[LIVE] Anthropic Distillation & How Models Cheat (SWE-Bench Dead) | Nathan Lambert & Sebastian Rasc Latent Space 2026-09-20 | [LIVE] Anthropic Distillation & How Models Cheat (SWE-Bench Dead) | Nathan Lambert & Sebastian Rasc
Beyond AlphaGo Latent Space 2025-07-23 | The hidden flaws of minimax training, the myth of generalizable strategy, and designing AI for real-world complexity
AlphaZero vs. ChatGPT Latent Space 2025-07-22 | What Strategic Games Taught Us About AI — and Why It Doesn’t Translate to Language
AI Agents That Pay for Their Own Tools? The Future of MCP Monetization Latent Space 2025-07-22 | Wild vision from Cline: agents that pay themselves for MCP tools 🤯Instead of juggling API keys for every service, imagine a unified payment layer where AI agents automatically handle transactions for each tool callStripe is exploring MCP payment abstractions, some are looking at stablecoins, others at HTTP 402 payment requestsStill early days but the infrastructure is taking shape. The future of AI commerce might be more autonomous than we thinkWhat's your take on self-paying agents?
Groks Secret Strategy: Why Theyre Holding Back Their Coding Model Latent Space 2025-07-22 | Grok's playing chess while others play checkers 🧠Word is they're holding back their coding model until it can beat a specific competitor. Strategic patience or missed opportunity?The real story: how X built a competitive AI from basically nothing. Data sourcing beyond Twitter, massive compute investment, and the right team.Most interesting part? Some AGI labs are frustrated with performance while others are celebrating. Same resources, different execution.The AI race just got more interesting 🚀
The Gathering Breaks AI Latent Space 2025-07-22 | Noam explains why Magic: The Gathering is so hard for AI—and what it teaches us about hidden information and reasoning in games
Why AI Will Replace Virtual Assistants Latent Space 2025-07-22 | Why AI assistants could outperform human virtual assistants—not just in speed, but in alignment with your goals and values.
MCPs Are Killing SaaS Products: Why Companies Are Cannibalizing Themselves Latent Space 2025-07-21 | MCPs are creating a fascinating paradox in the AI tooling spaceCompanies are literally cannibalizing their own paid products by offering free MCP integrationsExample: Sentry launched a paid AI agent for issue resolution, but their free MCP lets you do the same thing through Claude/other coding agentsThis raises huge questions about monetization strategies in the AI eraAre we heading toward MCP subscription models? Will companies fracture their pricing across different access methods?The economics don't make sense yet, but the technical capabilities are incredibleEarly adopters who understood MCP's potential are now seeing this play out in real-time
Does AI Need Intuition First? Latent Space 2025-07-21 | From Pokémon battles to chess, Noam dives into how much intuition (System 1) an AI needs before it can think deeply (System 2) about a game.
Arc Prizes Greg Kamradt: Reality is the Ultimate Eval Engine Latent Space 2025-07-21 | Greg Kamradt from Arc Prize drops a profound insight: "Reality is the ultimate eval engine" 🎯His take on AGI benchmarks is fascinating - Arc Prize doesn't aim for PhD-level impossible problems. Their constraint? Can humans do this?As long as they can create problems humans solve but AI can't, we don't have AGI yet.Arc 2 is still uncracked. Arc 3 coming 2026.Grok 4 hit 16% on Arc - everyone's celebrating but we're still far from the finish line.The path to AGI isn't about the hardest possible problems - it's about human-solvable problems that stump our best AI systems.
Why Small AI Models Can’t Reason Latent Space 2025-07-21 | What does it take for AI to "think" like a human? Noam explains why smaller models can't reason and why larger models unlock System 2 thinking
Why AI Models Arent Actually Intelligent (Yet) - The Physics Problem Latent Space 2025-07-20 | Humans RL on physics - the ultimate eval engine. But when AI uses manufactured environments, we're just injecting developer intelligence into the system.The real test? Interactive benchmarks where AI has to intuit goals, make plans, and adapt to novel situations - just like humans do.That's why static benchmarks won't declare AGI. We need environments that test true generalization, not memorized patterns.The shift from static to interactive evaluation is where the real intelligence emerges.
Why AI Inference Will Never Be Profitable (Cline founder Explains) Latent Space 2025-07-20 | 🔥 HOT TAKE: "Inference isn't a business" - Cline founder drops truth bombsKey insights from this clip:• Model costs trending toward zero (Gemini going free?)• Price transparency ) obfuscation • Developers need visibility into costs & models• ROI on coding agents is "almost hard to calculate"• Usage-based plans work when users trust the pricingThe real value isn't in running the models - it's in the tooling, transparency, and enabling developers to build things they never would have attempted before.Thoughts on inference commoditization? 👇
Why AI Safety Experts Loved Cicero Latent Space 2025-07-20 | Noam explains why Cicero was praised by the AI safety community and how steerability and controllability in AI systems could be a major step toward safer AI.
Microsoft Paint Inside a Notebook?! This AI Demo Will Blow Your Mind Latent Space 2025-07-20 | Microsoft Paint inside a Jupyter notebook? 🎨Akshay Agarwal just blew my mind with this Marimo demo. Draw something in MS Paint, it renders in real-time, then gets fed to a multimodal AI model that converts your drawing into a mermaid diagram.This is the kind of reactive notebook experience that makes you completely rethink what's possible with data workflows. The fact that you can chain together drawing → AI analysis → code generation seamlessly is wild.Marimo's execution model based on variable dependencies (like Excel) + wide screen layouts = notebooks that actually feel modern.The AI-native capabilities for model calls are just the cherry on top.
Why C++ Sucks for AI Programming Latent Space 2025-07-20 | Why build a new programming language when everyone already uses Python and C++? Chris Lattner explains the origin of Mojo and why C++ "sucks" for AI workloads.
Interactive Data Selection: Marimo vs Jupyter - Game Changer for Data Scientists Latent Space 2025-07-19 | The difference between Jupyter and Marimo is wild 🤯With Marimo, you can literally drag-select data points in a scatter plot and instantly see the underlying images. No complex setup, no painful workarounds - just pure interactivity.This is what modern data exploration should look like. The notebook reacts to your UI interactions in real-time, sending selected data back to Python automatically.Best part? Converting from Jupyter is as simple as copy-paste or using their CLI tool.Data scientists are about to have their minds blown 🚀
Max vs VLLM: What Makes Max Better? Latent Space 2025-07-19 | Chris Lattner breaks down what makes Modular’s Max engine radically different—and why it's faster, lighter, and more open than VLLM.
Cline AI Founder: Agent Memories Are Mostly Useless - The Surprising Truth About AI Code Agents Latent Space 2025-07-19 | Cline co-founder just dropped a reality check that's gonna shake up how we think about AI agents 🤯His take: "Agent memories are mostly useless"Why? Because forcing agents to remember specific project quirks or team workflows doesn't actually work the way we think it does. People don't want to manually manage what their AI remembers.The real insight: Those coding rules and team preferences? Better to put them in guidelines files than rely on agent memory.But here's the kicker - how do we capture that "tribal knowledge" that agents pick up naturally without making users think about it?This is the kind of nuanced AI engineering insight that separates the builders from the hype followers 💡
Scaffolds Are All Going To Die — Noam Browns Spicy AGI Prediction Latent Space 2025-07-19 | Noam Brown just dropped a spicy take: "Scaffolds are all going to die" 🔥While most AI engineers are building complex agent frameworks with memory systems, compression layers, and elaborate harnesses, Brown believes we're heading toward "thin agents" that won't need all this scaffolding.But here's the thing - the jury's still out. Current agents need scaffolds to handle memory, context compression, and multi-step reasoning. Will future models just absorb all of this into pure neural computation?The debate between thick vs thin agents is heating up, and honestly, I could argue both sides.What's your take? Are we overengineering agent architectures or building the foundation for true AGI?
Python Speed Without the Pain Latent Space 2025-07-19 | Mojo can replace Python for performance-critical code—without bindings or rewrites. Chris shares how Mojo makes Rust-level speed accessible to Python devs.
Engineering With AI Teammates Latent Space 2025-07-19 | Deep diving into next-gen development environments, tool-calling evolution, and how Zed is pioneering the integration of LLMs as first-class collaborators
Rust-Powered Editing at 120fps Latent Space 2025-07-19 | Inside Zed's architecture: game-engine rendering, CRDTs for collaboration, and why web tech couldn’t go the distance
Why Clines Founder Rejected $3B to Fork VS Code (The Hidden Pain Microsoft Creates) Latent Space 2025-07-18 | Why forking VS Code isn't the $3B opportunity you think it isCline's founder breaks down the brutal reality of maintaining VS Code forks - Microsoft makes it "notoriously difficult" with constant updates, API changes, and backend improvements that require reverse engineeringMeanwhile, building as an extension gives you:• Better distribution across editors• Less maintenance overhead• Focus on core product vs infrastructureSometimes the "obvious" path is a trap. Smart founders know when to build on top vs rebuild from scratch.
From Code Editors to Code Environments Latent Space 2025-07-18 | The evolution of developer tools through AI curation, multi-buffer workflows, and human-centered agent design
Don’t Reinvent—Remix Smart Ideas Latent Space 2025-07-18 | Chris Lattner explains how Modular decides what to build vs. what to adopt—and why composability and clean architecture win every time.
Code Without Leaving Your Tab: AI-Powered Analysis in Marimo Latent Space 2025-07-18 | This is exactly what we need more of in data science tooling 🔥@akshayagarwal shows how Marimo lets you "vibe-code" your entire analysis without ever leaving your notebook. The AI sees your variables, database connections, AND column names in real-time.No more copy-pasting to ChatGPT. No more context switching. Just pure flow state coding.The future of data analysis is conversational and contextual. This is it.
Why AI Language Learning Cant Handle Accents Yet (But Should) Latent Space 2025-07-18 | The psychological barrier to speaking a new language is HUGEPeople won't even practice pronunciation in front of teachers they're payingBut here's the thing - communication and pronunciation are almost completely separate skillsYou can have terrible pronunciation and still communicate effectivelyThe key? Just move your mouth and make sounds. Don't worry about perfectionAI tutors are removing that human judgment barrier that stops people from practicingInteresting insight from Andrew Struggs on language learning psychology
⚡️ARC-AGI-3: The Interactive Reasoning Benchmark Latent Space 2025-07-18 | Greg Kamradt, President @ ARC Prize Foundation, joins us for a sneak preview of the ARC AGI 3 benchmark, how they are approaching game design for AI, and why this is a better benchmark than traditional knowledge-based ones. The preview challenge of the new benchmark has a $10,000 prizOfficial website: three.arcprize.orgX: https://x.com/latentspacepodSubstack: https://latent.spaceChapters00:00:00 Introduction00:00:47 Background on ARC AGI and the Benchmark00:02:16 Defining Intelligence and Benchmark Philosophy00:04:20 Challenges in Measuring Intelligence & Generalization00:07:05 Introducing ARC AGI V3 and Interactive Games00:08:26 Game Demo: Locksmith & Mechanics00:12:30 Game Design, API, and Agent Competition00:23:32 Team Structure and Building the Benchmark00:24:43 Future Roadmap for ARC AGI Benchmarks00:27:05 Durability, Timelines & Benchmark Evolution00:29:17 The Role of ARC Prize in AGI Progress00:32:05 Grok 4 Livestream & Model Evaluation00:36:10 OpenAI, Competition, and the State of AI Labs00:38:47 Closing Remarks & Call to Action
DeepSeek’s GPU Hack Changed Everything Latent Space 2025-07-18 | The DeepSeek team shocked the world by going deeper than CUDA—and open sourcing it. Chris Lattner reacts to the moment that woke up the AI industry
⚡️The Future of Notebooks - with Akshay Agrawal of Marimo Latent Space 2025-07-18 | Akshay Agrawal joins us to talk about Marimo and their vision for the future of Python notebooks, and how it’s the perfect canvas for AI-driven data analysis.0:00 Introduction0:46 Overview of Marimo and Its Features2:33 Origin Story and Motivation Behind Marimo4:26 Demo: Classical Machine Learning with MNIST in Marimo6:52 Notebook Compatibility and Conversion from Jupyter7:42 Demo: Interactive Notebook with Custom UI and Layout10:08 AI-Native Utilities and Code Generation with Language Models11:36 Dependency Management and Integration with UV Package Manager13:00 Demo: Data Annotation Workflow Using a PS5 Controller15:51 Starting from Scratch: Blank Canvas AI Use Cases18:27 Context Formatting for AI Code Generation19:54 Chat Interface and Local/Remote Model Support21:01 WebAssembly Support and MoLab Cloud-Hosted Notebooks23:21 Future Plans and Breaking Out of Old Notebook Habits25:40 Running Marimo Notebooks as Scripts or Data Apps26:44 Exploring AI Agents and Community Contributions26:56 Call to Action: How to Get Started and Contribute
AI Understood What 4 Humans Couldnt - The Whisper Breakthrough That Changed Everything Latent Space 2025-07-17 | That moment when AI surpassed human ability wasn't gradual - it was instant.Andrew Hsu describes the exact moment his team realized speech AI had become superhuman. They played audio of a Korean English learner speaking. Four humans in the room closed their eyes and couldn't understand a word.Whisper got it perfectly.This was 2022. The pieces were finally clicking together - advanced speech models, ChatGPT launching, GPT-3.5 turbo dropping. Everything they'd predicted years earlier was suddenly real.The most interesting part? They saw it coming but were still shocked when it actually happened.
AI Engineer: Fast-Apply Models Will Be DEAD in 3 Months (Heres Why) Latent Space 2025-07-17 | Klein AI Engineer just dropped a reality check that's going to shake up the entire AI development spaceFast-apply models? Dead in 3 months, maybe lessHere's what's happening: The founders of companies built entirely around fine-tuning fast-apply models are having candid conversations about their own expiration dateCursor opened this window in July, but it's closing fastWhy? Models are getting dramatically better at large context and search/replace operations. RAG and FastApply were bandaids for when models couldn't handle these tasks wellNow they're just extra ingredients that can make things go wrongThe writing is on the wall - we're moving beyond these intermediate solutions faster than anyone expectedWhat does this mean for AI engineers? Time to adapt or get left behind
This Founders Business Model Answer Will Shock You (Open Source + Enterprise) Latent Space 2025-07-17 | Wild moment when a founder gets asked "how do you make money?" and just throws the question right back 😅The real answer? Enterprise adoption through open source trustWhen companies can control their data, choose their providers, and avoid sending code to unknown servers - that's when they're willing to payData privacy isn't just a feature anymore, it's the business model#AI #OpenSource #Enterprise #DataPrivacy #Startups
CTO: GPT-4 Changed Nothing Outside Silicon Valley - The Reality Check AI Needs Latent Space 2025-07-16 | CTO Andrew Struggs drops a reality check that'll make you think twice about AI hype."Outside the Bay Area, GPT-4 changed *nothing*" - and he's not wrong. While we're all talking about AI revolution, most people's daily lives remain unchanged.The real insight? We need more builders creating actual applications, not just talking about potential. Real world inertia is massive.This is exactly what we want though - slow takeoff gives us time to adapt. The tech is here, but the transformation takes time.What's your take? Are you seeing AI impact in your daily work or is it still mostly hype where you are?
Cline: the open source coding agent that doesnt cut costs Latent Space 2025-07-16 | Saoud Rizwan and Pash from Cline joined us to talk about why fast apply models got bitter lesson'd, how they pioneered the plan + act paradigm for coding, and why non-technical people use IDEs to do marketing and generate slides.Full writeup: https://www.latent.space/p/clineX: https://x.com/latentspacepodChapters:00:00 - Introductions 01:35 - Plan and Act Paradigm 05:37 - Model Evaluation and Early Development of Cline 08:14 - Use Cases of Cline Beyond Coding 09:09 - Why Cline is a VS Code Extension and Not a Fork 12:07 - Economic Value of Programming Agents 16:07 - Early Adoption for MCPs 19:35 - Local vs Remote MCP Servers 22:10 - Anthropic's Role in MCP Registry 22:49 - Most Popular MCPs and Their Use Cases 25:26 - Challenges and Future of MCP Monetization 27:32 - Security and Trust Issues with MCPs 28:56 - Alternative History Without MCP 29:43 - Market Positioning of Coding Agents and IDE Integration Matrix 32:57 - Visibility and Autonomy in Coding Agents 35:21 - Evolving Definition of Complexity in Programming Tasks 38:16 - Forks of Cline and Open Source Regrets 40:07 - Simplicity vs Complexity in Agent Design 46:33 - How Fast Apply Got Bitter Lesson'd 49:12 - Cline's Business Model and Bring-Your-Own-API-Key Approach 54:18 - Integration with OpenRouter and Enterprise Infrastructure 55:32 - Impact of Declining Model Costs 57:48 - Background Agents and Multi-Agent Systems 1:00:42 - Vision and Multi-Modalities 1:01:07 - State of Context Engineering 1:07:37 - Memory Systems in Coding Agents 1:10:14 - Standardizing Rules Files Across Agent Tools 1:11:16 - Cline's Personality and Anthropomorphization 1:12:55 - Hiring at Cline and Team CultureChapters00:00:00 Introduction and Guest Intros00:00:29 What is Klein? Product Overview00:01:42 Plan and Act Paradigm00:05:22 Model Evolution and Building Klein00:07:40 Beyond Coding: Klein as a General Agent00:09:12 Why Focus on VS Code Extension?00:11:26 The Future of Programming and Agentic Paradigm00:12:34 Economic Value: Programming vs. Other Use Cases00:16:04 MCP Ecosystem: Growth and Marketplace00:21:30 Security, Discoverability, and Trust in MCPs00:22:55 Popular MCPs and Workflow Automation00:25:30 Monetization and Payments for MCPs00:37:53 Competition, Forks, and Open Source Philosophy00:40:39 RAG, Fast Apply, and Agentic Simplicity00:50:11 Business Model and Enterprise Adoption00:57:04 Background Agents, Multi-Agent Systems, and CLI01:00:41 Context Engineering and Memory01:12:39 Team, Culture, and Closing Thoughts
The Second-Time Founders Latent Space 2025-07-16 | Quizzard to Unstuck AI: How leveraging institutional knowledge led to 1M users in 9 weeks, $6M ARR, and early profitability without major VC backing
Why AI Language Learning Apps Will Kill Duolingo Latent Space 2025-07-16 | The evolution of language learning is fascinating 🧠Gen 1: Rosetta Stone CDs at airportsGen 2: Mobile apps like Duolingo (gamified, casual)Gen 3: AI-native apps focused on functional fluencyAndrew Struggs breaks down how AI is transforming language learning - instead of vocab drills, it's about practicing real conversations with your Uber driver until speaking becomes automatic.The shift from gamification to practical fluency is huge. Language learning isn't just about points anymore - it's about actually being able to communicate naturally.What's your take on AI-powered language learning vs traditional methods?
Why AI Tool Calling Benchmarks Are Misleading (Galileo Labs Founder Explains) Latent Space 2025-07-15 | Reality check from @galileolabs: those shiny leaderboards showing "perfect" tool calling? They're lying.Prateek spent months finding where models actually break in real scenarios. Turns out they're still making critical mistakes that benchmarks miss.Time to recalibrate expectations. Models aren't as reliable as the numbers suggest.
The Two Minds of System Design Latent Space 2025-07-15 | Long-term strategy vs. spatial precision, and what Factorio teaches us about real-world complexity
Growth Loops in Metadata Latent Space 2025-07-15 | From visual rendering to conversion science — how teams are transforming RevenueCat into a growth experimentation engine
CTO: First-Token Latency Is a Vanity Metric - The Real Problem with AI Voice Apps Latent Space 2025-07-14 | First-token latency is a vanity metric.Speak CTO Andrew Struggs drops a reality check: everyone obsesses over how fast the first audio bytes come back, but they're missing the real bottleneck.The actual killer? Voice Activity Detection (VAD) - figuring out when someone stops talking. That can easily add another second if done poorly.Even worse for language learners who pause mid-sentence for 10+ seconds while thinking.The metric that actually matters: user stops talking → model starts responding.This is why building real-time AI feels so much harder than the benchmarks suggest. The devil is always in the details nobody talks about.
Architecting Intelligence Latent Space 2025-07-14 | Designing systems that reason, reflect, and remember—beyond prompt engineering
⚡️Ranking Agentic LLMs — Pratik Bhavsar, Galileo Latent Space 2025-07-14 | Grok 4 and Kimi K2 have upended the frontier LLM game last week. How do you know if you should swap your agent harness to use them?We recorded this pod with Pratik Bhavsar to discuss one of the leading approachest to Agent evals: huggingface.co/spaces/galileo-ai/agent-leaderboardread more:Chapters00:00:00 Introduction & Guest Welcome00:00:39 Shift from Model to Agent Evaluations00:02:02 Designing the Agent Leaderboard00:03:23 Methodology: Data Sets and Metrics00:06:42 Key Findings and Surprising Results00:11:42 Deep Dive on Benchmarks: TSQ, BFCO, XLAN, and Tau00:13:44 Tau Bench: Domain-Specific Evaluation00:16:38 LLM-as-Judge Metrics and Prompting00:23:59 Preview of Agent Leaderboard V200:26:41 Complexity in Realistic Evaluations00:30:33 Action Completion Metrics & Closing Thoughts
Building with Transformers Latent Space 2025-07-14 | How modern ML architectures, open ecosystems, and ethical questions are reshaping the future of software and research
Why AI Translation Will Never Replace Language Learning (Surprising Truth) Latent Space 2025-07-13 | Why won't real-time AI translation kill language learning?Andrew Struggs from Speak drops some fascinating insights:• Technical barriers are real - German verbs come at the end of sentences, creating unavoidable latency• But more importantly: people don't just want translation, they want CONNECTIONAsian users aren't looking for a translator - they want to look you in the eye and speak the same language. It's about becoming a better person and connecting with others.The human element always wins. Whether it's dating someone from Romania or learning Italian for your wife - language learning is deeply personal.Real-time translation will complement learning, not replace it.
![[LIVE] Anthropic Distillation & How Models Cheat (SWE-Bench Dead) | Nathan Lambert & Sebastian Rasc](https://i.ytimg.com/vi/7EBQ04OL-is/hqdefault.jpg)