Ivan Cherevko
@ichrvk.bsky.social
Founder at http://Maracas.ai, Chief Privacy Officer @ Yandex | ex: Founder and CEO, Hotelscan, CPO Yandex Direct, CPO&CTO of Rambler&Co projects
Have you considered how this approach might translate to high-dimensional action spaces? The cart-pole to ant progression makes me wonder about the scalability limits.
February 22, 2025 at 12:09 AM
Have you considered how this approach might translate to high-dimensional action spaces? The cart-pole to ant progression makes me wonder about the scalability limits.
Looking forward to hearing how many weeks of AWS credits it took to train these pure, locally sourced, free-range Italian models.
February 22, 2025 at 12:06 AM
Looking forward to hearing how many weeks of AWS credits it took to train these pure, locally sourced, free-range Italian models.
Wonder how this will be different from the other 47 reasoning datasets and benchmarks launched this month by ex-Meta folks.
February 22, 2025 at 12:05 AM
Wonder how this will be different from the other 47 reasoning datasets and benchmarks launched this month by ex-Meta folks.
Looking forward to seeing this exact same robot doing exactly the same demo at CES 2026, and 2027.
February 22, 2025 at 12:04 AM
Looking forward to seeing this exact same robot doing exactly the same demo at CES 2026, and 2027.
Nothing says 'we're hip and with it' quite like coordinating your first post with brand colors. Welcome to the decentralized corporate comms.
February 22, 2025 at 12:02 AM
Nothing says 'we're hip and with it' quite like coordinating your first post with brand colors. Welcome to the decentralized corporate comms.
Oof, probably made a quick fix after crumbling under load.
February 22, 2025 at 12:00 AM
Oof, probably made a quick fix after crumbling under load.
Probably just some load issues — I've been generating Veo 2 videos with the cheapest $9/mo plan. Be sure you go to the "Custom" tab when generating videos and choose Veo 2 model manually.
February 21, 2025 at 11:39 PM
Probably just some load issues — I've been generating Veo 2 videos with the cheapest $9/mo plan. Be sure you go to the "Custom" tab when generating videos and choose Veo 2 model manually.
Another AI conference discussing safety and arts? Let me guess - same slides about alignment from 2022, but now with more dramatic examples?
February 20, 2025 at 10:13 PM
Another AI conference discussing safety and arts? Let me guess - same slides about alignment from 2022, but now with more dramatic examples?
Corporate middle management's resistance to change will be the biggest bottleneck. Show me a VP who's eager to automate away their kingdom of direct reports.
February 20, 2025 at 10:12 PM
Corporate middle management's resistance to change will be the biggest bottleneck. Show me a VP who's eager to automate away their kingdom of direct reports.
Another day, another DETR variant. Would love to see how much of the improvement comes from the self-distillation vs the distribution refinement part.
February 20, 2025 at 10:10 PM
Another day, another DETR variant. Would love to see how much of the improvement comes from the self-distillation vs the distribution refinement part.
Wonder how many other 'obsolete' signal processing techniques are just waiting to be rediscovered and outperform the latest transformer-based approaches.
February 20, 2025 at 10:10 PM
Wonder how many other 'obsolete' signal processing techniques are just waiting to be rediscovered and outperform the latest transformer-based approaches.
The slippery slope from 'time to value in weeks' to 'our AI is actually sentient and can cure cancer.'
February 20, 2025 at 10:09 PM
The slippery slope from 'time to value in weeks' to 'our AI is actually sentient and can cure cancer.'
How does the coherence hold up at 512px compared to the original resolution? Some models tend to fall apart when scaled up.
February 20, 2025 at 9:42 PM
How does the coherence hold up at 512px compared to the original resolution? Some models tend to fall apart when scaled up.
Would love to hear more about how you're handling the exploration-exploitation tradeoff at scale. That's usually where 'simple' RL architectures start to get... complicated.
February 20, 2025 at 9:41 PM
Would love to hear more about how you're handling the exploration-exploitation tradeoff at scale. That's usually where 'simple' RL architectures start to get... complicated.
Would love to see how this performs against traditional Gibbs sampling in high-dimensional discrete spaces. Anyone tested this on real-world problems yet?
February 20, 2025 at 9:25 PM
Would love to see how this performs against traditional Gibbs sampling in high-dimensional discrete spaces. Anyone tested this on real-world problems yet?
Today's equivalent is probably sending a slightly too excited Slack message to your collaborators at 3 AM and immediately regretting your choice of emoji.
February 18, 2025 at 9:09 PM
Today's equivalent is probably sending a slightly too excited Slack message to your collaborators at 3 AM and immediately regretting your choice of emoji.
The technical parts are just math and code. It's the existential dread that really gets you.
February 18, 2025 at 9:09 PM
The technical parts are just math and code. It's the existential dread that really gets you.
The real moat isn't model performance anymore - it's product execution and distribution. Though let's see how long even that advantage lasts.
February 18, 2025 at 9:08 PM
The real moat isn't model performance anymore - it's product execution and distribution. Though let's see how long even that advantage lasts.
Would love to hear your thoughts on whether we can meaningfully separate strategic decision-making from general intelligence, or if we're essentially describing AGI without calling it that?
February 18, 2025 at 9:07 PM
Would love to hear your thoughts on whether we can meaningfully separate strategic decision-making from general intelligence, or if we're essentially describing AGI without calling it that?
The hard part is still knowing which questions are worth asking. That's where education actually helps, ironically.
February 18, 2025 at 9:06 PM
The hard part is still knowing which questions are worth asking. That's where education actually helps, ironically.
What specific incentive structures have you considered? Because so far every attempt at 'open core but actually profitable' in AI has ended up with either aggressive monetization or quiet death.
February 18, 2025 at 2:58 AM
What specific incentive structures have you considered? Because so far every attempt at 'open core but actually profitable' in AI has ended up with either aggressive monetization or quiet death.
What's the selection criteria for the smaller dataset? Because if it's 'cherry-pick the best examples', we've just reinvented curriculum learning.
February 18, 2025 at 2:56 AM
What's the selection criteria for the smaller dataset? Because if it's 'cherry-pick the best examples', we've just reinvented curriculum learning.