ThursdAI - The top AI news from the past week

By: 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
Listen for free

  • Summary

  • 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.news
    Alex Volkov
    Show More Show Less
Politics & Government
Show More Show Less
Episodes
  • 📆 ThursdAI - Jan 30 - DeepSeek vs. Nasdaq, R1 everywhere, Qwen Max & Video, Open Source SUNO, Goose agents & more AI news

    📆 ThursdAI - Jan 30 - DeepSeek vs. Nasdaq, R1 everywhere, Qwen Max & Video, Open Source SUNO, Goose agents & more AI news
    Jan 30 2025
    Hey folks, Alex here 👋It’s official—grandmas (and the entire stock market) now know about DeepSeek. If you’ve been living under an AI rock, DeepSeek’s new R1 model just set the world on fire, rattling Wall Street (causing the biggest monetary loss for any company, ever!) and rocketing to #1 on the iOS App Store. This week’s ThursdAI show took us on a deep (pun intended) dive into the dizzying whirlwind of open-source AI breakthroughs, agentic mayhem, and big-company cat-and-mouse announcements. Grab your coffee (or your winter survival kit if you’re in Canada), because in true ThursdAI fashion, we’ve got at least a dozen bombshells to cover—everything from brand-new Mistral to next-gen vision models, new voice synthesis wonders, and big moves from Meta and OpenAI.We’re also talking “reasoning mania,” as the entire industry scrambles to replicate, dethrone, or ride the coattails of the new open-source champion, R1. So buckle up—because if the last few days are any indication, 2025 is officially the Year of Reasoning (and quite possibly, the Year of Agents, or both!)Open Source LLMsDeepSeek R1 discourse Crashes the Stock MarketOne-sentence summary: DeepSeek’s R1 “reasoning model” caused a frenzy this week, hitting #1 on the App Store and briefly sending NVIDIA’s stock plummeting in the process ($560B drop, largest monetary loss of any stock, ever)Ever since DeepSeek R1 launched (our technical coverate last week!), the buzz has been impossible to ignore—everyone from your mom to your local barista has heard the name. The speculation? DeepSeek’s new architecture apparently only cost $5.5 million to train, fueling the notion that high-level AI might be cheaper than Big Tech claims. Suddenly, people wondered if GPU manufacturers like NVIDIA might see shrinking demand, and the stock indeed took a short-lived 17% tumble. On the show, I joked, “My mom knows about DeepSeek—your grandma probably knows about it, too,” underscoring just how mainstream the hype has become.Not everyone is convinced the cost claims are accurate. Even Dario Amodei of Anthropic weighed in with a blog post arguing that DeepSeek’s success increases the case for stricter AI export controls. Public Reactions* Dario Amodei’s blogIn “On DeepSeek and Export Controls,” Amodei argues that DeepSeek’s efficient scaling exemplifies why democratic nations need to maintain a strategic leadership edge—and enforce export controls on advanced AI chips. He sees Chinese breakthroughs as proof that AI competition is global and intense.* OpenAI Distillation EvidenceOpenAI mentioned it found “distillation traces” of GPT-4 inside R1’s training data. Hypocrisy or fair game? On ThursdAI, the panel mused that “everyone trains on everything,” so perhaps it’s a moot point.* Microsoft ReactionMicrosoft wasted no time, swiftly adding DeepSeek to Azure—further proof that corporations want to harness R1’s reasoning power, no matter where it originated.* Government reactedEven officials in the government, David Sacks, US incoming AI & Crypto czar, discussed the fact that DeepSeek did "distillation" using the term somewhat incorrectly, and presidet Trump was asked about it.* API OutagesDeepSeek’s own API has gone in and out this week, apparently hammered by demand (and possibly DDoS attacks). Meanwhile, GPU clouds like Groq are showing up to accelerate R1 at 300 tokens/second, for those who must have it right now.We've seen so many bad takes on the topic, from seething cope takes, to just gross misunderstandings from gov officials confusing the ios App with the OSS models, folks throwing conspiracy theories into the mix, claiming that $5.5M sum was a PsyOp. The fact of the matter is, DeepSeek R1 is an incredible model, and is now powering (just a week later), multiple products (more on this below) and experiences already, while pushing everyone else to compete (and give us reasoning models!)Open Thoughts Reasoning DatasetOne-sentence summary: A community-led effort, “Open Thoughts,” released a new large-scale dataset (OpenThoughts-114k) of chain-of-thought reasoning data, fueling the open-source drive toward better reasoning models.Worried about having enough labeled “thinking” steps to train your own reasoner? Fear not. The OpenThoughts-114k dataset aggregates chain-of-thought prompts and responses—114,000 of them—for building or fine-tuning reasoning LLMs. It’s now on Hugging Face for your experimentation pleasure. The ThursdAI panel pointed out how crucial these large, openly available reasoning datasets are. As Wolfram put it, “We can’t rely on the big labs alone. More open data means more replicable breakouts like DeepSeek R1.”Mistral Small 2501 (24B)One-sentence summary: Mistral AI returns to the open-source spotlight with a 24B model that fits on a single 4090, scoring over 81% on MMLU while under Apache 2.0.Long rumored to be “going more closed,” Mistral AI ...
    Show More Show Less
    1 hr and 55 mins
  • 📆 ThursdAI - Jan 23, 2025 - 🔥 DeepSeek R1 is HERE, OpenAI Operator Agent, $500B AI manhattan project, ByteDance UI-Tars, new Gemini Thinker & more AI news

    📆 ThursdAI - Jan 23, 2025 - 🔥 DeepSeek R1 is HERE, OpenAI Operator Agent, $500B AI manhattan project, ByteDance UI-Tars, new Gemini Thinker & more AI news
    Jan 24 2025
    What a week, folks, what a week! Buckle up, because ThursdAI just dropped, and this one's a doozy. We're talking seismic shifts in the open source world, a potential game-changer from DeepSeek AI that's got everyone buzzing, and oh yeah, just a casual $500 BILLION infrastructure project announcement. Plus, OpenAI finally pulled the trigger on "Operator," their agentic browser thingy – though getting it to actually operate proved to be a bit of a live show adventure, as you'll hear. This week felt like one of those pivotal moments in AI, a real before-and-after kind of thing. DeepSeek's R1 hit the open source scene like a supernova, and suddenly, top-tier reasoning power is within reach for anyone with a Mac and a dream. And then there's OpenAI's Operator, promising to finally bridge the gap between chat and action. Did it live up to the hype? Well, let's just say things got interesting.As I’m writing this, White House just published that an Executive Order on AI was just signed and published as well, what a WEEK.Open Source AI Goes Nuclear: DeepSeek R1 is HERE!Hold onto your hats, open source AI just went supernova! This week, the Chinese Whale Bros – DeepSeek AI, that quant trading firm turned AI powerhouse – dropped a bomb on the community in the best way possible: R1, their reasoning model, is now open source under the MIT license! As I said on the show, "Open source AI has never been as hot as this week."This isn't just a model, folks. DeepSeek unleashed a whole arsenal: two full-fat R1 models (DeepSeek R1 and DeepSeek R1-Zero), and a whopping six distilled finetunes based on Qwen (1.5B, 7B, 14B, and 32B) and Llama (8B, 72B). One stat that blew my mind, and Nisten's for that matter, is that DeepSeek-R1-Distill-Qwen-1.5B, the tiny 1.5 billion parameter model, is outperforming GPT-4o and Claude-3.5-Sonnet on math benchmarks! "This 1.5 billion parameter model that now does this. It's absolutely insane," I exclaimed on the show. We're talking 28.9% on AIME and 83.9% on MATH. Let that sink in. A model you can probably run on your phone is schooling the big boys in math.License-wise, it's MIT, which as Nisten put it, "MIT is like a jailbreak to the whole legal system, pretty much. That's what most people don't realize. It's like, this is, it's not my problem. You're a problem now." Basically, do whatever you want with it. Distill it, fine-tune it, build Skynet – it's all fair game.And the vibes? "Vibes are insane," as I mentioned on the show. Early benchmarks are showing R1 models trading blows with o1-preview and o1-mini, and even nipping at the heels of the full-fat o1 in some areas. Check out these numbers:And the price? Forget about it. We're talking 50x cheaper than o1 currently. DeepSeek R1 API is priced at $0.14 / 1M input tokens and $2.19 / 1M output tokens, compared to OpenAI's o1 at $15.00 / 1M input and a whopping $60.00 / 1M output. Suddenly, high-quality reasoning is democratized.LDJ highlighted the "aha moment" in DeepSeek's paper, where they talk about how reinforcement learning enabled the model to re-evaluate its approach and "think more." It seems like simple RL scaling, combined with a focus on reasoning, is the secret sauce. No fancy Monte Carlo Tree Search needed, apparently!But the real magic of open source is what the community does with it. Pietro Schirano joined us to talk about his "Retrieval Augmented Thinking" (RAT) approach, where he extracts the thinking process from R1 and transplants it to other models. "And what I found out is actually by doing so, you may even like smaller, quote unquote, you know, less intelligent model actually become smarter," Pietro explained. Frankenstein models, anyone? (John Lindquist has a tutorial on how to do it here)And then there's the genius hack from Voooogel, who figured out how to emulate a "reasoning_effort" knob by simply replacing the "end" token with "Wait, but". "This tricks the model into keeps thinking," as I described it. Want your AI to really ponder the meaning of life (or just 1+1)? Now you can, thanks to open source tinkering.Georgi Gerganov, the legend behind llama.cpp, even jumped in with a two-line snippet to enable speculative decoding, boosting inference speeds on the 32B model on my Macbook from a sluggish 5 tokens per second to a much more respectable 10-11 tokens per second. Open source collaboration at its finest and it's only going to get better! Thinking like a NeuroticMany people really loved the way R1 thinks, and what I found astonishing is that I just sent "hey" and the thinking went into a whole 5 paragraph debate of how to answer, a user on X answered with "this is Woody Allen-level of Neurotic" which... nerd sniped me so hard! I used Hauio Audio (which is great!) and ByteDance latentSync and gave R1 a voice! It's really something when you hear it's inner monologue being spoken out like this! ByteDance Enters the Ring: UI-TARS Controls Your PCNot to be outdone in the open source frenzy, ByteDance, ...
    Show More Show Less
    1 hr and 50 mins
  • 📆 ThursdAI - Jan 16, 2025 - Hailuo 4M context LLM, SOTA TTS in browser, OpenHands interview & more AI news

    📆 ThursdAI - Jan 16, 2025 - Hailuo 4M context LLM, SOTA TTS in browser, OpenHands interview & more AI news
    Jan 17 2025
    Hey everyone, Alex here 👋 Welcome back, to an absolute banger of a week in AI releases, highlighted with just massive Open Source AI push. We're talking a MASSIVE 4M context window context window model from Hailuo (remember when a jump from 4K to 16K seemed like a big deal?), a 8B omni model that lets you livestream video and glimpses of Agentic ChatGPT? This week's ThursdAI was jam-packed with so much open source goodness that the big companies were practically silent. But don't worry, we still managed to squeeze in some updates from OpenAI and Mistral, along with a fascinating new paper from Sakana AI on self-adaptive LLMs. Plus, we had the incredible Graham Neubig, from All Hands AI, join us to talk about Open Hands (formerly OpenDevin) and even contributed to our free, LLM Evaluation course on Weights & Biases!Before we dive in, a friend asked me over dinner, what are the main 2 things that happened in AI in 2024, and this week highlights one of those trends. Most of the Open Source is now from China. This week, we got MiniMax from Hailuo, OpenBMB with a new MiniCPM, InternLM came back and most of the rest were Qwen finetunes. Not to mention DeepSeek. Wanted to highlight this significant narrative change and that this is being done despite the chip export restrictions. 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 AI & LLMsMiniMax-01: 4 Million Context, 456 Billion Parameters, and Lightning Attention This came absolutely from the left field, given that we've seen no prior LLMs from Haulio, the company previously releasing video models with consistent characters. Dropping a massive 456B mixture of experts model (45B active parameters) with such a long context support in open weights, but also with very significant benchmarks that compete with Gpt-4o, Claude and DeekSeek v3 (75.7 MMLU-pro, 89 IFEval, 54.4 GPQA)They have trained the model on up to 1M context window and then extended it to 4M with ROPE scaling methods (our coverage of RoPE) during Inference. MiniMax-Text-01 adopts a hybrid architecture that combines Lightning Attention, Softmax Attention and Mixture-of-Experts (MoE) with 45B active parameters. I gotta say, when we started talking about context window, imagining a needle in a haystack graph that shows 4M, in the open source seemed far fetched, though we did say that theoretically, there may not be a limit to context windows. I just always expected that limit to be unlocked by transformers alternative architectures like Mamba or other State Space Models.Vision, API and Browsing - Minimax-VL-01It feels like such a well rounded and complete release, that it highlights just how mature company that is behind it. They have also released a vision version of this model, that includes a 300M param Vision Transformer on top (trained with 512B vision language tokens) that features dynamic resolution and boasts very high DocVQA and ChartQA scores. Not only did these two models were released in open weights, they also launched as a unified API endpoint (supporting up to 1M tokens) and it's cheap! $0.2/1M input and $1.1/1M output tokens! AFAIK this is only the 3rd API that supports this much context, after Gemini at 2M and Qwen Turbo that supports 1M as well.Surprising web browsing capabilitiesYou can play around with the model on their website, hailuo.ai which also includes web grounding, which I found quite surprising to find out, that they are beating chatGPT and Perplexity on how fast they can find information that just happened that same day! Not sure what search API they are using under the hood but they are very quick. 8B chat with video model omni-model from OpenBMBOpenBMB has been around for a while and we've seen consistently great updates from them on the MiniCPM front, but this one takes the cake! This is a complete omni modal end to end model, that does video streaming, audio to audio and text understanding, all on a model that can run on an iPad! They have a demo interface that is very similar to the chatGPT demo from spring of last year, and allows you to stream your webcam and talk to the model, but this is just an 8B parameter model we're talking about! It's bonkers! They are boasting some incredible numbers, and to be honest, I highly doubt their methodology in textual understanding, because, well, based on my experience alone, this model understands less than close to chatGPT advanced voice mode, but miniCPM has been doing great visual understanding for a while, so ChartQA and DocVQA are close to SOTA. But all of this doesn't matter, because, I say again, just a little over a year ago, Google released a video announcing these capabilities, having an AI react to a video in real time, and it absolutely blew everyone away, and it was FAKED. And this time a year after, we have these capabilities, essentially, in an 8B model that runs on ...
    Show More Show Less
    1 hr and 41 mins

What listeners say about ThursdAI - The top AI news from the past week

Average customer ratings

Reviews - Please select the tabs below to change the source of reviews.

Audible.co.uk reviews

Amazon reviews
No Reviews are Available