This Month in Tech: July 2025

Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad
A team from Google Gemini achieved a gold medal result on this year’s International Mathematics Olympiad (IMO), achieving the exact same score as OpenAI: 35/42. OpenAI claims its model had not been optimized for IMO questions. Google’s model was trained with a curated corpus of high-quality solutions to mathematics problems and general hints and tips on how to approach IMO problems. Neither team used tools or internet access to help solve problems.

The New Skill in AI is Not Prompting, It’s Context Engineering
Unlike prompt engineering, context engineering designs dynamic systems that deliver the right information and tools at the right time for LLMs to accomplish tasks. This means structuring various forms of context, like instructions, short and long-term memory, retrieved data, and available tools, into a properly detailed input. The quality of this context is necessary to better AI agents.

Six Principles for Production AI Agents
Building production-ready AI agents requires clear and detailed system prompts, strategic context management to avoid overwhelming the model, and carefully designed tools with straightforward interfaces. Robust feedback loops using domain-specific validation are also required to catch errors. LLM-driven error analysis is useful to identify system blind spots.

Qwen3-Coder: Agentic Coding in the World
The Qwen Team has announced Qwen3-Coder, their most advanced agentic coding model to date. It features a 480B-parameter Mixture-of-Experts model with 35B active parameters and extended context length. The model has state-of-the-art performance in Agentic Coding, Browser-Use, and Tool-Use, on par with Claude Sonnet 4. Qwen Code, a command-line tool adapted from Gemini Code, has also been released.

CTOs Reveal How AI Changed Software Developer Hiring in 2025
AI has changed software development, creating a need for developers who are especially good at critical thinking and problem-solving rather than just churning out code. CTOs are prioritizing skills like system design, business context understanding, and the ability to debug AI-generated code, which often looks perfect but contains subtle flaws.

Local-first software: You own your data, in spite of the cloud
”Local-first software” combines the benefits of cloud apps, like real-time collaboration and multi-device access, with the data ownership and control of traditional local applications. This approach prioritizes local storage and networks over centralized servers, making data modifications and access faster and independent of constant internet connectivity. Conflict-free Replicated Data Types (CRDTs) are a promising foundational technology for achieving these goals, but more research and exper…

SaaS vs AI-Native Software
Everyone claims to be AI-powered, but there’s a huge difference between SaaS with AI sprinkled on top and true AI-native software. The test is simple - does the product still work if you remove the AI? Salesforce works fine without AI features. Jasper dies instantly. The modern software stack now has five layers: cloud platforms (AWS), infrastructure (MongoDB), LLM providers (OpenAI), traditional SaaS (Zoom), and AI-native apps (Perplexity). AI compute demand will grow 1,000x this decade. The…

OpenAI Enters the Browser Wars With ChatGPT Agent
OpenAI combined deep research and Operator into ChatGPT Agent - it thinks like one and acts like the other. It was tested on 1,300 support emails, and in 15 minutes, it found customer complaints, searched LinkedIn for happy users, and built customer archetypes. The browser wars are heating up - Dia and Comet are building AI into browsers while OpenAI wants to skip browsers entirely. Whoever wins gets to stand between users and the entire web.

[Distribution - The Only Moat Left)
Google used to send you 1 visitor for every 2 pages it scraped. Now it’s 18 pages for 1 visitor. OpenAI scrapes 1,500 pages per visitor, Anthropic scrapes 60,000. AI is answering searches directly, keeping users from clicking through. HubSpot used to get 39% of customers from SEO - if that drops 80%, its acquisition costs explode. Companies are responding differently. Ramp sponsored the TBPN podcast where its audience lives. The winners own their distribution - newsletters, podcasts, controve…

Bundled Out of Business
Google just killed a $425 million company by making its product free. Marin Software spent 7 years building the perfect dashboard to manage ads across Google, Facebook, and Yahoo - agencies loved it, and it IPO’d in 2013. Then Google improved Ads Editor and Facebook launched Business Manager, both free. Marin’s revenue peaked in 2015 and slid every year until this week’s bankruptcy.

Qwen3 235B Thinking (Hugging Face Repo)
Alibaba’s latest open-source thinking model achieves 92.3% on AIME25 math competitions, matching OpenAI’s O4-mini performance while surpassing it on coding benchmarks with 74.1% on LiveCodeBench. The 235B parameter mixture-of-experts model activates only 22B parameters during inference, making it more efficient than traditional dense models.

OpenAI’s new ChatGPT Agent can control an entire computer and do tasks for you
ChatGPT Agent is a new tool from OpenAI that completes work on users’ behalf using its own virtual computer. It is powered by a new unnamed model that OpenAI developed specifically for the product. The model was trained on complex tasks that require multiple tools, like briefing users on upcoming client meetings, planning and purchasing ingredients for a meal, or creating a slide deck based on an analysis of competing companies. ChatGPT Agent is now rolling out to Pro, Plus, and Team users - …

Mirage, an AI-native game engine for real-time world generation
The first generative game engine lets players modify environments through natural language during gameplay, launching with GTA-style and racing demos that run entirely on AI-generated content at 16 FPS.

Sam Altman Slams Meta’s AI Talent-Poaching Spree: ‘Missionaries Will Beat Mercenaries’
OpenAI CEO Sam Altman has hinted that the company is evaluating compensation for the entire research organization in response to Meta CEO Mark Zuckerberg’s recent AI talent-poaching spree. He has dismissed the recruiting efforts, saying that Meta’s actions will lead to deep cultural problems down the road. Altman says that Meta failed to get OpenAI’s top people and that Meta had to go quite far down the list.