deko

Tokyo based
Software Engineer and Investor
Huge interest of Finance/Economy/Macro and using AI to turn this into real outcomes

[deko @andrerl]

zcode (glm desktop) is pretty great in linux, unfortunately I don't know if codex and claude desktop are comparably good as they don't have a linux version and don't seems that they will have

one more win for glm 5.2, whoever use linux, try it out

[イシナベ | AIプロトコル層に潜伏中 @ishinabegen]

結局これ日本版OpenRouter Fusionで、正直残念だった。
Sakanaの独自性は司令塔を7Bで強化学習した部分だけ。
これなら自前のほうが圧倒的につよい。

界隈の人はみんなやってると思いますが、これを機に自分の設計を書いてみます。
自前でマルチエージェントを作って肝だと思ったのは
・委譲モデル選定
・メインセッションの聖域化
の２つ。

まず委譲モデル選定では「系統距離」「逓減の見極め」あとは各モデルの発想を縛らないことが重要でした。

・系統距離。
OpenAI派生を3つ並べても盲点は同じ。
金融工学・学術NLP・コード特化・HPC、設計DNAが被らないモデルを選ばないと並列の意味がない。

・逓減の見極め
モデル数を増やせば精度は上がるが逓減する。
5→9は劇的改善、9→11は微増。闇雲に足してもコンテキストが死ぬだけ。
ただこの微増が侮れなくて、アイデアの種類が微増なら捨て置けるが、コードレビューで他の全モデルが見逃したバグを1体だけが見つけるケースがある。その1体を切れない。

・各モデルの発想を縛らない
呼び出しプロンプトでモデルの挙動を縛らない。
せっかく別DNAのモデルを呼び出してもフォーマットや観点を固定されては多様性が潰される。
自由に考えさせて、想定外の角度から殴ってもらうのが目的。

[deko @andrerl]

@NickOresme wondering same while my short gets pretty bad

[deko @andrerl]

@Impedro29 Num existe plano de tokens ilimitados

[deko @andrerl]

@sethrose @thsottiaux Yeah I found out I had to install codex desktop to see it as web doesnt have it, and I never saw it on cli as well

[deko @andrerl]

@sethrose @thsottiaux where I can find this screen?

[deko @andrerl]

@izzyz I've been doing this for a few months on my own project with mixed results.

Was clear for me there was potential and ways to leverage it but I would need time to optimize it to what I do in the right way. Still def going to do it optimized for my usecase someday.

[deko @andrerl]

@gpmartinelli84 @fernandoulrich E so p ter certeza, o preco final estimado é baseado no valor atual de energia dos eua ne? Num tem nenhum tipo de forma de corrigir o preco apos lancado certo?

[deko @andrerl]

Li tudo aqui, bacana, bacana. Tem q ver como vc força a ser so minerado por dram e se num tem workarounds

Mas tu continua beneficiando os mesmos players principais com dram, coreia do sul, us, japao, china, taiwan, os maiores produtores e q detem maior tecnologia de dram

Qual o problema q tu quer resolver e como tu quer resolver esse problema? Cada vez que entendo mais de fx e macro vejo menos sentido em algo do tipo

[deko @andrerl]

@ClaudeDevs Hey cmon I subscribed for claude max only because of fable

[deko @andrerl]

@gpmartinelli84 @jackprandelli This number right now pretty much will define brazil economy for the next decade

[deko @andrerl]

@GoogleResearch @LyalinDotCom This seems great

[Google Research @GoogleResearch]

🚀 Introducing Gemini-SQL2, our breakthrough text-to-SQL capability powered by Gemini 3.1 Pro! We've achieved state-of-the-art results on the highly competitive BIRD benchmark, translating natural language into execution-ready SQL queries. 🧵👇

[deko @andrerl]

@MiniMax_AI @Xinyu2ML 👏👏👏👏👏

[deko @andrerl]

I start to feel I know just too much about japanese stocks, I may start making some posts about them soon

Share a bit of knowledge, japan stock market is heating up and higher interest rates are coming, the opportunities are just too great to pass

No plans yet but is on my mind, maybe in semiconductors/AI as people love these rn

[deko @andrerl]

😂 haha nunca rolou comigo isso de ir dormir, mas ja ouvi falar. Mas horario japones e tranquilo, nunca tenho problema.

Quem sabe num e hora de passar uns tempos por esses lados aqui, alem de q a bolsa japonesa ta bombando e japao normalizando, iene vai fortalecer e bolsa jp daqui pra frente deve se manter como uma das mais fortes do mundo

Mas desde o lancamento de gpt 5.5 eu prefiro mto mais gpt q os modelos da anthropic. Vou rodar mta coisa com o fable hoje e ver se vale a pena, mas gpt custo/beneficio e bem melhor

[deko @andrerl]

Just dont rush and give us a model like opus that just uses too much tokens and destroy our quota.

GPT 5.5 was the first time I preferred a model over claude, I think is the best model overall OpenAI has ever done by generation, it does well in everything and best in quite a few areas and it doesnt burn tokens, you can feel the tokens are properly been used in relation to the quota size, I do not feel that with opus at all

There are tasks that people will want the best no matter how much tokens it uses, but 99% of the time users want the best cost/benefit and gpt 5.5 delivers this today

[deko @andrerl]

I don't want to explain why, and I have no idea why this is not common knowledge, but prob because uni economic courses are just full of bullshit

But anybody that thinks that Oil prices will go crazy high are not good to be followed, they clearly do not did their part in studying fx, or at least petrodollar

And yeah is quite a bit complicated and requires people to actual reason about things, so I guess will always be like this.

But understand both how it works and you will be front running almost everybody here

It will go crazy in a few currencies btw, it will not get crazy high tomorrow or in the next month. And ofc is all part of current us world macro, which is going great for them

And I guess I said enough

[deko @andrerl]

Hmm eu uso pra muita coisa e ainda num encontrei nada que faca sentido pagar api nao

Mas tb uso pra varias coisas como se fosse API, se tu usar via claude -m {model} -p {prompt} e como se fosse uma chamada de API direta numa nova sessao, e tu pode paralelizar isso se tu for executar varias coisas ao mesmo tempo

Num sei se e teu caso, tb uso mais no horario asiatico, horario de pico q e o americano deve ser oq mais trava mesmo