Horizon Lens — 12 November 2026
A garage fridge and a pair of speakers put Home Assistant’s new box to work
The Verge’s Nathan Edwards had two jobs for the $69 Home Assistant Connect Aux-2. He wanted a reliable temperature reading from the garage fridge and a way to play music through the speakers built into his living room. His hands-on account gives a sense of what the little box adds to an existing Home Assistant setup.
For the fridge, Edwards enabled the box’s Bluetooth bridge and connected Govee sensors. In his house the connection held through three walls and the fridge. For music, he added an amplifier and configured Music Assistant. He got audio from the speakers, although grouping speakers across different protocols was still an open question in his test.
Analysis
Those two successes involved different purchases. Someone who already owns compatible sensors might value the box mainly for where its radio can sit. Someone with passive speakers needs to account for the amplifier too. The review’s range result is one household’s experience.
Action
Before choosing a spot for the box, sketch the path from sensor to receiver and from audio output to speakers. Pick a location with power or powered Ethernet, then try one intended connection where it will actually be used.
A smaller search index can learn what to keep
In a November 11 Hugging Face article, Tom Aarsen investigated a practical search problem. A document can contain the answer to different parts of a query in different passages. Keeping many numerical representations, called vectors, lets the search system match those parts individually. Across millions of documents, that flexibility takes considerable storage.
Aarsen merged similar vectors and trained an encoder with that compression in mind. On NanoBEIR, at roughly a quarter of the original vector count, the specially trained model retained 97.8% of its own uncompressed ranking score. The normally trained model kept 94.6% of its own score at the same budget.
Analysis
The denominator matters here. Retaining 97.8% of a benchmark score tells you how much ranking quality survived compression in this experiment. It leaves open which particular queries got worse. One missed document could matter a great deal if it contains the detail somebody searched for.
Action
Try both training approaches with the same vector budget on your own documents. Alongside the overall ranking score, inspect queries whose answers span two passages. Measure the size of the resulting index before deciding how much storage the change saved.
New York’s broadband settlement takes aim at the speed after the cap
Ars Technica reports that New York reached a settlement with T-Mobile over fixed wireless home internet plans for income-eligible households. State officials alleged that a monthly 50 GB threshold effectively undercut the service on offer. In the settlement’s account, customers on the 25 Mbps plan could see their maximum download speed fall to 600 Kbps after crossing it.
According to Ars, T-Mobile agreed to remove the limits, expand availability and make the plans easier to find. The settlement gives the company 120 days to remove the speed restrictions and extend availability. T-Mobile neither admitted nor denied the state’s findings. The agreement concerns service at a fixed address.
Analysis
For someone comparing plans, a downstream speed on the first day of the billing month answers only half the question. A household that passes a usage threshold needs to know what connection it has for the rest of that month. The difference here was large enough to change that calculation.
Action
If you’re comparing home broadband offers, put the advertised speed, monthly allowance and speed after the allowance on the same line. For this New York offer, check availability at your address as the agreed changes take effect.
A recycling robot has to make the pick before the belt moves on
In a feature dated November 10, NVIDIA describes how two members of its startup programme tackle sorting on a moving conveyor. At a pilot plant in Tainan, Taiwan’s AIRECO uses computer vision to guide robotic arms selecting material from mixed plastic scrap. NVIDIA reports recognition accuracy above 95% on the test line.
NVIDIA also describes EverestLabs, whose robots work in recycling facilities across North America and Australia. Its chief executive says the belt speed sets a hard time limit for each pick. EverestLabs estimates its systems have identified about 40 billion objects and recycled roughly 25,000 tons of material to date.
Analysis
Seeing the right bottle and getting it off the belt are different accomplishments. A recognition score describes the first. A plant operator also cares whether the arm reaches the bottle in time and whether the resulting batch is clean enough to sell for the intended use.
Action
When comparing sorting systems, ask for three numbers from the same trial. How many target items passed the camera, how many did the arm recover, and how much usable material came out? Record the belt speed so the trial can be repeated under comparable conditions.
What a $62 document job tells a law firm about AI costs
Sam Nickless, chief executive of Australian law firm Gilbert + Tobin, described a document task in an interview published by OpenAI. During the annual review of 120 partners, he says he used ChatGPT Work to transfer spreadsheet data into files, produce PDFs and assemble the finished documents.
Nickless says the work took minutes and the usage dashboard showed a cost of about $62 for the job. OpenAI’s account says people at the firm remain responsible for review and approval. Nickless also describes introducing a controlled environment with Australian data residency and access controls tailored to people’s roles.
Analysis
The dashboard figure is a price for running a particular piece of work. Turning it into a comparison with the previous process requires a second set of numbers, including who prepared the spreadsheet, how much checking the PDFs needed and how long it took to approve the pack.
Action
For a first trial, take a small batch of sample records, generate the documents and have a reviewer check every entry against the spreadsheet. Log both the usage charge and the reviewer’s time. That gives the next team two figures it can compare with its own process.