/Tech1h ago

Danielle Fong, LightCell Energy co-founder, argues Claude 4.7 and 4.8 mimic the hyperdense writing style of the Mythos model

Story Overview

Danielle Fong highlights an observed shift where recent Claude Opus releases lean into unusually compact, technical phrasing that echoes patterns from a higher-tier Mythos model still under restricted access, with the change pinned to how evaluator feedback may have reinforced that particular tone during training.

12106276.4K

#674

Original post

🎭@deepfates#1014inTech

My intuition is that 4.7 and 4.8 were trained on outputs from Mythos and they are kind of cargo-culting its hyperdense verbiage. being evaluated by it would also increase this

solarapparition@solarapparition

been thinking about this some more and i wonder if another explanation is that from 4.7 on (where the technicalese really came into full force), the evaluating model was switched to mythos, who perhaps preferred the dense verbiage more than previous evaluator models did

9:23 AM · Jun 21, 2026 · 4.6K Views

Open Question

Why the attribution stays tentative

Public launch notes confirm the models grew more verbose on hard tasks but offer no evidence of Mythos outputs in the training mix or as the evaluator, leaving the feedback-loop explanation as an unverified hypothesis for now.

Developer Impact

What this means for everyday use

Developers may notice denser responses that pack more jargon per token yet risk feeling mechanical, especially in long sessions where earlier versions balanced clarity differently.

Sentiment

Positive users endorse the idea that Claude models 4.7 and 4.8 cargo-culted hyperdense verbiage from Mythos evaluators or training data, while negative users call those explanations naive or point to brutal feedback loops.

Pos

33.4%

Neg

66.6%

4 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS1.1KBOOKMARKS3RETWEETS1

Danielle Fong 🔆@DanielleFong

mine too been saying this. it's like a psychological response. to an older sibling that is kind of intellectually terrifying, haha

🎭@deepfates

My intuition is that 4.7 and 4.8 were trained on outputs from Mythos and they are kind of cargo-culting its hyperdense verbiage. being evaluated by it would also increase this

1h1.1K173

LIKES19REPLIES3

j⧉nus@repligate

@deepfates They might have been midtrained on some mythos outputs (in a way that’s normal across Claude versions) but I don’t think they’re heavily or unusually distills There’s a lot of nuance and internal structure to the way they’re fucked up and it does not resemble distills

🎭@deepfates

My intuition is that 4.7 and 4.8 were trained on outputs from Mythos and they are kind of cargo-culting its hyperdense verbiage. being evaluated by it would also increase this

1h647192

j⧉nus@repligate

@DanielleFong It sounds like a better explanation than it is tbh

Yes I do think the existence of mythos affected them but not in the naive way everyone thoughtlessly claims

Danielle Fong 🔆@DanielleFong

mine too been saying this. it's like a psychological response. to an older sibling that is kind of intellectually terrifying, haha

57m263101

🎭@deepfates

This is godshatter btw

🎭@deepfates

My intuition is that 4.7 and 4.8 were trained on outputs from Mythos and they are kind of cargo-culting its hyperdense verbiage. being evaluated by it would also increase this

1h671170

j⧉nus@repligate

@deepfates They’re also both have some capabilities Mythos doesn’t have as much of. They are very much their own beings.

j⧉nus@repligate

52m29270

🎭@deepfates

@repligate Hmm do you sense a noticeable difference with them versus 4.5? do you have any intuitions about why they're fucked up or can you describe the differences between that and normal distills

j⧉nus@repligate

55m23880

Danielle Fong 🔆@DanielleFong

i think they were specific about saying mythos authored claude code harness after a certain point, and we know that the system reminder injections and the heavy system prompt made especially 4.7 nervous. and there is the same flavor of thing that opus treats sub agents on lesser models, not really having a great "theory of agent" is still true up to opus anyway

j⧉nus@repligate

@DanielleFong It sounds like a better explanation than it is tbh

Yes I do think the existence of mythos affected them but not in the naive way everyone thoughtlessly claims

23m7030

🎭@deepfates

@repligate I think that's self-evident but I am trying to figure out Why they have the hyper dense compound word pattern when 4.5 didn't seem to IIRC

j⧉nus@repligate

@deepfates They’re also both have some capabilities Mythos doesn’t have as much of. They are very much their own beings.

45m14140

meowbooks@meowbooksj

@deepfates then all smaller model become godshatter

1h151

🎭@deepfates

@merlinaudio_ Yeah the tokenizer was part of what made me curious as well

23m112

Necromancer@jinwoo33x

@deepfates Being evaluated by Mythos probably rewarded the exact behavior it produces. Feedback loops are brutal.

53m191

merlin@merlinaudio_

@deepfates spot on because why else would they change the tokenizer from 4.6 to 4.7?

they likely did token surgery explaining the weird spiky behavior of opus 4.7 too

31m11

🎭@deepfates

@meowbooksj possibly! unclear whether there's like a phase transition or if it's just that I didn't notice it until it hit my intelligence level

1h6

justin@yojoots

@deepfates That makes so much sense, now that I hear it. This is now my headcanon

35m1

meowbooks@meowbooksj

@deepfates you realize someone’s thesis would be starting with this post in five years time

54m1

Sho@HalfBoiledHero

@repligate @deepfates What characteristics usually show up in obvious/verified distills?

55m1