April 16 - Codex uses your mac in the background, Opus 4.7 release not quite Mythos + 3 interviews

Episode Summary

Alex goes live EIGHT minutes early to cover breaking news: Claude Opus 4.7 drops mid-intro — 87.6% SWE-bench Verified, a 64.3% on SWE-bench Pro, and a brand-new 'extra high' reasoning effort that ends up feeling more like a Mythos teaser than a full release. Qwen 3.6-35B-A3B (Apache 2.0, 73.4% SWE-Verified with only 3B active params) and MiniMax M2.7 open weights keep the open-source train screaming, while OpenAI drops another breaking-news bomb during the show with a massive Codex update: native macOS background computer use, 90+ plugins, memory, gpt-image-1.5, and multi-terminal SSH. Three incredible interviews — Trevor Manz on Marimo Pair dropping coding agents into reactive Python notebooks, Kwindla on Gradient Bang (the multi-agent voice game that 'broke containment'), and Theodor Marcu on Windsurf 2.0 + Devin's agent command center. Plus Alex debuts the 'ZL Continuum' essay from AI Engineer Europe: do engineers still read code?

In This Episode

• 🔥 Pre-Show Banter & Opus 4.7 Breaking News
• 🧪 Opus 4.7 Evals & Benchmarks
• 📰 TL;DR — Weekly AI News Roundup
• 🧪 Opus 4.7 Live Testing — Martian Simulation
• 🔓 Qwen 3.6 Open Source Release
• 🔓 MiniMax M2.7 Open Weights
• 🧠 The ZL Continuum — Do Engineers Still Read Code?
• ⚡ AI Engineer Summit — Top 10 Themes
• 🎥 Pi Hard — Craziest AI Video Yet
• 🛠️ Interview: Trevor Manz — Marimo Pair
• ⚡ This Week's Buzz — Weights & Biases / CoreWeave
• 🔊 Interview: Kwindla Kramer — Gradient Bang & Google TTS
• 🛠️ Interview: Theodor Marcu — Windsurf 2.0 & Devin
• 🔥 Breaking News: OpenAI Codex Major Update
• 🎨 NVIDIA Lyra 2.0 & 3D World Generation
• 📰 AI for Normies — Robots & Allbirds Pivot

Hosts & Guests

Alex Volkov

Host · W&B / CoreWeave

@altryne

Trevor Manz

Founding engineer · Marimo

@trevmanz

Theodor Marcu

Product · Cognition (Windsurf)

@theodormarcu

Kwindla Hultman Kramer

Co-CEO Daily · Pipecat maintainer

@kwindla

Yam Peleg

AI builder & founder

@yampeleg

Nisten Tahiraj

Weekly co-host · AI engineer

@nisten

LDJ

Weekly co-host · AI researcher

@ldjconfirmed

Wolfram Ravenwolf

Weekly co-host · AI Evangelist at W&B/CoreWeave

@WolframRvnwlf

By The Numbers

🔥 Breaking During The Show

🔥 Pre-Show Banter & Opus 4.7 Breaking News

Alex and Yam go live 8 minutes before the official show start because Anthropic just dropped Claude Opus 4.7. No early access, no advance briefing — the crew opens the system card live alongside the audience.

• AI breaking news from the jump: Opus 4.7 drops with no prior access
• Anthropic ships the system card publicly, crew reads it in real time
• 4 cohosts on set: Yam, Wolfram, LDJ (and Nisten joining shortly)

Alex Volkov 0:07

"Anthropic made us go live, eight minutes before the official show start, because we got some breaking news."

🧪 Opus 4.7 Evals & Benchmarks

The crew walks through the Anthropic evals table live. Biggest jumps: SWE-bench Pro (+11), ScreenSpot Pro (+22% on computer use), new xhigh effort level. MRCR long-context drops from 78% → 32%, suggesting a new pre-trained base. The system card mentions 'Mythos' 331 times — this feels like an ad for the godlike version they haven't released yet.

• 331 mentions of 'Mythos' in the system card — 4.7 feels like the appetizer
• SWE-bench Pro jumps 11 points, passes GPT-5.4 on agentic coding
• MRCR 8-needle V2: 78% → 32% — LDJ thinks it's a new pre-trained base
• New 'xhigh' (extra high) reasoning effort gets best-in-class on HLE
• New tokenizer uses 1.0–1.35× more tokens per prompt — not a cash grab, probably multimodal reasons

Alex Volkov 2:10

"If you look for Mythos in the system card, there's three hundred and thirty-one mentions for Mythos. It does feel like a Mythos ad."

LDJ 6:59

"This kind of helps maybe confirm it is maybe a new pre-trained model from scratch — in MRCR it's performing much worse, but I expect in other areas it might be much better in dramatic ways."

Yam Peleg 9:07

"We're all cooked already at this point, everyone looking at the evals, who has the biggest number. If you do a good job and train a really good model, people like us are gonna look at the evals and say 'nah, this model is not good.' But probably the model is really good."

📰 TL;DR — Weekly AI News Roundup

The longest TL;DR in show history. CoreWeave signs Anthropic (multibillion), Meta ($21B expansion), and Jane Street ($6B cloud + $1B equity) — now serving 9 of the top 10 AI labs. Qwen 3.6-35B-A3B, MiniMax M2.7 open weights, Windsurf 2.0 + Devin, Warp's any-CLI-agent support, Claude Code Routines (cron-triggered agents on Anthropic's cloud), Marimo Pair, Gemma 4 live on W&B Inference, Gemini 3.1 Flash TTS, Baidu ERNIE-Image, Tencent HYWorld 2.0, Nvidia Lyra 2.0, a unitree humanoid breaking a 100m dash record, and Allbirds → NewBird AI.

• CoreWeave now backs 9 of the top 10 AI labs (Anthropic, Meta $35B+, Jane Street $7B)
• Qwen 3.6-35B-A3B: Apache 2.0, 35B MoE / 3B active, 262K→1M context, natively multimodal
• MiniMax M2.7 open weights: 230B / 10B active, 56.22% SWE-Pro matching GPT-5.3-Codex
• Gemma 4 now live on W&B Inference (CoreWeave) with LoRA inference support — code 'Gem Drop' for $20 credits
• Claude Code Routines: cron/GitHub-event/API triggered autonomous agents on Anthropic's cloud
• Super Gemma 4 26B Uncensored v2 by @songjunkr trending on HF — 0/100 refusals, fixed tool calls

Nisten Tahiraj 27:21

"People are going ham over this. It's like we're back to two years ago."

Yam Peleg 27:36

"Open source is so back."

🧪 Opus 4.7 Live Testing — Martian Simulation

Nisten puts Opus 4.7 through the infamous Martian simulation benchmark the crew has been running for months. Early impression: incremental over 4.6 but solid — 'not as much of a jump from 4.6 to 4.7 as it was from 4.5 to 4.6.' Vision / computer use, on the other hand, feels genuinely better.

• Nisten: 'It feels incremental this time' — not as big a jump as 4.6 over 4.5
• Vision + ScreenSpot Pro improvements are the real story
• New 'ultrareview' /command in Claude Code

Nisten Tahiraj 12:38

"It feels incremental this time. Like, it doesn't — 4.6 was smarter. 4.7 is not as much of a jump from 4.6 as it was from 4.5 to 4.6."

🔓 Qwen 3.6 Open Source Release

Alibaba Qwen drops Qwen 3.6-35B-A3B under Apache 2.0 the same morning as Opus 4.7 — 35B MoE with only 3B active parameters, 73.4% SWE-bench Verified, natively multimodal, 262K context extensible to 1M. After Junyang Ling left the team there were doubts about Qwen's open-source commitment; this release puts those to rest.

• Apache 2.0 — 35B MoE / 3B active, 73.4% SWE-bench Verified
• Natively multimodal, 262K context extensible to 1M
• Strongest mid-size LLM on nearly all benchmarks
• Confirms Alibaba Qwen's continued open-source commitment post-Junyang

🔓 MiniMax M2.7 Open Weights

MiniMax releases M2.7 open weights: 230B parameter MoE with only 10B active, 56.22% on SWE-Pro (matching GPT-5.3-Codex). Self-evolved via 100+ rounds of autonomous RL. Typical recent Chinese release cadence: ship the API first, then open-source the weights.

• 230B MoE / 10B active — matches GPT-5.3-Codex on SWE-Pro (56.22%)
• Self-evolved via 100+ rounds of autonomous RL
• Open weights released on HF after API launch

🧠 The ZL Continuum — Do Engineers Still Read Code?

Alex's essay from AI Engineer Europe: where are you on the Z–L spectrum? Ryan Lopopolo (OpenAI, token billionaire) says code is a liability, don't read it. Mario Zechner (creator of pyo, the harness powering OpenClaw) says slow the fuck down, read every line of critical code. Everyone else is somewhere in between. The crew + later guests weigh in: Nisten=Z, Yam=Z (with a brutal 'hidden-features accumulation' warning), Wolfram=L (mostly), LDJ=moving-to-L, Trevor=L-leaning, Kwindla=full-L ('my rule was not to read or write any code for the side project').

• Ryan Lopopolo: 'code is a liability' · Mario Zechner: 'read every line'
• Yam's warning: agents silently add hidden features (context truncation, etc.), accumulate, eventually no agent OR human can review the code
• Consensus: it's not per-person, it's per-task — critical code = read, throwaway = YOLO
• Poll opens live during the show

Yam Peleg 47:26

"You must read the code, man. Those things accumulate — the agent thinks 'oh, that's what the human wanted' so it maintains hidden features. You get to a point where the agents just can't do anything."

Wolfram Ravenwolf 50:48

"I trust the AI, I tell it what to do, I test it, and my most used command is review. I probably spend more time reviewing than actually coding."

⚡ AI Engineer Summit — Top 10 Themes

Alex's synthesized top-10 from AI Engineer Europe: (1) FMAT — Fear of Missing Agent Time, (2) the ZL Continuum, (3) everything is changing super fast (GitHub on track for 15B PRs, Vercel 60% agent traffic), (4) we're still early, (5) AGI is here and unevenly distributed, (6) 'just talk to your Clanker', (7) MCP is dead long live MCP (enterprises adopting faster than ever), (8) AI was supposed to make us work less — we work more, (9) MHC = Model / Harness / Context is the new ASL.

• FMAT: Fear Of Missing Agent Time — universal at the conference
• GitHub: 1B → 15B PRs YoY projected, 15× growth
• Vercel: 60% of traffic attributed to agents
• MHC framework — Model/Harness/Context is the new ASL for AI engineers
• MCP isn't dying — enterprises are adopting faster, especially with code-mode

Alex Volkov 59:03

"Model Harness Context is the new ASL. When somebody tells me 'OpenClaw is stupid', I have no idea how to react until they tell me if they use OpenClaw with Opus 4.7 or MiniMax 2.7."

🎥 Pi Hard — Craziest AI Video Yet

Alex plays the Pi Hard / Neil deGrasse Tyson / SBF AI trailer live. Even Alex's fiancée (who works with AI video daily) didn't clock it as AI until midway through. Seedance 2.0 is now everywhere; the crew agrees this is the craziest AI video they've ever seen.

• Multi-shot AI video production, Neil deGrasse Tyson deepfake
• Seedance 2.0 fully rolled out with video support everywhere
• Yam: 'That's the craziest AI video I've seen'

Yam Peleg 1:03:57

"That's the craziest AI video I've seen. That's... there's no competition."

🛠️ Interview: Trevor Manz — Marimo Pair

Trevor Manz, founding engineer at Marimo, on why reactive Python notebooks are suddenly very important for AI workflows — and on Marimo Pair, which drops Claude Code / Codex / OpenCode agents directly inside a reactive notebook. Trended on Hacker News this week. On the ZL continuum, Trevor is moving L-ward but focuses more on building verification systems around agents than on reading less code.

• Marimo = reactive Python notebooks (dependency-graph aware vs Jupyter)
• Marimo Pair: drop Claude Code / Codex / OpenCode agents in the notebook
• Trended on HN this week
• Trevor's take: shift burden of review onto better verification systems

Trevor Manz 1:07:46

"My job has shifted a lot to trying to build systems that I can have the AI tools verify their results and correctness of the programs. So trying to shift some of the burden of review onto just having better systems."

⚡ This Week's Buzz — Weights & Biases / CoreWeave

Marimo Pair (interviewed above) is the W&B/CoreWeave Buzz this week, plus Gemma 4 now live on W&B Inference with LoRA inference support. Reply to the W&B announcement post with code 'Gem Drop' for $20 in inference credits.

• Marimo Pair — the CoreWeave-family agent notebook integration
• Gemma 4 live on wandb.ai/inference with LoRA inference support
• Code 'Gem Drop' on X for $20 in free W&B inference credits

🔊 Interview: Kwindla Kramer — Gradient Bang & Google TTS

Kwindla Hultman Kramer (co-CEO of Daily, Pipecat maintainer) on Google's Gemini 3.1 Flash TTS (1,211 Elo, 70+ langs, fully promptable — but ~3s TTFT, so batch-only for now). Then the main event: Gradient Bang, his 'side project that broke containment' — a fully LLM-driven multiplayer voice-based space game inspired by Trade Wars. Built on a new Pipecat Sub-Agents library, uses Deepgram + GPT-4.1 voice agent + GPT-5.2 medium-thinking task agents + LLM-generated dynamic UIs. Kwindla's rule for his own side project: 'don't read or write any code.'

• Gemini 3.1 Flash TTS is fully promptable like an LLM (not fixed tags) — but has ~3s TTFT
• Gradient Bang: multi-agent voice space game inspired by BBS-era Trade Wars
• Pipecat Sub-Agents: new class-based event bus, works locally + over network
• Voice agent always runs (<1.5s response), task agents on GPT-5.2 medium thinking
• LLM-generated dynamic UI paradigm — React frontend rendered via JSON from LLM
• Open-sourced GB Benchmarks for evaluating agent task execution

Kwindla Hultman Kramer 1:23:09

"It's a side project that broke containment. I hacked together this game, we started playing it, and it became clear really quickly that a lot of things in voice AI — all these problems we were trying to solve actually are very general: how do you build AI-native software."

Kwindla Hultman Kramer 1:33:37

"Part of my goal for this, since it was a side project, was not to write or read any code. I've been doing that since November — and it's been painful in different ways, but also a great learning process."

🛠️ Interview: Theodor Marcu — Windsurf 2.0 & Devin

Theodor Marcu (product, Cognition) on Windsurf 2.0 — the first big post-acquisition launch. Headline: Agent Command Center (a Kanban-board mission control for dozens of agents), Spaces for task context switching, and full Devin integration inside Windsurf. Cognition's thesis: the future is managing a team of agents, both local (pair programmer) and cloud (end-to-end). Teodor also reveals Cognition's internal use has doubled since launching Managed Devins + Scheduled Devins.

• Agent Command Center = Kanban-board mission control for dozens of agents
• Spaces — switch contexts between parallel tasks, each with local + cloud agents
• Devin is now integrated directly inside Windsurf (Devin's desktop visible locally)
• Plan locally with a Socratic-method agent, hand off to Devin in the cloud for execution
• Internal Cognition usage doubled after launching Managed + Scheduled Devins
• 'Sub-Devins' — Devins managing Devins

Theodor Marcu 1:36:36

"The future of software engineering is managing a team of agents — both remote and local, that can work alongside you. Some of our best engineers are working with dozens of agents at a time."

Theodor Marcu 1:44:13

"A lot of folks on the team cannot go to sleep without starting at least a bunch of Devin sessions, sometimes multiple per task, so they can compare them in the morning."

🔥 Breaking News: OpenAI Codex Major Update

Second breaking news of the show: OpenAI drops a massive Codex update mid-conversation. Native macOS background computer use (with a separate cursor, so you can keep working), 90+ plugins, gpt-image-1.5 image generation + editing, in-app browser, memory ('learns from experience'), proactive work suggestions, multi-terminal SSH, and thread automations. Alex's hot take: Codex is becoming the super-app, not ChatGPT. Post-show, Alex streamed another hour of live testing — the background computer use in particular is much bigger than it looks on the landing page.

• Native macOS computer use — runs in a separate cursor, in the background
• 90+ plugins connecting Codex to external services
• gpt-image-1.5 image generation + editing inside Codex
• Memory preview — 'learns from experience', remembers corrections + preferences
• In-app browser closes the frontend feedback loop automatically
• Multi-terminal SSH into dev boxes + thread automations

Yam Peleg 1:48:33

"We just talked to Devin and Windsurf. And now we're talking about Codex. It's like it's a war, man. It's a war."

Kwindla Hultman Kramer 1:50:56

"'Learns from experience' is just a massive unlock if they can put the pieces together. All of us who do non-trivial stuff in coding agents — we're always trying to add this to the notes file, add this to the README, write down exactly how you got here so we don't get in this loop."

Alex Volkov 1:52:33

"The computer use happens in the background. They have another cursor — they don't use your cursor to click things, so you can actually ask it to do something and keep working on something else. This is the only experience like this I know of."

🎨 NVIDIA Lyra 2.0 & 3D World Generation

Quick hit on the 3D-world-from-single-image race: Baidu ERNIE-Image (8B DiT, #1 GenEval among open models), Tencent HYWorld 2.0 (editable 3D Gaussian Splats, Unity/Unreal/Isaac Sim ready), NVIDIA Lyra 2.0 (Apache 2.0, single image → explorable persistent 3D worlds). Essentially the open-source equivalents of what Fei-Fei Li's World Labs is building.

• Baidu ERNIE-Image — 8B DiT, #1 GenEval among open models
• Tencent HYWorld 2.0 — editable 3D scenes from single image, Unity/Unreal ready
• NVIDIA Lyra 2.0 — Apache 2.0, persistent explorable 3D worlds from one image

📰 AI for Normies — Robots & Allbirds Pivot

Unitree humanoid breaks the 100m dash world record at ~10m/s — faster than Olympic sprinters. And in the stupidest pivot of 2026, Allbirds (the shoe company) loses 99% of its value, rebrands to 'NewBird AI', raises $50M 'to buy GPUs', and the stock shoots up 600-800%. Alex: 'where are they buying those GPUs?'

• Unitree humanoid: ~10 m/s, world-record 100m dash
• Allbirds → NewBird AI: 600–800% stock pump after GPU-pivot announcement
• 'The more you buy, the more you save' — the entire new business model

Yam Peleg 1:57:15

"They literally implemented 'the more you buy, the more you save'. You just buy a bunch of GPUs and you print money. That's their new business model."