1. Is AI Already at the "Foothills of the Singularity," or Is It "Not Truly Intelligent"?

Demis Hassabis of Google DeepMind has said that humanity already stands at the "foothills of the singularity." Meanwhile, Meta's Yann LeCun countered that "current AI systems are not truly intelligent." Oriol Vinyals, co-lead on Gemini, staked out a middle ground between the two extremes, arguing that "this would have looked like AGI seven years ago, but it can neither learn from experience nor achieve a true breakthrough."

The fact that three of the AI industry's top researchers are split on how to assess the current state of AI capabilities is a stark illustration that the technology itself is still at a transitional stage.

2. AI Produces a Proof That Overturns a Mathematician's Conjecture

According to a report from Heise, AI has produced a proof that overturns a mathematician's conjecture. AI applied to mathematics has so far racked up successes in areas such as theorem proving, but a case in which AI logically overturns a human mathematician's intuitive conjecture suggests the field has entered a new stage.

The details of the paper are still being verified, but the broader trend — formal verification combined with AI reasoning reshaping the very methods of mathematical research — is well worth watching.

3. Hugging Face Brings PapersWithCode Back to Life — A Host of New Features Added in One Week

Niels Rogge of Hugging Face's open-source team announced a weekly update on the revived PapersWithCode. The main features added in a single week are as follows:

Multiple metrics support**: Benchmark leaderboards now handle multiple evaluation metrics — for example, WER and RTFx for ASR, or mAP and FPS for object detection, displayed side by side

Registering external papers**: Beyond arXiv, the site now supports GitHub repositories, blog posts, bioRxiv papers, and more

Paper lineage display**: Earlier versions and follow-up research are shown in banners (Mamba-3, DINOv2, GLM-4.5, and others)

New methods added**: Gated DeltaNet, Kimi Delta Attention, Mamba-2, and more

Leaderboard screenshots**: A copy-as-image feature for sharing leaderboards on social media

About 3,000 evaluations (evals) have already been registered, and coverage continues to grow, centered on models supported by the Transformers library.

4. "Constraint Decay" Plagues LLM Agents — Surfaces in Back End Code Generation

A paper posted to arXiv titled "Constraint Decay: The Fragility of LLM Agents in Back End Code Generation" is drawing attention. It reports a phenomenon dubbed "Constraint Decay": when LLM-based agents generate back end code, they initially honor the given constraints but begin ignoring them as the task progresses.

As autonomous coding agents move toward practical use, this problem — specifications degrading over long conversations and multi-step tasks — looks set to be a major challenge for production deployments.

5. Gemma 4 Quality Degrades After 30–40 Consecutive Inferences — Reported on a 4GB VRAM Setup

Reddit's LocalLLaMA community has reported quality degradation when running the Gemma 4 2B model continuously on a GPU with 4GB of VRAM. It works normally at first, but after 30–40 inferences the responses grow shorter, with missing fields in JSON output and empty responses occurring. Restarting llama-server restores normal behavior immediately.

A KV cache issue or VRAM fragmentation is suspected as the cause, and discussion continues over this as a stability challenge for running local LLMs over extended periods.

---