LLM links, 4/3
Bing, Bard, and Open AI; also Ethan Mollick; The Zvi; Ian Leslie's interview of Tom Chivers; Freddie de Boer
A reader writes that compared with Open AI’s version of GPT-4, Bing is terse in its responses. When the reader asked Bing to write about David Hackett-Fischer’s Albion’s Seed, the response was
Albion’s Seed by David Fischer is a history professor’s nine-hundred-page treatise on patterns of early immigration to the Eastern United States. The book describes four migrations: the Puritans to New England in the 1620s, the Cavaliers to Virginia in the 1640s, the Quakers to Pennsylvania in the 1670s, and the Borderers to Appalachia in the 1700s12. The book is not light reading but it explains everything about America according to some people who have read it3
OpenAI gave a more complete, satisfying response.
In another comparison, I gave the same prompt to write a children’s story to both OpenAI and to Google’s Bard, and Bard’s response was every similar to OpenAI’s, but flatter and duller. In general, Bard uses more recent information from the Web, but it is cautious and boring compared with OpenAI.
Ethan Mollick has a more well-informed comparison of the various models. For example, he writes,
Basic prompts result in boring writing. Getting good writing out of ChatGPT takes some practice, and here is a guide to doing that. ChatGPT-4 is much better at writing. Bing can be incredible at writing, but needs some convincing.
You will want to bookmark this comprehensive post.
A reader asks me about the idea of “pausing” AI. I refer you to coverage by The Zvi.
As per the position he staked out a few days prior and that I respond to here, Tyler Cowen is very opposed to a pause, and wasted no time amplifying every voice available in the opposing camp, handily ensuring I did not miss any.
My thoughts:
In terms of capability, the large language models are not close to being scary.
The speed at which the capability of the models improves is somewhat scary.
But my intuition, and it is little more than that, is that the LLM models will approach an asymptote that is well shy of Artificial General Intelligence. I think that requires more than word prediction plus Reinforcement Learning with Human Feedback.
The Zvi’s more regular long post is here.
You can also read Ian Leslie’s interview with Tom Chivers.
we [humans] have a theory of the world - we have a sense of how the universe functions that informs our linguistic parsing. And this, fundamentally, is a key difference between human intelligence and a large language model.
That might be a way of articulating my intuition that we will reach an asymptote with the current generation of AI based on LLMs.
Substacks referenced above:
@
@
@
@
@
Chat GPT seems like a those RAs that sometimes show up in 1st year grad programs--indifferent or willingly lying to get the job completed while minimizing effort. Test: have Chat search for some obscure data or technical, non-literary fact. It's not unusual for Chat to say the data doesn't exist or the fact isn't there. You can insist the data or fact is there, and sometime Chat will find it and apologize. So, be careful accepting Chat's factual claims--it's very lazy as a researcher. In contrast, it's helpful as an editor of prewritten text and amusingly successful in writing poetic forms and laudatory text for birthday and retirement congratulations. Being lazy, low cost and convincing, Chat and AI will dominate mass media, corporate and government writing.
Agreed: the more practical implications at this time (significant job displacement or adjustment, the ability to create fake videos etc) are the things I worry about most though these do not at all call for a pause.
I’m far from an expert on this topic but to the extent these models are trained on data that is ultimately from humans, it does seem that there should be an asymptote.
To me the asymptote of anything that is trained on human data is becoming the smartest person in every topic and having incredible speed at that. But I don’t see how that enables some AI take over.