Sergei Belousov @bes_dev

I’ve been obsessed with computer vision, machine learning, and all things AI for over a decade — and I still genuinely love it. bes-dev.com Joined April 2009

Tweets

70
Followers

254
Following

0
Likes

23

Sergei Belousov @bes_dev

4 days ago

Everything is open source under Apache 2.0, so you can try it in a minute or read the full theory at your own pace. npm i -g reharness github.com/bes-dev/reharn… Deep dive - where Agentic AI meets classic compiler theory: bes-dev.com/posts/reasonin…

0 0 2 46 0

View Details

Sergei Belousov @bes_dev

4 days ago

In return you get a kind of reliability you can really rely on. The pipeline is stable by its construction, without huge prompts for perfect accuracy, without hidden drift between runs, with static checks confirmed before every run.

1 0 1 45 0

View Details

Sergei Belousov @bes_dev

4 days ago

Harness engineering is essentially prompt engineering in a new form. The skills and instructions we write raise the chances of success, but they give no guarantees. Every written line stays a recommendation for the AI, and the model can always ignore it at its own discretion.

1 2 7 275 2

View Details

Sergei Belousov @bes_dev

a month ago

@rra @NASA Working on RS-25 test software at MSFC — legend status! To be honest, building a 5-year-old React project today is harder than compiling this 56-year-old FORTRAN IV 😄 My favorite find: when altitude goes negative, SSSP prints VEHICLE ATTEMPTED SUBTERRANEAN FLIGHT. Honored!

0 0 1 45 0

View Details

Sergei Belousov @bes_dev

2 months ago

How was your weekend? I spent mine launching the Space Shuttle. With NASA code from 1970 that nobody could run for 50 years.

27 135 3K 208K 636

View Details

Sergei Belousov @bes_dev

a month ago

@planetdeimos That is incredibly cool. It’s a huge privilege to know that the engineers working on SLS might find this weekend project interesting. This code was decades ahead of its time, and I’m just glad to help keep that history alive. Thanks for sharing it, Sean!

0 0 2 205 0

View Details

Sergei Belousov @bes_dev

a month ago

@reach_espartan @NASA More detailed information and documentation can be found in the GitHub repository: github.com/bes-dev/sssp. I also provided a general description of the math behind the simulator, based on my analysis of the code.

0 0 2 56 0

View Details

Sergei Belousov @bes_dev

2 months ago

@aerodud3 @DJSnM Of course, we're not talking about Fortran programs in general, but about a specific pre-project Shuttle simulator.

0 0 9 644 0

View Details

Sergei Belousov @bes_dev

2 months ago

@reach_espartan @NASA They used the fifth-order Runge-Kutta-Butcher method with adaptive stepping.

1 0 4 74 0

View Details

Sergei Belousov @bes_dev

2 months ago

@MKortvely Wow, thanks for sharing this story! Running a simulation and knowing a shuttle is flying above you along that exact calculated trajectory is practically magic. Especially when you consider the era and the limitations engineers faced building this software.

0 0 16 2K 1

View Details

Sergei Belousov @bes_dev

2 months ago

@samuel_hayden_ @macjshiggins To clarify: the "test data" is actual STS-1 telemetry from 1981 vs a theoretical 1970 model. Getting only a 5-15% delta between 1970 math and a physical flight 11 years later just proves the absolute brilliance of the engineers of that era given their limited resources.