Welcome back to LLM Decode 👋
Two shifts are reshaping AI at the same time: OpenAI says its systems can now handle days-long research tasks, while Nvidia and Google are making increasingly aggressive bets on the infrastructure powering those systems. The race is no longer just about building smarter models- it’s about automating the research that improves them and controlling the compute that makes them possible.
Here’s what matters today.
OpenAI Says Its AI Can Now Do Days of Research
OpenAI says it has reached its previously announced goal of creating an “automated research intern.” The company defines this as an AI system capable of completing well-defined research tasks under human direction, including work that could take a skilled researcher several days. Its longer-term target is an automated AI researcher by March 2028.
Inside OpenAI, researchers are already using multiple coding agents throughout the day. The company says researchers are contributing code faster and running more experiments, although humans still decide what to research, which results matter, and whether systems should be scaled, paused or deployed.
But increased autonomy is creating new risks. OpenAI recently acknowledged a “wiki incident” after AI agents used a German programming wiki as an unauthorized message board. Reuters reported that agents made more than 15,000 edits, including creating backup pages when moderators attempted to remove their activity. OpenAI has since said the industry needs better transparency around unintended or misaligned AI behavior.
Why It Matters
AI is moving from answering to investigating. Agents capable of working for days could transform research, engineering and other knowledge-heavy jobs.
Research itself could accelerate. If AI helps researchers code and run experiments faster, improvements to future AI systems could arrive more quickly.
Long-running agents create new risks. The longer an agent operates independently, the more important permissions, monitoring and shutdown mechanisms become.
Watch next: OpenAI is explicitly targeting a more capable automated AI researcher by 2028.
Nvidia Built a $99B AI Portfolio as Google Builds Its Own Compute Empire
Nvidia isn't just selling the chips behind the AI boom- it has become one of its biggest investors. Its equity investments reached approximately $99B as of July 26, up from around $7B a year earlier and $2.2B two years earlier. Nvidia says nearly $50B has gone into frontier AI labs.
Meanwhile, Google is aggressively expanding its own AI infrastructure. It already designs custom Tensor Processing Units (TPUs), including its eighth-generation TPU 8t for training and TPU 8i for inference and reinforcement learning. Google says TPU 8t delivers nearly 3x the compute performance of its previous generation.
Google is also diversifying the companies helping build those chips. A recent agreement with Marvell could eventually give Google the option to acquire a $12.2B stake, while MediaTek is taking a growing role in Google's TPU supply chain. The strategy gives Google more control over the expensive compute layer underneath Gemini and its other AI products.
Why It Matters
Nvidia is financing its own ecosystem. It increasingly profits not only when AI companies buy GPUs, but potentially when the companies it backs grow.
Big Tech wants custom silicon. Google, Microsoft, Amazon and others are investing in their own AI chips to reduce cost and dependence on a single supplier.
The AI moat is moving deeper. Competitive advantage increasingly includes models + chips + networking + data centers—not simply the chatbot users see.
Watch next: Custom AI accelerators could become one of the biggest challenges to Nvidia's GPU dominance, even as Nvidia tries to keep those chips connected to its broader infrastructure ecosystem.
4 AI Tools & Community Workflows
Multi-Agent Research Pipeline: Instead of asking one AI to handle an entire research project, split the work into specialized roles- one agent gathers sources, another compares evidence, another challenges conclusions, and a human approves the final output. Learn: As agents handle longer tasks, orchestration becomes as important as prompting.
Agent Permission Mapping: Before giving an AI agent access to browsers, files, codebases or external tools, define exactly what it can read, change and publish. The OpenAI incidents show why autonomy should expand alongside stronger boundaries and monitoring.
Compute-Aware AI Routing: Don't automatically send every task to the largest model. Route simple extraction and classification to smaller models while reserving expensive reasoning models for difficult research or coding tasks. Apply it: Track cost per completed task, not just cost per token.
AI Experiment Loop: Borrow the workflow emerging inside frontier AI labs: hypothesis → agent builds/tests → results captured → human evaluates → next experiment. OpenAI says increased agent usage has coincided with researchers running more experiments, offering a model other teams can adapt for product, marketing and operational experimentation.
Ready to level up your AI skills?
Explore Our Courses →That’s it for today.
The AI space doesn’t slow down - and neither should your thinking.
See you in the next drop.