@somewheresy IMO the only time that reading LLM output is tolerable right now is when it’s just been output at your request. No one wants to read anyone else’s generated content, or even their own content from previous sessions!
English
Mitchell Bosley
334 posts
Mitchell Bosley
@mitchellbosley
Postdoc at CENIA | Previously at the Schwartz-Reisman Institute at UofT | PhD from University of Michigan Poli Sci
Santiago, Chile Katılım Ocak 2009
1.4K Takip Edilen312 Takipçiler
@emollick This is also my experience. What’s not clear to me is whether this is the fault of inherent limitations of the tech, the fact that writing well is an oom+ harder than other tasks these models are trained for, or just that researchers haven’t modeled the reward function right yet
English
My experience so far with LLM fiction writing is that it takes advantage of our assumption that an author is writing things for a reason, so we are charitable to a book's quirks & do mental work to assign them real meaning. But the AI doesn't have a reason, its just bad writing.
English
Also supports my theory that the only person who gets anything out of LLM-content is the one generating it. Like, I’m sure this was a fun exercise for the OP, but my expectation is that no one will read this book end to end at any point
English
Yes it is basically unreadable lol, despite the noise that the OP made. You can scan any of the 200+ pages and it all has the same flat, formulaic tone, despite the human doing everything in their power to enable Claude to write well (multiple agentic passes, etc)
Ethan Mollick@emollick
I read a few dozen pages of this and it is not bad for LLM fiction, but also very very LLM-y, from the themes to the fact that there are lots of staccato conversations and meaningful silences and overwrought metaphors and very little differentiated character development.
English
The blurb from the article is *also* awful AI-inflected writing!
English
@angeris This post is a good case study: x.com/Houda_nait/sta…
this is clearly a smart person who is working on very consequential stuff at one of the best companies in the world to work at. but the fact that it is clearly co-written with gpt 5.x imo diminishes it
Houda Nait El Barj@Houda_nait
English
@mitchellbosley oh there's definitely an implicit signaling component to it sometimes (but not always!)
English
@angeris And so, when someone sends you an email/post/message clearly AI authored and passes it off as their own, the logical inference is that the sender is either not informed enough to tell the difference, or thinks that *you* are able to be duped
English
@angeris I think there’s a couple layers to this. First, pretty much every frontier model today stinks at writing (yes, even Claude!) in the sense that they are mode collapsed to a handful of writing patterns that are impossible to ignore (eg “It’s not X it’s Y” and all variants)
English
@angeris Then, once you recognize how superficially these systems model the act of writing, it becomes apparent pretty quickly that most of what feels like actual human-level intelligence is mostly just a parlour trick, and endeavour to do most of your important writing solo
English
@tszzl @nikitabier Sure but 99% of these articles are slop to begin with. Who will be served by slop summaries of slop articles?
English
@nikitabier you’re going to get hate for this but it’s obviously the right product choice. I’m sure you have the data but I assume 99% of people open any article and close it after seeing it’s longer than a paragraph
English
We’re rolling out summaries for Articles now. Just tap the Summarize button if you want to know if it’s worth your time to read it (or if your attention span is 12 seconds).
English
English
This is more evidence that current frontier models remain completely reliant on content-level memorization, as opposed to higher-level generalizable knowledge (such as metalearning knowledge, problem-solving strategies...)
Lossfunk@lossfunk
🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵
English
I agree with this. Even the best agentic systems are still very much “a bicycle for the mind” for me. They are really only useful insofar as I am actively providing direction, injecting taste, and so on. Maybe this changes, but I think it would require a capabilities step change
Ethan Mollick@emollick
We are back to the phase of the AI news cycle where people are underestimating how jagged the AI ability frontier is, as well as how much they still depend on expert human decision-making or guidance at key points in order to function well. Still far from "doing all jobs," today.
English
Mitchell Bosley retweetledi
We are back to the phase of the AI news cycle where people are underestimating how jagged the AI ability frontier is, as well as how much they still depend on expert human decision-making or guidance at key points in order to function well. Still far from "doing all jobs," today.
English
Great at building software and running experiments though!
English
the fastest way to disabuse yourself of the notion that GPT 5.4 is anywhere close to AGI is to try to get it to write a technical document for you about anything remotely novel. it is awful at modeling hierarchical depth in sections/subsections, quite a bit worse than Opus 4.6
English
I am honestly unsure whether this is because a) good technical writing is a uniquely complicated task or b) the post-training OpenAI is doing over-indexes on coding well in a harness rather than writing complex documents
English
@lVlarty @teortaxesTex I’m just saying he could have easily used a clanker to write out his screed like so many others on this site, but instead he sat down and typed out a wall of barely coherent text all by himself one word after another. I think that’s beautiful
English
@mitchellbosley @teortaxesTex It's semi-justified tech-psychosis. Sign of the times.
English
@alexolegimas x.com/mitchellbosley… I do think we will settle into an equilibrium shortly where AI content is mostly created for private immediate consumption, and human-generated content is preferred otherwise
Mitchell Bosley@mitchellbosley
@_aidan_clark_ My working hypothesis is that all else equal, humans enjoy the production of AI content more than consumption of already generated AI content due to the fact that the content generation loop is essentially a Skinner box
English
My prediction is that in the case of creative work, AI will be a complement to human creators. Much of the value of creative work is the human value, i.e., the fact that it was made by a person, which establishes a link between the consumer and the creator.
We see this in the data. AI-generated creative work is valued substantially less than human-generated work. And unlike other effects where "people will just get used to it", I think this will be fairly stable: we have way too many previous examples of creative automation to draw on.
AI will increase the scope of human creativity but it won't replace the creator. In fact, as other tasks get automated, it may actually increase demand for it.
papers.ssrn.com/sol3/papers.cf…
Ethan Mollick@emollick
One of the advantages of being an early user of LLMs is that I have seen The Curve with my own eyes (like in this post before ChatGPT or the term Generative AI). I notice recent AI users & companies adopting AI anchoring on recent capabilities as if they are stable. Probably not
English