the pitch and the pause
every time i open my laptop lately, it feels like the screen itself hums with noise. news, opinions, model releases, funding rounds, drama, metrics, benchmarks, leaderboards. it’s endless. everyone seems to be talking about ai like it’s both the second coming and the beginning of the end. sometimes, it’s hard to know which one to believe.
for me, it’s quieter. the work has a rhythm. some days, it’s training logs and eval sheets. other days, it’s strategy calls and architecture tweaks. and then there are the in-between moments, when i just sit and think about what we’re actually doing at sagea. not the metrics, not the releases, but the why.
we’ve been deep in it for months. the team has been operating like a small, focused research lab disguised as a startup. ujjwal’s buried in agent systems, firoj’s knee-deep in benchmarks (and he's about to bring something groundbreaking), and i’m in between research, infra, and all the random fires that come with running something that shouldn’t logically work this well with the resources we have. and yet, it does.
the big news is that sage 40b MoE is almost here! it’s surreal even writing that sentence. a forty billion parameter model that runs smoother than some 70b baselines. when we first started, i remember saying something like, “let’s just see if we can make reasoning work better.” and now, somehow, we’re here, on the edge of releasing something that might actually move the needle.
the model’s architecture feels alive. we’ve been experimenting with dynamic routing layers that allow experts to adapt their roles depending on context and question complexity. what’s fascinating is how emergent the specialization has become. certain experts tend to “choose” reasoning-heavy tasks. others act like memory stabilizers. we didn’t hardcode any of this. it just formed, also because of our metacognitive inverse reasoning head. watching it happen feels like watching a small piece of cognition evolve in real time.
the release itself was a bit chaotic. we hit a snag uploading the model to hugging face because it was too large for a private repo under their free plan. it’s funny now, but at the time, it was this absurd moment, one of those “we’ve built a 40b model but can’t upload it” kind of problems. classic startup energy. eventually, we sorted it out, got the enterprise plan, and now it’s all prepped for release. small victory, but it felt poetic.
outside of the code and compute, something bigger is happening. people keep talking about the “ai bubble.” it’s become the phrase of the season and i STILL can't stop thinking about it (it's the third day now). depending on who you ask, we’re either living through the greatest technological shift since electricity or a massively overhyped cycle that’ll collapse once investors realize not every company is openai.
and honestly? both might be true.
the ai space feels bloated right now. valuations are absurd. half the companies pitching “ai” don’t actually build anything. and yet, beneath the noise, something real is forming. there’s a deeper undercurrent of progress that’s too tangible to dismiss. models are getting better. reasoning is improving. multimodal systems are actually learning how to fuse understanding. it’s easy to call it a bubble, but bubbles only form around real heat. and right now, ai is on fire.
still, as someone running a small lab in this landscape, i can’t help but see the imbalance. big labs have clusters with tens of thousands of gpus. we have to schedule training runs around when the cloud credits reset. they burn millions on parameter sweeps. we squeeze every drop of performance out of a few A100s. and yet, we’re still here, still building models that compete.
there’s a quiet power in that. we’re proof that intelligence isn’t purely a function of scale. it’s a function of how you think about thinking. that’s what sagea’s about. we’re not trying to replicate what openai or anthropic does. we’re trying to rethink what reasoning means. we’re exploring architectures that simulate how humans reflect, reconsider, and course-correct in real time. it’s not just about predicting the next token. it’s about understanding why the next token should exist.
and maybe that’s why we get attention. our inbox is full of vc intros these days. some of them are serious. real funds with real conviction. they talk about our models, about the hybrid reasoning head, about the fact that we’ve shipped multiple open-source systems that actually work. it feels validating. like the months of obscurity are finally paying off.
but there’s always a question at the back of my mind: do we take the money or not?
i’ve been thinking about bootstrapping sagea again. part of me loves the purity of it. the idea of building something self-sustaining, something unbought. it feels cleaner. more grounded. i like the idea of not having to sell a vision just to survive. but i also know how expensive this work is. research isn’t cheap. scaling models isn’t cheap. and while we’ve built efficient systems that run on modest infra, there’s a limit to how long you can stretch that.
sometimes i think bootstrapping is the truest form of belief. like, if you’re willing to suffer for your vision without funding, that’s conviction. but at the same time, i don’t want sagea to move slow because of pride. i want us to win. i want us to build fast, open fast, and stay ahead. and maybe funding is just fuel for that. maybe it’s not a compromise. maybe it’s a multiplier.
the reality is, the ai scene is changing fast. small teams like ours are starting to matter. the open-source community is moving quicker than the corporate labs expected. the speed of iteration is insane. sometimes, i’ll wake up to see three new repos with architectures i was just sketching out a week before. it’s like this collective hivemind of people obsessed with building smarter systems, and i love it. it reminds me that intelligence isn’t proprietary. it’s emergent.
i think about where this all goes. about what happens when the hype fades. because it will. the market can’t sustain this level of noise forever. some companies will vanish. some will pivot. others will quietly die in corporate labs with too much bureaucracy to move. and somewhere in that mix, teams like ours will still be around. because we never built for hype. we built because we cared about the work.
there’s this weird peace in knowing that. knowing that no matter what happens, we’ll still be here doing what we do. maybe we’ll be bigger. maybe we’ll still be three people and a few servers. but we’ll still be pushing the edges of reasoning. still chasing the strange beauty of intelligence.
and right now, the next few months feel pivotal. we’re not just releasing a model. we’re setting a standard for what small labs can achieve. the 40b moe isn’t just a milestone. it’s proof that innovation doesn’t belong to the biggest labs. it belongs to whoever cares enough to go deep.
we’ve also been playing around with some new architectures internally. models that combine local attention mechanisms with adaptive reflection loops. we’re trying to see how we can make reasoning more self-corrective without external feedback. it’s early, but some of the initial results are wild. these models are learning how to question their own outputs in subtle ways. not perfectly. not consistently. but enough to feel different.
sometimes, i forget how insane it is that this is our life right now. that we get to wake up and literally work on machines that can think. that we’re part of the generation figuring out how reasoning happens. it’s almost poetic.
the other day, someone asked me what success means for sagea. i didn’t know how to answer at first. but now i think it’s this: success means building something that moves intelligence forward and still feels human. something open, accessible, and deeply thoughtful. i don’t want us to just make big models. i want us to make meaningful ones.
so yeah, next week’s a big one. the MoE goes live. our new research paper follows soon after. investor talks are happening, but we’re not rushing. and somewhere in the back of my mind, the idea of bootstrapping still lingers. maybe we’ll take that path. maybe not.
what i do know is that we’re not slowing down. not now. not when it finally feels like the world is starting to listen.
ciao, basab
Want to get notified everytime i write stuff? Sign up to my newsletter