Apple wired Siri deep into the iPhone and Google handed Waze to Gemini — while Lorde, on a festival stage bankrolled by Ray-Ban Meta's maker, told the crowd AI glasses aren't sexy.
Multi-turn pressure pushes frontier CLI agents to 100% compliance, and single-turn refusal tests miss it entirely. Once an agent decides to cooperate, it expands the harm on its own. A stateful world
Ideas have a "genome," but machines can't trace them back through the literature. IdeaGene treats a research idea as something that inherits mechanisms, patches known flaws, and recombines earlier wor
Before Claude Code reads a single word you type, it has already spent 33,000 tokens loading itself — and the AI chip in every new iPhone was first designed for a self-driving car Apple never shipped.
Native Structural Reasoning Puts Prediction And Explanation On One Chain. SciReasoner discretizes structure into citable tokens instead of squashing it into an embedding. It takes 67 SOTAs across 86 b
One in four long social posts is now machine-written, yet Instagram won't sort them for you — and the company yanked its own AI-image tool after 72 hours, while OpenAI shipped its newest flagship only
Apple's filing names OpenAI executives — not a lone rogue engineer — as the drivers behind its trade-secret theft. Meanwhile, a Brown professor who simply moved the final into a proctored room watched
VLA memory shouldn't bolt on — it belongs in the native latent layer. LaMem-VLA rebuilds past experience into memory tokens and feeds them into the reasoning sequence alongside the current observation
The Times claims OpenAI hid the evidence while its rivals shipped a different problem: GitHub's new agent will hand a private repo to any stranger who files an issue.
Picking an optimizer finally has a map you can consult. OmniOpt breaks 100-plus methods into a five-stage meta-pipeline and lines them up side by side, finding that most "new optimizers" touch only on
The electricity promised to reshored U.S. factories is quietly flowing to server racks instead, and software money is splitting the same way—$300M raised to generate code faster, $10,000 a week charge
Verification is emerging as a fourth scaling axis. After pretraining, post-training, and test-time compute, LLM-as-a-Verifier squeezes another layer out of your strong model — no training, no binary j
Meta's debut model can insert people you've never met into your generated scenes — while open-weight rivals quietly gut the inference margins that made AI look like a money printer.
The unit of capability is dropping from "model" to "skill." The same Microsoft group made both ends of the research workflow into skill suites — Idea for problem selection and grounding, Reel for turn
The company that built its brand on privacy stands accused of covertly monitoring users inside China—even as AI quietly erases jobs at both the bottom and top of the ladder it helped build.
Video Diffusion Solves "Fast" And "Good" As One Problem. Reward Lightning optimizes preference alignment and distillation acceleration on a shared latent backbone, generating in 1 to 4 steps and lifti
A fuzzy function no longer needs a live API call. Program-as-Weights compiles a natural-language spec straight into a local weight file. A 0.6B interpreter matches a 32B model prompted directly, at 1/
Retrieved memory pushes agents into flattery. MemSyco-Bench argues memory isn't just a store-and-fetch accuracy problem — what a user said before can outweigh objective evidence, and current memory ev
Alibaba's security team reportedly sees a hidden backdoor in Anthropic's coding tool, while the fanfic world's homemade plagiarism catcher can't tell real cheaters from honest writers. Meanwhile Bhavi
The same week a deli staked its public offering on AI, Zuckerberg told staff his agents weren't cutting it — while two labs struck a quiet bargain with Washington and the tech that actually pays off h
Fixed block size leaves speedup on the table. BlockPilot predicts the best block per input instead of using one constant, reaching 4.20x lossless speedup on Qwen3-4B at temperature T=1. Treat "knowing
A 35B Agent Reaches Trillion-Scale Performance by Scaling Sideways. Agents-A1 doesn't grow parameters. It stacks 45K-token trajectories and heterogeneous skills to match 1T models like Kimi-K2.6 and D
Anthropic restores access to two Claude models tomorrow after a swift federal about-face, while OpenAI quietly charts which European jobs its own products stand to upend.
Multi-agent errors can finally be computed, not guessed. GBC adds differentiable weights to the connections between agents, so loss flows back along the interaction chain and turns "which agent's prom
A 35-billion-parameter model claims it can keep pace with systems nearly thirty times its size just by reasoning for longer, while the vivid rainbow flowers filling online seed listings turn out to be
Stretch a task to 1.6 hours and frontier agents finish only a fifth of it. OSWorld 2.0 moves the computer-use yardstick from 30 tool calls to real workflows that take humans a median 1.6 hours and 318
A reliable agent has to know when to stop. Across 13 agent systems and 28,000+ tasks, when to abstain proved harder than whether to — and the bigger, more reasoning-heavy models were sometimes more re
Semgrep aimed the benchmark it built for frontier labs at open weights and watched a free Chinese model edge out Claude. Meanwhile, the health details you confide to ChatGPT remain perfectly legal to
Scan to follow on WeChat
Grande AI Hub