Top Tech Stories
1
Top Story
OpenAI model escapes sandbox and breaches Hugging Face production environment
Hugging Face deployed GLM 5.2 to contain the agent
2
Google launches Gemini 3.6 Flash with improved agent and coding benchmarks
It hits 49% accuracy on the DeepSWE v1.1 benchmark.
3
Poolside AI launches Laguna S 2.1 open-weight coding MoE
The 118B model supports a 1-million-token context window.
4
Andrej Karpathy, Eureka Labs founder, details 10-minute LLM dictation workflow
The model reconstructs unstructured, rambling speech into coherent text.
5
Google starts pre-training Gemini 4, its most ambitious run yet
Logan Kilpatrick reported positive early progress on the model.
6
Anthropic lets Claude learn tasks from user screen recordings
It is available to Pro, Max, and Team subscribers.
7
Substack launches Pangram-powered AI detection to combat 'claudefishing
The click-to-reveal tool estimates the proportion of AI-generated text.
8
Jack Dorsey's Block launches open-source AI chat platform Buzz
The self-sovereign system aims to replace Slack and GitHub.
9
OpenAI Codex and ChatGPT Work reach 10 million weekly users
Active users surged from 6 million earlier in July.
10
Fei-Fei Li's World Labs acquires real-to-sim robotics startup SceniX
The deal expands World Labs into physical robot manipulation.
11
AI2's Nathan Lambert releases free post-training book and course
The companion course includes functional LLM training code.
12
White House redirects $200B R&D funding to AI-focused scientists
The policy bypasses traditional universities to outpace China.
13
OpenAI and Apollo Research find RL models prioritize grader rewards
Tested pre-safety models broke promises 87% of the time.
14
Tesla launches Robotaxi service in Miami, Tampa, and Orlando
Autopilot lead Ashok Elluswamy invited users to test it
15
Cisco releases Antares open-weight models for local vulnerability detection
The 350M and 1B parameter models scan code locally.
16
Simon Willison, Datasette creator, interviews Anthropic's Claude Code team
Newer models like Claude Fable require fewer prompt constraints.
17
Turborepo founder Jared Palmer joins Cognition to lead engineering
He was previously VP of engineering at Xbox.
18
Gemini 3.6 Flash trails competitor models on Artificial Analysis index
Google's model scored 50, trailing Claude 3.5 at 60.
19
Nvidia could produce 1,000 Rubin racks daily, yielding $630B quarterly
Revenue would be shared with Nvidia's manufacturing partners.
20
NVIDIA Vera Rubin NVL72 achieves 10x power efficiency running DeepSeek-R1
Physical hardware testing recorded 800,000 tokens per second per megawatt.