ChatGPT has a new rival: Claude. Claude is built by Anthropic, a company founded by former OpenAI employees (The company that built ChatGPT). While both ChatGPT and Claude were built with similar functions in mind (i.e. conversational interaction), they are programmed and perform very differently. ChatGPT uses reinforcement learning with human feedback. To perform this type of learning, humans rank different potential outputs to the same prompt and ChatGPT learns from those rankings to make future decisions. Many people have noted that this style of learning means that ChatGPT may respond with vulgar, offensive or dangerous language. On the other hand, Claude uses an approach called Constitutional AI. In this approach, before learning anything, Claude is given a set of rules that it must abide by (like a Constitution). This makes it easier for Anthropic to ensure that Claude does not respond with offensive or dangerous language, even if the user asks for it. Since ChatGPT and Claude are programmed differently, they perform differently when given the same prompt, this article examines some of those differences and identifies some shortcomings of both algorithms. Check it out to learn more about Claude and the future of AI chatbots.
Rage-Inducing Robots: Newest Chess AI Takes Players by Storm
Check out the article here!
AI tends to play a behind-the-scenes role in most contexts. The end user probably doesn’t directly experience the model. In games, however, AI bots are directly interacted with as a training tool for players.
For Chess.com this month, a certain AI was the craze. Mittens, an incredibly strong but ironically cat-themed bot, dominated streaming content and player forums. Fascinatingly, this new computer isn’t as strong as other available chess engines like Stockfish! It plays ultra-aggressive chess while taunting its opponents with preprogrammed chat messages.
The effect of the limited-time cat bots on chess.com’s platform is undeniable — traffic spiked by 40% this month, with almost 30 million games being played each day of January. Perhaps Mittens being weaker than top chess engines made it seem beatable, or the meme-worthiness served as great content for top YouTubers. Maybe players just enjoyed experiencing all-out attacks of near-perfect accuracy. Regardless, chess.com’s campaigns and various computer difficulties available show that every AI model doesn’t necessarily need to be the most accurate or discover new technology. It’s purrfect for it to simply meet a specific task!
Want to check it out for yourself? Try it here!
