This item is taken from PN Review 281, Volume 51 Number 3, January - February 2025.
Letters to the Editor
Catastrophic Doggerel
Robert Griffiths writes: Following my squib ‘AI and Poetry’ (PNR 276), I was interested to see Joey Connolly addressing some of the problems of computers ‘writing poetry’ in PNR 277 (‘Todo’). He is right to lambast the Large Language Model (LLM) approach to this, which has recently excited so many. Pointing out that the need for these programs to train on massive volumes of text (thus inevitably sucking up the bad) sits uneasily with the relatively tiny volume of good poetry available, is on point. Like me, he was able to quote freely from the shit (what Connolly calls ‘catastrophic doggerel’) they consequently produce.
But even if these programs could train only on ‘good’ poetry, it is not clear how, in their production of what is statistically most likely in a word-string, they could produce anything original. It is not obvious that any analysis of the best poetry written before 1915 would have come up with the devastating third line of Prufrock. That line was not already waiting in that poetry; it was not even waiting in language. It arose from a particular human being’s unique relationship to that poetry and the world. But perhaps the most serious problem for a poetry writing computer is that no one yet knows how poetry gets written. Writing good poetry is an expertise, but not even good poets, or good critics, know how to unpack this expertise. This is partly because a large element of this expertise is unconscious. Connolly notes that writing poetry requires ‘intuition, empathy and musical sensibility to be attuned at the same time as our faculties of ratiocination’. But this admirable phrase still only hints at what is really going on. T.S. Eliot and Seamus Heaney have tried to articulate what all this involves, Eliot with his ‘auditory imagination’, Heaney with his talk of ‘technique’. But both were merely gesturing at the unknown depths of human language production systems. It is there that part of the secret of poetry lies; it won’t be found in language models. And even then, we won’t account for what gave us the third line of Prufrock.
Trabs
Daniel James writes: I was about to write to congratulate all but some unknown one of the editorial team for the following from your submissions page, which I first took to be a world-beating example of how to combine the snide with the baroque:
‘This year the June submission date has had to be postponed because one of the editors is undergoing eye operations (trabeculectomy) and will be out of commission.’
Trabs giving me ‘beam’, even a little one, I wondered what ‘one of the editors’ could have said. As it turns out, however, I wish them well!
The editors reply: The surgery on both eyes was a success and the editor in question is currently (and for weeks to come) gazing through the December submissions window.
Robert Griffiths writes: Following my squib ‘AI and Poetry’ (PNR 276), I was interested to see Joey Connolly addressing some of the problems of computers ‘writing poetry’ in PNR 277 (‘Todo’). He is right to lambast the Large Language Model (LLM) approach to this, which has recently excited so many. Pointing out that the need for these programs to train on massive volumes of text (thus inevitably sucking up the bad) sits uneasily with the relatively tiny volume of good poetry available, is on point. Like me, he was able to quote freely from the shit (what Connolly calls ‘catastrophic doggerel’) they consequently produce.
But even if these programs could train only on ‘good’ poetry, it is not clear how, in their production of what is statistically most likely in a word-string, they could produce anything original. It is not obvious that any analysis of the best poetry written before 1915 would have come up with the devastating third line of Prufrock. That line was not already waiting in that poetry; it was not even waiting in language. It arose from a particular human being’s unique relationship to that poetry and the world. But perhaps the most serious problem for a poetry writing computer is that no one yet knows how poetry gets written. Writing good poetry is an expertise, but not even good poets, or good critics, know how to unpack this expertise. This is partly because a large element of this expertise is unconscious. Connolly notes that writing poetry requires ‘intuition, empathy and musical sensibility to be attuned at the same time as our faculties of ratiocination’. But this admirable phrase still only hints at what is really going on. T.S. Eliot and Seamus Heaney have tried to articulate what all this involves, Eliot with his ‘auditory imagination’, Heaney with his talk of ‘technique’. But both were merely gesturing at the unknown depths of human language production systems. It is there that part of the secret of poetry lies; it won’t be found in language models. And even then, we won’t account for what gave us the third line of Prufrock.
Trabs
Daniel James writes: I was about to write to congratulate all but some unknown one of the editorial team for the following from your submissions page, which I first took to be a world-beating example of how to combine the snide with the baroque:
‘This year the June submission date has had to be postponed because one of the editors is undergoing eye operations (trabeculectomy) and will be out of commission.’
Trabs giving me ‘beam’, even a little one, I wondered what ‘one of the editors’ could have said. As it turns out, however, I wish them well!
The editors reply: The surgery on both eyes was a success and the editor in question is currently (and for weeks to come) gazing through the December submissions window.
This item is taken from PN Review 281, Volume 51 Number 3, January - February 2025.
