QUEST — An Open-Source Deep Research Agent Arrives in 2B–35B Sizes
Featured on Hugging Face's Daily Papers, "QUEST" is an open-source family of AI agents built to handle general-purpose deep research tasks. It is available in multiple parameter sizes from 2B to 35B, and excels at long-horizon search tasks such as fact-checking, citation grounding, and report synthesis.
Its most notable feature is that its training data consists entirely of synthetic tasks. This approach—training a frontier-level research agent on automatically generated tasks without relying on real data—points to a new direction for future agent development. Because it is released as an open model, validation and application by the research community are expected to follow.
talkie-1930-13b — A "Vintage LLM" Trained Only on Pre-1930 Texts
A remarkably unique language model has been drawing attention in Reddit's LocalLLaMA community. "talkie-1930-13b-it" is a 13B-parameter model pre-trained solely on 260B tokens of English text from 1931 or earlier.
The model was fine-tuned on instruction-response pairs extracted from pre-1930 reference books such as etiquette manuals, encyclopedias, and letter-writing guides, and its instruction-following ability was further improved through reinforcement learning with online DPO. Support has also been added in llama.cpp.
Born from the experimental motivation of "what would it be like to talk with someone from the past," the project stands as a fascinating case study in just how strongly the choice of LLM training data shapes a model's behavior.
Microsoft Asks Publishers: "Don't Block Our AI Bots"
According to AdExchanger, Microsoft is asking content publishers not to block its AI crawling bots.
With the standoff between AI companies and publishers over the use of web content still ongoing, Microsoft's stance is noteworthy. Publishers are concerned that unauthorized AI scraping leads to traffic loss and copyright infringement, while for AI companies, securing high-quality training data has become a make-or-break challenge.
The move is a telling example of how a platform operator positions itself between two competing demands: the growth of the AI industry and the protection of content creators' rights.
AI Experts Billing Wall Street $25,000 a Day — The Agentic Wave Reaches Finance
According to Bloomberg, AI experts are providing AI training to Wall Street banks at a rate of $25,000 (roughly 3.8 million yen) per day.
Behind this lies the "agentic shift" in the financial industry. Whereas earlier AI adoption was limited to analytics support, the rollout of AI agents that autonomously make decisions and take action is accelerating. Banks are now confronted with the need to redesign organization-wide workflows around AI agents, rather than simply deploying AI tools.
The trend shows that demand for AI talent is expanding beyond engineering into the consulting territory of business process transformation.
Alex Smola: "You Don't Need Every LLM Benchmark"
AI researcher Alex Smola has published his thoughts on how to select LLM benchmarks.
Numerous benchmarks exist today—MMLU, HumanEval, GSM8K, and many others—each measuring different capabilities. Smola argues that you should choose only the benchmarks you truly need for your use case. Chasing high scores on every benchmark, he notes, only drives up the cost and complexity of model evaluation.
As LLM evaluation methodology approaches maturity, this discussion offers an important perspective on reconsidering benchmarks from a practical standpoint.
PHP Institute Warns Against AI-Generated Fake Videos of Konosuke Matsushita
On May 26, the PHP Institute renewed its warning on its official X account, stating that fake videos that use AI to synthesize the image and voice of its late founder, Konosuke Matsushita, continue to circulate.
This is the same kind of issue as voice actor Kenjiro Tsuda's decision to file a lawsuit over TikTok, and cases of AI deepfake technology threatening the likeness and reputational rights of public figures are increasing in Japan as well. Official warnings from companies and bereaved families are expected to grow further going forward.