Monday-Thursday – ☕️ I’m in an Francisco, attending and moderating two sessions at HumanX – come say hi if you are there. ☕️

Wednesday / AI 101 series: Gemma 4 and why many OpenClaw users are now switching to it (basically, an Ode to Open Source!)

Friday / Our new series, The Org Age of AI, is getting a lot of traction. This time, we’ll discuss how small teams (1–10 people) should think about organizational questions.

From Hierarchy to Intelligence by Jack Dorsey. It’s a good reading along The Org Age of AI, we will be discussing this article in more detail in the future episodes.

Andrej Karpathy’s tweet “LLM Knowledge Bases“ has 16 million views. And for a reason. Here you can find both that post and Andrej’s update. Will be covering it in more details as well later.

Vibe Maintainer by Steve Yagge

OpenClaw – Now With a Studio
Very impressive update: OpenClaw 2026.4.5 arrives with built-in video and music generation, a real /dreaming feature, structured task progress, better prompt-cache reuse, and support for 12 more languages across its UI and docs. It also expands its model bazaar: image generation via Comfy, fal, Google, MiniMax, and OpenAI; music through Comfy, Google, and MiniMax; video from practically half the industry. A tidy reminder that “bring your own model” can become an empire if left unsupervised.

Meanwhile, Anthropic

• Declared War on the Harnesses

Users of third-party agent platforms – including OpenClaw – started receiving emails: subscriptions "weren't designed for these usage patterns," and harness usage will be billed separately at undisclosed rates. Boris Cherny confirmed the changes. The timing: OpenAI is actively courting the same developer audience. Anthropic introduced Claude Channels in March as the official integration path. Is this about compute costs or about controlling the agent layer?

• goes big on compute (very big)
  Fresh off throttling demand, Anthropic is doubling down on supply – signing a multi-gigawatt TPU deal with Google and Broadcom. With revenue reportedly surging past $30B (annualized) and enterprise adoption accelerating, this is less “capacity planning” and more “industrial-scale ambition.” The message is clear: if AI is the new electricity, Anthropic is buying the power plants early.

OpenAI

• War Chest, Fully Loaded
  OpenAI just secured a staggering $122B round at an $852B valuation. With $2B in monthly revenue and a fast-closing march toward 1B weekly users, this is less “startup” and more sovereign entity. The strategy: more compute, better models, tighter flywheel – repeat until intelligence becomes a utility.

• Buying the Narrative
  OpenAI acquires TBPN, the fast-rising tech talk show, in a move that blends media, influence, and distribution. Crucially, TBPN keeps editorial independence–credibility intact, megaphone amplified. OpenAI is trying to shape the conversation around them. In this cycle, controlling the story may be nearly as valuable as controlling the stack.

Hugging Face is filling the gap

Clem Delangue, HF’s CEO, started publishing production agent traces this week – real workflow logs from actual agentic tasks. This is the actual chain of reasoning a model uses when it plans, revises, and executes a multi-step task. OpenAI, Anthropic, and Google have been hoarding this data. Open-sourcing it changes the competitive dynamics in ways that won't show up in benchmarks for a while.

Perplexity – Taxes, but Make It Autonomous
Perplexity rolls out “Computer for Taxes,” aiming to tame the annual ritual of confusion and mild despair. With modular, up-to-date tax knowledge, it drafts IRS forms, audits professional returns, and even builds custom tools for complex scenarios. In testing, it caught costly human errors – politely proving accountants aren’t infallible. Filing season is here to prove it.

X Devs – Turning Twitter into an AI Action Layer
That’s a surprising announcements, because for me that was the only use of Grok – to parse X. But let’s see: The repo effectively turns X into something AI agents can use, not just read. Instead of scraping feeds, models can now post, search, DM, and analyze via structured tools. It is trying to make social platforms programmable surfaces for AI workflows →their GitHub

MemPalace – Milla Jovovich Walks Into AI
Yes, that Milla Jovovich. MemPalace, an open-source AI memory system she helped build, is suddenly one of the more unexpected names in the AI conversation. The project promises local-first memory, sharp benchmark results, and zero cloud dependence. In a field crowded with predictable launches from familiar labs, this one arrives like a plot twist—half engineering story, half cultural moment →tweet of her partner

Trinity-Large-Thinking by Arcee
Improves multi-turn reasoning and tool orchestration for long-running agent workflows, with a focus on coherence, instruction fidelity, and stability under real production constraints, while remaining fully open-weight for inspection, distillation, and self-hosting →get the model

Gemma 4 by Google
Expands the open model stack with multiple sizes and efficient architectures, emphasizing deployability, customization, and lower-cost inference for real-world products rather than centralized API usage → get the model

Qwen3.6-Plus
Pushes open models toward near-frontier performance on reasoning and agentic benchmarks, making open-weight systems increasingly viable as primary production backbones →get the model

By Microsoft:

• MAI-Transcribe-1
  Turns speech recognition into a native platform capability with production-grade accuracy and latency, positioning transcription as core infrastructure →read their blog

• MAI-Voice-1
  Enables controllable, identity-consistent voice generation, pushing voice toward becoming a primary software interface rather than an output format →read their blog

• MAI-Image-2
  Advances image generation quality and text rendering, integrating visual creation directly into productivity and enterprise workflows →read their blog

GLM-5V Turbo by Z.ai
Extends multimodal capability into interface-level reasoning, enabling models to operate across screens, files, and visual environments within controlled platform ecosystems →read their blog

A Survey of On-Policy Distillation for Large Language Models – Unifies distillation under an on-policy framework where models learn from their own generated trajectories, which connects directly to how agent loops and RL-style reasoning systems are actually trained →read the paper

The Latent Space: Foundation, Evolution, Mechanism, Ability, and Outlook

Meta-Harness: End-to-End Optimization of Model Harnesses – Formalizes harness engineering as a searchable optimization space by iterating over code, execution traces, and prior system behavior, showing that improving orchestration logic can outperform hand-designed agent pipelines across reasoning and coding tasks →read the paper

Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design – Centers verification as the core mechanism of research agents, which directly addresses the trust bottleneck in real deployments. →read the paper

Terminal Agents Suffice for Enterprise Automation – Argues that simple terminal-level agents can already cover a large portion of enterprise workflows, which is a very practical and slightly contrarian claim. →read the paper

MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome – Introduces evaluation that captures both process and results, which is essential if research agents are to be taken seriously. →read the paper

Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants – Simulates user-driven environments to test proactive agents, moving evaluation closer to real interaction loops. →read the paper

FIPO: Eliciting Deep Reasoning with Future-KL Influenced Policy Optimization – Applies RL to shape reasoning trajectories, making it structurally more important than prompt-level improvements. →read the paper

Reasoning Shift: How Context Silently Shortens LLM Reasoning – Shows that more context can degrade reasoning depth, which directly challenges how most RAG systems are built today. →read the paper

The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning – Exposes how shallow heuristics dominate reasoning even when deeper constraints are available, which is important for safety and reliability. →read the paper

Brevity Constraints Reverse Performance Hierarchies in Language Models – Demonstrates that evaluation conditions can flip model rankings, which raises uncomfortable questions about current benchmarking practices. →read the paper

Embarrassingly Simple Self-Distillation Improves Code Generation – Shows that re-training a model on its own sampled outputs improves code generation by reshaping token distributions, offering a lightweight alternative to RL and verifier-based pipelines while exposing the precision–exploration tradeoff in decoding →read the paper

Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models – Addresses temporal memory in video-based world models, which is a key limitation for persistent simulation. →read the paper

EgoSim: Egocentric World Simulator for Embodied Interaction Generation – Builds an egocentric simulation framework, pushing toward embodied agent training environments. →read the paper

VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward – Introduces reward shaping for consistency in generated worlds, which is a step toward controllable simulation. →read the paper

LongCat-Next: Lexicalizing Modalities as Discrete Tokens – Reframes multimodal modeling around a unified token interface, which is a strong architectural bet with broad implications. →read the paper

LatentUM: Unleashing the Potential of Interleaved Cross-Modal Reasoning via a Latent-Space Unified Model – Pushes latent-space unification for cross-modal reasoning, reinforcing a shift toward shared internal representations. →read the paper