AI spent this week speaking four languages: policy, models, robots, and markets, delivering a common message that the AI race is moving beyond spectacular demonstrations to address harder questions about distribution, embodiment, ownership, and economic returns. Industry leaders like Jensen Huang backed open-weight models as a necessary step for industrial strategy, while Chinese laboratories set the pace for this development. Google DeepMind's Gemini Robotics 2 pushed the frontier into controlling physical tasks with greater dexterity, highlighting architecture over answer quality. Big Tech earnings showed that AI investments must translate to customer revenue to succeed, with Meta receiving a harsher reaction due to accelerated infrastructure spending without clear monetization.
Hiring for recent engineering grads has fallen 25% since 2019, according to LinkedIn data, with English majors potentially seeing an increase in job opportunities. Boxed wine is rebranding itself as a high-end product, appearing in hotels and private clubs. Home-sharing platforms are emerging to match elderly individuals with roommates.
Seniority alone is no longer a reliable proxy for operating ability in AI hiring due to the rapidly changing nature of work and automation's impact on roles and career paths. The traditional "heads of X" management structure may not be sufficient to prepare candidates for the complexities of modern AI systems. Instead, companies should focus on finding candidates with relevant experience in AI-native areas, such as data decomposition, context design, and exception handling.
A series of Chinese technological advancements, including AI models, advanced robots, and specialty computer chips, have caused chaos in Silicon Valley and the White House. China's open-source AI models are free to download and use, threatening the dominance of proprietary products from OpenAI and Anthropic. The Trump administration is divided on how to respond to these new models, with some seeing revenue opportunities and others citing security risks.
A new AI approach, internal to OpenAI's Astra model family, has solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. The solution came at a cost of approximately $2,000 in inference. This achievement could significantly alter the economics of deciding which problems are worth pursuing.
Molt is a PyTorch-native agentic reinforcement learning framework developed by NVIDIA's NeMo team. It targets compactness in the codebase, with approximately 8.6K lines of RL code, making it roughly 7 times smaller than verl. Molt uses Ray for placement and asynchronous queues, vLLM for rollout, and NVIDIA AutoModel with FSDP2 for training.
A Time-series forecasting workflow is built using TimesFM 2.5, a deep learning library for time series forecasting. The model is configured, compiled, and tested on a multi-store retail dataset with trend, seasonality, pricing effects, promotions, holidays, temperature variations, and random demand noise. The model's performance is evaluated on various metrics such as MAE, RMSE, sMAPE, MASE, pinball loss, prediction-interval coverage, and batched inference, rolling-origin backtesting, and context ablation studies to determine its robustness to different forecasting horizons, covariates, and data preprocessing techniques.
AMD releases Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model trained on Instinct MI300X and MI325X GPUs with 2.8B active parameters per token. The model holds 16B total parameters but uses Gated Multi-head Latent Attention and FarSkip-Collective connectivity to achieve a 12.7% training speedup and 39.2% reduction in time to first token.
NVIDIA Transformer Engine accelerates transformer workloads by combining fused GPU kernels, BF16 computation, and hardware-aware FP8 execution. The engine is installed on a GPU, allowing the model to run at higher precision without sacrificing performance. The model is trained using a compact causal language model with te.TransformerLayer blocks for faster training times.
Open Weights and American AI Leadership, signed by 235 companies including Microsoft, NVIDIA, Amazon, and Y Combinator, argues against US government efforts to ban or limit open weight models due to safety concerns. The letter states that relying on closed models is not inherently safe and can lead to single points of failure, weakens competition, and concentrates advanced AI capabilities in the hands of a few providers. In contrast, open weight models enable a broad community to examine behavior, identify vulnerabilities, and improve them over time. Meanwhile, Anthropic CEO Dario Amodei signed a response opposing "industrial-scale distillation operations" and calling for government regulation of open-weights models.
Simon Willison has released the June edition of his sponsors-only monthly newsletter, which includes updates on AI model releases, accidental cyberattacks by OpenAl and Anthropic models, and a fireside chat with Cat and Thariq from the Claude Code team. The newsletter also features discussions on Al development and reignites Simon's interest in MCP (Meta-Productivity Copilot). A preview of the June newsletter is available, and subscribers can access the full issue for $10/month.
People don't like it when a coworker's AI chatbot contacts them for help with a task, even if they'd be happy to assist. This highlights how people value human relationships and want AI to enhance time together rather than create separation. Greg Brockman, OpenAI President and Co-Founder, notes this preference in his statement.
datasette-apps 0.2a0 has been released, improving Datasette Apps created and edited using the Datasette Agent. Two new tools have been added: app_debug(), which allows testing apps in an invisible iframe using JavaScript, and app_list(), which lists editable apps for the agent to edit.
Coding agents waste 70% of their context window on irrelevant code, which can be solved by building a Context Compiler that resolves dependencies, discards unreachable code, and keeps only the interfaces where implementation isn't needed. A three-pass pipeline written in pure Python cuts prompt sizes by 69–74% and reduces build time under 75 ms across two real Python repos.
A hybrid LLM application pattern combines a predefined workflow with adaptive agent behavior, allowing for clarity of control while leveraging flexibility in uncertain solution paths. In this case study, an LLM application is built to assist with hyperparameter tuning for classification problems by dividing the work into three stages: preparing the experiment, exploring model configurations, and summarizing the run. The first stage uses a structured LLM call to convert the modeling request into an experiment brief. The second stage employs an agent that dynamically chooses classifier configurations, evaluates them, and continues exploring until it finds a strong configuration. This is followed by a third stage where a single LLM call summarizes the completed run in a concise report.
Calibra v0.7.1 introduces a new integrity workflow to catch common dataset issues before training, including timestamp problems, jittery motion, duplicate frames, incomplete episodes. The goal is to make debugging easier and avoid wasted training runs caused by faulty datasets. Current supported datasets include LeRobot v1, HDF5/Isaac Lab, and robomimic.
A student with a background in ML & DL is seeking help finding resources to learn Reinforcement Learning (RL) and wants to practice with small projects, such as environments and agents with few actions. They are trying to follow DeepMind x UCL RL series but find it too theoretical.
A multilingual psychological coercion dataset with 100,000 rows was created from political dialogues in various languages to help train models detecting psychological coercion. The dataset is available on Hugging Face and includes labeled data for annotation methods at scale.
There is no content to summarize. The provided text only includes the title and author information, but does not include the actual article content.
There is no open-source implementation that directly compares RL/OPD/OPSD on small language models running on consumer-grade GPUs. However, some researchers and developers have implemented various deep learning frameworks such as TensorFlow or PyTorch to experiment with these algorithms. For example, the Hugging Face Transformers library offers a range of pre-trained models that can be used for research and development purposes.
Robert N. Charette, Editor in Chief of IEEE Spectrum, shares his parting thoughts with readers in an opinion piece titled "Fridays With Bob". He reflects on the natural world and describes a visit to a set of crab apple trees where he observed waxwings eating apples in late January.
A team of experts is working to aid people with AI-fueled delusions, providing a service where they call when AI breaks their brain. They are taking on the daunting task of addressing individuals struggling with artificial intelligence-related mental health issues. The company's founders aim to provide support and guidance for those affected by AI-driven hallucinations and false realities.
Australian secondhand book sellers have received suspicious orders from a Canadian company called Zoom Books, which buys books in bulk and then destroys them to feed AI technology. The companies behind the orders claim they only resell intact books, but some sellers suspect they may be part of the destructive scanning practice that has been reported by other AI companies. This raises concerns among book sellers who value the physical object as much as its content, with one seller describing it as "horrific" to think about destroying rare or unique books.