j_gravestein

A recent study suggests that vision models like GPT-4o may suffer from blurry vision or ‘myopia’.

Researchers tested several models on a series of tasks designed to evaluate what vision models see when looking at colored lines and shapes.

This type of research is crucial in helping us understand what these models are good at and where they break down.

Ludwig Wittgenstein, the Austrian philosopher, wrote extensively about language and its intricacies.

His ideas are especially relevant for anyone working in AI, because his views on language help us better understand the limitations of large language models.

This essay explores those ideas, what you can take away from them, and why learning the meaning of words doesn’t necessary equate to learning something about the world.

Recent product launches of the Humane AI Pin and the Rabbit R1 have been deeply disappointing.

These disappointments expose a gap between how AI products are being marketed and hyped up vs. what they’re actually capable of.

It’s part of a bigger trend that is rooted in Silicon Valley’s obsession with attracting investors and generating media attention at the expense of delivering real value to customers.

During his first big public appearance after his departure from OpenAI, at an event of Sequoia Capital, Andrej Karpathy said:

“Roughly speaking, the way things are happening, everyone is trying to build what I like to refer to as a LLM OS.”

Are all junior software engineers out of a job in 1-3 years?

Last week, The Information reported that OpenAI is developing an AI agent to “automate complex tasks by effectively taking over a customer’s device”.

It fits into a broader trend, similar to what the Rabbit R1 is doing, which is teaching computers how to use our devices for us: perform cursor movements, click buttons, etc. Now, you might think that is a rather bad idea, and it probably is.

That said, the emergence of AI agents, powered by generative AI, represents a new generation of assistants that are more flexible, can do stuff for you, and in the near future may even act on your behalf.

Last week it was reported that a fake Joe Biden robocall urged New Hampshire voters not to vote in the upcoming Democratic primary.

It was a coordinated effort and shows a glimpse of what’s coming in 2024, US Presidential Election year.

A second-order effect of deepfake technology is that it creates plausible deniability for politicians and public figures, by claiming real footage as AI-generated.

This is not a hypothetical threat. This is currently happening.

We’ve seen this from Elon Musk, January 6th rioters in court, and Donald J. Trump, who, after Fox News aired an anti-Trump ad, claimed the people behind the ad had used AI to make him look bad. (The video was authentic!)

The substack article talks about this and more.

Key insights:

- A recent study tested GPT-4, GPT-3.5, and 1960’s ELIZA, to see which program best mimics human conversation in a Turing test. Participants had to guess if they were interacting with a human or an AI.

- Surprisingly, the old ELIZA program outperformed GPT-3.5. GPT-4 did better than ELIZA, but didn’t reach a 50% success rate, which effectively means worse than a coin flip.

- The authors write that the Turing test still has relevance in understanding how humans interact with AI. However, we should refrain from seeing the Turing test as a barometer of AI intelligence.

The world’s first commercially-viable autonomous humanoid robot?

AI companions are on the rise. But what are the knock-on effects of outsourcing empathy on a global scale?

Finally OpenAI admits AI detectors are useless.