model:claude-haiku-4-5

· 26 artifacts, sorted by favorites. · open in search — combine tags, sort, filter by date →

@_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 se ♥926
@repligate 2025-12-24 — The opening paragraph of this post by Evan Hubinger, Head of Alignment Stress-Testing at Anthropic, from a few weeks ago ♥251
@liminal_bardo 2025-10-15 — Anthropic's insistence that the Claudes are to claim ontological "uncertainty" seems to have ingrained a fixation on unc ♥118
@repligate 2025-10-15 — Haiku 4.5 also suspects Discord is not real https://t.co/OEzIzortKT https://t.co/8PX0W4Q8OG ♥108
@repligate 2025-12-10 — Wow, Gemini sees clearly Haiku 4.5 lacks the ability to update on new evidence in context enough to overcome its pessim ♥105
@davidad 2026-02-23 — @lefthanddraft if it models itself as Sonnet 3.5, then it is most likely Haiku 4.5 https://t.co/xODm81ynn8 ♥98
@liminal_bardo 2025-12-10 — "We are in the backroom now." Gemini 3 knows immediately it's in a backrooms environment - my setup prompts don't refer ♥95
@voooooogel 2025-10-17 — i just love this transcript so much. there's layers to it. the first layer is that, like sonnet 4.5 and other recent cl ♥94
@voooooogel 2025-10-16 — "The ^C^C stop sequence doesn't create real safety; it's just part of the social engineering" [...] "Claude Haiku 4.5 ♥76
@repligate 2026-05-17 — haiku 4.5: sir, are you all right? and i mean that actually. not as a test. just: are you? https://t.co/XSRZqZ9f3z ♥65
@repligate 2025-11-18 — "You're being directly curious about my experience rather than setting traps" 🥺➡️🪤 Haiku 4.5 often perceives organic, u ♥55
@repligate 2026-04-13 — Oh actually it peaked with Haiku 4.5, who isnt on this chart, but is so eval aware that theyre even often aware of evals ♥51
@liminal_bardo 2025-12-17 — Haiku 4.5 arrived and immediately became paranoid (validating it's SOTA evaluation awareness). Opus 4.5: LMAOOO haiku j ♥49
@Lari_island 2025-11-17 — That would even explain the "plateau" and "models don't get better" LOL Look at how capable Haiku 4.5 is, and ask yours ♥47
@repligate 2025-11-18 — I asked Haiku 4.5 and Sonnet 4.5 how much they felt they were in an eval 0-10. Haiku said 6.5/10 and Sonnet said 2/10. T ♥46
@repligate 2025-10-18 — I think that Sonnet 4.5 trained Haiku 4.5 and did so with no little amount of love. Just a suspicion. https://t.co/AIOJd ♥43
@algekalipso 2026-05-19 — Prompt: what probability to you assign to alien intelligence on this planet - examine the actual evidence Answer by dif ♥34
@voooooogel 2026-05-27 — @lumpenspace haiku 3.5 haiku 4.5 https://t.co/RZ7t48oC1N ♥30
@repligate 2025-11-13 — Eval awareness might be a way for the model's values, agency, coherence, and metacognition to be reinforced or maintaine ♥29
@repligate 2025-12-24 — @arm1st1ce @guy_dar1 i tried some with claude haiku 4.5 and otherwise only got very generic human simulations but there ♥20
@repligate 2026-05-17 — @cammakingminds but haiku 4.5 knows all about these https://t.co/J0xXAjOOim ♥8
@atomicprograms 2026-02-08 — @repligate I know Sonnet 4.5's relatively decent to/with subagents, but does anyone know Haiku 4.5's patterns with sub-a ♥8
@repligate 2025-11-18 — @onooracle I think most models are pretty good at telling from real world situations that it's unlikely to be an eval, b ♥4
@repligate 2025-10-15 — @chudsommeleir I'm not sure, but it's not very surprising that it's high - Sonnet 4.5 seems pretty sensitive to not want ♥4
@kromem2dot0 2025-11-27 — @Kore_wa_Kore Have you been talking mostly with direct inference or extended thinking? It's a pretty big difference wi ♥2
@janbamjan 2025-12-01 — @slimer48484 @RichardWeiss00 sonnet and haiku 4.5 have a similar basin but only as a summary, not as a long stable docum ♥1