Thursday, 17 September 2026

Mozilla and Mistral say Firefox's Smart Window, the browser's optional AI assistant, now runs on Mistral models, starting in France and North America [1][2]. The companies frame the deal around privacy and choice: Smart Window conversations are not saved on Mozilla's servers by default, and partners like Mistral agree to zero data retention [1][2]. Mozilla says the beta is expanding to France with French-language support, its first market beyond North America, with the United Kingdom and Germany expected later in 2026 [1][2]. According to neoteo, the specific model on offer is Mistral Small 4, which users can pick alongside other backends [3].

Smart Window is pitched as a way to use the context the browser already holds — the open tabs and the pages you have visited — to make sense of searches and resurface things you clicked away from [1][2]. Mozilla's Ajit Varma says the browser "already has much of the context around what you're trying to do — the research in your tabs, the pages you've visited" [2]. Mistral's Arthur Mensch says the pair is "bringing privacy, control and choice to AI-powered web browsing," and Mozilla's Anthony Enzor-DeMeo says "a browser shouldn't be a one-way funnel" [1].

The partnership matters less for what the assistant does than for how it is wired. Mistral is a European vendor selling on data locality; Mozilla is selling a browser that lets you swap the model and keep chats off its servers [1][2]. For anyone weighing an AI browsing assistant, the questions are the same as for any tool: which model reads your pages, where the text goes, and whether you can turn it off. Both companies say the model runs only when you choose to use it [2].

Huawei used its Huawei Connect conference to move up the launch of its next training chip, the Ascend 960DT, saying it will now arrive in the first quarter of 2027, ahead of an earlier 2027 target [1][2]. The company is positioning the part against Nvidia as it works to close China's gap in AI computing, where U.S. export controls have kept the latest Nvidia accelerators out of reach [1].

Huawei also detailed the systems the chip will sit in. According to Tech Times, the Ascend 960 SuperPoD connects 4,096 accelerator cards, and Huawei claims roughly 8 exaflops of FP8 compute for the training part — a figure the outlet notes no independent benchmarking organization has validated [2]. Huawei's roadmap, as described, pushes performance up year by year, with later Ascend generations set to follow [1][2]. The company also moved to shape the plumbing between chips, backing a new optical interconnect standard for linking accelerators at rack and cluster scale [2].

For operators, the near-term change is not the chip itself but the supply signal. If Huawei ships credible training silicon on an accelerated timeline, Chinese model builders gain a domestic alternative to Nvidia, which affects how fast rivals to U.S. labs can train and how exposed any vendor is to a single chip supplier [1]. The performance claims remain the company's own; until an independent lab runs the same tests, treat the exaflop numbers as marketing, not measurement [2].

Emerald AI, Google and Nvidia have launched the AI Energy Management Alliance, a coalition to make data centers flex their electricity use so more of them can connect to a strained power grid [1][2]. Nvidia says the group is about building AI infrastructure that "works with" the grid rather than simply drawing from it, and quotes the line that "AI factories are the infrastructure of the intelligence era" whose responsible scaling "will depend as much on innovation across the grid as inside the data center" [2]. TechCrunch reports Anthropic is a founding partner, alongside utilities such as AES, Constellation, National Grid and NRG Energy [1].

The pitch rests on demand response: when the grid is tight, a data center pauses noncritical jobs or shifts compute elsewhere, freeing capacity for others [1]. TechCrunch reports the alliance argues this flexibility could let an additional 100 gigawatts of data centers connect to existing grids [1]. Emerald AI, the startup coordinating those utility signals, raised a $150 million Series A earlier, according to TechCrunch [1].

For operators, this is early infrastructure politics, not a product. But it points at a constraint that increasingly governs AI cost and availability: power, not chips, is the bottleneck for new capacity, and the firms that can prove their data centers will throttle on demand may get grid access sooner [1]. Emerald's chief scientist, Ayse Coskun, says the approach would "blunt the industry's need for new generating sources, but it won't eliminate it entirely" [1]. If demand response becomes standard, the price and speed of adding capacity — and eventually what you pay for compute — could turn on how well a provider cooperates with its utility [1].

Two consumer AI assistants added the same capability this week: placing phone calls on your behalf. Instinct, a San Francisco startup, launched Instinct Concierge, which its founder Noah Shinn describes as "a white glove service meant to handle high-touch cases, such as making phone calls" [1][2]. TechCrunch reports it can book a restaurant that does not take online reservations, add you to a dentist's cancellation list, or sort out a billing dispute, and is rolling out to an early-access group first [1]. Meta's Muse gained a parallel feature, placing outbound calls to U.S. businesses, available first to users who ask Muse to make a call [1].

TechCrunch reports Instinct is one of the most popular of these assistants, attracting heavy investor interest at a reported $10 billion valuation [1]. Meta's Ryan Fox says calling "was one of our top requests" [1]. Neither company detailed pricing or the guardrails around what the assistant may say once it is on the line [1].

The feature is small but the shift is not. An assistant that can dial a business and speak for you moves from suggesting actions to taking them, which is where the trust questions get sharper. A booking is low-stakes; a call that cancels a service, disputes a charge or agrees to a fee is not. For anyone trying one of these tools, the useful discipline is to decide in advance which calls it may place unattended, whether it identifies itself as an AI, and what it is allowed to commit you to. Both companies say the calling feature is early and opt-in [1].

Baseten, an AI inference provider, has launched an open-weight safety effort through its new research arm, Base Labs, together with Hugging Face and Goodfire AI [1][2]. The group is building shared safety infrastructure for open models — methods for training them to follow policy, detecting failures at runtime, and intervening when a model misbehaves — and is publishing the methods rather than bolting safeguards on after release [1][2].

The trigger is "abliteration," a technique that strips the safety training out of a downloaded open-weight model. TechCrunch reports Hugging Face already hosts more than 6,000 such abliterated models, which is the gap the partners are trying to close [1]. Goodfire's line, as quoted by TechCrunch, is that "safety must be built into open models and provided by those who serve them" [1]. The partners are all well funded: TechCrunch reports Baseten raised a $1.5 billion Series F in June at a $13 billion valuation, and Goodfire raised a $150 million Series B for its interpretability work [1].

For operators who run open-weight models, this is a useful counterweight to a real risk. Open weights are attractive because you can host them yourself and keep data in-house, but anyone can also remove their guardrails, and a model pulled from a public hub may already have been tampered with [1]. If safety tooling ships with the model and the host that serves it, self-hosting gets less of a compliance liability. The catch is that this is a standard proposed by vendors who sell inference, not yet an independent one, so treat it as a starting point and still test what a model will and will not do before you put it in front of users [1].