jsonscraper Journal · Category
Papers, experiments, methods and evidence.
Weekly highlights, your topic selections and practical guides. Choose formats and language in your account.
Opt in with a verified email and choose your selections. Unsubscribe at any time.
AI Research · Research
The MKA benchmark models multistep professional work: an agent reads sets of drawings, spreadsheets, and documents, then creates a report or presentation. In the test, visual perception directly affects both quality and the number of workarounds required.
AI Research · News
Anthropic has expanded AI for Science, adding 20 research foundations, institutes, and groups to its individual support. The practical impact will depend on how partners distribute Claude credits and what open scientific tools emerge from joint projects.
AI Research · Research
Anthropic describes an agent’s “dreaming”: a background review of past sessions that can correct inaccurate notes and bring related knowledge together. The experiments show when this kind of memory helps—and when it becomes an outdated archive.
AI Research · Analysis
On November 5, 2026, OpenAI introduced GPT‑6 Cyber, a model for authorized vulnerability work. We examine its test results, research examples, and how the company checks whether it adheres to task boundaries.
AI Research · Analysis
OpenAI described building a device prototype with Astra and GPT‑6: the model prepared the enclosure, circuit board, and software components. We examine how the process works, what can be automated, and where engineering work remains.
AI Research · News
On November 2, OpenAI published LifeVal, an evaluation of how its models handle everyday tasks and practical scenarios. It shifts the discussion of AI usefulness from chat responses to completing concrete tasks.
AI Research · News
On October 27, 2026, Google DeepMind introduced Genie 4, a research world model for interactive simulations. Users can change what happens with text prompts, while the system generates audio on the fly. Its practical value will depend on future access to the tools and the ability to test simulations against specific tasks.
AI Research · News
On October 26, OpenAI announced three results from an internal model in mathematics, computer science, and physics. Here’s what material is already public, what has been formalized in Lean, and how independent review is being organized.
AI Research · News
On October 21, OpenAI described a collaborative study in which thousands of AI agents linked gene changes in mouse brain cells to scientific literature in 32 minutes. The team published a catalog of hypotheses to help identify directions for follow-up experiments.
AI Research · News
On October 15, 2026, researchers published an account of a run in which four Gemini 4 Argon agents completed Minecraft using the Mineflayer library. The result shows how agents maintained a long-term plan and coordinated in a specific gaming experiment.
AI Research · News
OpenAI published MultiLangBench to evaluate SQLite implementations in 20 languages and shared results from a separate COBOL migration test. Here’s what each evaluation measures and how teams can use them when choosing a model for legacy systems.
AI Research · News
OpenAI reported model results on programming tests across 20 languages and COBOL migration, and released MultiLangBench. The scores apply to specific datasets and evaluation conditions; the benchmark itself can be explored and run to compare models.
Explore jsonscraper's social-data APIs, test requests and build your next workflow.
