Research — The Singularity Times

DeepMind · NeurIPS preprintHolds up

Test-time compute beats parameter count on hard reasoning

Chen et al.

A 32B model with 10× the inference compute matches a 540B baseline on AIME 2025 and codeforces — but only when the reward model is itself learned online.

AnthropicReplicated

Sleeper agents survive RLHF and adversarial training

Hubinger et al. (v2)

Backdoors planted at pre-training stay hidden through three rounds of safety fine-tuning, including red-team adversarial. A serious lifecycle concern.

MistralNew SOTA

A 2 trillion-parameter sparse MoE that fits on one rack

Mistral research

128 experts, 8 active per token. Routing learned with auxiliary-loss-free balancing; outperforms a dense 200B model at 1/9th the FLOPs / token.

OpenAIPromising

Deliberative alignment scales with model capability

OpenAI Alignment

Models that are taught to explicitly read and reason about the safety spec become more — not less — robust to jailbreaks as they get smarter. A rare good-news result.

Physical IntelligenceField shift

π-0.5: one VLA policy across 17 robot embodiments

Physical Intelligence

A single 7B vision-language-action model controls humanoids, quadrupeds and arms with zero per-embodiment fine-tuning. The 'foundation model for atoms' thesis gains evidence.

World LabsWatch

Persistent 3D worlds from a single image

Fei-Fei Li et al.

Generated scenes stay self-consistent for >10 minutes of free navigation. The first credible 'spatial' counterpart to large language models.