Horizon Lens — 10 November 2026
How OpenAI stops model output for one research job
OpenAI has built a way to cut off a research job’s access to its models even if the program running the experiment keeps going. Its Control Room assigns the job an identifier. Each time the program asks a model for more output, the engine checks whether that identifier still has permission.
OpenAI says output stops within 35 seconds once a decision to intervene has been made. Other authorised jobs can keep using the same engine. The company says its research infrastructure already uses the system and that losing the Control Room service itself is designed to stop output.
Analysis
The starting point for that clock matters. Someone or something first has to recognise a problem and decide to act. The figure describes how long enforcement takes from that point. A team investigating an incident would also want the time between the first worrying behaviour and the decision.
Action
Try the shutdown drill with two harmless jobs sharing an engine. Stop one, make it ask for another response, and check that the other still receives answers. Record the time of the decision and the time of the last output.
Which parts of a smaller language model should stay together?
Amazon researchers describe a puzzle for anyone trying to fit a large language model into memory. In one Qwen model, each layer picks just eight of 256 specialist components, or experts, for a given token. All the other experts still occupy memory so they can answer a later request.
Their method, HOPE, chooses which experts to remove by looking at pairs that work together. On the SWE-Bench Pro coding test, a smaller Qwen model with half its experts removed scored 42.3 using HOPE. The intact model scored 41.8 and a competing pruning method scored 36.2.
Analysis
It's tempting to read those first two scores as permission to discard half a model. The choice of what went into that half made a difference even on this one test. The authors also found a smaller advantage when they removed fewer experts.
Action
If memory is tight, pick a pruning rate that would fit your machine. Compare both selection methods with the intact model on examples from your own workload. Include the routine requests as well as the coding problems that take several turns.
A million fictional financial records for Japanese customer simulations
Elith and NVIDIA released a Japanese-language dataset for teams testing financial assistants. One million records each carry nine persona descriptions. Drawing on official statistics and public surveys, the authors generated fictional people with details about their households, work and finances. The records can be used commercially under a Creative Commons attribution licence.
Alongside income, the designers included household assets, familiarity with money matters and use of phones and the internet. They deliberately included life situations such as retirement and unemployment. The authors say the profiles were generated from distributions without drawing on any individual customer's records.
Analysis
That combination suggests a specific use in customer service. Imagine two customers with similar finances who ask for help with the same form. One regularly uses a smartphone and the other rarely does. A test can ask whether the assistant explains the next step clearly to both.
Action
Keep a record of which household situations you tested and where the assistant gave confusing directions. Treat those cases as prompts for investigation. To establish how often an actual customer gets stuck, you would need evidence of what actual customers do.
Forty tasks completed at least once tells only half the robot story
NVIDIA introduced GR00T 2, a model for controlling robots, at the Conference on Robot Learning. In the company's tests the model completed 40 of 50 tasks at least once without seeing a demonstration specific to each task. That wording leaves open how many attempts it took to get each success.
NVIDIA also reported an average success rate of 71% across 50 tasks after one demonstration per task, rising to 95% with five. One example going beyond the headline numbers comes from MicroAGI, which used additional reinforcement learning to get a robotic arm to swap Ethernet cables.
Analysis
The first measure tells an engineer how wide a range of jobs the model managed at least once. The later percentages answer a different question about repeat success after demonstrations. A buyer interested in one particular job still needs to know how the arm performed on that job.
Action
For a cable-handling trial, write down how many demonstrations the system gets, how many times it attempts the swap, and how many swaps it finishes. Keep any further training in the trial record so someone else can interpret the result.
What would cheaper filmmaking actually change?
Paramount Skydance chief executive David Ellison used an earnings call to make an argument about the kinds of films studios can afford to back. As The Verge reports, he pointed to The Terminator, saying a film that cost $4 million would cost $60 million to $70 million to reproduce shot for shot today.
Ellison said the company has been developing AI tools and expects them to speed up production. His argument for original stories is that a smaller budget makes taking a chance easier. His figures for The Terminator describe what he thinks a remake would cost. Assessing any savings from AI would require a separate production budget.
Analysis
There are two decisions in that argument. A producer can finish the same picture for less money. A studio can also use the lower expected cost to approve a picture it would otherwise decline. Only the second decision changes which stories reach an audience.
Action
When the studio offers a production example, look for the work that changed, its effect on the budget and whether the revised budget affected approval. Those details would let a filmmaker assess Ellison’s argument on its own terms.