Seven days that tell a single game. Frontier labs ask to slow down, sign joint appeals on oversight, then set up a body that writes the rules in place of legislators. In the same stretch of days, a model that does not produce text and costs a fraction of an LLM comes out, Perplexity brings the agent onto the PC disk, Google opens the home to agents via MCP. The tension between governance and product has never been so visible. And for those building on top of these platforms, the operational question remains the same: where workloads run, at what cost, with how much dependence on a single provider.

Image generated entirely with AI.

Who writes the rules now?

Sam Altman told Fortune that OpenAI is not going public in 2026. The official motivation is safety, and the key phrase is the one about the "unwise moment" to go public. The New York Times had already reported in June on a delay to an IPO potentially worth a trillion dollars.

Operationally, the move has a simple reading. Delaying the listing means a roadmap that still answers to technical goals rather than the quarterly report. For those integrating the APIs, longer timelines translate into fewer compatibility breaks, deprecations announced in advance, changelogs readable before updating a production workflow. Useful, concretely.

The comparison with Anthropic is the most honest part of the story. Dario Amodei writes an essay asking to slow down, Anthropic starts promoting its IPO in mid-October, Nvidia ready to put 10 billion into the listing. Two companies preaching the same principle while one runs toward capital.

On September 14 the joint call arrives. Altman, Musk, Hassabis agree with Amodei on independent oversight, external evaluations conducted by parties who do not answer to whoever pays their salary. On September 16 OpenAI confirms it is working with Anthropic and Google on an AI standards body, based on the FINRA model proposed by Hassabis. Testing of the most powerful models before release, financed largely by industry.

Aidan Gomez, CEO of Cohere, calls things by their proper name: under the label of safety, the dominant labs are forming a "cartel" that writes the entry rules. If those who build the models also certify the models, the conflict of interest is structural. Those who enter the club write the rules, those who stay outside pay for compliance.

The most interesting technical counterargument comes from Peyman Milanfar of Google. Recursive self-improvement is a naive hypothesis, feedback loops are unstable by construction, and the natural brake is already there. In real workflows, an agent that self-optimizes without constraints degrades in two weeks. Damping is needed, a margin that seems wasted is needed.

A body with access to weights and logs is worth ten signed press releases.

The practical consequence for those building automations is to keep a multi-provider plan B and treat open weights as a portable asset. Dynamic routing between models remains the simplest defense against a regime of layered licenses with mandatory audits, where the cost of compliance falls entirely on the small players.

Does local compute change costs?

Perplexity brings Portable Computer to Windows. The agent that once lived in the cloud now runs on the machine, accelerated by NVIDIA RTX GPUs, and plans multi-step tasks without leaving the device.

The technical point is simple: local models have reached sufficient capability for real orchestration. Inference on the PC zeroes out per-token costs and network round-trip, two items that weighed on every agentic automation. Sensitive data stays on disk, a detail that matters in regulated contexts such as finance, healthcare, and legal.

Hardware becomes the real enabler. A local agent that plans consumes VRAM and memory bandwidth, so the high end of RTX becomes the bottleneck. With an adequate GPU you get low latency and zero marginal cost. It is the calculation that pushes companies toward local hardware, and the trend did not start today.

The timing remains curious. While labs ask to slow down, NVIDIA brings compute to users' desks. The game is played on where models run and how much they cost, not on statements.

Is the home becoming an agentic system?

Google has opened early access to the MCP server for Google Home. Agents that support MCP - Claude, ChatGPT, Hermes, OpenClaw, Antigravity - can connect to home devices securely and read event history. Natural-language instructions, clip summaries, custom-built dashboards.

The setup goes through a Google Cloud project configured to use Home MCP. The path is for tinkerers disguised as a consumer feature: configuration, permissions to grant by hand, distribution limited to the USA and Google Home Premium Advanced subscribers at 20 dollars per month. The free tier and the segment that experiments most stay out.

The operational value is that event history becomes a queryable dataset in words. Camera clips and home events are structured data, and an agent that reads them can generate daily summaries, conditional alerts, dashboards. It is the same pattern entering business workflows, applied to the living room.

Image generated entirely with AI.

The direction is clear: the agent becomes the orchestration layer above devices, and MCP leaves the developer niche. Those who already have an MCP environment in the company can replicate the pattern without waiting for the extension in Italy.

Is shadow AI already in companies?

An employee opens ChatGPT, pastes a contract, generates a summary, uploads a CSV to a tool found online ten seconds earlier. No ticket, no vendor assessment, no trace in the record of processing activities. It is shadow AI.

It grows for a banal reason: it takes a browser and a credit card. The promise of productivity beats any internal policy written six months earlier. Risks concentrate on five fronts: personal data ending up in others' training sets, GDPR compliance that is hard to demonstrate after the fact, unevaluated software supply chain, reputation, cybersecurity. The cost of inaction here is not theoretical: without internal oversight, every prompt pasted into an unapproved tool remains invisible until it becomes an incident to report to the Italian data protection authority. In the projects I follow, the starting point is always an audit of AI flows in your company, to map where data exits before choosing controls. Without that picture, you build the gateway in the dark.

Banning works zero. If the approved tool is slow, expensive, and without the features needed, people use the pirated one from their personal phone. The operational response is an internal LLM gateway with SSO, logging of every prompt, cost tracking per user, self-hosted models for sensitive data. In the stack this translates into four pieces: an egress proxy to intercept calls to external APIs, DLP on outbound payloads, a catalog of approved tools labeled "safe for customer data", and telemetry that makes the phenomenon visible before it becomes an incident.

If the approved tool is more convenient than the pirated one, shadow AI empties itself. If it is not, no policy stops it.

Does Jev rewrite inference economics?

Diogo Almeida left OpenAI after contributing to ChatGPT and inventing RLHF, the technique that made the LLM era possible. His conclusion is blunt: we are optimizing for human language while computers speak another language. Two years ago he founded TypeSafe AI, this week he released Jev, a transformer that is not an LLM and does not produce text.

Jev returns calibrated probabilities. Output tokens are free, inputs are paid by the billion, not the million. It cannot hallucinate, because outputs are defined upstream by the user. Vercel replaced an OpenAI model used as a safety classifier and got responses 5 to 18 times faster with higher accuracy. The CTO of Bryo AI compared Jev with Gemini on email classification: Gemini slightly more precise, but 10 to 20 times more expensive.

Jev's natural place is the guardrail. Monitoring LLM agent traces, preventing jailbreaks, doing real-time model routing: activities where a classic LLM burns budget for no reason. A real confidence score changes how policies are written. 50% is a coin flip, 95% is an action, and the threshold shifts accordingly.

The Jevons paradox applied to intelligence says that the collapse in the price of a resource multiplies its use. It is the same dynamic behind the collapse in inference costs over recent months and the new economic calculus of agents. When a decision costs almost zero, you stop choosing which processes to automate and automate them all.

Skepticism remains warranted: closed architecture, 100% synthetic data, "System One" positioning. In a narrow domain this combination works very well, and that is enough.

The tools of the week

• Jev (TypeSafe AI): non-LLM transformer for calibrated decisions. Free output tokens, guardrails at near-zero marginal cost.
• Perplexity Portable Computer: local agent on Windows, multi-step tasks with data staying on the device.
• Google Home MCP Server: early access to control lights, thermostats, and cameras in natural language from Claude, ChatGPT, and other agents.
• Kimi K3 on Amazon Bedrock: open-weight with native vision, one-million-token window, and explicit prompt caching.
• SageMaker HyperPod Inference Gateway: Kubernetes add-on to route inference requests based on real GPU signals.
• Claude Code Projects: parallel cloud threads coordinated by a supervisor agent that opens pull requests and runs tests on shared memory.
• TimesFM-3: 330-million-parameter forecasting from Google Research, fills future points in one shot.
• Ellf (beta): Explosion's platform for agentic NLP, looking for beta partners for information extraction and pipelines.

The thread of the week is the same on every front. Those who decide where models run and at what cost control the game, much more than those who sign a press release about safety. Governance slows on paper, but compute moves to the disk and the price of a decision collapses. That is where things are built, not in press releases.