『ThursdAI - The top AI news from the past week』のカバーアート

ThursdAI - The top AI news from the past week

ThursdAI - The top AI news from the past week

著者： From Weights & Biases Join AI Evangelist Alex Volkov and a panel of experts to cover everything important that happened in the world of AI from the past week
無料で聴く

このコンテンツについて

 Every ThursdAI, Alex Volkov hosts a panel of experts, ai engineers, data scientists and prompt spellcasters on twitter spaces, as we discuss everything major and important that happened in the world of AI for the past week. Topics include LLMs, Open source, New capabilities, OpenAI, competitors in AI space, new LLM models, AI art and diffusion aspects and much more.

sub.thursdai.newsAlex Volkov 政治・政府
エピソード
  • 📆 ThursdAI - July 17th - Kimi K2 👑, OpenAI Agents, Grok Waifus, Amazon Kiro, W&B Inference & more AI news!

    📆 ThursdAI - July 17th - Kimi K2 👑, OpenAI Agents, Grok Waifus, Amazon Kiro, W&B Inference & more AI news!
    2025/07/17
    Hey everyone, Alex here 👋 and WHAT a week to turn a year older! Not only did I get to celebrate my birthday with 30,000+ of you live during the OpenAI stream, but we also witnessed what might be the biggest open-source AI release since DeepSeek dropped. Buckle up, because we're diving into a trillion-parameter behemoth, agentic capabilities that'll make your head spin, and somehow Elon Musk decided Grok waifus are the solution to... something.This was one of those weeks where I kept checking if I was dreaming. Remember when DeepSeek dropped and we all lost our minds? Well, buckle up because Moonshot's Kimi K2 just made that look like a warm-up act. And that's not even the wildest part of this week! As always, all the show notes and links are at the bottom, here's our liveshow (which included the full OAI ChatGPT agents watch party) - Let's get into it! ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.🚀 Open Source LLMs: The Kimi K2 RevolutionThe New Open Source King Has ArrivedFolks, I need you to understand something - just a little after we finished streaming last week celebrating Grok 4, a company called Moonshot decided to casually drop what might be the most significant open source release since... well, maybe ever?Kimi K2 is a 1 trillion parameter model. Yes, you read that right - TRILLION. Not billion. And before you ask "but can my GPU run it?" - this is an MOE (Mixture of Experts) with only 32B active parameters, which means it's actually usable while being absolutely massive.Let me give you the numbers that made my jaw drop:* 65.8% on SWE-bench Verified - This non-reasoning model beats Claude Sonnet (and almost everything else)* 384 experts in the mixture (the scale here is bonkers)* 128K context window standard, with rumors of 2M+ capability* Trained on 15.5 trillion tokens with the new Muon optimizerThe main thing about the SWE-bench score is not even just the incredible performance, it's the performance without thinking/reasoning + price! The Muon MagicHere's where it gets really interesting for the ML nerds among us. These folks didn't use AdamW - they used a new optimizer called Muon (with their own Muon Clip variant). Why does this matter? They trained to 15.5 trillion tokens with ZERO loss spikes. That beautiful loss curve had everyone in our community slack channels going absolutely wild. As Yam explained during the show, claiming you have a better optimizer than AdamW is like saying you've cured cancer - everyone says it, nobody delivers. Well, Moonshot just delivered at 1 trillion parameter scale.Why This Changes EverythingThis isn't just another model release. This is "Sonnet at home" if you have the hardware. But more importantly:* Modified MIT license (actually open!)* 5x cheaper than proprietary alternatives* Base model released (the first time we get a base model this powerful)* Already has Anthropic-compatible API (they knew what they were doing)The vibes are OFF THE CHARTS. Every high-taste model tester I know is saying this is the best open source model they've ever used. It doesn't have that "open source smell" - it feels like a frontier model because it IS a frontier model.Not only a math geniusImportantly, this model is great at multiple things, as folks called out it's personality or writing style specifically! Our Friend Sam Paech, creator of EQBench, has noted that this is maybe the first time an open source model writes this well, and is in fact SOTA on his Creative Writing benchmark and EQBench! Quick ShoutoutsBefore we dive deeper, huge props to:* Teknium for dropping the Hermes 3 dataset (nearly 1M high-quality entries!) (X)* LG (yes, the fridge company) for EXAONE 4.0 - their 32B model getting 81.8% on MMLU Pro is no joke (X)🎉 This Week's Buzz: W&B Inference Goes Live with Kimi-K2! (X)Ok, but what if you want to try Kimi-K2 but don't have the ability to run 1T models willy nilly? Well, Folks, I've been waiting TWO AND A HALF YEARS to say this: We're no longer GPU poor!Weights & Biases + CoreWeave = Your new inference playground. We launched Kimi K2 on our infrastructure within 3 days of release! Sitting behind the scenes on this launch was surreal - as I've been covering all the other inference service launches, I knew exactly what we all want, fast inference, full non-quantized weights, OpenAI API compatibility, great playground to test it out, function calling and tool use. And we've gotten almost all of these, while the super cracked CoreWeave and W&B Weave teams worked their ass off over the weekend to get this shipped in just a few days! And here’s the kicker: I’m giving away $50 in inference credits to 20 of you to try Kimi K2 on our platform. Just reply “K2-Koolaid-ThursdAI” to our X launch post here and we'll pick up to 20 winners with $50 worth of credits! 🫡It’s live now at api.inference.wandb.ai/v1 (...
    続きを読む 一部表示
    1 時間 45 分
  • 📆 ThursdAI - Jul 10 - Grok 4 and 4 Heavy, SmolLM3, Liquid LFM2, Reka Flash & Vision, Perplexity Comet Browser, Devstral 1.1 & More AI News

    📆 ThursdAI - Jul 10 - Grok 4 and 4 Heavy, SmolLM3, Liquid LFM2, Reka Flash & Vision, Perplexity Comet Browser, Devstral 1.1 & More AI News
    2025/07/11
    Hey everyone, Alex hereDon't you just love "new top LLM" drop weeks? I sure do! This week, we had a watch party for Grok-4, with over 20K tuning in to watch together, as the folks at XAI unveiled their newest and best model around. Two models in fact, Grok-4 and Grok-4 Heavy. We also had a very big open source week, we had the pleasure to chat with the creators of 3 open source models on the show, first with Elie from HuggingFace who just released SmoLM3, then with our friend Maxime Labonne who together with Liquid released a beautiful series of tiny on device models. Finally we had a chat with folks from Reka AI, and as they were on stage, someone in their org published a new open source Reka Flash model 👏 Talk about Breaking News right on the show! It was a very fun week and a great episode, so grab your favorite beverage and let me update you on everything that's going on in AI (as always, show notes at the end of the article) Open Source LLMsAs always, even on big weeks like this, we open the show with Open Source models first and this week, the western world caught up to the Chinese open source models we saw last week! HuggingFace SmolLM3 - SOTA fully open 3B with dual reasoning and long-context (𝕏, HF)We had Eli Bakouch from Hugging Face on the show and you could feel the pride radiating through the webcam. SmolLM 3 isn’t just “another tiny model”; it’s an 11-trillion-token monster masquerading inside a 3-billion-parameter body. It reasons, it follows instructions, and it does both “think step-by-step” and “give me the answer straight” on demand. Hugging Face open-sourced every checkpoint, every dataset recipe, every graph in W&B – so if you ever wanted a fully reproducible, multi-lingual pocket assistant that fits on a single GPU, this is it.They achieved the long context (128 K today, 256 K in internal tests) with a NoPE + YaRN recipe and salvaged the performance drop by literally merging two fine-tunes at 2 a.m. the night before release. Science by duct-tape, but it works: SmolLM 3 edges out Llama-3.2-3B, challenges Qwen-3, and stays within arm’s reach of Gemma-3-4B – all while loading faster than you can say “model soup.” 🤯Liquid AI’s LFM2: Blazing-Fast Models for the Edge (𝕏, Hugging Face)We started the show and I immediately got to hit the #BREAKINGNEWS button, as Liquid AI dropped LFM2, a new series of tiny (350M-1.2B) models focused on Edge devices.We then had the pleasure to host our friend Maxime Labonne, head of Post Training at Liquid AI, to come and tell us all about this incredible effort! Maxime, a legend in the model merging community, explained that LFM2 was designed from the ground up for efficiency. They’re not just scaled-down big models; they feature a novel hybrid architecture with convolution and attention layers specifically optimized for running on CPUs and devices like the Samsung Galaxy S24.Maxime pointed out that Out of the box, they won't replace ChatGPT, but when you fine-tune them for a specific task like translation, they can match models 60 times their size. This is a game-changer for creating powerful, specialized agents that run locally. Definitely a great release and on ThursdAI of all days! Mistrals updated Devstral 1.1 Smashes Coding Benchmarks (𝕏, HF)Mistral didn't want to be left behind on this Open Source bonanza week, and also, today, dropped an update to their excellent coding model Devstral. With 2 versions, an open weights Small and API-only Medium model, they have claimed an amazing 61.6% score on Swe Bench and the open source Small gets a SOTA 53%, the highest among the open source models! 10 points higher than the excellent DeepSwe we covered just last week!The thing to watch here is the incredible price performance, with this model beating Gemini 2.5 Pro and Claude 3.7 Sonnet while being 8x cheaper to run! DevStral small comes to us with an Apache 2.0 license, which we always welcome from the great folks at Mistral! Big Companies LLMs and APIsThere's only 1 winner this week, it seems that other foundational labs were very quiet to see what XAI is going to release. XAI releases Grok-4 and Grok-4 heavy - the world leading reasoning model (𝕏, Try It) Wow, what a show! Space uncle Elon together with the XAI crew, came fashionably late to their own stream, and unveiled the youngest but smartest brother of the Grok family, Grok 4 plus a multiple agents swarm they call Grok Heavy. We had a watch party with over 25K viewers across all streams who joined and watched together, this, fairly historic event! Why historic? Well, for one, they have scaled RL (Reinforcement Learning) for this model significantly more than any other lab did so far, which resulted in an incredible reasoner, able to solve HLE (Humanity's Last Exam) benchmark at an unprecedented 50% (while using tools) The other very much unprecedented result, is on the ArcAGI benchmark, specifically V2, which is designed to be very easy for humans and ...
    続きを読む 一部表示
    1 時間 50 分
  • 📆 ThursdAI - Jul 3 - ERNIE 4.5, Hunyuan A13B, MAI-DxO outperforms doctors, RL beats SWE bench, Zuck MSL hiring spree & more AI news

    📆 ThursdAI - Jul 3 - ERNIE 4.5, Hunyuan A13B, MAI-DxO outperforms doctors, RL beats SWE bench, Zuck MSL hiring spree & more AI news
    2025/07/03
    Hey everyone, Alex here 👋Welcome back to another mind-blowing week on ThursdAI! We’re diving into the first show of the second half of 2025, and let me tell you, AI is not slowing down. This week, we’ve got a massive wave of open-source models from Chinese giants like Baidu and Tencent that are shaking up the game, Meta’s jaw-dropping hiring spree with Zuck assembling an AI dream team, and Microsoft’s medical AI outperforming doctors on the toughest cases. Plus, a real-time AI game engine that had me geeking out on stream. Buckle up, folks, because we’ve got a lot to unpack!ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.We had incredible guests like Michael Luo from Agentica, dropping knowledge on RL coding agents, and Ivan Burazin from Daytona, revealing the infrastructure powering the agent era. We had an incredible episode this week, with over 8,000 views for the live show (as always, Links and Show notes in the end, and the YT live video is here for your convienience if you'd prefer watching) Open Source AI & LLMs: The Chinese Powerhouse WaveMan, if there’s one takeaway from this week, it’s that Chinese companies are absolutely dominating the open-source LLM scene. Let’s break down the heavy hitters that dropped this week and why they’ve got everyone talking.Baidu’s ERNIE 4.5: A Suite of 10 Models to Rule Them AllBaidu, a giant in the Chinese tech space, just flipped the script by open-sourcing their ERNIE 4.5 series. We’re talking 10 distinct models ranging from a whopping 424 billion parameters down to a tiny 0.3 billion. With an Apache 2.0 license, 128K context window, and multimodal capabilities handling image, video, and text input, this is a massive drop. Their biggest Mixture-of-Experts (MoE) model, with 47B active parameters, even outshines OpenAI’s o1 on visual knowledge tasks like DocVQA, scoring 93% compared to o1’s 81%! What’s wild to me is Baidu’s shift. They’ve been running ERNIE in production for years—think chatbots and more across their ecosystem—but they weren’t always open-source fans. Now, they’re not just joining the party, they’re hosting it. If you’re into tinkering, this is your playground—check it out on Hugging Face (HF) or dive into their technical paper (Paper).Tencent’s Hunyuan-A13B-Instruct: WizardLM Team Strikes AgainNext up, Tencent dropped Hunyuan-A13B-Instruct, and oh boy, does it have a backstory. This 80B parameter MoE model (13B active at inference) comes from the legendary WizardLM team, poached from Microsoft after a messy saga where their killer models got yanked from the internet over “safety concerns.” I remember the frustration—we were all hyped, then bam, gone. Now, under Tencent’s wing, they’ve cooked up a model with a 256K context window, hybrid fast-and-slow reasoning modes, and benchmarks that rival DeepSeek R1 and OpenAI o1 on agentic tasks. It scores an impressive 87% on AIME 2024, though it dips to 76% on 2025, hinting at some overfitting quirks. Though for a 12B active parameters model this all is still VERY impressive.Here’s the catch—the license. It excludes commercial use in the EU, UK, and South Korea, and bans usage if you’ve got over 100M active users. So, not as open as we’d like, but for its size, it’s a beast that fits on a single machine, making it a practical choice for many. They’ve also released two datasets, ArtifactsBench and C3-Bench, for code and agent evaluation. I’m not sold on the name—Hunyuan doesn’t roll off the tongue for Western markets—but the WizardLM pedigree means it’s worth a look. Try it out on Hugging Face (HF) or test it directly (Try It).Huawei’s Pangu Pro MoE: Sidestepping Sanctions with Ascend NPUsHuawei entered the fray with Pangu Pro MoE, a 72B parameter model with 16B active per token, and here’s what got me hyped—it’s trained entirely on their own Ascend NPUs, not Nvidia or AMD hardware. This is a bold move to bypass US sanctions, using 4,000 of these chips to preprocess 13 trillion tokens. The result? Up to 1,528 tokens per second per card with speculative decoding, outpacing dense models in speed and cost-efficiency. Performance-wise, it’s close to DeepSeek and Qwen, making it a contender for those outside the Nvidia ecosystem.I’m intrigued by the geopolitical angle here. Huawei’s proving you don’t need Western tech to build frontier models, and while we don’t know who’s got access to these Ascend NPUs, it’s likely a game-changer for Chinese firms. Licensing isn’t as permissive as MIT or Apache, but it’s still open-weight. Peek at it on Hugging Face (HF) for more details.DeepSWE-Preview: RL Coding Agent Hits 59% on SWE-BenchSwitching gears, I was blown away chatting with Michael Luo from Agentica about DeepSWE-Preview, an open-source coding agent trained with ...
    続きを読む 一部表示
    1 時間 36 分

ThursdAI - The top AI news from the past weekに寄せられたリスナーの声

カスタマーレビュー：以下のタブを選択することで、他のサイトのレビューをご覧になれます。

Audible.co.jp

Amazon.co.jp
レビューはまだありません。