Salvatore Sanfilippo argues that AI is already useful in code, still crude in music and audiobooks, and only works in art when taste and selection stay human.
<p>Sanfilippo’s central claim is blunt: AI is already good enough to change programming, but it is still nowhere near replacing the human judgment that makes art work. He says the difference is structural, not sentimental. Code has huge training sets, clear tests, and a domain where reinforcement learning can reward something that functions. Music, writing, illustration, and voice work, by contrast, still depend on taste, selection, and craft that the models do not reliably supply.</p>
Sanfilippo begins by arguing that programming is the field most exposed to AI, precisely because language models are unusually good at it. The reason, he says, is practical: code is abundant online, reinforcement learning can be used, and the output can be checked for whether it works. That makes the gains harder to dismiss as novelty.
The result of using these tools is working code, often even better than what one could otherwise obtain.
1:06
We programmers are, right now, in the worst condition of all.
5:00
He ties that vulnerability to the nature of programming itself. It is a language task, he says, but also one with a vast training corpus and a feedback loop that can reward correctness in a way art often cannot. In his view, that combination makes software the cleanest case for machine substitution, even if the best results still come with human prompting and supervision.
The programming domain is partly verifiable, so reinforcement learning has worked in a superior way there than in the other fields.
5:32
Once he moves into music and writing, Sanfilippo says the balance changes sharply. Generative music tools, in his view, produce rubbish, while writing models can be pushed into decent prose but not into good stories. The missing ingredient is not output volume but discernment: someone still has to know what is worth keeping.
They produce the purest trash.
1:42
They can write good prose, if deeply guided, but they cannot write good stories.
2:13
His larger point is that AI does not remove the need for taste, because taste is what decides whether a line of dialogue is merely passable or actually alive. He says the machine can propose material, but a human still has to select, reject, and revise it until it becomes coherent. Without that filter, he argues, the result is just a mix of bad, decent, and good fragments that never adds up to anything.
At the base there must still be a deep capacity for selection by the human being.
2:58
Otherwise what you get is a mix of bad things, decent things, and good things that, without selection and reworking, still ends in nothing good.
3:07
Sanfilippo is skeptical that AI will wipe out musicians or artists, but he is equally skeptical of the cultural economy that already surrounds them. He says mainstream music has long rewarded a tiny set of hits and familiar faces, while audiences are trained by repetition into liking what they are fed. On that view, AI is not the original threat; the market was already built to flatten taste.
People, on average, have zero critical aesthetic capacity.
3:48
The songs from Sanremo impress me because they seem like things you could invent in three seconds.
3:59
He extends that skepticism to the economics of art. Musicians already struggle to earn a living, he says, and writers almost never did so in the first place. For that reason, he argues, the idea that AI is uniquely destroying artistic livelihoods misses a harder truth: the business was already hostile to most artists.
Music is a subsection of a subsection, you can never live off it.
4:42
Writing has never been something one lived off.
4:51
Where Sanfilippo sounds most exasperated is on audio generation. He says text-to-speech remains far behind what is needed for a convincing audiobook, even when the technical benchmarks look impressive. He tested Gemini 3.1 Flash TTS on one of his own stories, segmented it with Claude, and rendered it with Python, only to find the result technically polished but dramatically wrong.
The quality of the audio is excellent, the intonation is not at all robotic, but reading an audiobook well is something else.
8:03
All the voices I tried were voices like a damned bad Boris actress.
8:34
What bothers him most is the industry’s impatience. He says companies are pushing automatic audiobook production to avoid paying actors and narrators, then treating the cheap result as proof that the technology is ready. The outcome, he argues, is slop, and the blame lies less with neural networks than with organizations that use them before they are good enough.
They have taken the step too far in order not to pay the artists, the actors, the actresses.
9:16
This is a reason to hate capitalism, not neural networks.
9:38
Sanfilippo ends by separating technical capability from cultural consequence. He says image models are improving, illustrators can use them as tools, and even some visually polished output may help people with taste but without resources. Yet he keeps returning to the same line: the human artist is still the selector, the editor, the one who decides what counts.
These are very interesting tools, they allow those who have taste to make the cover of the record by themselves.
6:21
I do not see the disappearance of musicians on the horizon.
4:28
Still, his conclusion is not that AI leaves art untouched, only that it has not yet overrun the part of art that matters most to him. He points to microtonal band Yin as an example of new music that can still feel specific, even if it draws from older forms and may irritate conservative ears. For him, that is evidence that art is not dying, only being sorted into better and worse uses of new machines.
Art is certainly not at the end because of AI.
12:17
The problem is not the networks.
9:39
Why does Sanfilippo think programmers are most exposed?
He says programming gives AI the best mix of abundant data, language-like structure, and verifiable outputs. That makes it the field where models can produce working results fastest.
Does he think AI can write stories?
Not well, according to Sanfilippo. He says some models can generate good prose if deeply guided, but they still cannot write good stories on their own.
What is his view of AI in music?
He is dismissive of current automatic music tools. In his view, they mostly generate terrible results and cannot replace the aesthetic judgment that makes a track worth hearing.
Why is he so angry about AI audiobooks?
He thinks companies are using automation to avoid paying narrators and actors. The result, he says, is cheap, bad audio sold as if it were a solved problem.
Does he believe AI will end art?
No. He says AI may change workflows and increase output, but it will not make artists disappear because taste and selection still belong to humans.
AI-assisted summary of Salvatore Sanfilippo's podcast, verified against the original transcript.