• Join Simtheory: https://simtheory.ai
• "Is This The End" now on Spotify: https://open.spotify.com/album/2Py1MyADUFqJFVUISI2VTP?si=oT3PWyJYRA2BspOmzT_ifg
• Register for the STILL RELEVANT tour: https://simulationtheory.ai/16c0dationtheory.ai/16c0d1db-a8d0-4ac9-bae3-d25074589a80

Two new models dropped this week — Gemini 3.1 Pro and Claude Sonnet 4.6 — and honestly? We're struggling to care. In this episode, we break down why Gemini went from being our daily driver to a model we barely touch, the "tunnel vision" hallucination problem that killed the Gemini 3 series for us, and whether 3.1 Pro actually fixes it. We put Gemini 3.1 Pro head-to-head against Claude Opus building a Geoffrey Hinton Doom Center, debate whether anyone can actually tell the difference between Sonnet 4.5 and 4.6, and make the case that smaller models running in agentic loops are secretly beating the frontiers. Plus: OpenAI acquires OpenClaw and we ask why a $100B company couldn't just build it themselves, DHH calls out the AI pricing bubble, Mike compares AI models to cheap wine hangovers, and Sam Altman refuses to hold Dario's hand at the India AI Summit. The model wars are getting weird.

CHAPTERS:

0:00 Intro & "Is This The End" Now on Spotify
1:10 Gemini 3.1 Pro: Thinking Controls & The Medium Mode Fix
3:14 The Speed vs Intelligence Trade-Off in Agentic Work
5:10 Why Multitasking With AI Agents Made Us Anxious
6:34 Solid Updates: The Real Goal of Agentic Coding
7:45 Gemini's Fall From Grace: From Daily Driver to Dead Model
10:08 The Tunnel Vision Problem That Killed Gemini 3
13:35 Mixed Reactions: Fanboys vs Reality on Gemini 3.1 Pro
15:06 Side-by-Side Test: Gemini 3.1 Pro vs Claude Opus (Hinton Doom Center)
17:39 Why File Manipulation Accuracy Matters More Than Context Windows
19:27 The Context Window Debate: 1M Tokens vs Smart Sub-Agents
22:05 DHH on Token Pricing: "If There's a Bubble, It's This"
24:11 Should Models Ship as Agent vs Chat Variants?
28:43 Claude Sonnet 4.6: A $2 Discount on Opus?
31:44 The Model Mix: Why One Model Won't Rule Them All
34:40 Anthropic Is Winning — But Can Anyone Tell the Difference?
38:58 OpenAI Acquires OpenClaw: Why Couldn't They Just Build It?
44:18 The Silicon Valley Moment: Sam vs Dario at India AI Summit
47:05 Will Smaller Models Win the Enterprise? The Cost Reality Check
51:27 The End of Single-Shot: Why Agentic Loops Change Everything
55:48 Final Thoughts & Gemini 3.1 Pro Gets One More Week

Thanks for listening. Like & Sub. Links above for the Still Relevant Tour signup and Simtheory. Two models dropped on a week again. What a time to be alive. xoxo

Join Simtheory: https://simtheory.ai

Register for the STILL RELEVANT tour: https://simulationtheory.ai/16c0d1db-a8d0-4ac9-bae3-d25074589a80

GLM-5 just dropped and it's trained entirely on Huawei chips – zero US hardware dependency. Meanwhile, we're having existential crises about whether we're even needed anymore. In this episode, we break down China's new frontier model that's competing with Opus 4.6 and Codex at a fraction of the price, why agentic loops are making 200K context windows the sweet spot (sorry, million-token dreams), and the very real phenomenon of AI productivity psychosis. We dive into why coding-optimized models are secretly winning at everything, the Harvard study confirming AI doesn't reduce work – it intensifies it, and the exodus of safety researchers from XAI, Anthropic, and OpenAI (spoiler: they're not giving back their shares). Plus: Mike's arm is failing from too much mouse usage, we debate whether the chatbot era is actually fading, and yes – there's a safety researcher diss track called "Is This The End?"

CHAPTERS:

0:00 Intro - Is This The End? (Song Preview)
0:11 Still Relevant Tour Update & NASA Listener Callout
1:42 AI Productivity Psychosis: The Pressure of Infinite Capability
4:25 GLM-5 Breakdown: China's New Frontier Model on Huawei Chips
7:24 First Impressions: GLM-5 in Agentic Loops
9:48 Why Cheap Models Matter & The New Model War
14:09 Codex Vibe Shift: Is OpenAI Winning?
16:24 Does Context Window Size Even Matter Anymore?
22:27 The Parallelization Problem & Cognitive Overload
27:27 Mike's Arm Injury & The Voice Input Pivot
31:17 Single-Threaded Work & The 95% Problem
35:06 UX is Unsolved: Rolling Back Agentic Mistakes
38:45 Harvard Study: AI Doesn't Reduce Work, It Intensifies It
44:01 How AI Erodes Company Structure & Why Adoption Takes Years
50:14 My AI vs Your AI: Household Debates
50:43 The Safety Researcher Exodus: XAI, Anthropic, OpenAI
56:49 Final Thoughts: Are We All Still Relevant?
59:04 BONUS: Full "Is This The End?" Diss Track

Thanks for listening. Like & Sub. Links above for the Still Relevant Tour signup and Simtheory. GLM-5 is here, your productivity psychosis is valid, and the safety researchers are becoming poets. xoxo

Join Simtheory: https://simtheory.ai

Register for the STILL RELEVANT tour: https://simulationtheory.ai/16c0d1db-a8d0-4ac9-bae3-d25074589a80

It's the model same-day showdown of 2026. Opus 4.6 and Codex 5.3 dropped within minutes of each other, and we're breaking down what this means for the future of AI work. In this episode, we unpack Opus 4.6's million-token context window (if you've got billies in the bank), why Codex's pricing makes it nearly impossible to ignore for agentic loops, and the real cost of running agents for 24 hours ($10K, apparently). We dive deep into why coding-optimized models are secretly crushing it at non-coding tasks, the mental fatigue of managing AI workers, and whether the chatbot era is actually fading or just evolving. Plus: Chris accidentally books three real pig grooming appointments, we debate whether you need a "life coach agent" to manage your agent swarm, and yes – there's an Opus 4.6 diss track that goes unreasonably hard.

CHAPTERS:

0:00 Intro - Opus 4.6 Diss Track Preview
0:09 The Model Same-Day Showdown: Opus 4.6 vs Codex 5.3
0:50 Opus 4.6 Breakdown: Million Token Context & Premium Pricing
2:31 Token Bill Shock: $10K Research Bills & Extended Context Costs
5:04 Codex Pricing: Why It's Nearly Free for Agentic Loops
6:42 Why Coding Models Are Secretly Crushing Non-Coding Tasks
10:14 Tool Fatigue: Too Many Models, Too Many Workflows
12:47 Opus 4.6 First Impressions: "Solid" and "Faultless"
13:48 Chris Accidentally Books Three Real Pig Grooming Appointments
16:01 Unix Tools & Why Code-Optimized Models Win at Everything
19:59 The Agentic Retraining Imperative: Chat to Delegation
22:16 Agent Swarms & The Master Thread Architecture
24:51 OpenAI vs Anthropic: The Enterprise Battle
27:09 Corporate Espionage 2.0: Stealing Skills & The Open Source Threat
31:19 The UX Problem: Why Delegation Isn't Solved Yet
34:24 The Stress of Hyper-Productivity & Managing Agent Swarms
37:07 Coordination: The Next Layer of Abstraction
40:09 The Fantasy vs Reality of Autonomous AI Businesses
44:37 Is the Turn-by-Turn Chatbot Era Actually Fading?
49:23 Tokens as Spice: Turning Compute Into Money
52:08 Reduce Cognitive Overload: The Real Goal of AI
55:07 Still Relevant Tour Announcement
55:39 BONUS: Full Opus 4.6 Diss Track

Thanks for listening. Like & Sub. Links below for the Still Relevant Tour signup and Simtheory. The model wars are heating up, and your token bill is about to get interesting. xoxo

Join Simtheory: https://simtheory.ai
Register for the STILL RELEVANT tour: https://simulationtheory.ai/16c0d1db-a8d0-4ac9-bae3-d25074589a80
---

The hype train is 2026 knows only Moltbot (RIP Clawdbot). In this episode, we unpack the viral open-source AI assistant that's taken over the internet what it actually does, why everyone's losing their minds, and whether it's worth the $750/day token bills some users are racking up. We dive deep into why locally-run skills and CLI tools are beating computer-use clicking, how smaller models like GPT-5 Mini are crushing it in agentic workflows, and why the real magic is in targeted context - not massive swarms. Plus: Kimi K2.5 drops as a near-Sonnet-level model at 1/10th the price, we debate whether SaaS is dead, and yes – there are TWO Kimi K2.5 diss tracks. One made by Opus pretending to be Kimi. It might just slap?

CHAPTERS:

0:00 Intro - Still Relevant Tour Update
0:48 What is Moltbot? The Viral AI Assistant Explained
3:57 Token Bill Shock: $750/Day and Anthropic Bans
5:00 The Dream of Digital Coworkers on Mac Minis
6:52 Why CLI Tools & Skills Beat Computer-Use Clicking
10:57 Why This Way of Working Is Genuinely Exciting
14:47 Smaller Models Crushing It: GPT-5 Mini & Targeted Context
17:30 Wild Agentic Behavior: Chrome Tab Hijacking & Auto-Retries
20:10 Security Architecture: Locked-Down Machines & Enterprise Use
24:01 AI Building Its Own Tools On-The-Fly
27:08 The Fear & Overwhelm of Rapid Progress
29:10 2026: The Year of Agent Workers
31:43 The Challenge of Directing AI Work (Everyone's a Manager Now)
37:24 Skills Will Take Over: Why MCPs & Atlassian Can't Stop Us
40:38 Real-World Use Cases: Doctors, Lawyers & Accountants
46:28 Cost Solutions: Build Workflows Around Cheaper Models
52:58 Kimi K2.5: Sonnet-Level Performance at 1/10th the Price
1:00:55 The "1,500 Tool Calls" Claim: Marketing vs Reality
1:05:23 The Kimi K2.5 Diss Tracks (Opus vs Kimi)
1:08:08 Demo: Black Hole Simulator & Self-Trolling CRM
1:12:55 Is SaaS Dead?
1:14:30 BONUS: Full Kimi K2.5 Diss Tracks

Thanks for listening. Like & Sub. Links below for the Still Relevant Tour signup and Simtheory. The future is open source, apparently. xoxo

Join Simtheory: https://simtheory.ai

Reserve your seat on the STILL RELEVANT tour: https://simulationtheory.ai/16c0d1db-a8d0-4ac9-bae3-d25074589a80
----
Two episodes in one week? We're either above average or completely unhinged. In this one, we dive deep into the new phenomenon of "AI exhaustion" – that fried feeling you get after multitasking across six agent tabs all day. We share our breakthroughs with AI-assisted presentations (20 minutes vs several hours), why browser-use on your local machine bypasses every anti-scraping technique known to man, and how enterprise context sharing could be the real unlock for organizations. Plus: OpenAI announces ads for ChatGPT (even on paid tiers), their CFO floats taking cuts from drug discoveries (seriously), and Google publicly dunks on them for it. Also – the Still Relevant Australia Tour is coming, and our LinkedIn group hit 200 members (we're basically LinkedIn influencers now too).

CHAPTERS:

0:00 Intro - Still Relevant Tour Announcement + LinkedIn Milestone
2:08 AI Exhaustion: The Cognitive Overload of Multitasking with Agents
4:14 Why Single-Tasking with AI Beats Parallel Agent Chaos
7:02 The Problem with "I Spun Up 70,000 Sub-Agents" Twitter Posts
10:03 Mike's Presentation Workflow: From Hours to 20 Minutes
14:06 Why Isn't Copilot Doing This Already?
16:54 Old Models + Great Context = Still Amazing Results
21:14 What's Actually Changed? It's the Software Layer
25:22 Enterprise Context Sharing & Organizational IP
31:22 Skills, Sub-Agents, and Role-Based Knowledge
35:22 Security Concerns: Can You Hack an Agent with Malicious MD Files?
38:23 Cloud Providers Have a Bigger Moat Than the Labs
43:16 Browser Use: The Ultimate Context Gathering Weapon
48:25 Rethinking SaaS: Software That Actually Thinks
53:08 Smart Paste, Smart CC – Why Isn't All Software Like This?
56:32 OpenAI's Desperate Moves: Ads, Age Verification & Drug Royalties
1:03:03 Google Says "No Plans for Gemini Ads" (Shots Fired)
1:07:24 Is OpenAI Okay? The Vibes Are Definitely Off
1:10:35 Capitalism Won't Give You Free Time, Just More Demands
1:11:20 Outro + Still Relevant Tour Details

Thanks for listening. Like & Sub. Links below for the Still Relevant Tour signup and Simtheory. xoxo

Join Simtheory: https://simtheory.ai
---
Join the most average AI LinkedIn group: https://www.linkedin.com/groups/16562039/

It's 2026 and everyone's having an existential crisis. In this episode, we unpack the two camps dominating AI C/Twitter: hype boys claiming "Claude Code can do my washing" vs. software developers doom-scrolling themselves into career panic. We put the agentic hype to the test and discover that no, you can't actually run 8 agents recreating your local business ecosystem while you sleep. Plus, we reflect on why MCP is exhausting, why Gemini 3 Pro is somehow worse than Gemini 2.5 Pro, and why Geoffrey Hinton would rather write his book than answer questions in Tasmania. Also featuring: the $200,000/month enterprise AI problem, why SaaS isn't dead (but it's scared), and our prediction that AI workspaces will become the everything app.

CHAPTERS:

00:00 Intro - Unpacking the 2026 AI Vibes
02:21 Putting Claude Code and Agentic Hype to the Test
05:57 Why Twitter AI Demos Never Show the Receipts
07:03 Honest Assessment of Where Frontier Models Are At
11:19 Building the Everything App with Email, Calendar and Files
16:47 Collaborative Mode vs Agentic Delegation in Practice
21:29 The Real Cost of Enterprise AI at Scale
24:32 Why Cheaper Models Like Haiku and Gemini Flash Matter
29:25 Is SaaS Actually Dead or Just Disrupted
38:11 The Future of AI Platforms, SDKs and App Stores
43:35 The Untapped Opportunity in Paid Proprietary MCPs
51:21 Geoffrey Hinton Refuses to Take Questions in Tasmania
55:05 2026 Plans and the Still Relevant Tour Announcement

Thanks for listening. Like & Sub. xoxox

The Gift of Simtheory: https://simtheory.ai
---
2025 Model Timeline: https://simulationtheory.ai/5fd0e964-4c41-4f9a-bbb3-2a398d8500f0

It's the long-anticipated holiday special... except Mike and Kris forgot to prepare so it's just a normal episode. 🎅 This week: Gemini 3 Flash drops and it's actually incredible - cheap, fast, and weirdly smarter than Gemini 3 Pro at tool calling. We put GPT Image 1.5 head-to-head against Nano Banana Pro using hobo photos (spoiler: Google wins again). Plus, FireCrawl Agent is the research tool we've been waiting for, Anthropic launches Skills as an open standard, and we do a full 2025 model timeline recap. Also featuring: Best and Worst Model of the Year awards, 2026 predictions where Mike bets on OpenAI (controversial), and the full holiday musical outro where AI sings about what an "average" year it's been.

CHAPTERS
00:00 Intro - Holiday Special That Isn't
00:55 Shipping Gemini 3 Flash While Looking Like a "Sophisticated Programming Hobo"
02:52 Gemini 3 Flash Review: Cheap, Fast, Surprisingly Smart
06:31 The Unreliable Frontier Model Problem
10:45 GPT Image 1.5 vs Nano Banana Pro Showdown
17:04 FireCrawl Agent: Research That Actually Works
25:56 Gemini Deep Research Agent Deep Dive
31:57 Skills vs MCPs: The New Paradigm
43:35 Enterprise Skills: Codifying Business Procedures
49:57 2025 Model Timeline Recap
59:53 Best & Worst Model of 2025 Awards
1:04:58 2026 Predictions: Mike Bets on OpenAI
1:14:09 Final Thoughts & Holiday Thank Yous
1:19:35 🎄 Holiday Musical: "A Very Average Christmas"

Have a great Christmas/Holiday/New Year, see you in 2026! xox

Join Simtheory: https://simtheory.ai

GPT-5.2 is here and... it's not great. In this episode, we put OpenAI's latest model through its paces and discover it can't even identify a convicted serial killer when the text literally says "serial killer." We compare it head-to-head with Claude Opus and Gemini 3 Pro (spoiler: they win). Plus, we reflect on the "Year of Agents" that wasn't, why your barber switched to Grok, Disney's billion-dollar investment to use Mickey Mouse in Sora, and why Mustafa Suleyman should probably be fired. Also featuring: the GPT-5.2 diss track where the model brags about capabilities it doesn't have.

CHAPTERS:

00:00 Intro - GPT-5.2 Drops + Details
01:25 First Impressions: Verbose, Overhyped, Vibe-Tuned
02:52 OpenAI's Rushed Response to Gemini 3
03:24 Tool Calling Problems & Agentic Failures
04:14 Why Anthropic's Models Just Work Better
06:31 The Barber Test: Real Users Are Switching to Grok
10:00 The Ivan Milat Vision Test (Serial Killer Edition)
17:04 Year of Agents Retrospective: What Went Wrong
25:28 The Path to True Agentic Workflows
31:22 GPT-5.2 Diss Track (Yes, Really)
43:43 Why We're Still Optimistic About AI
50:29 Google Bringing Ads to Gemini in 2026
54:46 Disney Pays $1B to Use Mickey Mouse in Sora
56:57 LOL of the Week: Mustafa Suleyman's Sad Tweets
1:00:35 Outro & Full GPT-5.2 Diss Track

Thanks for listening. Like & Sub. xoxox

Join Simtheory: https://simtheory.ai/

OpenAI has declared "Code Red" as ChatGPT faces growing competition from Gemini and other rivals. In this episode, we break down OpenAI's 6% market share decline, why their ad strategy is on hold, and what they need to do to reclaim the AI crown. We also explore DeepSeek V3.2's impressive capabilities as a cheap open-source alternative, Meta's new policy grading employees on AI skills, and the crisis facing higher education as AI fluency becomes essential. Plus, Fatal Patricia hits #1 on our Spotify charts, and Tesla's Optimus robot is running like a slightly unfit human.

CHAPTERS:
00:00 Intro - OpenAI Code Red & Market Share Crisis
07:03 ChatGPT's Failure to Go Deeper Into Users' Lives
16:33 What OpenAI Needs to Win Back the Crown
26:46 Chris's Wishlist for an OpenAI Comeback
31:22 DeepSeek V3.2 - The Open Source Threat
39:34 Meta Grading Workers on AI Skills
46:29 The University & Education AI Crisis
56:25 Fatal Patricia Hits #1 & WTF of the Week

Thanks for listening. Like & Sub. xoxox

Join Simtheory: https://simtheory.ai (Use coupon BLACKFRIDAY15 for $15 USD off any subscription).
----
Simtheory Discord: https://discord.gg/Ar6GeQnAR7
This Day in AI Discord: https://discord.gg/TVYH3HD6qs
LinkedIn Group: https://www.linkedin.com/groups/16562039/
Spotify: https://open.spotify.com/artist/28PU4ypB18QZTotml8tMDq?si=FPaJU2NRSnOSNPmnsfwA_g
---
CHAPTERS:
00:00 Intro & Fatal Patricia Update
01:40 Promotions (Discord, Black Friday, LinkedIn)
04:36 Claude 4.5 Opus - Best Anthropic Model Ever?
31:17 Computer Use API Updates
36:14 Will AI Replace 57% of Jobs? (McKinsey Report)
1:00:52 Claude 4.5 Opus Demos (Christmas Hut & Diss Track Preview)
1:07:13 Microsoft Farah 7B - Moose Porn Refusals
1:21:51 Why ChatGPT's MCP-UI Apps Are a Bad Idea
1:42:01 🎵 Claude 4.5 Opus Diss Track (Full Song)
---
Thanks for listening. Like & Sub. xoxox

Anthropic just dropped Claude 4.5 Opus and it might be the best AI model of 2024. In this episode, we compare Claude 4.5 Opus vs Gemini 3 Pro vs GPT-5.1, breaking down the new API features including effort parameters, context management, and computer use updates. We also test Microsoft's new Farah 7B parameter model for computer use - with hilarious refusal results. Plus, we react to McKinsey's controversial report claiming AI agents could automate 57% of US jobs by 2030. 

We dive deep into Anthropic's pricing (3x cheaper than Opus 4.1), why Claude is now beating Google and OpenAI on agentic coding benchmarks, and whether MCP-UI apps in ChatGPT are a step backwards for AI workflows. Is Claude 4.5 Opus the new king of AI coding assistants? Should enterprises be worried about AI job replacement? And why did Microsoft's Farah model refuse to draw a moose? All this plus an AI-generated diss track roasting Sam Altman, Elon Musk, and Sundar Pichai.