Snap puts Salesforce and Nvidia inside its $2,195 AI glasses
The machine is the easy part. Today's news is about who the AI is wired into — a workplace, a broadcast regulator, and a loop of startups selling the tools to automate AI research itself.
Snap will ship its Specs augmented-reality glasses this fall at $2,195 with Salesforce's Agentforce, Nvidia's vision stack and Amazon Web Services' assistant built in, and a cellular charging case that takes the bundle to $2,395. Salesforce is putting its AI agent platform on the device so workers can reach company data and automate tasks through the glasses; Nvidia handles the part that reads a worker's surroundings and pulls back the relevant company data; AWS supplies voice-driven tasks like checking product availability. Trifork and Hololight round out the partner list, and telecoms including Verizon, Orange and BT's EE provide connectivity through the case. Snap did not disclose financial terms, and shipping starts later this fall in the US, UK and France, with a Los Angeles retail trial in October.
The same event produced the more interesting product. Specs Intelligence is Snap's AI assistant, launching alongside the hardware with an iOS preview live Wednesday and a Mac waitlist behind it — and it is pitched as anticipatory rather than reactive: it builds a model of your goals, relationships and routines from the accounts you connect, then surfaces what needs attention before you ask. Snap's own examples are mundane and specific — a meeting brief with the decisions to make and questions to ask, a trip where flight, hotel and dinner plans arrive together with the work deadline that falls inside the visit. It is the same bet the rest of the field is making on proactive agents, sold through hardware instead of a chat window; we covered the Meta version in Meta ships Muse: a consumer agent inside a sealed VM.
Why the enterprise framing is the honest one: a $2,195 headset does not have to win the consumer market to pay for itself. Aimed at factory floors, retail counters and field service — environments where a desk was never an option — the value case is hands-free access to company data, and the glasses only need to beat a ruggedized tablet. That is a much lower bar than beating a phone, and it is the first time Snap has named partners who can actually put the device in front of workers.
China's broadcast regulator will require content labels on anything AI-generated and human sign-off at key production stages, with AI alteration of existing programming prohibited outright. The National Radio and Television Administration announced the rules at a State Council press conference on Wednesday, framing them as the boundary line for the next five-year plan: keep applying AI across planning, production, review and distribution, but label it and keep a person accountable for the finished program. It is the cheapest control a regulator can impose — no model approval, no compute limits, just disclosure and a name on the file — and it lands after AI-made drama reached Chinese prime time, which we covered here — China's prime-time TV slot now belongs to an all-AI drama.
The clearest evidence that recursive self-improvement is now a funded category is who just raised money to sell the tooling: Inherent, founded by former DeepMind researchers Edward Hughes and Louis Kirsch, and Recursive Superintelligence, led by Richard Socher and Tim Rocktäschel. Recursive holds more than $650 million at a reported $4 billion valuation, and Rocktäschel frames the goal as breaking what he calls the information barrier — the point where new knowledge arrives faster than humans can integrate it. Both companies are selling the same missing piece: a self-improvement loop needs experiment design, evaluation harnesses and reproducible replication before it can spin without a person in the middle. We covered Inherent's first benchmark claim in August — DeepMind alumni's Inherent says its Faraday agent beats GPT-5.5 at reproducing research — and the vocabulary this debate was missing arrived with the five-level RSI ladder in Today in AI — September 16, 2026, which sorted 491 papers and found under 6% at the top rung.
What to watch: whether Snap's enterprise partners convert into shipping volume, how the broadcast labeling rule is enforced once AI-generated drama is the cheap option, and whether Inherent or Recursive Superintelligence publishes a result an outsider can check.
Is an agent that acts before you ask the feature that makes wearable AI work — or the one that makes it impossible to audit? Tell us in the comments.
