Silico, the platform for ambitious AI research, is publicly available today. AI is advancing fast. The tools to understand it need to advance even faster. Silico lets you interpret and train your models at frontier scale. Learn more + get access 🧵

Silico is available now. Get access here: More about how Silico works in the thread 🧵

Silico plans and executes long-horizon experiments. It develops a plan, runs the work in parallel, monitors progress, and returns results you can inspect and build on.

Silico uses Goodfire’s frontier interpretability techniques. Opening the black box lets Silico trace harmful behaviors to their root and intervene.

Silico orchestrates compute, so you no longer need to babysit jobs. Silico coordinates experiments across GPU clusters, monitors every training run, and keeps long-running work moving without constant supervision. This lets Silico interpret Kimi K3, at 2.8T parameter scale.

Examples of what you can do with Silico: → map your model with frontier interpretability techniques → diagnose failure modes in your dataset + architecture → train models w/ SFT, DPO, and RL → replicate + extend papers

Long-horizon experiments are expensive. Silico’s pricing reflects that. We’re working to reduce costs over time. We’re offering 50% discounts for early signups and grants for researchers in critical areas, including AI safety + life sciences. Get Silico:

silico is so fun! great work from the team

Interpretability is underrated! Congrats on the launch

+1 to that! Thanks Alex!

Had a chance to try this. It’s pretty damn cool

Thanks Andy!

@SrikarVaradaraj We’ve loved previewing this the past few months and actually made some new discoveries with our bio models! Can’t wait to see what others use Silico for

@SrikarVaradaraj Thanks Ravi! We're excited for more discoveries!

It was very cool getting a sneak peak of this at @spc, excited for the world to start working in Silico!

@spc we're very excited too :)

Congrats on the launch! Great product.

Thanks Max!

really awesome product - love using it for automated post training!

thanks Vincent! we’re really excited about what you guys are doing with Silico!

OBSESSED

thanks Chris!

cc @DrDatta_AIIMS you should try it, happy to introduce you to the folks behind it for generous access.

Congratulations!

Thank you, Nina!

This is a huge deal, congrats!

Thanks Zach!

Congrats team Goodfire!

Super exciting launch, congrats team @GoodfireAI

Thank you Kevin!

I used an early version of this, and can personally say I loved this tool. Colleague sent me an interesting paper, I asked silicon to replicate it. A few hours later, it gave me the result that the finding held, but the effect was 3x smaller if the researcher had properly used early stopping. I looked a lot smarter than I actually was that day!

@thomas_fel_ Glad you liked it! Silico makes paper replication way easier

Congratulations!

Thanks Jorge!

🤘🥳

@b1shtream the platform you wanted are launching

O M G! I want to try this so bad

How can I as a student get access to this?

As AI systems become more capable, understanding why they behave the way they do is becoming just as important as improving their performance.

Looks great but $1,000/mo is expensive

Can’t wait to try it out!

Interpreting and training at frontier scale, that's a big ambition to tackle

Interpretability is going to become a core infrastructure layer for the AI ecosystem.

Beautiful

Whats the data retention policy?

Frontier tools for frontier models. Huge step

congrats on the launch! excited to try it

@grok 这是一家什么公司？创始人和主要投资人背景

Hey, looks cool, quick question. Do I correctly understand that you can sell my prompts, agent interactions etc to third parties (as anonymized derivatives or so)? Not saying this is a deal breaker but the TOS are not super clear to me.

watched interpretability pass clean evals, then miss failures on drifted production traces

The inspectable experiment output is the detail that makes long-horizon research feel usable. A plan is only valuable if the team can audit what happened afterward.

The operational layer is becoming a differentiator. Managing distributed compute reliably is increasingly as valuable as the models themselves.

Congratulations!