The AI conversation split between new capability and new accountability. StepFun introduced a 600-billion-parameter mixture-of-experts model aimed at agentic and professional work, while new research highlighted cheaper self-improving coding agents and the performance gains available from inference-time scaling.
Policy and safety moved in parallel. A lawsuit alleges that major AI labs coordinated to slow development, while industry figures argued over whether stronger oversight, AI-powered cyberdefense, or direct government evaluations offer the better path.
The practical layer keeps broadening: model-powered 3D creation, app-connected assistants, and workflow harnesses are moving AI systems from chat toward execution. The recurring business lesson is that model quality alone is no longer the whole product; orchestration, evaluation, cost, and domain context increasingly determine whether an agent is useful.
A new federal lawsuit alleges that Anthropic, OpenAI, SpaceXAI, and Google made an illegal agreement to slow AI development. The claim puts industry safety coordination under antitrust scrutiny and could force courts to distinguish legitimate risk management from prohibited coordination among competitors.
TechCrunch examined two viral AI-safety conversations and the difficulty of separating verified incidents from exaggerated or incomplete claims. The episode matters because policy and public trust increasingly depend on evidence quality, especially when dramatic model-behavior stories spread faster than technical details.
StepFun introduced Step 5 Preview, a 600-billion-parameter mixture-of-experts model with 27 billion active parameters and a one-million-token context window. The company positions it for software engineering and professional knowledge work, particularly finance, adding another serious entrant to the market for long-context agentic models.
Researchers from MIT and Sakana AI presented Self-Improvement via Fast Tree-search, or SIFT, for improving coding agents. The reported result reaches 35.1% on Polyglot with o3-mini while using one-tenth of the CPU hours of DGM, pointing toward more compute-efficient agent self-improvement.
White House AI and crypto adviser David Sacks argued that AI-powered cyberattacks should be met with AI-powered cyberdefense. The statement captures a central policy divide: whether frontier risk is best addressed by slowing capabilities, creating stronger oversight, or accelerating defensive systems alongside offensive ones.
Selected posts from the edition, with links to the originals.
“Introducing Step 5 Preview: Advancing the Pareto Frontier. Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance.”
View post on X ↗“We asked for your wildest 3D builds with GPT-6 Astra. You delivered.”
View post on X ↗“The solution to AI-powered cyberattacks is AI-powered cyberdefense.”
View post on X ↗“Inference scaling part 1. Starting with a modded text generation function to generate diverse outputs for self-consistency and best-of-N, improving answer accuracy by more than 2x.”
View post on X ↗“AMD found a business opportunity in dumping US military chips for 1/4 of the cost in China. US list price is $36k for this chip, with $4–5k to US companies at volume, but it's getting quoted $1k in China.”
View post on X ↗“Seems like one of the most important research problems for CS academics is LLM-supervised peer review. If we don’t solve it the academic institutions are toast.”
View post on X ↗“One future is that we muddle through on most risks and the positive impacts that early data hints at — higher productivity, more scientific discovery, higher entrepreneurship, employment stable or growing — continue.”
View post on X ↗“You can now connect apps from the Computer homepage composer. Live now on the web for all Computer users.”
View post on X ↗“Databricks' Ali Ghodsi says, ‘Right now the existential risk is close to zero.’ Why he thinks the public doom talk is irresponsible.”
View post on X ↗“A great example of practical AI solving real-world problems! Faster than ever. Thanks for building with Qwen, @cerebras.”
View post on X ↗