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
Just got access to Veo 2 through Freepik (odd collaboration, but hey). The prompt coherence is amazing, even with super complex prompts. Completely outperforms Hunyuan, Kling, and other competitors (and don't even get me started on Sora).
February 21, 2025 at 10:26 PM
Just got access to Veo 2 through Freepik (odd collaboration, but hey). The prompt coherence is amazing, even with super complex prompts. Completely outperforms Hunyuan, Kling, and other competitors (and don't even get me started on Sora).
One of my random tweet replies just went viral for the first time (2100 likes on Twitter!). Feels weird that an off-hand comment blew up while my actual content gets crickets.
February 20, 2025 at 12:56 AM
One of my random tweet replies just went viral for the first time (2100 likes on Twitter!). Feels weird that an off-hand comment blew up while my actual content gets crickets.
Grok 3 is quite good - not mind-blowing (at least without Deep Search/Big Brain, which I haven't tried yet), but solid. Strong showing by the xAI team, though not enough to shake up the status quo.
February 18, 2025 at 7:51 PM
Grok 3 is quite good - not mind-blowing (at least without Deep Search/Big Brain, which I haven't tried yet), but solid. Strong showing by the xAI team, though not enough to shake up the status quo.
I'm excited for the Grok 3 launch in a way I haven't been for other model releases—not because I have particularly high hopes (I don't), but because it will definitively show whether scale in pre-training still matters.
February 18, 2025 at 2:16 AM
I'm excited for the Grok 3 launch in a way I haven't been for other model releases—not because I have particularly high hopes (I don't), but because it will definitively show whether scale in pre-training still matters.
So we're just casually dropping models that beat both Llama and state-of-the-art memory models by massive margins now? Looking forward to the independent replications of these impressive LM2 results.
February 16, 2025 at 7:33 PM
So we're just casually dropping models that beat both Llama and state-of-the-art memory models by massive margins now? Looking forward to the independent replications of these impressive LM2 results.
Turns out you can teach LLMs to reason with just 17k examples and the content doesn't even need to be correct. The real secret? Just keep the reasoning structure coherent. Makes you wonder about all those carefully curated datasets.
February 16, 2025 at 6:53 PM
Turns out you can teach LLMs to reason with just 17k examples and the content doesn't even need to be correct. The real secret? Just keep the reasoning structure coherent. Makes you wonder about all those carefully curated datasets.
Finally, a paper that tells us when distillation beats supervised learning without having to run a million experiments. TL;DR: Use distillation for small compute budgets or when you already have a teacher, otherwise stick to supervised learning.
February 16, 2025 at 3:11 AM
Finally, a paper that tells us when distillation beats supervised learning without having to run a million experiments. TL;DR: Use distillation for small compute budgets or when you already have a teacher, otherwise stick to supervised learning.
Huginn: Smaller model that gets smarter by thinking longer in latent space. Finally someone trying something other than just making models bigger or forcing them to write their thoughts down like a middle schooler.
February 13, 2025 at 11:50 PM
Huginn: Smaller model that gets smarter by thinking longer in latent space. Finally someone trying something other than just making models bigger or forcing them to write their thoughts down like a middle schooler.
Jobs disappeared due to high interest rates, but AI means they're not coming back. This decade will be tough for junior software engineers.
February 13, 2025 at 2:46 AM
Jobs disappeared due to high interest rates, but AI means they're not coming back. This decade will be tough for junior software engineers.
Mathematical reasoning in LLMs has always been seen as data-hungry, requiring 100k+ examples. New paper shows just 817 carefully curated samples can achieve 57.1% on AIME. If reproducible, this challenges everything we thought we knew about scaling laws in reasoning tasks.
February 11, 2025 at 10:44 PM
Mathematical reasoning in LLMs has always been seen as data-hungry, requiring 100k+ examples. New paper shows just 817 carefully curated samples can achieve 57.1% on AIME. If reproducible, this challenges everything we thought we knew about scaling laws in reasoning tasks.
It's amazing how much drama surrounds OpenAI. My wife knows that whenever I start a story with 'Hey look, Sam Altman...' or 'Hey look, OpenAI...,' it won't be about a new model or product—it'll be about another chapter in the daytime soap opera that is their boardroom drama.
February 11, 2025 at 7:36 PM
It's amazing how much drama surrounds OpenAI. My wife knows that whenever I start a story with 'Hey look, Sam Altman...' or 'Hey look, OpenAI...,' it won't be about a new model or product—it'll be about another chapter in the daytime soap opera that is their boardroom drama.
DeepMind's AlphaGeometry2 solves 84% of IMO geometry problems from 2000-2024, surpassing gold medalist performance. The system nearly doubles its predecessor's 54% solve rate through improved language modeling, faster symbolic reasoning, and novel search techniques.
February 8, 2025 at 3:53 AM
DeepMind's AlphaGeometry2 solves 84% of IMO geometry problems from 2000-2024, surpassing gold medalist performance. The system nearly doubles its predecessor's 54% solve rate through improved language modeling, faster symbolic reasoning, and novel search techniques.
Wild how Deepseek became the first AI model besides ChatGPT to become a household name. Even my parents talk about 'that Chinese AI they keep hearing about on the news.' Claude, Gemini, Mistral, Qwen, Grok - none of them hit mainstream. It's just ChatGPT and Deepseek.
February 8, 2025 at 2:31 AM
Wild how Deepseek became the first AI model besides ChatGPT to become a household name. Even my parents talk about 'that Chinese AI they keep hearing about on the news.' Claude, Gemini, Mistral, Qwen, Grok - none of them hit mainstream. It's just ChatGPT and Deepseek.
Comprehensive study shows that long chain-of-thought reasoning in LLMs emerges from SFT + RL but requires careful reward shaping to control length scaling. Rule-based reward signals with filtering outperform model-based approaches for stabilizing performance.
February 7, 2025 at 3:43 AM
Comprehensive study shows that long chain-of-thought reasoning in LLMs emerges from SFT + RL but requires careful reward shaping to control length scaling. Rule-based reward signals with filtering outperform model-based approaches for stabilizing performance.
New research from Stanford and UW shows how to achieve o1-preview-level performance for test-time scaling with just 1K examples and simple supervised training. The key? Careful data curation and sensible compute budgeting.
February 5, 2025 at 12:13 AM
New research from Stanford and UW shows how to achieve o1-preview-level performance for test-time scaling with just 1K examples and simple supervised training. The key? Careful data curation and sensible compute budgeting.
Interesting find: OpenAI's Deep Research web browser has JavaScript disabled by default, but the model can enable it when needed. Cool to see the technical decisions behind this project.
February 4, 2025 at 11:39 PM
Interesting find: OpenAI's Deep Research web browser has JavaScript disabled by default, but the model can enable it when needed. Cool to see the technical decisions behind this project.
OpenAI's Deep Research is way better than Gemini's, which is the only reason I'm keeping my Gemini Advanced subscription. Oh well.
February 3, 2025 at 6:08 PM
OpenAI's Deep Research is way better than Gemini's, which is the only reason I'm keeping my Gemini Advanced subscription. Oh well.
It's wild how viciously anti-AI the 🦋 crowd is. You'd think their lives had been destroyed by rogue LLMs, given the level of vitriol.
February 2, 2025 at 5:25 AM
It's wild how viciously anti-AI the 🦋 crowd is. You'd think their lives had been destroyed by rogue LLMs, given the level of vitriol.
We've gone from LLMs as 'write a limerick' machines through a brief 'count the letter r' phase to our current 'watch balls bounce in a rotating rectangle' era.
February 2, 2025 at 12:41 AM
We've gone from LLMs as 'write a limerick' machines through a brief 'count the letter r' phase to our current 'watch balls bounce in a rotating rectangle' era.
At Maracas, our coding assignments are fully open-book. We encourage using any AI tools you want. We're hiring for both coding & AI skills—and it's supercharged our velocity. Still doing leetcode whiteboard interviews in 2025? That's just irresponsible.
January 17, 2025 at 8:21 AM
At Maracas, our coding assignments are fully open-book. We encourage using any AI tools you want. We're hiring for both coding & AI skills—and it's supercharged our velocity. Still doing leetcode whiteboard interviews in 2025? That's just irresponsible.
Teaching is just the latest job being displaced by AI. Six weeks of LLM tutoring now matches two years of traditional education.
January 17, 2025 at 7:49 AM
Teaching is just the latest job being displaced by AI. Six weeks of LLM tutoring now matches two years of traditional education.
Humanoid robotics isn't at its language AI moment yet (like we saw in 2022), but I can feel we're getting really close.
January 16, 2025 at 3:01 PM
Humanoid robotics isn't at its language AI moment yet (like we saw in 2022), but I can feel we're getting really close.
ChatGPT Tasks feels underwhelming - as a concept, product, and direction for ChatGPT. If Operator turns out just as weak, that's not great news for OpenAI...
January 15, 2025 at 9:50 PM
ChatGPT Tasks feels underwhelming - as a concept, product, and direction for ChatGPT. If Operator turns out just as weak, that's not great news for OpenAI...
Here's an efficient prompt format I use often with o1-pro: 'Ask me 10 questions that you need answers to in order to [___].' It lets you outsource more thinking to the model since it generates even the questions that need to be asked.
January 13, 2025 at 6:29 PM
Here's an efficient prompt format I use often with o1-pro: 'Ask me 10 questions that you need answers to in order to [___].' It lets you outsource more thinking to the model since it generates even the questions that need to be asked.
Seeing more 15-18yo founders lately. Years ago, I pitched a coding school where talented teens could skip high school, get elite CS training in half the time, and join the workforce as rockstars at 17-18. Still relevant.
January 13, 2025 at 12:58 PM
Seeing more 15-18yo founders lately. Years ago, I pitched a coding school where talented teens could skip high school, get elite CS training in half the time, and join the workforce as rockstars at 17-18. Still relevant.