Geomar Abreu
229 posts
Geomar Abreu
@_TheGAC
Building quantitative ML models & trading strategies — Building WRK
Internet Katılım Haziran 2026
28 Takip Edilen35 Takipçiler
The feeling of vibe coding your first app...
Ahh, so great.
But I surely will be learning programming and know what the heck the AI is doing.
My planned career requires it.
But some experimentation is useful too.
Soon I will post an article about it.
English
@karpathy @WisprFlow is great for it.
I don't use voice, but it's getting better each day.
English
One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.
English
@kimmonismus Something tells me benchmarks are just a suggestion.
If it works for you, it's great, so no need to switch.
Though it's cheaper than Sonnet 5, and that's what I use.
Should I switch? Nah, I'm good.
Good competition will make everything better overall.
English
Its release day. Gemini 3.6 Flash
- It's cheaper than Gemini 3.5 Flash ($7.50 output instead of $9.00 output).
- It outperforms Gemini 3.1 Pro in almost every benchmark.
It's being compared to GPT-5.6 Luna and Sonnet 5. Priced between these two models, it generally performs better.
So, it's a solid mid-tier release.
leo 🐾@synthwavedd
Gemini 3.6 Flash benchmarks are out, and it's... beaten by other models on code tasks, and is only really consistently SoTA on vision and context benchmarks. But hey, 3.1 Pro is now so old 3.6 Flash outperforms it across the board 😭
English
@Da7_Tech It's not a bad idea, but there's one problem:
You can't buy Kimi right now; it's sold out.
TLRD: Their GPUs can't handle the success they've achieved. So they are focusing on their existing users.
English
FABLE 5 IS STAYING ON CLAUDE MAX. FOR GOOD.
Thank you Kimi K3 and GPT 5.6 for making this happen.
Claude@claudeai
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.
English
@robj3d3 Bro, announcing a usage limit time increase 3 TIMES, isn't good communication...
English
They definitely have Fable 6 now.
There’s no way their communication suddenly got this good.
Claude@claudeai
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.
English
@AlexFinn Bro, I only found out about Kimi just a day ago.
Are we going Chinese for AI too? Will wait for better confirmation...
GIF
English
They did it. They announced Fable 5 will be included in subscriptions forever
This is the first time I've ever seen Anthropic bow to pressure
They won't admit it, but this is 100% because ChatGPT 5.6 sol is excellent for a fraction of the price
I will say this tho: Fable is still the token for token king
It's unbelievably smart and focused. 5.6 sol is very smart but like a rabid hyaena. A little too unfocused
Each has their place though
- Fable 5 for planning/super complex tasks
- ChatGPT 5.6 sol medium for simple tasks
- ChatGPT 5.6 sol ultra for long running /goal
Kimi K3 is very good, easily in the same league as these two, but the pricing isn't good enough to warrant moving off the Codex/Claude Code harnesses
Moment they come out with a quant compatible with Mac Studios tho I'm on it
Today is a great day. When labs compete, we win
And the more we get to use Fable 5, the more we win
Now we wait for the Opus 5 announcement over the next week
Claude@claudeai
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.
English
@claudeai I didn't even use it once. I knew today was the last day for me to build something with it.
And now this? Well, I'm on Pro, so I won't be able to use it anyway...
GIF
English
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits.
Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit.
Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.
English
The more I live, the more I realize how many people just don't want to work hard enough to get it.
They *want* it, not *get* it.
English
ICT has become the most popular trading model we've ever encountered.
How well it performs usually depends on "discretion".
My advice? If you can, just code some of your favorite ICT strategies and see what happens.
But I'm not saying they work. Just do an experiment.
English
The sooner you realize the power of prop accounts, the sooner you will generate huge capital for "free".
They don't follow a symmetrical risk factor. You either win big or lose, at most, 100 bucks.
You don't need a market-breaking strategy. Just a good enough one.
Code it.
English
Code every day. Code your ideas. Code your businesses. Code your strategies. Code without vibe coding. If you vibe code, ideas don't come up right. Either you code, or you code with AI. You do not vibe code. You always code.
English
@aether_uiux Second one. Much clearer to see what you're getting.
The first one seems too much at once. Nothing clear to read or see.
English
@brenzhills Se siente bien ver más hispanos en X.
¡Más si es sobre cómo usar Claude!
Español
NO PAGUES EL PLAN DE 100€ DE CLAUDE (TODAVÍA).
Estos 21 trucos hacen que el de 20€/mes sea suficiente:
(guárdalo para mañana)
1. Subes PDFs a lo loco. Una página = 3.000 tokens.
Solución: Pega el texto en un Google Doc. Descárgalo como .md. Menos de 200 tokens.
2. Empiezas a crear archivos en Cowork demasiado pronto.
Solución: Planifica primero en Chat. Pasa a Cowork solo cuando sepas exactamente qué quieres.
3. Escribes prompts de 500 palabras que se repiten cada vez.
Solución: Escribe 29 palabras: "Quiero [tarea] para [objetivo]. Hazme preguntas antes de empezar."
4. Dices "rehazlo todo" para arreglar la sección 3.
Solución: "Rehaz solo la sección 3. Mantén todo lo demás. Sin comentarios. Solo el resultado."
5. Mandas 3 mensajes separados para 3 tareas.
Solución: Un mensaje, tres tareas. "Resume esto, lista los puntos, sugiere un titular."
6. Escribes "No, me refería a..." apilando sobre el historial.
Solución: Dale a 'Editar' en tu mensaje original. Corrígelo. Regenera.
7. Reescribes los prompts desde cero cada vez.
Solución: Crea una biblioteca de prompts. Misma estructura, cambias la variable.
8. Usas Opus para una simple corrección de ortografía.
Solución: Sonnet para tareas rápidas. Opus + Pensamiento Extenso para trabajo profundo. Y la novedad, Fable 5: el modelo más inteligente que existe, pero devora tus límites — resérvalo para estrategia y para lo que los demás no consiguen resolver.
9. Tu archivo "sobre mí" tiene 22.000 palabras (demasiado).
Solución: Recórtalo a menos de 2.000 palabras. Al final de cada sesión: "Escribe un .md de notas de sesión."
10. Nunca reinicias y sigues apilando chats larguísimos.
Solución: Cuando Cowork se descontrola, dale a "Reiniciar la conversación desde aquí" en un mensaje anterior.
11. Nunca resumes antes de que la conversación se haga eterna.
Solución: Cada 15-20 mensajes → resume, copia el brief, abre una sesión nueva.
12. Re-subes los mismos archivos a cada chat nuevo.
Solución: Usa Proyectos. Subes una vez. Todos los chats dentro lo referencian sin quemar tokens otra vez.
13. Metes 50 archivos en Cowork "por si acaso".
Solución: Solo incluye lo que necesita esta tarea. Cero carpetas para tareas rápidas como redactar emails.
14. Mantienes 3 temas en 1 chat. Claude se tiene que releer todo.
Solución: Tema nuevo = chat nuevo. Siempre. El contexto muerto son tokens muertos.
15. Dejas la búsqueda y los conectores activados por defecto.
Solución: Todo desactivado por defecto. Activa las funciones por tarea, no por cuenta.
16. Ejecutas el mismo informe a mano cada semana.
Solución: Usa /schedule. "Cada lunes a las 7am, crea mi resumen semanal."
17. Dejas que Claude Code explore todo tu repositorio.
Solución: Sé específico. "Crea un gráfico de barras con este CSV. Guárdalo como chart.png."
18. Te saltas las Preferencias Personales y pierdes setup.
Solución: Ajustes → Preferencias Personales. Define tu tono y estilo una vez.
19. Escribes prompts vagos tipo "hazlo mejor."
Solución: Dicta tus prompts con wispr.ai. Más contexto de golpe. (Es gratis)
20. Te fundes todo el límite en una mañana.
Solución: Claude funciona con una ventana móvil de 5 horas. Repártelo.
21. Usas Claude para cosas que no puede hacer.
Solución: Conoce tus herramientas.
Imágenes → Gemini.
Búsqueda en tiempo real → Grok.
Bonus. ¿Empezando con Claude Code?
El curso de abajo te lo deja funcionando en una tarde: proyectos, automatizaciones y ejemplos reales 👇
Si te sirvió, compártelo con tu red para ayudarles a ahorrar tokens.
AI Spanish Community@aiscwork
CURSO COMPLETO para EMPEZAR con CLAUDE CODE Aprende paso a paso a crear proyectos, automatizar tareas y aprovechar todo su potencial. Con ejemplos reales y prácticos. (Guárdalo antes de que sea tarde)
Español
POV: You first heard the term "loopmaxxing"
Where the hell are we going fam.
GIF
English
@sickdotdev I used to avoid AI completely...
Then I bought Claude Pro, and boom.
Never going back
(I have a M4 MacBook Air — best ROI on a laptop IMO)
English
