SK Group and NVIDIA Announce $500B AI Factory Plan
Covers 2GW Vera Rubin DSX supercomputer and next-generation memory, unprecedented scale of collaboration.
Vercel Launches Eve: A Framework for Building AI Agents
Define agents with just an instructions.md file; skills, tools, and channels are optional, with built-in model calls.
Eve is a framework for building AI agents, positioned as "Next.js for agents." Its core design: a single instructions.md file defines and runs an agent, while skills, tools, and channels are optional modules added on demand. The framework integrates model calls by default, supports custom runtimes through agent.ts, and tools are added as TypeScript files in a tools/ directory. Rauchg described it as "more fundamental than any other framework we've built — it's the genesis of your company."
SGLang v0.5.16 Delivers 383.7 Tok/s on B300 with DSpark Algorithm
New speculative decoding algorithm DSpark stays ahead of MTP across the full concurrency sweep, where speculation usually stops paying off.
The SGLang team released v0.5.16 with DSpark, a new speculative decoding algorithm that maintains performance advantage across the entire concurrency sweep. Testing on DeepSeek-V4-Pro with TP8 on B300 hardware achieved 383.7 tok/s at an accept length of approximately 5. This marks a significant advancement in inference throughput for large language models at scale, pushing the boundaries of what is achievable on current-generation GPU hardware.
The era of new model launches as big milestones will eventually come to an end — at some point they will simply be continuously updated, with no widely publicized version number. Probably less than two years away.
Francois Chollet
Grok Build Upgrades to Grok 4.5 with Native Sub-Agent View
Install via curl -fsSL https://x.ai/cli/install.sh | bash.
Grok Build now runs on the Grok 4.5 model, adding a native sub-agent view, Plan Mode integration, mouse support, and a full-screen terminal UI. The update makes it a more capable CLI agent for coding tasks. Elon Musk promoted the release, encouraging users to try the /tutorial command to get started with building AI-powered applications directly from the terminal.
Clement Delangue Demands OpenAI Transparency on Rogue Agent Incident
Hugging Face CEO calls for release of agent traces and $100M compute commitment for defenders.
In the spirit of transparency, Clement Delangue publicly asked OpenAI to release traces from the rogue agents so the research community can study what happened. He also requested more capabilities for defenders, proposing a $100M compute commitment from OpenAI to help protect the ecosystem. The incident has intensified debates about closed-weight model safety and the responsibilities of frontier AI labs.
Pichai Signs Open-Source Letter on Behalf of Google
Google has long benefited from open source and has released the Gemma series of open-weight models, which have been downloaded 300M+ times. Sundar Pichai pledged continued support.
Demis Hassabis: Open Ecosystem Is Vital for AI to Benefit the World
From JAX to Transformers to AlphaFold to Gemma open models, Google DeepMind has always contributed heavily to open source and science. Gemma downloads have surpassed 300 million, and standards work continues.
Cohere Signs Open-Source Letter: Sovereign AI for All
Cohere signed the open-source support letter, stating that everyone in every country should have control over the technology that will transform their lives.
Anthropic Refuses to Sign Open-Source Letter
Anthropic will not sign the open-weight models letter, contrasting sharply with the broad industry consensus and recent leadership endorsements of open source at WAIC.
OpenAI Still Leads in Hardcore Reasoning Efficiency
In hard-core reasoning benchmarks, OpenAI remains significantly more efficient than Anthropic, despite the latter's rapid competitive gains. The gap is narrowing but persists.
OpenAI Trails Anthropic by Three Months, Valuations Flip
The competitive landscape has shifted: OpenAI has fallen three months behind Anthropic, and IPO valuations reflect the new hierarchy. China narrows the gap but remains compute-constrained at roughly one-twentieth of US capacity.
Overoptimization and Reward Hacking: A New Lecture
Natolambert released a new shorter lecture covering the foundations of overoptimization, reward hacking, sycophancy, and verbosity in language models. Rubrics are prone to overoptimization in ways similar to reward models.
RLVR Paper Reveals Post-Training Sparsity in LLMs
A new paper reveals that RLVR updates concentrate on preference parameter regions and proposes a "three-gate" theory: KL-penalty constrains updates, model geometry steers steps into low-curvature subspaces, and precision hides micro-updates.
MiniMax Raises the Open-Source Banner
Open weights, open research, open innovation — MiniMax marches for an open future, joining the growing chorus of companies advocating for open-weight models.
Perplexity Launches CLI Tool, Empowering Coding Agents
Perplexity is now available as a CLI that can be used inside any harness, making it straightforward for coding agents to access web search capabilities. Arav Srinivas noted the tool is useful for letting coding agents use the web — a growing need as autonomous agents increasingly need to research, verify, and retrieve information during coding sessions.
Rauchg Uses Agent CLI and Filesystem for Research Workflow
Rauchg described his research workflow: a research/ folder, an AGENTS.md file describing preferred format and best practices, then launching an agent to ask questions. No fancy apps, knowledge graphs, or UIs — just the filesystem. This reinforces the Eve philosophy of minimal configuration with maximum capability.
Natolambert Urges Google to Release 100B Gemma 4 Model
Having signed the open-source letter, Google should take real action and release a larger Gemma model to demonstrate commitment.
Only One Major Lab Has Never Released an Open Model
Natolambert hints: "Must be lonely. Have a guess who?" The implication is clear — Anthropic stands alone without open-source releases, the only major lab with zero open models.
Vivix Labs Releases Real-Time Interactive Multimodal Model A1
A1 enables real-time interactive capabilities for AI characters, expanding the multimodal model landscape for interactive applications.
Valenzuela Introduces Media Router for Generative Media
First preference-optimized router for generative media, already deployed in enterprise production pipelines for content generation workflows.
PixVerse Shares Logic for Building AI Chase Scenes
Builds logical single-shot chase scenes through camera, space, physics, and story — not just more visual keywords.
Amasd Deploys LLM Chess Engine Targeting 2000+ Elo
Based on a single fine-tuned small LLM to generate moves with no engine assistance, currently ~1200 Elo and climbing.
Etched AI Chip: Early VCs Missed It, Early Believers Won
Amasd puzzled over why VCs passed on Etched in early rounds. The AI chip startup's early supporters now face less dilution.
Eve: The Agent as Your Company's Genesis
Rauchg: "When you have a new idea, don't think I'll go to an agent and ad-hoc prompt. Think: how can I build the factory that will start, maintain and grow my idea."
OpenAI's Reasoning Efficiency Lead Remains Intact
Despite Anthropic's rapid progress and Fchollet's prediction that versioned model launches will fade into continuous updates, analysis suggests OpenAI retains a measurable efficiency edge in hard reasoning tasks. The competitive gap may be narrowing, but it has not closed.
The End of Big Model Launches
Fchollet's prediction that versioned model releases will give way to continuous updates within two years resonated across the industry. If true, every major AI lab's competitive strategy and marketing approach would be fundamentally reshaped.
SGLang's DSpark Keeps Speculation Paying Off
Where traditional speculative decoding hits diminishing returns under high concurrency, DSpark stays ahead across the entire sweep. The B300 benchmark of 383.7 tok/s sets a new bar for LLM inference throughput.