DausnArt

Data Points: Gemini update adds a new reasoning tier Why some AI models give worse answers to non-native English speakers. First Proof, a frontier math challenge OpenAI may have partially solved...

The study shows that experienced users grant Claude more independence by enabling auto-approve more frequently. However, these users also interrupt the process more often, indicating a shift from reviewing every action to monitoring and intervening when necessary.
www.deeplearning.ai/the-batch/ge...

Anthropic Agent Autonomy Study: Anthropic analyzed millions of Claude interactions and found that the duration of the longest Claude Code sessions has nearly doubled. This suggests that models can handle more autonomy.

It operates in both "Speedy Mode" and "Quality Mode" and incorporates the first large-scale reinforcement learning framework for diffusion language models.

LLaDA2.1 Diffusion Model: Ant Group released LLaDA2.1, a discrete diffusion language model with a unique token-to-token editing mechanism that corrects errors dynamically during the generation process.

According to OpenAI, at least five of its proof attempts have a high probability of being correct, based on expert feedback.

OpenAI's Frontier Math Proofs: An unreleased OpenAI reasoning model solved all ten problems in the research-level First Proof competition, which requires complete end-to-end proofs.

These chatbots also sometimes use condescending language when interacting with users who have lower English proficiency, less formal education, or non-U.S. origins.

Chatbot Bias Study: Researchers from MIT found that older AI chatbots, such as GPT-4, Claude 3 Opus, and Llama 3, tend to provide less accurate responses and refuse to answer questions more frequently.

It is trained to detect complex vulnerabilities that rule-based static analysis tools often miss, such as flaws in business logic. Using Claude Opus 4.6, the Anthropic team reportedly identified over 500 previously undetected vulnerabilities in open-source code bases.

Claude Code Security: Anthropic has introduced a limited research preview of Claude Code Security. This tool is designed to analyze code the way a human security researcher would.

Gemini 3.1 Pro achieved a score of 77.1% on ARC-AGI-2, surpassing Claude Opus 4.6 and GPT-5.2. It also demonstrated substantial progress on agentic benchmarks.

The update includes a medium thinking level and an enhanced high setting that operates similarly to a lightweight Deep Think reasoning model.

Gemini 3.1 Pro Update
Google released Gemini 3.1 Pro, which features a new three-tier thinking system. This system enables developers to scale their reasoning efforts, providing options ranging from quick responses to multi-minute deep analyses within a single model.

In the past three decades, the share of the world population living in extreme poverty has declined more rapidly than ever before.
ourworldindata.org/data-insight...

The European View of Debt, Deficits, and Inflation Italian economist Bernardo Ferrero joins Ryan McMaken to discuss the state of European politics over taxes, spending, inflation, and fiscal and monetary policy.

Italian economist Bernardo Ferrero joins Ryan McMaken to discuss European politics, focusing on taxes, spending, inflation, fiscal policy, and monetary policy.

Do Europeans care about deficits and debt as much as Americans do?

Webpage
mises.org/podcasts/rad...

File
dts.podtrac.com/redirect.mp3...

Billions of people have access to far less electricity per day than an air conditioner uses in just one hour.
ourworldindata.org/four-minutes...

How AI Will Create Jobs | AI News & Insights Will AI create new job opportunities? My daughter Nova loves cats, and her favorite color is yellow.

Technology: It uses a combination of a convolutional neural network (CNN), a transformer and an LSTM (long short-term memory) network.
Significance: This provides a potential early warning system for serious diseases, enabling preventive action to be taken.
www.deeplearning.ai/the-batch/ho...

Data: The system analyses a person’s vital signs and activity signals during one night of sleep, including brain, heart, respiratory and leg muscle activity.

Sleep Signals Predict Illness (SleepFM):

Researchers have developed SleepFM, an AI system capable of classifying over 130 medical conditions, including Alzheimer's disease, Parkinson's disease and stroke, up to six years before symptoms manifest.

Limitations: The model struggles with hallucinations, so Liquid AI recommends using it for 'agentic tasks, data extraction and retrieval-augmented generation' rather than 'knowledge-intensive tasks and programming'.

Performance: It matches or exceeds the performance of larger models, such as Qwen3-1.7B, on most reasoning benchmarks while running at roughly twice the speed.

Technology: It uses a hybrid transformer-convolutional neural network architecture that processes adjacent tokens, thereby reducing computation and memory usage.

Faster reasoning at the edge (LFM2.5-1.2B-Thinking):

Liquid AI has launched a small-scale model called LFM2.5-1.2B-Thinking. Designed to run efficiently on small devices, it requires less than 900 MB of RAM.

While this has created a streamlined regulatory environment, it could make it more difficult for smaller companies that are not involved in the lobbying effort to do business.

This spending has been associated with favourable policy shifts for the industry, such as limiting state-level AI regulation, supporting the construction of large-scale data centres (e.g. OpenAI's Stargate project), and lifting the ban on selling advanced AI chips to China (e.g. Nvidia).

The top spenders were Meta ($26.29 million), Amazon ($17.89 million), Alphabet ($13.1 million) and Microsoft ($9.36 million). Nvidia's budget increased sevenfold to $4.9 million.

Big AI's Political Influence

In 2025, top tech and AI companies collectively spent over $100 million on political lobbying for the first time, exceeding that figure.

It was pre-trained on 28.5 trillion tokens and uses Z.ai's open-source reinforcement learning software, SLIME. DeepSeek sparse attention is also used to efficiently process long contexts (up to 200,000 input tokens).

Performance: It achieved the highest performance of any open-weights model on Artificial Analysis’ Intelligence Index and other benchmarks. However, it still lags behind proprietary models such as Claude Opus 4.6 and GPT-5.2.

The new open-weights leader: GLM-5

Z.ai has released the GLM-5: a Mixture-of-Experts Transformer model comprising 744 billion parameters (40 billion of which are active per token).
Purpose: Designed for long-running agentic tasks.