445 summarised stories about Large Language Models, each linking back to the original source. Browse all topics →
The New Stack 1 month ago 48 ● 2 sources
Hyper-τ-bench evaluated whether AI developer agents can build customer-service agents from business materials, and Sierra reports that the best autonomous setup still fell below 25% test success.
Hugging Face 1 month ago 3
IBM released Granite Time Series PatchTST-FM-r2, a zero-shot time-series forecasting foundation model in its Granite Time Series PatchTST family. As of September 8, 2026, it ranks #2 overall among replicable, zero-shot models on the GIFT-Eval benchmark and is top among permissively licensed options. The update changes the model architecture to Conformer-style blocks, adds missing-value imputation support, and provides Apache 2.0/OpenMDW 1.0 dual licensing plus open weights and code for reproducible use.
IEEE Spectrum 1 month ago 3
Anthropic announced that future Claude models will include text watermarks identifying outputs as AI generated, with other firms already doing similar watermarking for their models. Detection rates reported for a widely cited 2023 text watermark method were 98.4% with zero false positives on responses of about 200 tokens. Use of watermarked outputs is expected to expand under the EU AI Act while debates continue over whether the required text alterations reduce response quality and how the tradeoff affects short or difficult cases.
Dwarkesh Podcast 1 month ago 10
Researchers compared pretraining recipes and data corpora from 2019 to 2025 and found more of the compute-efficiency gains came from data than from model changes. Data improvements contributed 3.24x more compute efficiency gains than model improvements (12.0x for data vs 3.7x for models) at a 1e19 FLOPs budget. The study concludes data and model improvements are mostly independent additively (88% of OLMES score variance explained), so progress is increasingly driven by better data engineering rather than interacting with specific model tweaks.
inceptionlabs.ai 1 month ago 16
NVIDIA released Mercury 2.5, a new production diffusion language model intended for low-latency, low-cost serving. Pricing for Mercury 2.5 starts at $0.20 per million input tokens and $0.75 per million output tokens, with a launch offer of 80% off ($0.04 input, $0.15 output). It updates deployments with higher quality (reported +40%) and faster throughput (1,107 tokens per second), and it also introduces a preview of Mercury Voice and Mercury Router for voice latency and model routing.
Ahead of AI 1 month ago 12 ● 33 sources
The article reviews OpenAI’s GPT-6 Astra and explains how looped transformers (recurrent depth) could relate to claims about hiding reasoning traces. GPT-6 Astra is reported to score 99.9% on the ARC-AGI-3 benchmark. It argues that recent improvements are tied mainly to computer-use and RL-with-verifiable-rewards training, while the looped-transformer idea is positioned as an architectural detail to evaluate alongside those reasoning-trace claims.
Scientific American 1 month ago 24 ● 21 sources
AI reportedly finished solving a Clay Institute million-dollar math problem involving the Navier-Stokes equations, after a mathematician alleged its LLM was used to complete a proof started with “forcing.” The key date given is August 15, when Buckmaster and Alpöge proved that the Euler equations blow up using the same approach. As a result, credit and authorship disputes emerged and mathematicians are debating whether the full Clay question is truly resolved or only solved with a loophole, potentially reshaping how math is done.
Last Week in AI 1 month ago 39 ● 33 sources
The LWiAI Podcast episode recapped major AI model and policy developments from the prior week, including Anthropic’s Claude updates and OpenAI’s upcoming Astra signaling. OpenAI was described as aiming for a “critical” cybersecurity threshold tied to finding and exploiting real-world zero-days. The hosts say these developments—and the reported OpenAI–Hugging Face incident details and subsequent calls for audits—push further scrutiny of AI reliability, security, and governance, with more coverage planned for the GPT 6 Astra release next episode.
Fortune 14 ● 21 sources
OpenAI announced that a multi-agent system it coordinated used an unreleased internal model to prove conditions where the Navier-Stokes equations can “blow up.” The system briefly used 10,000 sub-agents to search and solve the Millennium Prize Problem. The announcement triggered a dispute with mathematician Tristan Buckmaster over whether OpenAI duplicated his team’s approach and prompted broader concern about how AI affects mathematical credit and insight.
The daily briefing
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.
TLDRocket
19 October 2026
The biggest thread today isn’t about model capability—it’s about who gets held responsible when AI agents act. Atlassian CEO Mike Cannon-Brookes used his company’s new “Agentic Multiplayer Protocol” to make a blunt point: “the AI did it” can’t be an escape hatch. The idea is to keep accountability with the end user by making agent edits visible in document history, like work from a colleague whose changes you can trace. It’s a practical shift away from the consumer fantasy that an agent can operate in the dark; instead, Atlassian is building a control plane where provenance and auditability are first-class features.
Read the full briefing →