The Pipeline

Jan Verhulst
August 01, 2025

Good morning,

First of all : it’s our birthday !

Two years ago - this newsletter saw the light of day as something completely different at the time. In all that time we missed only 1 week - Christmas of 2023.

The very first newsletter went out to 14 readers and very gradually and organically we grew to a readership well into the 4 figures.

I can’t stress this enough but to all you reading this today : THANK YOU !

I find tremendous joy in producing this every week and I try to deliver a product that every week is a little better than the one you got the week before. In this attention economy having more than 1000 loyal readers is a big deal. My goal is to make you the smartest person in the room, when it comes to this new technology (without trying to picture you all together inside of a sports arena - which is frankly terrifying).

Which brings us flawlessly to this week’s post. I am looking to do something different this week.

Keeping track of all developments in this space is near impossible.

This week we’re going to press the pause button and not jump on the latest and newest thing that was released, launched, introduced or unveiled.

No, this week we’ll take a look at the 6 most prolific names in the industry and very quickly give you an overview of what their latest product release is and especially what they have in the pipeline.

You will notice that slowly and steadily these companies are carving out specific niches for themselves. It’s good to take a helicopter view of things and see where everyone is at.

Here goes.

OpenAI

This month, August 2025, we will see the long-awaited release of GPT-5. GPT-5 represents a move toward a single unified AI system—combining language understanding, reasoning, and multimodal processing into one engine. That means no more switching between different models depending on the task. It also introduces tiered access: all ChatGPT users will be able to use GPT-5, but how advanced it gets depends on whether you’re on the Free, Plus, or Pro plan. Pro users will get access to the most powerful capabilities. On top of that, OpenAI has released a new “ChatGPT Agent” that can run independently for hours, using virtual tools to browse the web, analyze data, and perform multi-step tasks. Their Codex platform, which helps with collaborative coding inside Visual Studio Code and GitHub Actions, is now generally available. They’ve also launched a new API called Responses that gives developers more control over how the model behaves in apps. If GPT-5 lives up to the hype this might be another acceleration vector in the race for AGI.

xAI

Over at xAI, Elon Musk’s team has released Grok 4, which comes in two flavors: one general-purpose and one optimized for heavy multi-agent reasoning. This means it can think through complex problems by breaking them into parts and working on them in parallel. xAI is planning rapid updates: a specialized coding model is coming in August, multimodal capabilities in September, and enhanced reasoning agents in October. Elon has even claimed Grok could invent new technology or discover new physics by the end of the year. Whether that happens or not, xAI is clearly pushing fast and hard into areas that blend language with action and logic.

Google

Google, meanwhile, is steadily building out its Gemini ecosystem. The latest versions—Gemini 2.5 Flash-Lite and Pro—are optimized for speed and reasoning tasks, respectively. Google is integrating AI into search, into its development tools, and now even into the command line via something called Gemini CLI. On the infrastructure side, they’ve launched their seventh-generation TPU, called Ironwood, which powers many of these upgrades. Google’s focus is on embedding AI deeply into their existing products and platforms, with a strong emphasis on multimodal abilities—like image generation, code interpretation, and even camera and screen sharing from mobile devices. No real announcements yet, but expect updates to come hard and fast.

Anthropic

Anthropic has carved out a niche with Claude 4, especially in software development. The Claude Opus model has set new records on important coding benchmarks and can perform complex, multi-hour reasoning tasks without losing accuracy over time. It integrates natively with developer tools like VS Code and JetBrains, and supports background task execution through GitHub. Claude models are also designed to use tools as they think, switching between reasoning and action in real-time. They include memory systems that help the models retain knowledge across sessions, improving performance the more they’re used. Also here , no real announcements but constant improvements are obviously expected.

Microsoft

Microsoft has gone all-in on integrating AI throughout its product suite. Its Copilot tools are now available across Word, Excel, PowerPoint, Teams, and Outlook, and extend into Azure and GitHub as well. Copilot is becoming smarter, too, with features like memory of user preferences, the ability to take actions online, and voice/vision support. Microsoft has also launched a full development environment for building and managing agents within Azure. Their approach is clearly focused on enterprise use cases—making AI a seamless part of workplace tools and infrastructure.

DeepSeek

Lastly, China-based DeepSeek may not be as well-known outside its notorious emergency at the beginning of this year, but it’s become one of the most important open-source players in the field. Their R1 model performs near the level of top-tier models from OpenAI and Google, but it can run on a single GPU. That’s a huge deal for researchers and startups. It’s released under an MIT license, meaning anyone can use it commercially. Recent updates have significantly boosted reasoning accuracy and reduced hallucinations. DeepSeek is planning more improvements this year, including better support for multilingual tasks, options to self-host models privately, and lightweight versions that can run on phones or edge devices.

So that’s where we are. The whole field is pivoting toward longer context, deeper reasoning, and autonomous agents. Every vendor is moving in that direction, whether through chat interfaces, developer tools, or cloud platforms. The race isn’t just about who has the best language model but rather about who can integrate that intelligence into the real world—and who can keep it under control once it’s there.

Consider this your status update.

—Jan

AI News

OpenAI is expected to launch GPT-5 in August, combining advanced language and reasoning into one system that CEO Sam Altman called a “here it is moment.” While it won’t include the math Olympiad-level abilities seen in recent experiments, it represents a major step forward, alongside OpenAI’s first open-weight model since 2019, set to arrive by late July. With massive hype and talk of mind-blowing results, this next release could significantly raise the bar for AI capabilities.
Researchers in Denmark developed an AI-powered platform that designs custom proteins to help T cells hunt down and destroy cancer cells in weeks instead of years. Using multiple AI models, including AlphaFold2, the system can target specific cancer markers and filter out unsafe designs before lab testing. This points to a faster, more personalized future for cancer treatment and drug development.
Microsoft analyzed 200,000 Bing Copilot chats to identify which jobs are most exposed to AI, with roles like content creators, office staff, and sales showing the highest overlap. Physical jobs like nursing assistants and mechanics remain largely unaffected, at least for now. The findings offer a clearer view of where AI is already reshaping work — and where it hasn’t reached yet.
China unveiled a global AI action plan at the World Artificial Intelligence Conference, promoting international cooperation, open-source development, and support for developing nations. The proposal contrasts sharply with the U.S. strategy, which focuses on deregulation and competition. By pushing a collaborative message, China may be positioning itself as an alternative AI leader for countries left out of Western alliances.
Meta named former OpenAI researcher Shengjia Zhao as chief scientist of its new Superintelligence Labs, adding to a roster of high-profile recruits. Zhao, a key contributor to GPT-4 and several OpenAI reasoning models, will shape Meta’s research direction alongside Alexandr Wang. With this hire, Meta’s ambitious plan to build a top-tier AI lab from scratch appears fully in motion.
Runway introduced Aleph, a next-gen video model that can edit and transform existing footage with text prompts — changing lighting, adding elements, or even generating new camera angles. It builds on Runway’s Hollywood partnerships and pushes AI post-production beyond random generations toward full creative control. Aleph could mark a turning point in how films and video content are made.
Chinese startup Z.ai released GLM-4.5, a powerful open-source AI model that rivals top systems like o3 and Grok 4 in reasoning, coding, and agentic tasks — all while undercutting competitors on price. The model hits a 90% success rate in tool use and comes with an open training framework, adding fuel to China’s fast-growing, open-source AI surge. It’s yet another sign that Chinese labs are rapidly closing the gap with frontier models.
Microsoft just rolled out Copilot Mode in its Edge browser, adding AI tools that can search across tabs, handle tasks, and offer proactive suggestions directly in your browsing flow. The feature is free for now and will eventually allow deeper access like auto-booking with user permission. This pushes Edge into the competitive race to build the first truly “agentic” browser experience.
Alibaba’s Tongyi Lab launched Wan2.2, a next-gen open video model with strong cinematic control and motion quality for both text-to-video and image-to-video generation. Using dual specialized AI “experts,” it beats rivals like Sora and Seedance in detail, camera control, and style — and offers deep customization over video outputs. It’s a major step forward in China’s broader push to lead in open AI tools beyond language.
Mark Zuckerberg outlined Meta’s new AI vision: building “personal superintelligence” assistants designed to empower individual goals, not just automate work. While Meta previously championed open models, Zuck now says safety concerns may lead to keeping advanced systems closed. The pivot comes as Meta leans into smart glasses and a more personalized, device-driven AI future — diverging from rivals and China’s open-source momentum.
Amazon-backed Fable just launched Showrunner, a platform that lets users generate custom animated TV episodes with text prompts and even upload themselves as characters. The tool aims to redefine animation as interactive, remixable entertainment, with plans for creator revenue sharing. As Hollywood wrestles with AI, Showrunner may spark a new genre of personalized, playable storytelling.
Google DeepMind introduced AlphaEarth, an AI model that fuses massive satellite data into detailed, real-time maps of the Earth’s changing surface. It can track things like deforestation and ecosystem shifts far faster than traditional methods. By unifying scattered data sources, AlphaEarth offers a powerful new tool for environmental monitoring and planetary-scale decision-making.

Quickfire News

Elon Musk revealed that X is working on bringing back Vine “in AI form,” using the video app’s intellectual property now owned by X.
Similarweb reported that OpenAI’s ChatGPT still holds 78% of total AI platform traffic, with Google in second at 8.7%.
Alibaba launched Qwen3-MT, a translation model supporting over 92 languages with high benchmark performance.
Google released Opal, an experimental tool that turns natural language prompts into editable AI mini apps with customizable workflows.
HiDream released HiDream-E1.1, a new image editing model that ranked first among open-weight models on Artificial Analysis' Image Editing Arena.
Figma made Figma Make generally available, allowing users to convert design prompts into interactive prototypes and working code.
OpenAI CEO Sam Altman clarified that ChatGPT conversations aren’t legally confidential, and user chats could be used in legal cases if required.
Alibaba updated Qwen3-Thinking, making it competitive with top models like Gemini 2.5 Pro, o4-mini, and DeepSeek R1 across reasoning, coding, and knowledge tasks.
Tencent released Hunyuan3D World Model 1.0, an open-source tool for generating editable 3D worlds from text or image prompts.
Hallwood Media signed Suno music creator Imoliver to a record deal, marking the first time a “music designer” from the platform has joined a label.
Vogue drew criticism after Guess featured an AI-generated model in a full-page ad in the magazine’s August issue.
Alibaba unveiled Quark AI glasses, a smart glasses line launching later this year, powered by its Qwen model.
Google DeepMind CEO Demis Hassabis shared that Google AI systems processed 980 trillion tokens in June—more than double the total from May.
Anthropic introduced weekly usage limits for Claude Pro and Max users due to high demand, affecting fewer than 5% of users.
Tesla signed a $16.5 billion chip manufacturing deal with Samsung to produce its next-gen AI6 chips, which Elon Musk called “strategically critical.”
Runway partnered with IMAX to bring AI-generated films from its 2025 AI Film Festival to big screens in 10 U.S. theaters this August.
Anthropic published new research on using AI agents to audit models for alignment issues, aiming to detect subtle misbehaviors that humans may miss.
Anthropic is reportedly raising $5 billion in a new funding round led by Iconiq Capital, boosting its valuation to $170 billion—almost triple its March figure.
OpenAI announced Stargate Norway, its first European data center project, built in partnership with Aker and Nscale.
YouTube is introducing AI moderation tools that estimate a user’s age based on viewing habits to better identify and protect minors.
Amazon is paying $20–25 million annually to license New York Times content for AI training and integration into its platforms.
Neo AI launched NEO, a machine learning engineering agent powered by 11 sub-agents, claiming top scores on ML-Bench and Kaggle tests.
An Associated Press study found that AI is most commonly used for information searches, with younger users also relying on it for brainstorming ideas.

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter, please pay it forward !

Thank you for being here !

Reply

or to participate.