Left open, stopped short
Ten questions built from 10 records: a top-tier model under the loosest license, agents that stop at the last step, and AI coming down to the desk.
Class quiz
Q1
Reflection released its open-weight model Beam under which license?
Beam is a sparse mixture-of-experts model with 501 billion total and 23 billion active parameters per token, put out under Apache 2.0 with a 1 million-token context.
See the recordOct 16, 2026 · Reflection launches Beam, an open-weight modelQ2
In FDE-Gym, a benchmark modeled on an engineer posted to a customer, how many of the 95 tasks did the best setup solve?
The authors reported that where the model has to find the constraints itself, even the best configuration averaged only 0.56 points and solved 14 of the 95 tasks. They concluded that finding the unstated constraints is the bottleneck.
See the recordOct 15, 2026 · FDE-Gym releasedQ3
Where does Anthropic's midterm election plan get the official voting information it links into answers?
It links official voting information from Democracy Works into answers and shows the count on election day. It also adds classifiers aimed at misuse split across several steps, fake accounts and fake news sites.
See the recordOct 14, 2026 · Anthropic sets out its preparations for the 2026 US midterm electionsQ4
Cohere grouped coding agents into teams. What happened to the cost compared with a single agent?
Company-reported, a single Sol agent went from 52.1% to 61.8% once it worked as a team, but teams cost 3 to 7 times more than a single agent.
See the recordOct 14, 2026 · Cohere publishes experiments on grouping coding agents into teamsQ5
Compared with running screen control on its own, by how much did OpenGCUA cut the steps per task on average?
The authors reported that on OSWorld-v2 the average partial score of the 8 models evaluated rose by 19.6 points and that steps per task fell by 53.0% on average.
See the recordOct 15, 2026 · OpenGCUA releasedQ6
How does IBM's Granite Switch 4.1 handle chat, judging and speech transcription?
Because these are activated LoRAs, the KV cache built by the base model can be reused for the judgment that follows, and IBM said GPU memory with all three parts added stays close to the base model alone.
See the recordOct 15, 2026 · IBM releases Granite Switch 4.1Q7
What does NVIDIA's RTX Spark laptop start at?
They are Windows laptops that pair Blackwell RTX graphics with up to 128GB of unified memory to run models on the device, selling from $2,399.
See the recordOct 16, 2026 · NVIDIA launches RTX Spark laptopsQ8
Cohere agreed to buy the Czech company AISLE. What does AISLE build?
AISLE said that in its first year out of stealth it was assigned more than 500 CVEs across more than 100 codebases, and Cohere said it will make Prague a base for AI security.
See the recordOct 15, 2026 · Cohere announces plans to acquire security company AISLEQ9
Anthropic signed an MOU with Singapore's government and opened an office there. Which office in Asia-Pacific was it?
The signing was timed to Singapore International Cyber Week, and company-reported figures put Singapore at No. 2 of 121 countries in Claude.ai use per capita.
See the recordOct 14, 2026 · Anthropic opens a Singapore office and signs an MOU with the governmentQ10
How does xAI's Grok Voice 3.0 answer speech?
Company-reported, response latency is under 1 second, and xAI said about 70% of roughly 20,000 calls a day in Tesla vehicles and Starlink support end without a person.
See the recordOct 14, 2026 · xAI releases Grok Voice 3.0