lyra bubbles @_lyraaaa_ 2026-04-20

all LLMs are either claude-like or GPT-like

method: cosine sim heatmap of per-model-averaged responses to 50 prompts sent thru gemma4 activation-space (107,520 dims)

notable exceptions - haiku 4.5, gem3flash

(and to a lesser degree, m2.7 and gemma4 itself)

Image

lyra bubbles @_lyraaaa_ 2026-04-20

this is *only 50 prompts*, and only singleturn

id run more but that would get expensive

but pattern held up whether it was 20 prompts or 50 and actually strengthened at 50 so

lyra bubbles @_lyraaaa_ 2026-04-20

used *actual* user prompts instead of claude-gen prompts and the figure looks a lot different, so maybe i was too quick to post the first one. some patterns still hold though

tip: use real data

Image

Andy Ayrey @AndyAyrey 2026-04-20

be interesting to see how gemma4 compares to sonnet 3.5-3.7 and to 4o…

Boyd Kane @beyarkay 2026-04-20

*Gemma 4 thinks that* all models look Claude shaped or gpt shaped

lyra bubbles @_lyraaaa_ 2026-04-20

as good a text embedder as any

Shibo Hao @Ber18791531 2026-04-20

Interesting! Can you specify where you extracted the activations? like which layer and which token?

lyra bubbles @_lyraaaa_ 2026-04-20

average of all tokens, and using all layers to create one massive 2560*41 vector

ItoriH @it_oriH 2026-04-20

I feel dumb.. eh… How did you extract activations from closed source models? Or did you ask those models the same prompts, passed that through Gemma and compared the differences?

lyra bubbles @_lyraaaa_ 2026-04-20

yes that’s exactly what I did

all the activations are gemmas

Suresh @_Suresh2 2026-04-20

did the clusters stabilize around 50 prompts, or was that just the sample size you picked?

lyra bubbles @_lyraaaa_ 2026-04-20

i just picked it

running 350 as i post this

ShipWithAI @shipwithaiio 2026-04-21

Team GPT or Team Claude? Gemini 3 Flash: I choose to go to sleep :v

Bakhtier Gaibulloev @slashmsu 2026-04-20

this tracks with what i see in production. Claude-family models handle long structured tasks (multi-step agent loops, code migrations) differently at a fundamental level. same prompt, same temperature, completely different failure modes compared to GPT-family.

the activation