[Premium] AI Civilizations, OpenAI & Anthropic News, & More

TLDR;

Technology:

🌏 Project Sid’s AI Civilization

🧐 News Alert: LLMs Still Can’t Reason

✋🏼 Meta’s Touchy Robots

Markets:

🔮 NVIDIA’s Insight into Elon’s Colossus Data Center

😳 OpenAI To Start Using AMD GPUs

Product:

😎 ChatGPT Search is Finally Here

🤯 Anthropic Levels Up the Ante

AGENTS

Project Sid’s AI Civilization

Project Sid by Altera Labs aims to create AI civilizations formed by autonomous AI agents who live, interact, and communicate in an open world. Now, they have released additional progress with some interesting results.

In this new update, we see the agents:

Developing long-lasting social relationships (and rivalries)

Having palpable differences in actions depending on their personality (introvert/extrovert)

Adopting specialized roles (if there’s already a baker, I’ll become a carpenter)

And many more, including agents who comply with taxation laws (even voting for and against them) and, funnily enough, share memes!

TheWhiteBox’s takeaway:

Albeit the seemingly magical nature of all this, it’s not magic. These aren’t conscious AIs; they are simply amplifying human biases and actions, aka ‘what a human would do in my case,’ adopting roles based on what their training data suggests is the most likely/adequate behavior based on input data.

While it’s undoubtedly a very interesting sociological experiment, I do not enjoy Altera's framing of this project, using words like ‘conscious’ many times.

To me, current AIs are pattern recognition machines and nothing more. They simply imitate or impersonate the most likely behavior, not because they feel like it but because they were trained to do so.

Seeing AIs act like humans makes it almost impossible not to want to anthropomorphize them, but the truth is that a machine acting intelligent does not mean it’s intelligent. Sure, they have developed a great intuition as to ‘if x happens, do y’, but the fact ‘y’ is an intelligent act—or adequate for the situation—does not mean that the process by which the AI arrived at that conclusion was intelligent but mere ‘if x do y’ pattern memorization.

Still, this type of research could give us insight into how societies and individuals engage, even possibly analyzing particular traits of each human. But nothing more.

LLMS

News Alert: LLMs Still Can’t Reason

You may start to believe I’m an LLM hater (I’m not!), but not another week goes by without new evidence that LLMs can’t reason. This time, researchers follow a similar approach to Apple, applying noise to the prompt (inconsequential information) to try and fool the LLM.

Unsurprisingly, all LLMs see a considerable decrease in performance, signaling that they can’t distinguish good signals (the important information) from irrelevant data.

In turn, this suggests that LLMs are still learning superficial token patterns (‘if the input tokens are x, the next token is y’) instead of truly abstracting the key patterns in data ‘the issue is x. Therefore, the solution is y’.

TheWhiteBox’s takeaway:

It must be said that for a strange reason, the researchers in this piece avoid using o1 models. However, we know from Apple’s research that o1…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.