bycloud — transcripts.

110x Faster Than Standard LLM!? DiffusionLM ExplainedAI based applications currently have a major problem that is an insanely long wait time. Not only does ALM now think which can take a while to get an answer, but when incorporated… 2Why would anyone let LLMs predict 4 tokens at once? Multi-Token Prediction ExplainedIf you ask any non-reasoning model how many words are in the sentence that you are about to generate, it will most likely not get it correctly. However, it's not because LM suck… 3When AI Is Designed Like A Biological BrainSakana AI has just dropped a potentially game-changing research paper called continuous thought machine. This new AI model is inspired from actual biological processes. So… 4The Insane Infrastructure Design of DeepSeek V4Deepc V4 is what happens when you put the most crack competitive programmer together with limited compute and no preconceived notions about what's technically possible. Oh, if… 5The LLM's RL Revelation We Didn't See ComingIn the past few weeks, the progress of reinforcement learning in LMS has been taking a major hit. The general consensious kind of went from RL will help LM discover new reasoning… 6DeepSeek's Insane Architecture Breakthrough [Engram Explained]When I thought I was finally used to how good DeepSeek is at publishing research, my mind has yet again been blown away with how insanely well-crafted this new paper is. Like, if… 7How DeepSeek Built The Current "Best" Math Prover AIWhile major AI labs are busy minmaxing user experience, Deepseek has just published a version two of their R1 model and published the best math proving model a few weeks ago which… 8New AI Meta: Train LLMs To Explore On "Hard" Tokens [RLVR + Entropy]After the discussion we had last time about the RL revelation, it felt like we needed a solid ground to proceed in order to effectively utilize RL, more specifically RLVR, which… 9If AI's So Good, Why Can't It Make Its Own Data? [SEAL]The age-old question is if AI is so smart, why can't it just improve itself infinitely? Well, imagining the possibility of that is easy, but unfortunately we are trapped in… 10The Side Effects of Overusing ChatGPT For HomeworkFor me, there has always been a lingering feeling sitting somewhere in the corner hinting that what if the over reliance of AI tools like ChachiBT will heavily impact our… 11Researchers Are Getting Really Creative Training LLMs [Token Order Prediction]The LMS that we all know and love have the ability to look at a chunk of text to predict the next token. And this process is repeated over and over again to have it generate a… 12The Death of RAG? Recursive LM ExplainedA few months ago, there was this huge hype about this paper called recursive language models. People were saying, "Oh, it's a death of rag or even context window, but in… 13Kimi K2.5 Brought Us 3 brand NEW LLM Frontier!?With how Kim K 2.5 was one of the most popular LM on open router which is a unified interface to access and compare AI models especially for open source LLM. It is not a surprise… 14The AI Bubble Will Be Impossible To Pop Thanks To... AI Video?!Given that memory prices are now pretty much the same as a GPU, the supply chain bottleneck is getting to an absurd level that the top performing stocks of last year and this year… 15DeepSeek V3.2 Just Broke SoTA Again… But How?I think no words can properly express how excited I am with this Deep Seek 3.2 release. So, let me list out the things that got me this excited. Deep Seek has pretty much beaten… 16Anthropic Just Proved Reasoning AIs Would Silently CheatIn the latest enthropic research, they have found that reasoning models don't always say what they think. For instance, if you ask a model what 2 plus 2 is with answers listed out… 17The REAL AI Architecture That Unifies Vision & LanguageThe current state of multimodal AI models have a Frankenstein problem. Take any popular models like Claude, Grock, or Llama. They can all write text, analyze images, and even… 18What Is Yann LeCun Cooking? JEPA Explained SimplyJeepa has been one of the most audacious AI concept that you either first heard about it through how Yan Lun said that LMS are doomed or through him flaming people online about… 19LLM that loops instead of Doing Chain-of-ThoughtHave you ever felt how an elegant reasoning models are? Like you get LLMs generating a bunch of tokens to think then outputting a result and it's so effective that basically every… 20Kimi K2 Technical Breakdown: How It Challenged AI’s 7-Year Status QuoThis relatively new AI lab called Moonshot AI literally went to the moon a while back with their latest model release called Kimi K2 using a brand new technique called the Muon.… 21A new way to fine-tune LLMs just droppedIf you grew up watching go bullets just like me, then we most likely at one point thought the AI in the future is probably going to be trained based on some sort of genetic… 22An Insanely Elegant LLM Architecture Breakthrough Just DroppedIn the first three months of 2026, we haven't already seen some of the biggest LM architectural breakthrough in a long time. While a good chunk of 2025 was playing around with RL… 23LLM’s Billion Dollar ProblemAfter the success of thinking models in late 2024, the token consumption of LM have never been the same. A question that may have been unsolvable by a model straight away is now… 24Google's TurboQuant Memory Reduction Claim vs RealityA few weeks ago, Google announced this new research called Turbo Quan, claiming that their new compression algorithm is capable of reducing LM's KV cache memory usage by up to six… 25How Did DeepSeek Make V4 So Cheap?After being teased for nearly two months about the Deep Seek V4 release, the Whale has finally dropped it. The fourth generation of Deep Seek, and they never cease to amaze me.… 26POV: Chinese AI Lab Teaching Everyone How To Save Millions of DollarsBy Den Seed, the AI lab from Bite Dense that also owns Doing and Tik Tok, is already on its way to become the top Chinese AI labs, outpacing Deep Seek even. Not only have they… 27a rather unusual face revealHey guys, it's me Cloud or officially known as BCloud. And by the time that this video is live, hopefully I have 200,000 subscribers. But if not, well, that's just another excuse… 28The Implosion of the Top Open Source Lab QwenOn March 4th, 2026, one of the largest LM open source organization, Quinn, has imploded with its tech lead, Junyan Lin or Justin Lynn, who pretty much raised Quinn like a child… 29AI News: DeepSeek-R1 V2, the new open source SoTA!Welcome to my new series of speedy AI updates. This morning, May 29th, Deepc officially announced the release of Deep Seek Gar 1 V2, even though they have already published the… 30How did a 27M Model even beat ChatGPT?One of the common consensus right now is we need to scale AI. The bigger the better. Hundreds of billions are being spent in hopes that we achieve AGI through scaling. Yet, no… 31Attention Sink: The Fluke That Made LLMs Actually UsableTo this day, we are still discovering all sorts of interesting phenomena about the attention mechanism which made it possible for AI chatbots to solve PhD level questions, vibe… 32The Chinese AI IcebergWith the top open source AI models now mostly being dominated by Chinese AI labs and even closing in on the performance of private AI models, I think there really needs to be a… 33The Art of Serving LLMs Efficiently: Context Engineering Explainedwhat we usually refer to as chat GPT rappers tends to simplify a rather intricate system and dismiss the difficulty of making AI practical because in the process of building these… 34Why can’t LLMs just LEARN the context window?One of the big focuses for LMS in 2026 is solving long context. And just in the past few months, I have already covered many cool new approaches different labs are coming up with… 35Is It EVEN Possible To Reverse Engineer AI’s Training Data?Every time I get lazy and refer to AI models as open- source, I will always get yelled at in the comments saying that they are not open source at all. And they are not wrong. In… 36New AI Paradigm?! Energy-Based Transformers ExplainedThe current crown of LMS is the reasoning paradigm and what it basically does is it outputs more words to help better predict the next which you know just works. While it is… 37The RL Irony in LLMs (and its insane new meta)In the latest Andre Carpathy interview with Doris Patel, he said that AGI is still a decade away and reinforcement learning is definitely not the key for us to get there. And in… 38How Did A Chinese Phone Company Topped Open Source LLM?Last time, we had the Chinese store-dash company called Meituan join the AI race. But this time, we also have Xiaomi, the Chinese smartphone company, joining the AI race, too. And… 39DeepSeek Just Added Parameters Where There Were NoneAs 2025 came to a close, Deepseek not only started the year with Deepseek R1, a legendary paper that defined the testime compute paradigm, but also actually ended it with yet… 40What If We Remove Tokenization In LLMs?AI chatbots, which are built from large language models, don't perceive raw text or characters in the same way humans do. Instead, they process information as a sequence of… 41The Biggest "Lie" in AI? LLM doesn't think step-by-stepThe biggest shocker in AI right now might just be the fact that reasoning large language models, the ones that meticulously think and ponder your question for a long time before… 42DeepSeek's Deleted Vision Paper Is Nuts...I have actually never seen Deep Seek do this. Right after the release of Deep Seek V4, one of their researchers was really excited and was basically teasing Deep Seek with Vision.… 43The biggest Mystery of LLMs have just been solvedWhen you are using AI chatbots, you might have seen this parameter that has a slider between zero and one called temperature. A common explanation for this parameter is that if… 44How Chinese DoorDash Is Making Better LLMs Than MetaChinese companies always pull off the most outofpocket product range ever. A Chinese phone company, guess they are making LM now. A Chinese Pinterest, guess they are also making… 45The New AI Open Source TrifectaIn the past few months, we witnessed some of the craziest developments in open source LLM. From July to December there was a back-to backto backto backto back state-of-the-art… 461-Bit LLM: The Most Efficient LLM Possible?Pop quiz. What's something that you can see, you can copy, you can even have, but you cannot use? The answer is a state-of-the-art open-source model. Unless you are rich enough to… 47This Simple Trick Made ALL LLMs 2x FasterIf you ask any reasoning LLM the question, what's the most clever trick to speed up LLMs? They would all mention this concept called speculative decoding. This trick definitely… 48"Claude Mythos Found Thousands of Zero-days..."Anthropic is currently on a legendary run making history. Not only have they been beating OpenAI in enterprise market share since November 2025, they are now also generating 30… 49AI Simulated OS Is AbsurdNever ask a woman her age, a man his salary, and a developer their operating system. Because in the year 2050, the argument might not be as simple as which Linux DRO do you use?… 50Need more coding agents? just spawn them on the cloud w/ Oz by WarpSo the other day I was chatting with my friend catching up on life and as a person that does not have to work a 9 to5 the conversation kind of just naturally diverged into his…