Inference serving's next leap is not a faster chip — it is using less memory per token
Preprints propose broadcasting weights instead of storing them, early-exit decoding at 2.2x speed, and agent-aware serving that cuts latency up to 4.1x.
29 stories, 121 Manifests cited.
Preprints propose broadcasting weights instead of storing them, early-exit decoding at 2.2x speed, and agent-aware serving that cuts latency up to 4.1x.
In preprints, subliminal learning shifts student preferences by up to 10 points, a dormant backdoor hits 97% success, and unlearning leaves 'forgotten' facts behind.
V4.1-Flash brings web search, file parsing and safety checks to developers, while Verda, NVIDIA and Intel each push a different way to run giant models affordably.
Profound tracked 279,741 AI responses across 25 retail categories: AT&T's phone visibility rose 645%, Wayfair's furniture visibility hit 85%.
OpenAI launched ChatGPT Wallet with tokenized payments and a biometric check per purchase, while agentic orders across Forter's network rose 6.8x since mid-August.
Cyera argues agents operate on over-scoped access that needs no adversary to exploit. Harvey, Airia and LangWatch show what replaces it.
Google says /teamwork proved the 2013 Courtade-Kumar conjecture, while a preprint framework diagnoses when weather models break the laws of physics.
Surge AI's 60-task benchmark found Claude models ahead on all eight dimensions, while a one-line cache setting cut Claude Code bills 10-15%.
OpenRouter's benchmark shows a cross-encoder reranker rescuing near-miss retrieval, while preprints tackle cost, privacy and when to stay silent.
The US data-center power shortfall is projected to reach the mid-30 gigawatt range by 2028, with memory the biggest component bottleneck.
Viam lets developers train robot policies with no physical hardware, Ropedia captured 4,230 egocentric recordings, and Zoox began robotaxi operations in Atlanta.
In a preprint, researchers show OverServe amplifying API requests up to 93.88x. New defenses answer by checking what the agent meant to do.
In preprints, LLM judges miss forbidden database writes that rule-based checks catch, while judging each step beats predicting success.
The Series C values LangChain at $3 billion. The money is chasing the same question Databricks and Airia are answering: who owns, governs and pays for the agents.
In preprints, ACE-RL trains one model to act, reflect and curate playbooks, and Evidence-Governed Evolution gates every self-change on statistics.
A peptide model, an efflux correction, a synthesizability score and an automated synthesis platform all point the same way: check what the metric actually measures.
Three preprints posted on 13 November show the same split: simulations that reproduce a battery's headline number while misrepresenting what the ions are actually doing.
After the November midterms, markets priced lower inflation and higher rates at the same time. The October CPI report is why, and the Fed's December meeting is where the contradiction gets resolved.
Hyperscalers still pay from cash flow, but the bill has grown past what cash flow covers. Bond markets, private credit and GPU-backed loans are filling the gap.
LangChain's Series C at a $3 billion valuation caps a quarter in which venture capital returned 12.2% on the back of AI. But deal value collapsed from the first quarter, and distributions to investors are still running well below normal.
The Reserve Bank of India fined a payments firm for letting money move to unregistered beneficiaries. South Africa's Prudential Authority sanctioned two banks for anti-money-laundering failures. In the US and Europe, regulators extended comment periods and simplified reporting instead.
Fundraising surged in the second quarter, but returns trailed public markets, small-borrower defaults ran hot, and a $3.4 billion collapse in Australia put the whole model under the microscope.
The Bank of England, the Monetary Authority of Singapore and the Bank of Thailand showed that foreign exchange can settle simultaneously across different systems. In the same week, four Southeast Asian central banks agreed to use their own currencies more.
Across four preprints posted on 13 November, continuous ancestry beats categories at placing people, explains gene regulation, and sharpens disease timing.
A meta-analysis of 57,156 people with stage 5 kidney disease finds fewer heart-failure admissions and heart attacks — but the drugs do not switch on PANK1.
Boston paid for Bauers' breakout season, Milwaukee banked a compensation pick, and San Diego and New York left the GM Meetings with pitching at the top of their lists.
The proposal would refocus pay disclosure on the top two officers, simplify the narrative, and replace a framework untouched since 2006. It is the second step in a three-part rewrite of public company disclosure.
Two Senate bills from Rachel May would speed zoning decisions and publish preapproved home designs statewide. An Assembly bill from Linda Rosenthal would make the housing voucher pilot permanent, and a fourth bill sets discharge rules for supportive housing. None of them is law yet.
Precipitation supplies nearly half the Amundsen Sea's freshwater, one flood model beats another by metres, and a treaty suspension carries a price tag.