The robot's world model is getting better. Its planning still isn't.
Thinker-WM unifies vision, depth, and action in one transformer; decoupled world-model designs prove most robust; but the best VLM still plans at 75%.
31 stories, 124 Manifests cited.
Thinker-WM unifies vision, depth, and action in one transformer; decoupled world-model designs prove most robust; but the best VLM still plans at 75%.
TreeGraft mixes drafters by cost, FOLD fixes long-context drafting without retraining, and ReaDy mines recurrent states for better drafts.
Two-thirds of executives believe their company has already suffered an AI data breach. Only 14.4% of organizations deploy agents with full security approval. A vector database patched an authorization hole the same week.
Moral reframing jailbreaks with high single-trial success, judge-guided attacks hit 89.8% ASR in 10 queries, and the LLM judge itself silently refuses to score safety items.
NOMOS turns written policies into verified tool-call gates, AgenticRV monitors any formalism across frameworks, and SecureClaw keeps secrets behind a trusted gateway.
Meta released the weights for an 816B-parameter coding model and NVIDIA opened a 120B-parameter agent model the same week. Intel, Modal, and AI21 had plans ready before the ink was dry.
PrismQuant aligns outliers for low-bit quantization, GLUESTICK recovers pruned robot models without retraining, and LayerTrim prunes MLPs while keeping attention intact.
Liquid AI released d1-ultra, a multimodal decision model priced by input token. Two new benchmarks measure whether decision models are calibrated and whether they survive adversarial prompts.
JSSA claims 6.43x decoding throughput at 128K context. LatentLoop claims 19% compute savings on looped MoE. Two more papers cut video distillation and code-editing costs.
SWEeper-Bench finds frontier agents fail at noticing bugs. Doc2Harness finds they fail at building the tooling around themselves. PreciseCoder trains them to edit less, not more.
Answer-extraction tricks swing scores from 20.8% to 93.9%, answer variants collapse DROP to 33.3%, and a task-specific rubric is the only lever that aligns judges with humans.
Right-size the model for the task, fix the silicon encoding, and reuse the KV cache: efficiency work stops asking the model to do less and starts asking the system to waste less.
114 of 135 agent-written papers failed integrity checks, while a rival orchestration pipeline got a paper accepted at ALIFE 2026. The difference is verification, not the base model.
Four results point the same way: the bottleneck in agent training is not the model, it is the environment the model trains in and how the training loop is engineered.
LangChain's October update adds semantic trace search and red-teaming. Zep v4 stores multi-entity facts as hyperedges. Google's adk-recipes unifies agent resources in three layers.
Agents are getting governed like employees: authorization enforced before each action, behavioral monitoring for drift, and OAuth tokens that cannot be replayed if stolen.
Five preprints posted Oct. 15 to the ESS Open Archive measure motion nobody can watch directly: melt shifting inside Mars, plasma thickening at Mercury, oceans read via magnetism on Europa and Ganymede, solar-wind turbulence tied to the averaging window, and meltwater skewing glacier radar.
Five preprints this week share one concern: chemistry that holds up outside the flask — designed photoenzymes, a directing-group-free amination, a diagnosed photocatalyst failure, gold recovered from e-waste, and PFAS recycled into a fluoride source.
September's inflation surprise put a December hike back on the table. Apollo's Torsten Slok says the deeper reason rates won't fall is structural: AI is expected to return more than 10% against a 4% funds rate.
Agents went from spectators to actors in commerce in three months. The battle is over who becomes the default credential provider, while wallets add investing, attackers target custody, and merchants rebuild the data layer.
PNC rode commercial loan growth to a $2.1B quarter and Schwab posted record revenue of $7.2B, even as both banks warned the easy part may be over. Hamilton Lane and Wells Fargo, meanwhile, rebuilt their leadership around technology and AI.
EQT is buying $16.5B of real assets for $375M, Accel is backing robots for job sites, and Brookfield is making the transport-infrastructure case. The money is moving to the things you can touch.
September inflation spiked rate-hike odds and sank equities, but the dollar gave back its gains as AI-debt fears took over. The Swiss franc is the safe haven of the moment.
Three finance podcasts this week converged on the 2026 midterms: the House as the Democrats' to lose, AI policy drifting to the states, and data centers showing up on local ballots.
Six preprints posted October 15 add working tools to the bench — photoactivated labeling chemistry, covalent mitotic inhibitors, an effector-free RAB27A structure, a computational amylase candidate, a dual-use photosensitizer, and a formal fix for silent mass-spec errors.
Five bioRxiv preprints map the cell from the inside out: subcellular proteomes of a marine flagellate and an animal relative, a PIEZO1-YAP1 mechanosensing axis in joint fibrosis, an Ncf1/Igfbp2 repair switch in airway progenitors, and a fix for genotype bias in single-cell methylome atlases.
The Brewers lead the NLCS 3-0 after a 6-2 Game 3 victory. The Dodgers are 3-for-32 with runners in scoring position, and only one team in history has climbed out of a 3-0 hole.
The White House tallied its anti-gang operation, ICE charged a Bahamian national with voting in 2024, New York's 204th Engineer Battalion mobilized for the border, and the CEA priced Democratic immigration policies at $49 trillion.
A Federal Reserve speech calls for sharper, more accountable bank supervision. Trading assets at the largest banks have doubled since 2019. And researchers warn the energy lenders left behind by securitization are holding worse collateral.
West Virginia won $9M in federal funds for college prep, Montana secured $4.8M for employer-led training, California posted record graduation rates, and Alabama awarded a bridge contract.
Four preprints posted this week put new numbers on climate costs: a methane pledge gap, a proposed waste-heat driver of uneven warming, a constant marginal cost of CO2, and two opposing Himalayan snow futures. None have been peer-reviewed.