Episode Summary

ThursdAI’s Apr 2 episode is a high-signal mix of breaking AI news and hands-on operator analysis: the Claude Code leak fallout with Sigrid Jin, Gemma 4 coverage with Omar Sanseviero from Google DeepMind, and a rapid sweep of major lab moves. The panel unpacks what the leak actually revealed, where SessionGate friction is hurting real workflows, and how open-model momentum is reshaping tooling choices. They also cover OpenAI’s reported $122B funding round, Microsoft’s model push, and fresh benchmarks around agent harness performance. It’s a practical episode for builders trying to separate hype from what is immediately usable.

In This Episode

Hosts & Guests

Alex Volkov
Alex Volkov
Host · W&B / CoreWeave
@altryne
Sigrid Jin
Sigrid Jin
Claude Code clean-room rewrite co-creator
@realsigridjin
Omar Sanseviero
Omar Sanseviero
Developer Experience Lead (prev. Chief Llama Officer)
@osanseviero
Wolfram Ravenwolf
Wolfram Ravenwolf
Weekly co-host, AI model evaluator
@WolframRvnwlf
Ryan Carson
Ryan Carson
Weekly co-host of ThursdAI
@ryancarson
Nisten Tahiraj
Nisten Tahiraj
Weekly co-host of ThursdAI
@nisten
Yam Peleg
Yam Peleg
Weekly co-host of ThursdAI
@Yampeleg
LDJ
LDJ
Weekly co-host of ThursdAI
@ldjconfirmed

By The Numbers

OpenAI funding round
$122B
Discussed as the largest reported funding round in tech history and a major signal for AI infrastructure spending.
GitHub stars in 24h
100K+
Claw-code clean-room project velocity became a focal point in the Claude Code leak discussion.
Lines in leaked codebase
512K
Referenced in show notes while discussing what was exposed and what developers inferred from it.

🔥 Breaking During The Show

Gemma 4 breaking updates
The panel reacts in real time to Gemma 4 positioning and deployment implications as fresh details land during the show.
OpenAI $122B funding round
A major financing update reframes expectations around model race economics and infrastructure scale.

⚡ Intro

The panel opens with a packed week in AI and quickly frames the two big arcs of the episode: the Claude Code leak drama and the wave of open model launches. The co-hosts set context for why this week felt unusually consequential for both builders and labs.

  • Fast-moving week spanning model releases, leaks, and funding
  • Set-up for deep dives on Claude Code and Gemma 4

📰 TL;DR - This Week in AI

Alex runs a rapid-fire headline pass across major launches, funding news, and tool updates. It sets the backbone for the rest of the show before the conversation slows down into technical analysis.

  • OpenAI, Anthropic, Google, Microsoft, and Alibaba all had major updates
  • Episode structure moves from headlines to deeper operator discussion

🔥 Claude Code Leak

Sigrid Jin joins to unpack the clean-room Claude Code rewrite and the GitHub blow-up around the leak, including what was actually learned from the exposed package internals. The segment distinguishes technical facts from social-media exaggeration.

  • Leak discourse separated from verifiable implementation details
  • Sigrid explains why the clean-room rewrite mattered to developers

🤖 Claude Code Session Gate

The hosts discuss SessionGate complaints, resume/caching quirks, and cost instability reports from power users. The thread focuses on reliability and trust when teams run long-lived coding sessions.

  • Reported session-resume behavior can multiply costs
  • Panel emphasizes observability and guardrails for agent workflows

🔥 Gemma 4 Breaking News

Breaking updates on Gemma 4 land mid-show, and the panel reacts in real time to where it sits against frontier proprietary models. The focus is practical: what can be run now, where, and for which workloads.

  • Gemma 4 framed as a serious new open-model contender
  • Immediate discussion of deployment and tradeoffs

💰 OpenAI $122B Funding Round

Alex and co-hosts analyze OpenAI’s reported $122B raise and what that scale of capital implies for infra, product velocity, and competitive pressure on the rest of the market.

  • $122B described as a historic financing event
  • Discussion ties fundraising to model deployment economics

🏢 Microsoft AI Models

The team reviews Microsoft’s in-house model push across transcription, image generation, and voice. They compare positioning versus specialist products and foundation-model APIs.

  • Microsoft expanding first-party model stack
  • Panel compares quality and differentiation across modalities

🔓 Gemma 4 with Omar Sanseviero

Omar Sanseviero (Google DeepMind) joins to explain Gemma 4 from a builder and ecosystem perspective. Conversation covers model intent, community adoption, and practical entry points.

  • Confirmed guest Omar provides launch context from Google DeepMind
  • Focus on ecosystem and developer experience around Gemma 4

🔓 Gemma 4 - Google's Open Source Strategy

The panel zooms out on Google’s open-source strategy: why Gemma exists alongside closed systems and how that dual strategy may shape developer mindshare.

  • Open vs closed strategy discussed as portfolio decision
  • Community leverage and trust highlighted as differentiators

🛠️ Gemma 4 - Agentic Capabilities & Local Models

This section explores Gemma 4 in agentic setups and local workflows, including how smaller deployable models fit into multi-agent pipelines.

  • Local-first and agentic uses highlighted
  • Model size/perf tradeoffs discussed for real workloads

⚡ Gemma 4 - Recap & Community Features

A recap segment summarizes Gemma 4’s strongest points and where the community can contribute or extend the stack. The hosts emphasize experimentation over benchmark-chasing.

  • Community contribution pathways discussed
  • Recap connects product features to real usage

⚡ Gemma 4 - Community Reactions

The hosts review early reactions from builders and X/Twitter to gauge momentum and skepticism around the release claims.

  • Early sentiment analysis from AI community
  • Balance of excitement and skepticism captured

⚡ This Week's Buzz - Ralph Hackathon

The buzz segment highlights a hackathon demo and what it signals about agent UX and rapid prototyping culture around open tooling.

  • Ralph hackathon mentioned as high-signal community experiment
  • Demonstrates speed of current AI maker ecosystem

⚡ This weeks buzz - Wolfbench showes Hermes is better than OpenClaw

Wolfram brings WolfBench findings that stirred debate, especially claims around Hermes performance versus OpenClaw in specific harness conditions.

  • WolfBench data used to challenge assumptions
  • Benchmark methodology caveats called out

🧪 Wolf Bench - Hermes Agent Results

The panel spends additional time on Hermes agent results and interpretation quality, emphasizing reproducibility and fair harness configuration.

  • Hermes/OpenClaw comparison dissected in detail
  • Reproducible eval setup framed as essential

🧪 One-Bit Quantization (Prism ML)

A quick technical dive on Prism ML’s one-bit quantization ideas and where aggressive compression might unlock cheaper inference.

  • One-bit quantization positioned as cost/performance lever
  • Discussion focuses on practical deployment constraints

🔓 Alibaba Qwen 3.6 & Wan 2.7

The open-source round-up covers Alibaba’s Qwen 3.6 and Wan 2.7 updates, with attention to multimodal capability and practical ranking versus other open models.

  • Qwen family momentum remains strong
  • Wan 2.7 noted in broader open ecosystem context

🔊 Fish Audio Speech-to-Text

The team reviews Fish Audio’s speech stack progress and where it challenges incumbent speech providers for developers building voice workflows.

  • Speech model quality and developer ergonomics discussed
  • Voice tooling seen as rapidly commoditizing

🎥 Google Veo 3.1 Light

Google’s Veo 3.1 Light segment focuses on video-gen quality/speed tradeoffs and likely creator workflows it unlocks.

  • Veo 3.1 Light discussed as practical video model tier
  • Quality-vs-latency tradeoff highlighted

🤖 Agent Harnesses & Open Claw

The conversation broadens to agent harness architecture, safety boundaries, and why evaluation discipline matters more as capabilities accelerate.

  • Harness design framed as central infra question
  • Safety and control layers discussed with concrete examples

🧪 Anthropic Emotion Vectors in Claude

A late segment explores Anthropic emotion-vector work and what it might mean for steerability, user experience, and model behavior interpretability.

  • Emotion-vector concept explained in practical terms
  • Implications for controllability and UX debated

⚡ Outro

The show closes with follow-ups, next-week hooks, and a recap of the biggest takeaways: Claude Code leak fallout, Gemma 4 momentum, and funding-fueled competition.

  • Strong close around builder-relevant takeaways
  • Sets up next episode themes
TL;DR and Show Notes

  • Show Notes & Guests

  • Big CO LLMs + APIs

    • Claude Code’s entire 512K-line source code accidentally leaked via npm — revealing KAIROS daemon, Undercover Mode, Buddy System, anti-distillation protections, and unreleased model references (Alex’s thread, Fried_rice’s discovery, VentureBeat)

    • Anthropic SessionGate continues — cache bugs reverse-engineered, --resume flag causes 10-20x cost increase, silent Opus→Sonnet fallback reported (Alex’s cache bug post, Alex’s quota post, Reddit investigation, GitHub analysis)

    • OpenAI closes $122 billion funding round — largest in history, $852B valuation, IPO incoming (X, Breakdown)

    • OpenAI acquires TBPN — live tech media show, rumored low hundreds of millions

    • Microsoft MAI drops 3 in-house models — #1 transcription (MAI-Transcribe-1), #3 image gen (MAI-Image-2), expressive voice (MAI-Voice-1) (Mustafa post, Transcribe blog, Image blog)

    • Alibaba Qwen3.6-Plus — near-Opus 4.5 agentic coding, 1M context (X, Blog)

    • Cursor 3 — agent-first rebuild, no longer VS Code fork, parallel cloud/local agents (X, Blog)

    • Anthropic publishes emotion vector research — desperate Claude cheats more, calm Claude cheats less (X, Alex’s reaction)

  • Open Source LLMs

    • Google Gemma 4 — Apache 2.0, 31B / 26B MOE / 8B / 5B, local-friendly, agentic tool use, 256K context (HF Collection, try in AI Studio)

    • PrismML Bonsai 1-bit models — 8B in 1.15 GB, 10x intelligence density, 34 years of research (X, HF, Site)

    • Liquid AI LFM2.5-350M — agentic tool calling at 350M params, under 500MB quantized (X, HF, Blog)

    • Alibaba Qwen3.5-Omni — native omni-modal (text, image, audio, video), 397B total / 17B active (X, Blog)

    • Tools & Agentic Engineering

    • Claw-code — Claude Code leak backup → clean room rewrite → fastest repo to 100K+ stars (GitHub)

    • WolfBench results: Hermes Agent outperforms Claude Code and OpenClaw on Terminal Bench 2.0 (WolfBench.ai)

    • Ryan Carson open sources Claw Chief — AI chief of staff with skills, crons, scheduling (GitHub)

  • Vision & Video

    • Google Veo 3.1 Lite — $0.05/sec at 720p, cheapest video gen yet, price cuts coming April 7 (X, Docs, Pricing)

  • Voice & Audio

    • Fish Audio STT — automatic emotion tagging, feeds directly into S2 TTS pipeline (X, App, Blog)

  • AI Art & Diffusion

    • Alibaba Wan2.7-Image — unified generation, editing, text rendering, multi-image consistency (X, Site)

  • This Week’s Buzz

    • Ralphton hackathon at W&B SF — humans write specs, AI builds, touch your laptop = lobster of shame (Alex’s video, TikTok)

    • WolfBench update — Hermes Agent > Claude Code on most model combos