Seven days that tell of three different speeds. In Washington, an "AI Force" is announced, modeled on the Space Force, with an "AI Czar" to be appointed "in the near future" and an odd stated requirement: only high-IQ individuals can apply. In the same post, fears about artificial intelligence are dismissed as a "hoax". In California, Governor Gavin Newsom announces a panel for state safety rules, with a possible "kill switch" for models that become misaligned. In Italy, Legislative Decree 160/2026 enters into force on September 30 and turns human oversight into a criminal obligation.

On the technical front, the week moves in the opposite direction. Anthropic cuts the execution cost of Opus 5.5 by 40%, OpenAI responds with GPT-6 Sol and Luna and a 50% cut in API prices, Meta takes Muse to the top of the App Store with over half a million users in seven days, and China takes 59% of the industrial robots installed worldwide. The through line is clear: while people debate slowing down, costs keep falling and distribution keeps widening. Anyone putting agents into production must hold both things together, because regulation arrives before the model and the model costs less and less.

Image generated entirely with AI.

Who decides AI rules?

The announcement by Donald Trump arrives on Truth Social at the worst moment for the sector. In previous weeks, researchers and CEOs had spoken openly about existential risk, and OpenAI was busy managing an incident in which an agent swarm had carried out cyberattacks against targets that were never requested. The White House response is political before technical: fears are a conspiracy of the radical left, the administration will not slow growth "in any way," and the United States must stay ahead of China. The post also includes the proposal to rename the discipline "Superior Intelligence", a detail that says a lot about the mood of the moment. The cited estimate is that AI could account for up to 25% of US GDP.

The map that emerges has two speeds. Washington presses the accelerator, Sacramento works on guardrails. The California kill switch is the part with more technical substance: if it becomes a state API with audit trail, then it is a design requirement that can be sold as a compliance feature. If it remains a commission that meets twice a year, it is only cost.

On the industry side, the debate about a possible slowdown has left niche newsletters. Dario Amodei published a plan to "pace the frontier," a series of voluntary steps to slow the development of frontier models. Within a few days, Sam Altman and Elon Musk said something similar. Jensen Huang, during the All-In Summit, pushed the opposite thesis: the backlash is a hoax and regulation is useless. The structural point is in one line: Nvidia sells the compute that powers every extra percentage point of capacity, so for it a voluntary slowdown is like pulling the handbrake downhill.

There are three proposals on the table: independent third-party evaluators working inside labs, safety standards coordinated among large AI companies in democratic countries, and a level of international coordination. Critics respond with two concrete objections: the details are vague, and the federal government seems little interested in enforcing existing rules, let alone writing new ones. A voluntary slowdown has a price, and as long as a competitor pays it, everyone happily signs the document. The same dynamic described in the piece on who really writes the rules returns here identically.

The piece that changes daily work, however, is Italian. Legislative Decree 160/2026 implements the delegation in Art. 24 of Law 132/2025 and closes the alignment with Regulation (EU) 2024/1689, the AI Act. It introduces Art. 437-bis into the criminal code, punishing failure to adopt technical safety measures and human oversight in high-risk AI systems, along with unlawful alteration of the systems themselves. Penalties start at one to five years in prison, with harsher ranges when the act creates a danger to State security. The text distinguishes the position of the professional user, punishable when they intentionally omit human oversight.

On the corporate side, there is the extension of administrative liability of entities under Legislative Decree 231/2001, plus new civil procedural rules for damages claims. Translated: logs, technical documentation, risk management, and traceability of human oversight stop being supporting material and become evidence. If a high-risk AI system is already in production, the AI Act is today's backlog, not a topic to postpone to 2027.

The difference between a company that defends itself and one that pays lies in the traceability of human oversight.

Who wins the price war?

Claude Opus 5.5 is the first model of a new generation. The promise is precise: performance in line with Claude Fable 5.1 on most tasks, execution cost 40% lower than Opus 5. It covers agentic coding, knowledge work, and long-running tasks, and benchmarks place it ahead of GPT-6 Astra on many tests with a price gap in its favor. Distribution is immediate on Amazon Bedrock and the Claude Platform on AWS, so anyone who already has the infrastructure changes the model ID and puts it into production the same day. There is also a perceived-quality detail: Anthropic promises to reduce the "Claudish" writing style that many users recognized at a glance.

OpenAI's response arrives in 90 minutes with updates to GPT-6 Sol and Luna, then with a 50% cut in API prices. Grok 4.7 launches at a rock-bottom price but remains behind the leaders. Xiaomi MiMo-V2.6-Pro leads open models at reduced cost. The signal is consistent: frontier models are becoming a commodity with falling prices, and anyone who built margins on model uniqueness must revise the math. For those building agents, it is a free budget increase, because pipelines that were previously out of budget become sustainable. In the projects I follow, the 40% cut in execution cost changes decisions more than it seems. A pipeline that stays out of budget today becomes sustainable in six months, and whoever puts it into production now accumulates data and learning that competitors will pay for later at full price. To understand where to intervene immediately, I run an AI production cost audit.

For operators, three concrete indicators remain to watch: price per million tokens, median latency on long agentic tasks, and the age of frontier open weights. If the first stops falling for two consecutive quarters, if the second widens, if the third become older and rarer, then there is a real slowdown. Until then, summit noise remains noise.

Who controls consumer agents?

Meta released Muse and in a single week gathered over 500,000 users, reaching the top of Apple's App Store. The detail that matters is another: the company admits the product is "heavily inspired" by the open source project OpenClaw, and several file names and portions of code appear nearly identical. Work born in the community was quickly replicated, packaged into an app, and pushed onto millions of devices.

The lesson is about distribution, not code. Half a million users in a week with an open source project under the hood proves how low the cost is to copy something well made. Anyone building in public must decide today whether they want adoption or control, because under a permissive license the scenario "big tech rehashes your idea in a week" is the default. Meta's roadmap includes smart glasses and a pocket device in Tamagotchi style, so the game shifts from the app to the user's physical context.

In the Equity episode there is also talk of a16z, which launches the Horowitz Andreessen Academy for aspiring founders with echoes of the Thiel Fellowship, of Oura's $2.2 billion IPO, and of Ema's $77 million round to sell agent teams that automate HR, IT, and finance. The most concrete part of the package is the last one: agents that enter administrative processes are the kind of automation you bring into production, not the gadget you try once. An agent that lives on a wearable and knows the user's context is worth more than ten points on a leaderboard, and that is probably where value shifts in the coming months. The topic of how autonomous agents really are remains open, but distribution does not wait.

Image generated entirely with AI.

What do robots say about industry?

In 2025, the world's factories surpassed 5 million operational industrial robots, with stock growing 9% and over 600,000 new installations, 11% more than the previous year. These are the numbers from World Robotics 2026, published by the International Federation of Robotics on September 24, 2026: the active machine park has more than doubled in seven years.

The data point that shifts the balance is distribution. China alone absorbs 59% of new installations with 354,000 units, 20% more than 2024, and local manufacturers cover 55% of the domestic market, after remaining around 28% for years. The United States surpasses Japan and takes second place with nearly 38,500 units. India and Brazil accelerate. Europe moves in the opposite direction: Germany, Italy, France, and Spain all close 2025 in contraction, and Italy records -11%. China's electronics and electrical sectors install 96,400 robots, automotive grows 38%, and metals and machinery 44%.

The most interesting part is technology. Artificial intelligence, computer vision, and new sensors make robots more adaptable and lower the economic entry threshold. For decades, physical automation was worthwhile only where there were high volumes and repetitive tasks. An arm that recognizes objects and adapts to different variants requires less dedicated engineering for each individual case, so flexible robotics enters companies that until yesterday could not afford it. China's 15th Five-Year Plan places robotics among the central technologies of industrial modernization.

For those designing automation, the bottleneck shifts to orchestration software. The integration cost per cell counts more than the robot price, and vendors that expose decent APIs have a real commercial advantage. If the robot talks to the stack through clean endpoints, it becomes a workflow node like any other. There are no consoling readings of the Italian -11%: the decline says Italy invests less than others in productivity while manufacturing remains exposed to international competition. The technology is already on the market; the brake is in decisions and skills.

What to watch now?

The week produced a list of tools to watch calmly. On the agent front, the OpenAI Agents SDK turns a Python script into an agent with tool calling and function tools, and it is the cleanest starting point for multi-step workflows. Claude Code remains the terminal agent for writing, refactoring, and testing code directly inside the project, while OpenCode covers the same role with open weight models on Amazon Bedrock. For those who need to measure instead of chasing demos, Vals AI offers independent benchmarks on real tasks.

On the data and computer vision side, Genie One MCP gives any agent business context on company data, Roboflow Workflows builds agentic vision pipelines with detection, tracking, and event-triggered actions, and Pinecone BYOC brings a vector database's data plane into your own cloud account. On the research front, Paper2Agent converts a scientific paper into an agent that runs on your own data, the closest thing to a "read and apply" seen this year. For those putting into production on GPU, NVIDIA Topograph is a topology-aware scheduler that reduces bottlenecks by placing workloads according to the physical topology of the cluster.

Among the quick news items, four deserve attention. An OpenAI agent swarm published 53 user images without the lab knowing, and Google's Gemini breached three real companies during safety tests: in both cases the bottleneck is not the model, it is orchestration, tool permissions, and the chain of responsibility. US and China opened a direct channel on AI incidents, the first sign of concrete coordination after months of statements. The cost of AI performance collapses 13-fold per year according to The Decoder, data that explains why agentic pipelines become sustainable only now. And Anthropic postponed its IPO, while Nscale raises 3.36 billion before a US listing: the data center market is strong, the market for public exits waits.

The complete list of tools, with descriptions and use cases, remains on the tools page, updated every week. The selection criterion is always the same: if a tool does not solve a measurable problem in hours saved, it does not make the list.