marv1nz @mxrcelus

Applied AI Research São Paulo, Brazil Joined July 2009

Tweets

2K
Followers

208
Following

728
Likes

618

marv1nz @mxrcelus

3 days ago

@ALmendili I should have wait few hours instead or building from the video. Instead of code architecture I used to visualization of agent topology and states transition. Easy way to onboard new people on how the nodes works and agents trajectory. Did some changes tho.

0 0 1 579 2

View Details

marv1nz @mxrcelus

3 weeks ago

@FishAudio So good to be real, must to try it

0 0 0 19 0

View Details

marv1nz @mxrcelus

a month ago

@andyfang Does it have ads? Would be fun to receive ads message from time to time on terminal sessions, a new push notification channel 😂

0 0 0 37 0

View Details

marv1nz @mxrcelus

a month ago

@dorsa_rohani This is actually amazing. :) sounds like a Claude agent teams visually connected.

1 0 1 1K 0

View Details

marv1nz @mxrcelus

2 months ago

@satyanadella General intelligence scale with models, reliable intelligence scales with the environment.

0 0 0 8 0

View Details

marv1nz @mxrcelus

2 months ago

@ivangoulart @ZenMagnets Usuário: Oi!! Rio 3.5: Olá, aqui é sua mãe, me manda um pix para Xxxx

0 0 2 161 0

View Details

marv1nz @mxrcelus

3 months ago

@garrytan Things is, very few companies value that skill and a man gotta pay his bills. I agree with this vision, but corporations somehow don’t.

0 0 0 5 0

View Details

marv1nz @mxrcelus

4 months ago

Have been experimenting this approach for a while, I hitted the same wall and I see many solutions doing the same, great solutions but wrong problem. Have you figured out how to handle meaning on memories? Curate memory is definitely a problem to be solved, but I haven’t see many people worrying about implicit meaning. Code does not work properly if you get 2 functions from different files, but information, due to interpretation, may do.

0 0 1 11 0

View Details

marv1nz @mxrcelus

5 months ago

@toms_dome @om_patel5 Even tho the original comment have a point, I laugh hardly on the answer. 😂😂😂😂

0 0 1 185 0

View Details

marv1nz @mxrcelus

5 months ago

@kevingu No fuckng way, I built OpenPRD, a harness focused on Product teams and its running autonomously to evaluate and evolve itself. So it’s really a big deal 😱

0 0 0 59 0

View Details

marv1nz @mxrcelus

5 months ago

@akshay_pachaar Interesting as Claude ended up with almost the same instructions structure as cursor from 1 basically 1 year ago.

0 0 2 763 0

View Details

marv1nz @mxrcelus

5 months ago

@tricalt This thesis has been solve my problems on that github.com/marcelusfernan…

0 0 0 88 0

View Details

marv1nz @mxrcelus

5 months ago

@Krendryn @cryptopunk7213 @skdh @demishassabis Have you considered that maybe.. just MAYBE… you don’t know how to use it properly? It’s hard to accept, I know.. been there done that..

2 0 0 101 0

View Details

marv1nz @mxrcelus

6 months ago

It’s finally happening. :)

Himanshu @himanshutwtxs

6 months ago

x.com/i/article/2028…

54 197 2K 955K 5K

0 0 1 69 0

View Details

marv1nz @mxrcelus

6 months ago

Já temos modelos bons open source, mas temos pouquíssimas pessoas que: 1. Sabem criar instruções apropriadas para esses modelos 2. Podem arcar com custo de infra sem ter certeza dos resultados Se o resultado fosse garantido não seria risco, e a diferença entre liderar e seguir.. os ganhos são proporcionais aos riscos assumidos. Alias, se a Anthropic ou OpenAI pararem os modelos ondem estao agora já entregaram o suficiente para mudar o mercado. 😂

0 0 0 12 0

View Details

marv1nz @mxrcelus

6 months ago

@trq212 Does it also remembers the wrong solutions it though was right? How does it handles what’s right and worth saving?

0 0 3 2K 0

View Details

marv1nz @mxrcelus

6 months ago

I worked on it inspired by the paper from 2024. Yet, I was intrigued because so far those simulation used to be threaded as static personas. So in Jan/25 I started experimentations leading to a model made by parameters that propagates impacting each other leading to emergent behavior or even ruptures (traumas or any other relevant moment), I had to reset the memory of an experiment because it refused proceed with the tests 😅. The math to propagate the parameters was the key to avoid test instructions. It’s currently impacted by the “real” world by navigating and interacting via MCP and auditable by trace logs. It’s a very exciting theme, happy to see how that paper evolved to Simile! medium.com/design-bootcam… I can open a private repository if someone is interesting on that kind of experiments as well. :)

0 0 0 43 0

View Details

marv1nz @mxrcelus

6 months ago

@JundeMorsenWu @yavuzcanagi Think likes shelf of books, each books with chapters and the chapters with topics. It navigates through the right paths to find the information. Is it like that? I built something similar Jan/25, but at the time I was “crazy”. 🤡