Horizon Lens — 7 November 2026
BioBasicsBench asks when the lab evidence runs out
OpenAI and biotech company Ligo built BioBasicsBench using raw experimental data from working scientists. On November 6 they described 100 tasks that include reading images, spotting faulty controls and deciding when a conclusion is unjustified. Answers earn partial credit against expert-written criteria. At the xhigh reasoning setting, GPT‑6 Astra earned 66.7% of available credit on average, compared with 43.8% for GPT‑5.6 Sol.
One example explains why that scoring detail matters. Astra noticed the wrong antibiotic on a control plate and rejected the control. In another test it correctly distinguished two groups of bands in a gel, then chose the wrong colonies because it assigned the groups the wrong identities. A model can earn credit for what it sees while reaching a conclusion that would send the next experiment in the wrong direction.
Analysis
A scientist reviewing an assistant’s answer needs to see how it connected its observations to the controls. A plausible description of an image can survive even when the inference drawn from it fails.
Action
Ask for the observation, the relevant control and the conclusion separately. Include cases where the right response is to ask for more evidence.
ToyBlocks makes a failed robot move easier to diagnose
Ai2’s ToyBlocks simulates a robot arm searching, turning and stacking marked blocks. A planner assembles demonstrations automatically. Researchers can change the mix of training examples and reserve unfamiliar tasks for evaluation. They can also replace either the planner or the motion controller with a reference component to investigate which part caused a failure.
In one study, an agent using Claude Opus 5.5 solved all 30 classical stacking tasks when observing the scene through a camera after each move. Searching for letters to spell a word was harder. Half the scenes lacked a required letter. The agent either built the right stack or correctly recognised an impossible task in 63% of cases. Look-alike letters and incorrect orientations accounted for many errors.
Analysis
Including impossible cases lets researchers measure whether an agent can recognise that something essential is missing before declaring success.
Action
When comparing robot demos, ask which failures the score includes and whether a system gets credit for correctly stopping.
Two accents expand a private speech recognition test
Hugging Face and Liva AI have added conversational Egyptian-accented and Nigerian-accented English to the Open ASR Leaderboard. Their November 6 post describes 1.34 hours of Egyptian speech and 1.55 hours of Nigerian speech. The reference transcripts retain punctuation, capitalisation and disfluencies, including the rough edges people naturally leave in conversation. The recordings expand a private evaluation collection that already included scripted and conversational speech.
Developers can now inspect a US versus non-US comparison, as well as a grouping that puts India, Nigeria and Egypt together. The authors say private material helps reduce the opportunity to tune directly to benchmark examples. The sample durations also set a practical limit on interpretation. Each addition represents recordings from a particular collection, and the broader group includes speakers from several countries.
Analysis
For a transcription service, a single overall score can hide the result that matters to the people using it. A wider range of test speech gives buyers a better starting point for investigating that gap.
Action
Compare performance on the accents your users speak. Listen to samples of spontaneous conversation and check how the transcript handles hesitations and corrections.
Wireless research measures what a robot gets done
In a November 6 roundup, NVIDIA describes a King’s College London experiment that sent compact descriptions of objects and their relationships to help a robot decide what to do. The comparison used periodically transmitted images. In fruit-packing and cube-stacking trials with deliberately introduced faults, the researchers cut task completion time by up to 52.6% and improved success, according to NVIDIA. That percentage refers to completing the robot task under the tested conditions.
A University of Leeds team tackled a different question. It forecast demand ten minutes ahead, checked whether neighbouring stations could cover devices and put spare capacity to sleep. In a portable 5G test system supported by a digital twin, the team reported a 45.6% reduction in radio access network power against a full-power baseline while keeping devices connected.
Analysis
Both experiments work backwards from a service requirement. One asks how much information the arm needs to recover from an error. The other asks how much capacity devices need to stay connected.
Action
When reviewing network efficiency claims, identify what was measured and what service had to keep working throughout the test.
Sequoia ends a route from operating a startup to investing
Sequoia is closing the scout programme through which selected founders and operators could invest on the firm’s behalf, TechCrunch reports. The decision forms part of a wider reorganisation that includes closing its London office. Sequoia framed the move as concentrating its attention. For founders who first encountered the firm through another operator writing a cheque, one route to an early investor is now closing.
Roxanne Varza, who scouted for Sequoia from 2020 to 2025, told TechCrunch that such programmes helped aspiring fund managers build a track record and relationships. A comparison from earlier this year shows how firms can take different approaches. Andreessen Horowitz partner Gabriel Vasquez said his firm had reduced the number of scouts while giving those who remained more resources and expecting greater involvement, including co-hosting dinners.
Analysis
For an aspiring investor, organising introductions and making investment decisions build different kinds of experience. A programme’s value depends partly on which responsibility it actually delegates.
Action
Before approaching a scout, establish what they can commit and who makes the investment decision. The same title can come with different budgets and expectations.