Horizon Lens — 13 November 2026
Who decides whether to reroute a flight?
The US Federal Aviation Administration is extending SMART, its traffic management advisory system, to the airspace covered by all 22 of its centres. Ars Technica reports that the FAA announced the expansion on November 10. The system brings together weather, flight paths, traffic and controller staffing to recommend changes such as a different route or arrival order.
The FAA says SMART helped reroute 180 flights after its September introduction over Washington, DC. The agency estimates average savings of ten minutes and nearly 80 gallons of fuel per rerouted flight. Specialists in Washington check the recommendations before forwarding them. Local controllers then decide whether to accept them.
Analysis
Those 180 flights give the pilot results a denominator. A national roll-out raises another question about what happens to all the suggestions along the way. Knowing which ones specialists reject, and which ones controllers decline, would help explain where the tool earns its keep.
Action
When the FAA reports nationwide results, look for the number of proposed reroutes alongside the number flown, then the time saved on those flights. That would let readers compare the pilot with later operations without assuming every recommendation was used.
Training a 3D model to get the soles of shoes right
A team writing on Hugging Face noticed that TRELLIS.2 could generate shoes that looked convincing until someone inspected their soles. The authors used a set of 188 footwear objects from Google Scanned Objects to investigate whether they could improve the shape. They trained small additions to the model components responsible for structure and geometry, leaving the texture component fixed.
The authors compared the results with scans of nine objects held out from training. The mean F-score, a measure of how closely the generated and scanned surfaces match, rose from 0.276 to 0.803. All nine scored better after training. The source gives the matching tolerance as 2% of the diagonal of each scan’s bounding box.
Analysis
This is a good example of choosing a test that can see a particular mistake. An attractive view of a shoe can conceal a misshapen sole. Comparing the generated surface with a scan gives the team a way to tell whether extra training corrected the geometry.
Action
For a trial on a different object, identify one feature whose shape matters to its intended use. Keep some scans out of training, choose a tolerance in advance and inspect that feature as well as the average score.
IBM gives a model separate labels for the task and its input
IBM’s Granite team released two experimental models trained to read LLMON, a markup language that identifies the parts of a request. In the team’s example, a request to write about IBM has a subject, an instruction and two writing constraints. Each part has an identifier, so the model can tell which material the instruction applies to.
The team built the previews from the three-billion- and eight-billion-parameter Granite models. It also released tools to convert JSON into its format and back without losing information. During training, some blocks were deliberately left unreferenced. The model learned to leave those blocks alone while completing the assigned task.
Analysis
That distinction matters when an assistant reads something written by a third party. A document can contain a perfectly grammatical instruction that the reader merely wants summarised. IBM describes the model’s ability to keep those roles apart as learned behaviour and recommends the preview for evaluation.
Action
Try a harmless editing task with a paragraph, a word limit and a punctuation requirement. Check whether the result follows both constraints and still reflects the paragraph. It gives a team evaluating the markup a concrete example it can score by hand.
Muse opens business calls to paying US users
Meta has broadened access to a feature that lets its Muse assistant call US businesses on a customer’s behalf. TechCrunch reports that an initial group tested the feature in September. Meta’s AI chief, Alexandr Wang, gave a familiar example of where it might help, making a dinner reservation at a restaurant that takes bookings by phone.
As of the November 12 report, Muse plans started at $20 a month. A paid US subscriber can now ask for a call without having to request early access. The business still decides how to respond. TechCrunch notes that some people answering business phones hang up when they recognise an automated caller.
Analysis
For someone organising dinner, the value is in knowing whether the table is booked. A service could place the call correctly and still leave that question open if nobody picks up, the restaurant declines, or the requested time is unavailable.
Action
For a first reservation, specify the party size, time and whether an alternative is acceptable. Look for confirmation of the booking details before telling everyone where to meet. Count a call that merely connected as an unfinished task.
What changed when Amplitude personalised its homepage suggestions?
On Amplitude’s homepage, people could start a conversation with an AI agent by picking a suggested question or writing their own. In a customer study published by OpenAI, Amplitude says that out of 100 recorded sessions, 80 users wrote their own questions and 12 picked a suggestion. Its Wave agent proposed tailoring suggestions to recent activity.
The team reviewed the plan, sent it to a coding agent and ran an A/B test. The change kept free typing available and used general suggestions where personalisation was impossible. Amplitude reports that, after one day, messages sent via the cards were 160% higher than in the control group.
Analysis
That result is about how conversations began. If someone who previously typed a question now clicks a card, card usage rises even if the overall number of conversations stays the same. The distinction matters when deciding whether a homepage change helped more people get started.
Action
When testing a shortcut like this, track both the shortcut and the task it leads into. Compare total conversation starts in the two groups as well as starts from cards, and check whether the result holds beyond the first day.