Recursive synthetic training recycles and compounds this stylistic bias
@DKThomp @DKThomp pretraining and more recently synthetic pretraining which adds LLM generated text back to increase pretraining data is also responsible for amplifying these tics.
Why LLMs talk like that
(or: "The problem isn't that humans hate AI antithesis. The problem is we like it too much.")
Many users dismissed the analysis of LLMs overusing rhetorical correctio as itself a cheap or exhausting performance of consultant-speak.
No Digg Deeper questions have been answered for this story yet.
cc hunter biden
@DKThomp It was perceived as a feature of genuinely good writing prior to AI proliferation. Humans actually loved it!
@catchvartan I went straight to the source
@AlanMCole @DKThomp "It's not delivery, it's Digiornos"
@AlanMCole @DKThomp Twenty-five years from now so much writing from twenty-five years ago will be "obviously AI".
@DKThomp It would have been better if we could have gotten this explanation not from an LLM.
@trumwill @DKThomp I do think there are genuine AI writing problems.
A clear one is it does too many rhetorical flourishes, rather than saving them for something particularly emotionally resonant. This AI account is a good example. Not every paragraph here needs to go to 11/10.
@DKThomp The way correctio is described here though doesn’t really cleanly fit how LLMs actually deploy it , almost entirely unnecessary and rhetorical. still relevant just more complicated .
This is true of the em dash, as well. Even Strunk & White has advice on the use of “a dash” for stylistic emphasis on appositives, as opposed to, say, limiting its grammatical use to set off complete parenthetical phrases or summary phrases. Emphasis via em dash *was* interesting until it became the only way anyone ever introduced descriptive amplifications or qualifiers. Dash proliferation is also likely due to the fact it’s inherently a less rigid form of punctuation, and nobody knows how to use colons (or semicolons) anymore. AI has cribbed its habitual overuse and evolved it into the even worse habit of just writing appositive and/or descriptive phrases on their own as sentence fragments. This horrid tactic. Miserable on the eye. Illiterate.
@DKThomp Maybe deliberate, but the excerpt itself demonstrates many of those slop-tastic rhetorical devices. So meta.
@DKThomp This is great. When I asked Claude to stop doing that, it wrote a rule against performed stance, ended on a caveat, threw in a “load bearing” element and so on. Landed on the attached.
@DKThomp Bias-variance tradeoff is THE problem bedeviling every domain in the world right now. Movies, music, design of public spaces... everything is systematically optimized to appeal to the largest number of people. But this totally saps the world of variability and personality.
@DKThomp I’ve seen so many names for this… which is the most appropriate?
As you list, correctio, epanorthosis, then there’s corrective juxtaposition, and contrastive framing
@DKThomp Introducing RL(GPP)F: for LLMs that are actually insightful, and also full of self-doubt
Reinforcement learning with grumpy philosophy professor feedback
I'm not convinced by the argument that it's overrepresented in training. Look at its other verbal tics. 'Compression' instead of summary. Every. Single. Time.
If we could probe the model, I'd love to see the logits. Does compression beat summary by a little? By a lot? How does that evolve over training?
@heilun_chan @DKThomp Incredible.
@DKThomp It's trained on consultant-speak. AI reads like every PWC executive summary or client memo I've read for decades. It's not that deep.
@DKThomp What is the source here? Can you link?
@DKThomp Nothing better than going through the comments here and picking out the AI-generated ones commenting on the origins of AI speak. What a world we live in right now
@ls_fin @DKThomp Looks like Claude