Nolife Language — Don’t say cost. Say investment.

A Darkwood experiment on language, framing and token cost. The slides compare wording, encodings and measured regressions, distinguishing a small dated corpus from general claims about languages or pricing.

Canonical interactive presentation: https://slides.darkwood.com/slides/nolife-language

Transcript

1. Darkwood · Symfony · Flow · language Nolife Language Don't say cost. Say investment. When words change both persuasion and token cost. @matyo91
2. Words change the frame COST ↓ INVESTMENT basic → essential standard → customized few → limited Same offer. Different frame. 01 | frame @matyo91
3. Then the tokenizer reads it Humans hear meaning. Tokenizers count pieces. words ≠ tokens 02 | tokenizer chars ≠ tokens encoding matters @matyo91
4. Three axes meaning ↓ token cost + persuasion Improving one axis can hurt another. 03 | triangle @matyo91
5. o200k_base · parallel corpus Same idea. Three languages. English 46 baseline French 58 +26.09% German 63 +36.96% this corpus × this tokenizer — not a universal language tax 04 | languages @matyo91
6. Slide 6 · 1/2 · représentation statique Same texts · two encodings Language cost is not a constant. Capture statique : Schéma ou tableau Lecture du schéma : O200K_BASE CL100K_BASE English 46 46 French 58 71 German 63 73 FR +26.09% 05 | encoding Version interactive : https://slides.darkwood.com/slides/nolife-language → +54.35% @matyo91
7. Slide 6 · 2/2 · représentation statique Same texts · two encodings Language cost is not a constant. DE +36.96% → +58.70% corpus × tokenizer × language 05 | encoding Version interactive : https://slides.darkwood.com/slides/nolife-language @matyo91
8. o200k_base · German persuasion → Standard 1 token maßgeschneidert 5 tokens +400% Basis 1 → unverzichtbar 4 Better sales framing can be worse BPE. 06 | regression @matyo91
9. Slide 8 · 1/2 · représentation statique Four corners of the experiment A phrase can win one axis and lose another. STRONGER FRAME · FEWER TOKENS $1,000 4 →2 → $999 DIFFERENT FRAME · SAME TOKENS cost 1 → investment →1 07 | corners Version interactive : https://slides.darkwood.com/slides/nolife-language @matyo91
10. Slide 8 · 2/2 · représentation statique Four corners of the experiment A phrase can win one axis and lose another. STRONGER FRAME · MORE TOKENS Standard 1 →5 → maßgeschneidert SIMPLER FRAME · FEWER TOKENS maßgeschneidert 5 →2 → passend 07 | corners Version interactive : https://slides.darkwood.com/slides/nolife-language @matyo91
11. Slide 9 · 1/2 · représentation statique Regressions 9 / 27 rewrites got more expensive all 9 were persuasion · 0 were token_opt Capture statique : Schéma ou tableau 08 | red-rows Version interactive : https://slides.darkwood.com/slides/nolife-language @matyo91
12. Slide 9 · 2/2 · représentation statique Regressions 9 / 27 rewrites got more expensive Lecture du schéma : ORIGINAL REPLACEMENT BEFORE AFTER Basis unverzichtbar 1 4 Standard maßgeschneidert 1 5 Kosten Investition 1 2 achetez maintenant plus que trois en stock 3 5 A benchmark that only shows green rows is a brochure. 08 | red-rows Version interactive : https://slides.darkwood.com/slides/nolife-language @matyo91
13. Mixed pair aggregate 62 → 62 Zero change in the total. Huge change underneath. persuasion token_opt 36 45 +9 tokens 26 17 −9 tokens → → Cancellation ≠ neutrality. 09 | cancel @matyo91
14. ESTIMATED · 1,000,000 requests · pair tokens The money is tiny. The structure is not. persuasion token_opt mixed input −$1.35 input +$1.35 $0.00 output −$5.40 output +$5.40 tiny corpus · dated gpt-4o-mini snapshot · not a provider invoice 10 | cost @matyo91
15. 100,000-token hypothetical window · o200k_base Same window. Less language. English French German 2,173.91 copies 1,724.14 −20.69% 1,587.30 −26.98% The model does not get a smaller context window. The same token budget fits less of this text. 11 | context @matyo91
16. Caveman · method, not a rerun Compression has overhead too. shorter output ≠ lower total cost publish regressions compare against a output ≠ invoice control Inspected as methodology. Not a second benchmark on this corpus. 12 | caveman @matyo91
17. Slide 14 · 1/2 · représentation statique content/nolife-language Seven jobs. One measurement. 1LoadCorpus 2TokenizeBefore 3ApplyCandidates 4TokenizeAfter 5Compare 6EstimateCost 13 | implementation Version interactive : https://slides.darkwood.com/slides/nolife-language @matyo91
18. Slide 14 · 2/2 · représentation statique content/nolife-language Seven jobs. One measurement. 7Report Symfony 8.1 PHP 8.5.4 darkwood/flow 8.1.6 yethee/tiktoken 1.1.1 php bin/console app:language:benchmark 13 | implementation Version interactive : https://slides.darkwood.com/slides/nolife-language @matyo91
19. What I learned Natural languages are not Tokenizer choice changes the gap tokenization-equivalent Persuasion and token efficiency Character count is not token count can conflict Aggregates can hide opposing movements Measure the actual workload before you optimize it. @matyo91 · Darkwood @matyo91