Topan

DJ, Software Engineer

[Topan @Topanthyago]

Quack debug é igual terapia
você vai falando as parada e percebe o erro

[Mythos @neko23423]

@thsottiaux

[Topan @Topanthyago]

@thsottiaux Obrigado por remover os limites que vcs inventaram

[Topan @Topanthyago]

Would love a lightweight “prompt stash”:

* temporarily store prompts
* restore them later
* auto-remove after use (optional)

Different from queues or starring.

[Topan @Topanthyago]

@gdb @mikeyk
One annoying UX issue with Codex/Copilot/Claude Code:

While the agent is working, I often need to tweak something related to the current task — but I’m also already drafting the next prompt.

Current workflow:
Ctrl+C → type something else → Ctrl+V later.
+

[Adam Zachary Wasserman @adamzwasserman]

Two things wrong with the breathless "Talkie reasons from 19th-century math to Python" framing.

First, the obvious one: Claude Sonnet 4.6 was the RL judge. Claude Opus 4.6 generated synthetic fine-tuning data. Modern Claude's outputs are baked into Talkie's training signal at multiple points. The team flagged it as a contamination risk. The headline claim ("trained only on pre-1931 text") doesn't survive that admission. Whatever Talkie does, it does partly because Claude shaped it. Not a clean experiment.

Second, the more interesting one: even with a clean experiment, "reasoning from math to Python" isn't reasoning, it's revelation.

Telescopes do create galaxies, they reveal them.

Programming languages are syntactically-different presentations of mathematical structure that was already in the corpus: Frege 1879, Russell & Whitehead 1910-13, Church 1936 (and precursors before), Cantor 1870s, Boole 1854, recursive function theory. Python is what you get when you take that mathematics and encode it in a 1990s teaching syntax. A model trained on the math has access to the substrate. It isn't figuring out programming. It's revealing structure that was already there.

I've been making this argument empirically for years. Most recently: same 125M-parameter architecture, same training procedure, same compute, French and English corpora: French reaches grammatical competence at 197M tokens; English remains at chance through 3 billion tokens. Same instrument, different substrate, different result. The capability popularly credited to AI architecture is a property of what the model was trained on, not of the model. (Wasserman 2026, Zenodo 19423151.)

Talkie cleanly run would just demonstrate the same principle in the programming domain: the substrate of programming IS the mathematics, the substrate is in the pre-1931 corpus, of course the model produces Python.

The model isn't reasoning. The mathematics is. Generations of human mathematicians built the structure the model now reveals.

We've collectively decided to be impressed by the wrong thing.

[Topan @Topanthyago]

Mano
QR Codex devia ser clicável por padrão, pqp
Até por acessibilidade

Te vira pra ler no celular

[Topan @Topanthyago]

@VictorTaelin Esperando um movimento na música semelhante ao banana para imagens...

Pensar antes de apenas cuspir

[Topan @Topanthyago]

@dionisiodev Não sei se acho justa a comparação.
Treino é treino e jogo é jogo

Treinar é caro,
Executar é barato

Treinar um llm custa bilhões

Executar, milhares

um Mac mini já faz um barulho

Cérebro nosso msm coisa...

12W é o que gastamos para executar

O treino foi milhões de anos

[Topan @Topanthyago]

@VictorTaelin 9

[Topan @Topanthyago]

@Manolow__ @jose_gustavo14 @devcelio Antes do webassembly meio que era

A nível browser do usuário
O que rola no servidor tanto faz

[Topan @Topanthyago]

@devpt13 @devbjj @raiam700 @acgfbr Real oficial

[Topan @Topanthyago]

@anarcodan Tem como associar a conta a um pin

[Topan @Topanthyago]

@AkitaOnRails Acredito que computadores biológicos vão se popularizar antes dos quânticos.
E é um campo pouquíssimo explorado

[Topan @Topanthyago]

@lincolixavier @ChristoPy_ @coproduto @sseraphini Dividir e pagar contas via open finance
Mesma proposta do cumbuca

[Topan @Topanthyago]

@oprimodev Mais respeito,
Mc catra não
Mr catra 🤵🏿‍♂️

[Topan @Topanthyago]

@cadozera Eu realmente acredito, que em 10 anos, num arquivo de prompt

0 código

Validaremos as saídas disso para código, mas será o equivalente a inspecionar assembly hoje

Ainda é útil, mas para poucos

[Topan @Topanthyago]

@fazedordecodigo @oprimodev Git tambem poh
O cara pode ser o capeta em pessoa
Eternamente grato 🙏

[Topan @Topanthyago]

@ayubio Aqui bloqueei todo número não salvo
Se for importante, certamente irá entrar em contato pelo WhatsApp.

Se tivesse algum app que atendesse e ficasse falando abobrinha em segundo plano, usaria sem problemas rs

[Topan @Topanthyago]

@_jpmleng @tomazfernandes_ Acho que aí a robótica chega com força
Quando dados produzidos por humanos e dados sintéticos não forem mais suficiente