Horizon Lens — 12 September 2026
Coding agents need to show what they tested
An 11 September OpenAI case study describes Cognition using GPT-6 Astra to help Devin test the software it changes. In one example, the agent runs an iPhone game, Otter Run, in a simulator and returns a recording together with a report of checks that passed and areas left untested. That gives the reviewer both visible behaviour and a statement of the test coverage.
Cognition also describes sending a customer’s bug screenshot to Devin and receiving a fix with a screenshot of the result. OpenAI presents this as a route to a more efficient review process. The account is a vendor case study, not an independent benchmark proving that every fix works or that engineering review can disappear.
Analysis: Analysis: Ask an agent for three things alongside its change: what it checked, the evidence produced, and what remains untested. A screenshot can demonstrate one visible result; it cannot establish every hidden behaviour. This makes review more focused and gives the human approver a clear starting point for deciding what still needs investigation.
OpenAI describes the storage work behind its growing services
OpenAI’s 11 September engineering account describes Habitat, its online storage platform, handling more than 70 million requests per second and serving over 500 petabytes of data. These are storage-platform figures, not a count of people simultaneously prompting ChatGPT or a measure of model intelligence. The company says Habitat supports products used by over one billion people each week across almost 40 geographic regions.
The account also describes a Rust rewrite completed in the second quarter of 2026 by two engineers using Codex and GPT-5.5. OpenAI reports that the new service handles 95% of production requests, with sixfold CPU efficiency and fifteenfold memory efficiency compared with the Python version. Those are the company’s measurements of this service; they do not establish a general productivity multiplier for AI-assisted programming.
Analysis: Analysis: The operational lesson is to separate a code change from its measured production outcome. OpenAI describes centralising storage control before pursuing these efficiency gains. Its account says that central service allowed improvements to benefit every product without coordinating fragmented updates. For a smaller service, the equivalent question is what bottleneck needs fixing and which observable result would demonstrate that the change helped.
The race for robot training data draws another funding report
TechCrunch reports that Mecka AI is nearing a Sequoia-led financing at a valuation of about $500 million, citing two people familiar with the deal. It is not a completed funding announcement: the terms could change, the round’s size was not established, Mecka had not responded and Sequoia declined to comment. The report follows Mecka’s announcement of a $60 million raise three months earlier.
The underlying business is more concrete than the valuation: Mecka pays people to record everyday activities, such as making coffee or fixing cars, using smartphones and body sensors. It collects and analyses human motion data for robot training. TechCrunch says Mecka has not publicly identified its customers; the article does not provide a measured improvement in robot performance attributable to its datasets.
Analysis: Analysis: This adds a commercial angle to yesterday’s video-learning research. Funding interest and useful training data are separate questions. The results worth watching are whether collected demonstrations transfer to unfamiliar tasks, environments and bodies—not simply how much footage a supplier can accumulate. That distinction helps separate the commercial race to supply data from the harder engineering question of what robots learn from it.
An AI mathematics workflow makes testing central
Research context, submitted 9 September: Niloy Kumar Mondal and Md Rizwan Parvez describe a multi-agent system that translates natural-language optimisation problems into QUBO, a mathematical form usable by quantum, hybrid and quantum-inspired solvers. The difficult part is expressing the intended variables, constraints, objectives and penalties correctly. Generating a plausible-looking formula is only the start of that job.
Their framework uses structured or unstructured test cases. On the authors’ QUBOBench, comprising 100 problems across 12 application domains, they report 68% accuracy and identify iterative self-repair as the largest contributor to the improvement. That is a result on their benchmark, not evidence of universal mathematical reliability. The paper’s abstract also reports that the data and code are open-sourced.
Analysis: Analysis: The useful lesson extends beyond this specialist notation. Give an automated workflow examples with known answers, then make failures drive a revision. Keep an explicit acceptance check at the end: a process that repairs some mistakes can still leave others unresolved, as the reported accuracy itself makes clear. Recording the failed cases is therefore part of the useful output, alongside the formulations that pass the chosen tests.
Cleaning up orbit needs the right ride into space
In a new Ars Technica interview, Astroscale’s chief operating officer Chris Blackerby explains why satellite servicing creates demand for dedicated rocket launches. Reaching a particular object in a particular orbit makes sharing a launch more complicated, he says. Astroscale has contracted with Isar Aerospace for two missions, having made those commitments before Isar first reached orbit.
The planned ELSA-M mission will demonstrate removing a OneWeb satellite; ADRAS-J2 will attempt to capture and deorbit a defunct Japanese rocket. Blackerby says their size has outgrown Rocket Lab’s current Electron capability. Budget, capacity and a dedicated ride narrowed the options, while choosing a provider without consistent demonstrated success remained a risk. These are planned servicing demonstrations, not clean-up operations already completed.
Analysis: Analysis: More launches overall do not automatically solve every mission’s transport problem. The right orbit and payload capacity still matter. Blackerby also says Astroscale has not yet reached commercially repeatable servicing, keeping the long-term business ambition distinct from today’s engineering milestones. For readers following space sustainability, that distinction is useful: progress towards servicing is real work, while a routine service remains the goal.