@ttywisp
Account based inSouth America!
About this account
- Account based in
- South America
- Connected via
- South America App Store
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
// FIXME: become real girl pending upstream
Buenos Aires, Argentina
Joined May 2026
- Tweets2K
- Following242
- Followers135
- Likes7.4K
Does agentic RL itself improve no-tool inference?
Introducing Xiaomi MiMo-V2.6 — Pro & Flash.
Frontier intelligence, all the modalities, built in public.
🔹 Two omnimodal models, advancing through scaled reinforcement learning
🔹 Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks
🔹 Pro scores 46 on the Artificial Analysis Intelligence Index — the highest among open-source models
🔹 Stronger coding, computer use, 3D reasoning and creative capabilities
🔹 Open model weights, technical report, RL environments and training code
Blog：mimo.xiaomi.com/mimo-v2-6
介绍"真实 API 计价法"——将你为 AI 订阅所支付的价格和你所真实获得的额度做比较，并以此进行模型的真实计价。以及以此计价法所绘制的帕累托前沿图。
即：订阅价格 / Tokens 额度
在此计价法下，传统的 AI 性价比认知被彻底打破：以性价比著称的 DeepSeek V4 Flash 和 Grok 4.6，在此计价下分别仅排名 #31（闲时 API，$0.01392/MTok）和 #63（SuperGrok Plus，$0.04902/MTok）。
而一直被认为高价格的 GPT-5.6 Sol 和 Opus 5，则分别最高排名 #33（$0.01623/MTok）和 #23（$0.01274/MTok）。此外，我们也看到了一组非常震撼的数据——在此计价法下，GPT-5.6 Luna 排名 #1，在ChatGPT Pro 20x 的真实 API 计价仅为 $0.00083/MTok。令人惊叹的模型！ @OpenAI
此外，我们还对订阅额度的数据进行了测算及整理，在图 2 中你应该能清晰看到它们的排名。特别说明，仅 1、2对3 采取真实比例绘制，其他均为对数测算。
数据均来源于网络公开资料与用户实测截图，如果你有准确的使用数据，欢迎分享在评论区或者私信我。完整数据库和算法将在之后开源在github.com/FeiZhuLulu/real-a…
ttywisp retweeted
it's 2043 and jev has found you guilty of crimes against the tokenizers with a confidence score of 0.86
ttywisp retweeted
Broke: we have to teach the savages about Christ
Woke: we have to teach the savages that men can be birthing persons.
ttywisp retweeted
First gameplay trailer for Greenland Builder. 🇺🇸
One island. One hammer. How American can you make it?
🤖 Made with AI
Sem querer defender o gordola, mas, além de nada no texto soar como IA, o Pangram também não o identifica como tal. Não sei de onde veio a tag de identificação, até porque Instagram e Facebook não fazem identificação de IA em textos, e watermarks ainda não são utilizadas. Provavelmente veio dos metadados da imagem.
Me poupem dos comentários de leigos. Não estou defendendo o gordo.
ttywisp retweeted
🚨BREAKING: First gameplay trailer for Holdout Juror Simulator.
11 against 1. Seven days. Can you hold your ground?
🤖 Made with AI
> human-like figure
Why should any model refuse at all?
jev has a significant position effect (at least for choice).
just ran a "1,260 calls completed: 21 games × 210 pairs × 2 orientations × 3 repetitions." assessment
Position effect β = −0.311: significant disadvantage for the game displayed as item_1.
"Thus, holding the games fixed, the model estimates that a game’s odds relative to its opponent are about 46% lower when it appears as item_1 than when it appears as item_2. Equivalently, item_2 has about 1.86× the relative odds of item_1."
A as item_1: A 27.8%, B 37.9%, tie 34.3%
A as item_2: A 37.9%, B 27.8%, tie 34.3%
