I almost have trouble believing this survey result - that is incredibly high adoption. But it suggests that by not providing guidance, support, and access to frontier models, companies are not avoiding AI at work, they are getting secretive, bad AI at work. An urgent issue, now.

To be fair, everyone did cheer when I asked.

Regulation is a completely reasonable response to the potential risks and harms of AI. But this is not going to be the right way to do it. And work should be done to create nimble policy in response to emerging harms, rather than trying to guess all the possible uses & missuses.

One helpful high-value uses of AI is to get a first pass at legal documents and form adhesive agreements that you never would have paid a lawyer to help you with. I wouldn't trust it with anything too serious, but it does a good first pass/second look. Claude 3 Opus does well.

I showed Google Gemini 1.5 the first part of Apple's 'hydraulic press' ad for iPad. I think it kind of nailed it: 'The ad could be seen as sending a mixed message or even a negative one, suggesting that the new iPad might lead to the destruction of other valuable things.'

Interesting

The evidence across multiple papers is pretty strong: don’t use an LLM for hiring.

They exhibit many of the same biases as humans, but we don’t understand their magnitude or well how to mitigate them.

There is likely no national security organization (possibly outside of China) that can build their own frontier LLM. They have supercomputers, but the wrong kind for training models. They are either going to modify Llama/Mistral or buy one from one of the few firms with compute.

Groks “explore” Twitter trending topic summary headlines are wishy-washy, sometimes have hallucinations, and are open to manipulation. Surprisingly, though, I am finding them quite useful & much better than the old non-AI methods, once you know the limitations. AI in a nutshell.

Even after only a dozen uses, it is clear im-a-good-gpt2-chatbot is full of ghosts.

I mean this in the same way that GPT-4 and Claude 3 Opus and Gemini 1.5 are full of ghosts/sparks/whatever - they are occasionally uncanny. Seems to be an emergent feature of frontier LLMs.

I was randomly assigned the mysterious OpenAI im-a-good-gpt2-chatbot in the LLM Arena, so I naturally asked it how a superhero would make fudge and to make an ASCII control panel for a time machine. Then it timed out before I could do anything serious.

Very good answers, though.

For those of you not following the drama, gpt2-chatbot, which mysteriously appeared on an AI leaderboard site, then disappeared, has now reappeared, is definitely from OpenAI, and seems to be quite good, not sure how good

Whether this is a preview release of GPT-4.5/5 is unknown

Prompt: Produce 2 million pages of credible documents for my DCA application for my nuclear plant.

Should save $499.9M & 2M hours of work. (Whether you think this is a good or bad use of AI, it is definitely going to be a way people use it. Regulators should think about that.)