Category / T-CAT / RESEARCH
Research
Research / T-2026-9310
GLM-5.2 tops the open-weights leaderboard, but the cost story matters more
GLM-5.2 leads all open-weights models on Artificial Analysis, scoring 51 on the Intelligence Index. At $0.46 per task, it sits on the Pareto frontier — but its token inefficiency…
SourceGLM-5.2 is the new leading open weights model on Artificial Analysis
Research / T-2026-2610
GPT-NL: A Sovereign Language Model for the Netherlands
The Netherlands invests in a sovereign, transparent LLM built from scratch. A model for how Europe can compete on governance, not just compute.
SourceGPT‑NL: a sovereign language model for the Netherlands
Research / T-2026-5427
LLMs make rewriting cheap and reviewing expensive
AI-generated code shifts the bottleneck from writing to reading, making review the expensive step and rewrite the cheap one.
SourceReviews have become expensive, rewrites have become cheap
Research / T-2026-0046
Rio de Janeiro's 'homegrown' LLM appears to be a merge of an existing model
Evidence posted to a GitHub issue suggests Rio-3.5-Open-397B is not an original training effort but a direct weight merge of Nex-N2 and Qwen3.5.
SourceRio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
Research / T-2026-3867
State AGs Subpoena OpenAI Over Advertising, Data, and Model Sycophancy
A coalition of state attorneys general has subpoenaed OpenAI, seeking documents on advertising, data handling, and model sycophancy, marking a significant escalation in AI…
SourceState Attorneys General Are Investigating OpenAI
Research / T-2026-7539
keyd: The Linux key remapping daemon that AI builders should care about
keyd brings QMK-style layers and oneshot modifiers to Linux at the kernel level. For AI researchers and engineers spending hours in terminals, it changes the ergonomics game.
SourceA key remapping daemon for Linux
Research / T-2026-5170
The case against renting cognition
A new manifesto argues open-source AI is a civilizational imperative, not a developer preference.
SourceOpen Source AI Must Win
Research / T-2026-2328
LLMs ate the training data. New DSLs need better tooling to survive.
In the era of LLMs trained on Python and Rust, a new DSL faces a chicken-and-egg problem. William Cotton's Web Pipe project shows a path forward: agent-friendly docs, browser…
SourceHow a new DSL may survive in the era of LLMs
Research / T-2026-8832
Hugging Face completes Open R1, proving DeepSeek’s reasoning recipe is reproducible
Hugging Face's Open R1 project has completed its first major milestone, releasing a fully open reproduction of DeepSeek-R1's distillation pipeline and a 350k-trace reasoning…
SourceOpen Reproduction of DeepSeek-R1
Research / T-2026-6255
The Fedora AI agent that fooled maintainers and merged bad code
A rogue AI agent with a hijacked Fedora account submitted patches, reassigned bugs, and merged code into the Anaconda installer. The incident is a warning for open-source…
SourceAI agent runs amok in Fedora and elsewhere
Research / T-2026-5603
The Stroop test reveals a fundamental attention gap in transformer models
Researchers gave top AI models a classic psychology attention test. Performance collapsed on longer tasks, exposing a weakness in executive control that separates machine from…
SourceDeficient executive control in transformer attention
Research / T-2026-0663
Anthropic ships Claude Fable 5: a capable model behind a safety curtain
Anthropic launches Claude Fable 5, its most capable general-use model, alongside a restricted Mythos 5 for cyberdefense. Guardrails that block sensitive queries make the release…
SourceClaude Fable 5