The Sora app had 627,000 downloads compared to ChatGPT’s 606,000 in its first week. It also reached No. 1 on the US App Store
TLDR AI 2025-10-09
Headlines & Launches
Sora’s iOS downloads in its first week was nearly as big as ChatGPT’s launch (2 minute read)
Despite being invite-only, the Sora app had 627,000 downloads compared to ChatGPT’s 606,000 in its first week. It also reached No. 1 on the US App Store by its third day.
An Interview with OpenAI CEO Sam Altman About DevDay and the AI Buildout (33 minute read)
Sam Altman expects that OpenAI’s revenue will eventually pay for its trillion-dollar infrastructure effort - but to lock in hardware agreements, he’s willing to finance investments ahead of revenue. He portrays this as a better-than-average interest rate loan for key partners. OpenAI’s success thus far comes down to two main surprises: ChatGPT far exceeded expectations at launch, and big tech companies have fumbled catching up since then.
OpenAI’s Nick Turley on transforming ChatGPT into an operating system (13 minute read)
Third parties can now reach ChatGPT’s 800 million users during their everyday conversations. Apps are now part of the core ChatGPT experience. Developers can build interactive experiences in ChatGPT beyond just chatbots connected to their company’s data. This article contains an interview with OpenAI’s Nick Turley on how ChatGPT will distribute AGI to the masses.
Deep Dives & Analysis
State of LLMs in Late 2025 (17 minute read)
The AI landscape has evolved into a hyper-specialized ecosystem where each model has distinct strengths. The challenges emerging include diminishing returns on scaling, massive energy consumption, and the rise of smaller specialized models. This guide explains the technical foundations that make each model different so developers can choose the right tool for their task. Success now means understanding each model’s strengths, testing on specific use cases, routing intelligently, and staying current.
The massive spending on AI data centers is almost single-handedly propping up the US economy. It drove 92% of GDP growth in the first half of this year. The economy could have flatlined with just 0.1% growth without the investment. The surge stems from an escalating AI compute arms race. The spending raises concerns about an unsustainable bubble masking weakness elsewhere.
Engineering & Research
AI-Generated Tests are Lying to You (8 minute read)
Developers have shifted to getting AI to generate unit tests over the last year. While this makes everything look good, often, developers are actually just replacing validation with transcription. The more developers automate testing without understanding it, the more they risk turning existing bugs into features. The future of software engineering belongs to those who automate the right things and keep thinking about the rest. AI can amplify good engineering discipline, but it can’t replace it.
Less is More: Recursive Reasoning with Tiny Networks (14 minute read)
A single Samsung researcher developed Tiny Recursive Model, which uses a 2-layer network that recursively improves its answer up to 16 times. It outperforms DeepSeek R1 and o3-mini-high on ARC-AGI by 10% to 30% despite having 0.01% of the parameters. The approach simplifies the Hierarchical Reasoning Model by eliminating complex mathematical theorems and dual-network hierarchies, and suggests that smaller networks with deeper recursion generalize better than larger ones without recursion.
Meet LLMc: Beating All Compression with LLMs (4 minute read)
LLMc uses AI models to compress natural language. It serves as a high-capacity probabilistic reference system, enabling compression based on token prediction and rank encoding. It efficiently identifies high-likelihood token sequences and encodes them into a compact form using LLMs’ in-context capabilities. Benchmarks show that LLMc outperforms traditional compressors such as ZIP and LZMA across a range of datasets.
Miscellaneous
5 things Nvidia’s Jensen Huang said about the state of the AI race with China (4 minute read)
Nvidia CEO Jensen Huang stressed the US is “not far ahead” of China in AI and urged a nuanced strategy to stay competitive. China’s rapid energy production and AI adoption, coupled with local tech giants like Huawei advancing chip systems, pose significant competition. Huang warned of isolating American tech, advocating for global diffusion to maintain leadership in the AI race.
Why California’s new AI safety law succeeded where SB 1047 failed (1 minute read)
California’s new AI safety law, SB 53, requires major AI labs to disclose their safety protocols, ensuring greater transparency. The law’s success where SB 1047 failed is due to its focus on “transparency without liability” and includes whistleblower protections and safety incident reporting. This move could set a precedent for other states to adopt similar regulations.
Quick Links
Gemini CLI Extensions (1 minute read)
Gemini CLI Extensions allows 3rd party tools like Figma and Stripe to integrate directly into its command-line AI coding tool.
AI Mode can now help you search and explore visually (4 minute read)
Google’s updated AI Mode for Search now allows users to find visual results using conversational queries.
A high-performance, open-source physics engine and simulation platform for robotics and embodied AI that offers ultra-fast, cross-platform physics, photorealistic rendering, and natural-language-driven generative data scene creation.
Sora Extend allows users to seamlessly generate extended Sora 2 videos beyond OpenAI’s 12-second limit.