Ivan Yamshchikov

Here is my third English episode of the podcast. Some of us are lucky to meet exceptional people at the right time of our life. Jürgen happened to be such a person for me. I am honored to share his thoughts on mathematics and science here. youtu.be/QY5ZG7mwDa8

@bougakov епппп

@kr0niker Огнище 🔥

this is the best repost (and thread) you will read tonight

Glad to see this called out so clearly. Fine-tuning infra is maturing, but the real challenge is everything around the run: data quality, iteration speed, deployment path, and how fast you can turn a model update into a production system. That is the gap Nebius Token Factory is built to close 👇

A cool project by @pleiasfr co-founder and CTO @Dorialexander and @ynckdrt11496 @nebiustf check that out!

The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons. We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views. We’re speaking up now because once a misleading narrative spreads, it becomes much harder to correct. We’ve written a public comment on openreview (openreview.net/forum?id=tO3AS…). We would greatly appreciate your attention and help in sharing it.

interesting times will continue until morale improves

Taido Purason presents our paper “Teaching Old Tokenizers New Words: Efficient Tokenizer Adaptation for Pre-trained Models” at #EACL2026 Findings today. aclanthology.org/2026.findings-…

The $46 billion vape industry hired flavor chemists to solve a specific problem: nicotine tastes bitter. The human tongue has T2R receptors that detect nicotine and trigger aversion. So the industry reverse-engineered the problem. They loaded e-liquids with fruity aromatic compounds like farnesol, farnesene, and ethyl butyrate that suppress the bitterness signal and activate sweetness perception through the orbitofrontal cortex instead. Turns out they built something more powerful than a nicotine delivery system. A 2023 study in the Journal of Neuroscience found that green apple vape flavorants, with zero nicotine present, independently fire dopamine neurons in the ventral tegmental area and increase dopamine release in the nucleus accumbens. The same reward circuitry that nicotine hijacks. The flavor chemicals alone were producing reward-seeking behavior in mice. A separate study found strawberry additives significantly increased nicotine vapor sampling, meaning the fruit smell made subjects inhale more of the drug without any conscious decision to do so. 95% of vape users choose flavored products. The industry has always framed this as “consumer preference.” The neuroscience says the flavors are pharmacologically active compounds that directly alter brain reward circuits and increase drug intake. Now look at what those compounds do once the device gets tossed. The UK was discarding 5 million disposable vapes per week before the ban. The aromatic volatiles don’t stop broadcasting once the device is empty. Orthonasal olfaction, the same pathway that makes you smell a strawberry from across a room, works identically in mammals. A squirrel’s olfactory system processes fruity volatiles through the same receptor families humans use. The signal reads as food. Nobody saw squirrels gnawing on Marlboro butts for 60 years. Vapes show up and suddenly there’s footage from London, Philadelphia, and Wales. The animals are chewing on lithium batteries wrapped in candy-scented plastic because the flavor engineering worked exactly as designed on a nervous system it was never tested on. The lithium in those discarded vapes equals 5,000 electric vehicle batteries per year. Oxford researchers found the cells inside can cycle 450+ times, but the product is built to be used once and thrown in a park. A squirrel holding a blueberry vape on a Brixton fence is the most honest product review the flavor chemists have ever received.

@Dorialexander @V4ldeLund both. The workshop is more realistic though.

@kr0niker @V4ldeLund the workshop or deutsche bahn working?

While I like the way Evaluation is reframed, but we really should have a separate Data research track (not just "datasets").

@V4ldeLund @Dorialexander it’s doable

@Dorialexander Looking forward, if accepted I will definitely try to attend ! (even despite the fact that I need to use Deutsche Bahn to get to Paris)

I can't get over the fact that Dune is about an oppressed people fighting for their homeland, waging a jihad to bring down a hegemonic empire by threatening to cut off the flow of their most precious commodity after the empire had assassinated their religious leader's father.

Congrats to the @cursor_ai team on the launch of Composer 2! We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support. Note: Cursor accesses Kimi-k2.5 via @FireworksAI_HQ ' hosted RL and inference platform as part of an authorized commercial partnership.

🚨NEWS: Cursor’s $50B “in-house model” is literally Kimi K2.5 with RL on top. Got caught in 24 hours >be Moonshot AI >spend hundreds of millions training Kimi K2.5 >1 trillion parameters, 15 trillion tokens, agent swarm architecture >beat GPT-5.2 and Opus 4.5 on real benchmarks >open-source it because you believe in the ecosystem >one condition: display “Kimi K2.5” if you make over $20M/month from it >Cursor takes the model >runs RL on coding tasks >ships it March 19 as “Composer 2” >blog post: “continued pretraining + scaled reinforcement learning” >zero mention of Kimi K2.5 >“our in-house models generate more code than almost any other LLMs in the world” >publishes benchmark chart >Composer 2 against Opus 4.6 and GPT-5.4 >uses the chart to justify raising at $50 billion! >less than 24 hours later >kimi dev intercepts the API response >model ID: kimi-k2p5-rl-0317-s515-fast >they didn’t even rename it >Moonshot head of pretraining runs tokenizer test >confirms: identical to Kimi’s tokenizer >publicly tags Cursor’s co-founder: “why aren’t you respecting our license?” >two more Moonshot employees post confirmations >all three posts deleted within hours >legal is now involved >but it gets worse >Cursor had Kimi K2.5 listed as a FREE model in their UI just weeks ago >users were openly using it >Feb 9: “K2.5 was in my model list. I updated and it vanished” >it vanished because Cursor pulled it from the picker, and relaunched it as their own model >Moonshot valuation: $4.3B >Cursor valuation: $50B Absolute state of Cursor.

@ashvardanian is doing God’s work.

@kukutz этого я не знал. Спасибо за информацию.

@kr0niker But before long, the conversation moved in a different direction as Kent nodded to conspiracy theories that pro-Israel forces were behind the assassination of conservative activist Charlie Kirk. washingtonpost.com/politics/2026/…

big, if true. A really nice hypothesis on the hormesis effect.

@kukutz «всё ок» в данный момент мировой истории — это оч странное утверждение, как по мне. Примерно всё не ок, но мне кажется, что это важное событие. А ты почитал про intelligence loop? Там оно на сайте Кнессета есть?

@kr0niker так. и? ты прочитал заявление об увольнении целиком? тебе кажется, что все ок?

@kukutz вот тут должен репорт лежать, но у меня с мобильного почему-то не открывается. knesset.gov.il/committees/eng….

@kr0niker так ты не прочитал его заявление, которое ретвитнул? «war manufactured by Israel», и речь даже не про Иран, а про Ирак, etc.

An investigation into Israel’s failure to provide accurate intelligence on Iraq’s weapons capabilities found that Israeli intelligence agencies suffered from a closed “information loop,” as well as other failures. The conclusions are the result of an eight-month investigation by the Knesset Foreign Affairs and Defense Subcommittee for Intelligence and Secret Services that questioned intelligence officials, military officers, and Cabinet members. On March 28 2004, the subcommittee released an 81-page declassified report pending the completion of a longer, classified version. The report criticized Israeli intelligence agencies for a number of failures, concluding that the agencies overestimated Iraq’s ability to strike at Israel directly. Likud Knesset member Yuval Steinitz, who headed the investigation, cited an “escalation” between 1998 and the beginning of the war in the number of missiles believed to be possessed by Iraq for which there was “no explanation.” The subcommittee found that Israeli intelligence agencies used information from foreign intelligence services without recognizing that the other states obtained the data from Israel in the first place. The result, according to Steinitz, was that speculation was passed in circles “without any substantiation from the field.” However, Steinitz rejected suggestions that the Israeli agencies intentionally misled the United States and others in hopes of encouraging them to go to war against Iraq, a longtime enemy of Israel.

@kukutz а если ты про чувака, что уволилися, то мне кажется это довольно важная информация сама по себе. А он антисемит что ли?

@kr0niker слушай, а вот ты Кента с его антисемитскими теориями заговора репостнул — можешь плиз пояснить, в каком смысле? "Смотрите какое странное" или "я совершенно согласен"?