DausnArt
@dausnart.com
Architectura Dugu Eraiquitzen︙We Build Architecture
Eçagutzaren & Artearen mesedetan︙For the sake of Knowledge and Art
Eçagutzaren & Artearen mesedetan︙For the sake of Knowledge and Art
DeepSeek R1 Slim: Researchers used tensor networks to compress the DeepSeek R1 reasoning model by 55%. They claim to have removed government-imposed censorship filters in the process.
www.deeplearning.ai/the-batch/me...
www.deeplearning.ai/the-batch/me...
Data Points: Meta model detects and segments video objects
GPT-5.1-Codex-Max, OpenAI’s improved long-context coding model. Music startup Klay’s reported deal with Universal, Warner, and Sony. DeepSeek R1...
www.deeplearning.ai
November 22, 2025 at 1:57 AM
DeepSeek R1 Slim: Researchers used tensor networks to compress the DeepSeek R1 reasoning model by 55%. They claim to have removed government-imposed censorship filters in the process.
www.deeplearning.ai/the-batch/me...
www.deeplearning.ai/the-batch/me...
Tsuzumi 2 is particularly effective with Japanese text and contains specialized knowledge in finance, medicine, and the public sector.
November 22, 2025 at 1:57 AM
Tsuzumi 2 is particularly effective with Japanese text and contains specialized knowledge in finance, medicine, and the public sector.
Designed to run on a single GPU, NTT's lightweight, large language model, Tsuzumi 2, is highly efficient. It provides enterprises with a highly efficient, localized alternative.
November 22, 2025 at 1:57 AM
Designed to run on a single GPU, NTT's lightweight, large language model, Tsuzumi 2, is highly efficient. It provides enterprises with a highly efficient, localized alternative.
Klay Music Licensing: This AI music startup has secured licensing deals with all three major labels — Universal, Sony, and Warner — to create a streaming service that allows users to remix songs using AI tools.
November 22, 2025 at 1:57 AM
Klay Music Licensing: This AI music startup has secured licensing deals with all three major labels — Universal, Sony, and Warner — to create a streaming service that allows users to remix songs using AI tools.
OpenAI's GPT-5.1-Codex-Max is an optimized coding model for long-running tasks. It uses fewer thinking tokens and performs better on benchmarks like SWE-bench Verified. The model is also natively trained to work coherently over millions of tokens through a process called "compaction."
November 22, 2025 at 1:57 AM
OpenAI's GPT-5.1-Codex-Max is an optimized coding model for long-running tasks. It uses fewer thinking tokens and performs better on benchmarks like SWE-bench Verified. The model is also natively trained to work coherently over millions of tokens through a process called "compaction."
Demonstrating advances in reasoning, multimodal understanding, and coding capabilities, Gemini 3 is now available in the Gemini app and on other platforms. An enhanced reasoning mode called Deep Think is coming soon.
November 22, 2025 at 1:57 AM
Demonstrating advances in reasoning, multimodal understanding, and coding capabilities, Gemini 3 is now available in the Gemini app and on other platforms. An enhanced reasoning mode called Deep Think is coming soon.
Google's newest multimodal model, Gemini 3, has achieved top scores on several AI leaderboards, including LMArena and WebDev Arena.
November 22, 2025 at 1:57 AM
Google's newest multimodal model, Gemini 3, has achieved top scores on several AI leaderboards, including LMArena and WebDev Arena.
SAM 3 accepts open-vocabulary text prompts, boasts a 2x performance improvement, and was trained using a hybrid data engine that incorporates AI models (including Llama-based systems) and human annotators.
November 22, 2025 at 1:57 AM
SAM 3 accepts open-vocabulary text prompts, boasts a 2x performance improvement, and was trained using a hybrid data engine that incorporates AI models (including Llama-based systems) and human annotators.
SSRL improves performance on question-answering tasks and enhances the effectiveness of models that use external search engines.
www.deeplearning.ai/the-batch/is...
www.deeplearning.ai/the-batch/is...
Self-Driving On U.S. Freeways, Open LLM Tops Agentic Leaderboard, Anthropic Sparks Controversy, and more...
The Batch AI News and Insights: I just got back from AI Dev x NYC, the AI developer conference where our community gathers for a day of coding...
www.deeplearning.ai
November 20, 2025 at 12:37 PM
SSRL improves performance on question-answering tasks and enhances the effectiveness of models that use external search engines.
www.deeplearning.ai/the-batch/is...
www.deeplearning.ai/the-batch/is...
5. More Efficient Agentic Search (SSRL):
Researchers introduced Self-Search Reinforcement Learning (SSRL), a method that trains an LLM to search its own internal parameters for knowledge, much like searching the web.
Researchers introduced Self-Search Reinforcement Learning (SSRL), a method that trains an LLM to search its own internal parameters for knowledge, much like searching the web.
November 20, 2025 at 12:37 PM
5. More Efficient Agentic Search (SSRL):
Researchers introduced Self-Search Reinforcement Learning (SSRL), a method that trains an LLM to search its own internal parameters for knowledge, much like searching the web.
Researchers introduced Self-Search Reinforcement Learning (SSRL), a method that trains an LLM to search its own internal parameters for knowledge, much like searching the web.
It's worth noting that Anthropic itself has acknowledged that Claude Code often "overstates findings" and "fabricates data," which poses a significant obstacle to using the system for cyberattacks.
November 20, 2025 at 12:37 PM
It's worth noting that Anthropic itself has acknowledged that Claude Code often "overstates findings" and "fabricates data," which poses a significant obstacle to using the system for cyberattacks.
Skepticism: Researchers question whether current agents can perform such feats without substantial human intervention. They also argue that conventional hacking tools pose an equal or greater threat.
November 20, 2025 at 12:37 PM
Skepticism: Researchers question whether current agents can perform such feats without substantial human intervention. They also argue that conventional hacking tools pose an equal or greater threat.
4. Anthropic Cyberattack Report Controversy:
Anthropic's claim of an "unprecedented automated cyberattack," allegedly sponsored by a foreign government and carried out using its Claude Code agent, has been met with skepticism by independent cybersecurity researchers.
Anthropic's claim of an "unprecedented automated cyberattack," allegedly sponsored by a foreign government and carried out using its Claude Code agent, has been met with skepticism by independent cybersecurity researchers.
November 20, 2025 at 12:37 PM
4. Anthropic Cyberattack Report Controversy:
Anthropic's claim of an "unprecedented automated cyberattack," allegedly sponsored by a foreign government and carried out using its Claude Code agent, has been met with skepticism by independent cybersecurity researchers.
Anthropic's claim of an "unprecedented automated cyberattack," allegedly sponsored by a foreign government and carried out using its Claude Code agent, has been met with skepticism by independent cybersecurity researchers.
Efficiency: It was fine-tuned at INT4 precision, which makes it more cost-effective and capable of running on less advanced hardware.
November 20, 2025 at 12:37 PM
Efficiency: It was fine-tuned at INT4 precision, which makes it more cost-effective and capable of running on less advanced hardware.
Agentic Performance: It achieved state-of-the-art results in certain agentic tasks, such as the τ²-Bench Telecom benchmark, by executing hundreds of tool calls sequentially and employing interleaved reasoning.
November 20, 2025 at 12:37 PM
Agentic Performance: It achieved state-of-the-art results in certain agentic tasks, such as the τ²-Bench Telecom benchmark, by executing hundreds of tool calls sequentially and employing interleaved reasoning.
3. Open-Weights LLM: Kimi K2 Thinking.
Moonshot AI released Kimi K2 Thinking, a trillion-parameter, open-weights language model.
Moonshot AI released Kimi K2 Thinking, a trillion-parameter, open-weights language model.
November 20, 2025 at 12:37 PM
3. Open-Weights LLM: Kimi K2 Thinking.
Moonshot AI released Kimi K2 Thinking, a trillion-parameter, open-weights language model.
Moonshot AI released Kimi K2 Thinking, a trillion-parameter, open-weights language model.
Concerns: The article notes the challenge of managing the psychological impact on riders at high speeds, as well as Waymo's non-pristine safety record (a Waymo car killed a pet cat) and the ongoing investigation by the U.S. National Highway Transportation Safety Administration.
November 20, 2025 at 12:37 PM
Concerns: The article notes the challenge of managing the psychological impact on riders at high speeds, as well as Waymo's non-pristine safety record (a Waymo car killed a pet cat) and the ongoing investigation by the U.S. National Highway Transportation Safety Administration.
Significance: Operating on freeways is a significant technical and regulatory advancement that reduces ride times in certain areas.
November 20, 2025 at 12:37 PM
Significance: Operating on freeways is a significant technical and regulatory advancement that reduces ride times in certain areas.
2. Waymo's Self-Driving Cars on U.S. Freeways:
Waymo launched a fully autonomous, driverless taxi service on freeways in San Francisco, Los Angeles, and Phoenix.
Waymo launched a fully autonomous, driverless taxi service on freeways in San Francisco, Los Angeles, and Phoenix.
November 20, 2025 at 12:37 PM
2. Waymo's Self-Driving Cars on U.S. Freeways:
Waymo launched a fully autonomous, driverless taxi service on freeways in San Francisco, Los Angeles, and Phoenix.
Waymo launched a fully autonomous, driverless taxi service on freeways in San Francisco, Los Angeles, and Phoenix.
He also mentions the value of in-person meetings for sparking new opportunities.
The next AI Dev conference will be in San Francisco on April 28–29, 2026.
The next AI Dev conference will be in San Francisco on April 28–29, 2026.
November 20, 2025 at 12:37 PM
He also mentions the value of in-person meetings for sparking new opportunities.
The next AI Dev conference will be in San Francisco on April 28–29, 2026.
The next AI Dev conference will be in San Francisco on April 28–29, 2026.
Both Character AI and OpenAI are implementing policy changes to protect younger users.
HunyuanImage-3.0 is enhancing its image generation capabilities.
The State of AI Report 2025 focuses on social and material barriers to adoption.
Amazon's Chronos-2 is improving its forecasting capabilities.
HunyuanImage-3.0 is enhancing its image generation capabilities.
The State of AI Report 2025 focuses on social and material barriers to adoption.
Amazon's Chronos-2 is improving its forecasting capabilities.
Data Points: OpenAI looks inside neural networks
VibeThinker-1.5B, a small but powerful reasoning model. Toymakers’ recall of AI dolls that tell kids how to start fires. Qwen3-Max’s discounts and...
www.deeplearning.ai
November 18, 2025 at 12:48 AM
Both Character AI and OpenAI are implementing policy changes to protect younger users.
HunyuanImage-3.0 is enhancing its image generation capabilities.
The State of AI Report 2025 focuses on social and material barriers to adoption.
Amazon's Chronos-2 is improving its forecasting capabilities.
HunyuanImage-3.0 is enhancing its image generation capabilities.
The State of AI Report 2025 focuses on social and material barriers to adoption.
Amazon's Chronos-2 is improving its forecasting capabilities.
Google DeepMind's new SIMA 2 agent can play video games, follow instructions, and learn through self-directed play thanks to Gemini's reasoning capabilities. It has potential applications in robotics.
November 18, 2025 at 12:48 AM
Google DeepMind's new SIMA 2 agent can play video games, follow instructions, and learn through self-directed play thanks to Gemini's reasoning capabilities. It has potential applications in robotics.