tokenstrom.com

Sunday 15 November

Minus 19 or Plus 12 Percent? Why AI Studies on Entry-Level Jobs Disagree So Much

AI-generated summary Recent studies from the Stanford Digital Economy Lab, the IAB, the Reserve Bank of Australia, and Ramp and Revelio Labs report conflicting figures regarding the effect of artificial intelligence on entry-level employment. The analyses differ because they rely on distinct data sets, age groups, definitions of exposure, and measurement methods, such as tracking potential task automation versus company spending on tools.

PlayStation 4 can now run 9B AI model on its GCN GPU at up to 39 tokens per second

AI-generated summary Developer Leonardo Alegría created an open-source application called Garlic LLM that enables the original PlayStation 4 to run the Qwen3.5-9B AI model. The application reportedly achieves 31 to 39 tokens per second on the 2013 console by utilizing custom GPU kernels and quantization to fit within the system memory limits.

Also covered byTokenstrom

Search APIs for AI agents: every vendor wins its own benchmark

AI-generated summary Tokenstrom reports on various search API providers for AI agents, including Brave, Exa, Parallel, Tavily, Firecrawl, SerpApi, cloro, and DeepSeek, noting that public benchmarks are frequently produced by the vendors themselves. Google is suing SerpApi in federal court in California over bot protection and copyright issues, with the court dismissing the amended complaint while granting another opportunity to amend.

AI peer reviewers find most planted errors – but omissions remain their blind spot

AI-generated summary Tokenstrom reports on how artificial intelligence peer reviewers perform when evaluating scientific papers containing planted errors. Research shows that detection rates vary based on models, prompts, and tools, while omissions remain a difficult challenge.

Earlier:

Europe's Trillion-Parameter EUROPA: Enough Compute, but Only to Match 2025

AI-generated summary The EUROPA consortium described how it got a language model with nearly a trillion parameters running on the JUPITER supercomputer at Forschungszentrum Jülich. Domyn and the Fraunhofer institutes IAIS and IIS make up the consortium behind the project, which won compute time through the Frontier AI Grand Challenge.

Also covered byHIMIAO Horizon

YouTuber aces Korea's tax accountant exam with AI glasses: 400 out of 400

AI-generated summary YouTuber Techmong posted a video showing how he used AI glasses from Rokid and an Anthropic Claude model to score 400 out of 400 points on the first round of the tax accountant exam. Top scorer Han Won-seok confessed to the police that he had used AI to cheat on the exam after his score of 397.5 points raised suspicions.

Earlier:

Korea's top tax exam scorer confesses he cheated using AI and a hidden camera

AI-generated summary Han Won-seok confessed to using generative AI and a hidden camera to achieve the top score on the national certified tax accountant exam in South Korea. The Seoul Metropolitan Police Agency received a written confession from the 34-year-old after he initially denied the allegations in an interview.

Also covered bySouth China Morning PostYonhap News AgencyNakoda News

Saturday 14 November

tiny-voca: A 4.5M-parameter text-to-speech model that runs entirely in your browser

AI-generated summary Developer nato_nob has showcased tiny-voca, a text-to-speech model with 4,537,291 parameters that runs locally within web browsers. The model uses a source-filter synthesizer instead of a neural vocoder and was trained on synthetic data generated by Qwen3-TTS.

Also covered byTensor Journal

Thousands of Lovable apps expose passports, chats and payment data without a login

AI-generated summary Lovable uses an AI agent to build apps and connect databases in a process called vibe coding. An investigation by Crissy Field and Handelsblatt found that thousands of these apps exposed private data without requiring a login.

Also covered byYESODIThe Kernel PostHIMIAO Horizon