Today's AI news: Minecraft agent chaos, visual reasoning benchmarks, image transformer advances, and agentic financial workflows.
AI Agent Swarms Collapse Under Their Own Bureaucracy in Minecraft
A striking experiment reported by Tom's Hardware saw teams of up to 50 AI agents unleashed inside Minecraft with the goal of building a thriving civilisation — and failing spectacularly. Rather than collaborating efficiently, the swarms fell into what researchers described as paralysing bureaucracy: agents got stuck coordinating, issuing conflicting instructions, and repeating tasks rather than making progress. The experiment is a vivid real-world demonstration of a core challenge in multi-agent AI systems — scaling the number of agents does not automatically scale useful output. It is one of the most-discussed AI stories on Hacker News today, with 123 points at time of writing.
Can Frontier Models See Past a Decoy? Visual Reasoning Still Has Gaps
A new piece from Latent Space examines Elorian's work on visual reasoning, specifically a decoy-counting benchmark designed to test whether frontier models can reason about what they see — or merely pattern-match. The findings reveal a meaningful gap between seeing and reasoning: models that perform well on standard visual tasks struggle once a deliberate distractor is introduced. Elorian's tool, Glimpse, attempts to close this gap by encouraging models to think with images rather than just about them. The research adds to a growing body of evidence that visual understanding in current AI systems is shallower than benchmark scores alone suggest.
New Research Challenges the Neural Thickets Hypothesis
A paper published on arXiv takes aim at the Neural Thickets hypothesis — the idea that pretrained language models sit in dense neighbourhoods of task-specific experts in weight space, discoverable through random Gaussian perturbations. The researchers reproduce the originally reported gains, but then show the geometric explanation is not required. Perturbing a single decoder block recovers most of the improvement, and keeping the model entirely frozen while adding a random bias to its output logits matches or exceeds full-weight perturbations on most reasoning tasks. The authors formalise this as Perturb-Select-Vote (PSV): generate perturbed predictors, select high-performers on a small validation set, and aggregate predictions. Crucially, the same logit-level approach also improves performance under domain shift in vision tasks, including Earth observation — all without any training. The work reframes inference-time ensemble methods and connects them to self-consistency and verifier-guided selection.
SpecSlot: Faster, Cheaper Image Generation with Spectral Tokenisation
Researchers have introduced SpecSlot, a pixel-space generative transformer that decouples image detail from the length of the global attention sequence — a longstanding bottleneck in image generation models. The key insight is that low-frequency image structure benefits from global communication between tokens, while high-frequency detail can be handled locally. SpecSlot expands each spatial anchor into multiple slots, each assigned a separate band of DCT (frequency) coefficients; only the lowest-frequency slot participates in the expensive global attention pass. On ImageNet-512, SpecSlot improves FID by 2.84 over a compute-matched baseline while using 3.9% fewer FLOPs. On text-to-image generation, it surpasses a comparable baseline on DPG-Bench at 40.6% lower forward cost. The result is a meaningful efficiency gain for high-resolution image synthesis.
Agentic AI for Long-Horizon Financial Work
Latent Space published an in-depth conversation with Arca's head of AI about using agentic workflows for financial analysis tasks that span multiple days. The discussion covers persistent memory — allowing agents to carry context across sessions — the advantages of a custom agent harness over off-the-shelf frameworks, and the practical realities of deploying AI in a domain where accuracy and auditability matter enormously. It is a useful case study for anyone thinking about where autonomous AI agents can deliver genuine value in high-stakes professional settings, and where the current limits lie.
AI's Hidden Cost in JavaScript Frameworks
A blog post gaining significant traction on Hacker News (243 points) examines the AI cost of JavaScript frameworks — specifically, how the complexity and volume of framework-specific syntax affects how much token budget AI coding assistants consume when generating or analysing front-end code. The piece argues that heavier frameworks impose a real, measurable cost when AI tools are in the loop, a consideration that is increasingly relevant as AI-assisted coding becomes standard practice.
Frequently Asked Questions
What happened when AI agent swarms tried to build a civilisation in Minecraft?
Teams of up to 50 AI agents were tested in Minecraft with the goal of building a thriving civilisation but failed. Instead of collaborating effectively, the swarms fell into paralysing bureaucracy — agents issued conflicting instructions and repeated tasks rather than making progress, highlighting core challenges in scaling multi-agent AI systems.
What is the Perturb-Select-Vote method in the new Neural Thickets research?
Perturb-Select-Vote (PSV) is a method proposed by researchers who challenged the Neural Thickets hypothesis. It works by generating perturbed predictors, selecting high-performing ones on a small validation set, and aggregating their predictions. The researchers showed that simply adding a random bias to a frozen model's output logits matches or exceeds full-weight perturbations on most reasoning tasks, making PSV a training-free alternative to weight perturbation ensembles.
How does SpecSlot improve image generation efficiency?
SpecSlot is a pixel-space generative transformer that splits image tokens into spectral slots, sending only the lowest-frequency slot through expensive global attention while handling high-frequency detail locally. On ImageNet-512 it improves FID by 2.84 over a compute-matched baseline using 3.9% fewer FLOPs, and on text-to-image tasks it outperforms a comparable baseline at 40.6% lower forward cost.
What did Elorian's visual reasoning research find about frontier AI models?
Elorian's decoy-counting benchmark found a meaningful gap between seeing and reasoning in frontier models. Models that perform well on standard visual tasks struggle when a deliberate distractor — a decoy — is introduced, suggesting that visual understanding in current AI systems is shallower than benchmark scores alone indicate. Elorian's Glimpse tool aims to address this by encouraging models to think with images rather than just about them.