AI for Anything Daily Brief: Thu · 25 Jun 26

The AI news you can actually use — decoded daily.

!AI for Anything Daily Brief — Thu · 25 Jun 26

☕ The 60-second version

• OpenAI and Broadcom unveiled Jalapeño, a custom LLM-optimized inference chip designed to make AI faster and cheaper at scale.
• Google's Gemini 3.5 Flash now has computer-use capability — it can click, type, and navigate your screen like a human operator.
• Anthropic accused Alibaba of illicitly extracting Claude model capabilities, while the NSA lost access to Anthropic's Mythos tool amid a separate dispute.

🔥 Today's big story

OpenAI Drops Jalapeño — Its First Custom AI Inference Chip, Co-Built with Broadcom

• Jalapeño is purpose-built for LLM inference (not training), meaning OpenAI is optimizing the speed and cost of running its models at scale — not just building smarter ones.
• Custom silicon means OpenAI reduces dependence on Nvidia and gains direct control over latency, throughput, and unit economics for every ChatGPT and API call.
• For learners and builders: cheaper inference = lower API costs, faster responses, and more room for agentic multi-step tasks that previously hit rate limits or latency ceilings.

💡 How to use this today: If you build on the OpenAI API, bookmark your current latency and cost benchmarks now. Jalapeño-driven infrastructure will roll out gradually — watch for speed improvements in GPT-4o and o-series models. Use this as your baseline to measure gains. If you're studying for AI engineering certs, add 'custom inference silicon' to your vocabulary — it's increasingly appearing in system design interview questions. OpenAI Blog · TechCrunch

📰 Also today

Gemini 3.5 Flash Now Does Computer Use — Click, Type, Navigate, Automate

• Google's Gemini 3.5 Flash joins Claude and GPT-4o in the computer-use category, letting the model interact with real GUIs — browsers, apps, desktops — without custom integrations.
• Flash is Google's fast, cost-efficient model tier, which means this capability is now accessible for high-volume, low-latency automation tasks — not just premium use cases.

💡 Practical move: Test Gemini 3.5 Flash computer use on a repetitive browser task you currently handle manually — form filling, data extraction from a web UI, or navigating a legacy tool. Compare it against Claude's computer use on the same task for speed and accuracy. Document your findings — this kind of benchmark is exactly the content that lands well in AI practitioner communities. Google DeepMind Blog · HN Discussion

Anthropic vs. Alibaba: Model Extraction Accusation — And the NSA Lost Access Too

• Anthropic alleges Alibaba illicitly extracted Claude's capabilities — a rare public accusation of model theft between major AI players, with serious IP and legal implications.
• Separately, the NSA lost access to Anthropic's Mythos intelligence tool amid an unrelated dispute, highlighting how enterprise AI access can be revoked instantly — a risk for any org that doesn't own its model stack.

💡 What to learn from this: If you're building workflows or products on third-party AI APIs, understand your dependency risk. This week's exercise: audit one workflow where losing API access would break your pipeline. What's your fallback — open-source model, alternative provider, or cached outputs? This is a real engineering skill, not just a theoretical one. Reuters — Alibaba · NYT — NSA/Mythos

GitHub Copilot Drops Manual Model Selection for Free & Student Plans — Auto Mode Only

• Copilot Free and Student users can no longer pick their model — GitHub's 'auto' selection now dynamically chooses the best model per task, removing user control in exchange for optimized routing.
• For students learning AI-assisted coding, this shifts the skill from 'pick the right model' to 'write prompts that work well regardless of model' — a fundamentally different and more durable skill.

💡 Skill shift: Since you can't pick the model on free Copilot tiers, double down on prompt quality. Practice writing highly specific inline comments and docstrings before triggering Copilot — that context is what auto-routing uses to serve better completions regardless of which model fires. GitHub Changelog

🛠️ Use this today — Test Computer-Use AI on a Real Browser Task in 15 Minutes

Gemini 3.5 Flash now supports computer use. Here's your 15-minute hands-on:

Pick a repetitive browser task: scraping a table from a website, filling a web form, or navigating a SaaS UI to pull a report.

Write a plain-English instruction: 'Go to [URL], find the pricing table, extract all plan names and prices, return as JSON.'

Run it via the Gemini API with computer-use enabled (see Google's docs for the tool config).

Time the run. Note where it succeeds and where it halts.

This gives you a real benchmark for when computer-use AI is faster than writing a scraper — and when it isn't. That judgment is now a billable skill.

⚡ The feed

Models

• NVIDIA and Hugging Face published a guide on accelerating Transformer fine-tuning using NeMo AutoModel — useful for anyone running custom training runs.

Agents

• OpenAI published research showing AI agents are handling longer, more complex multi-step tasks — expanding productivity across professional roles beyond simple Q&A.

Business

• Qualcomm is acquiring AI startup Modular — the company behind the MAX platform and Mojo language — a major on-device AI inference play.

Tools

• Haystack, the open-source Python framework for production RAG and agentic pipelines by deepset, is worth a look if you're building beyond toy LLM apps.
• RubyLLM launches as a clean Ruby framework for integrating all major AI providers — a solid option for Rails shops that have avoided LLM tooling due to Python-heavy ecosystems.

Research

• Alibaba's Qwen team released Qwen-AgentWorld — a language world model for training general-purpose agents that can reason about environment state, not just text.
• Microsoft's recent 'quantum leap' claim is under scrutiny — a researcher alleges basic Python errors in the benchmark code invalidate the result.

Other

• The Economist reports big AI labs are hiring moral philosophers at scale — alignment, ethics review, and policy reasoning are becoming specialized AI roles worth tracking.

📈 Tip of the day

When an AI agent gets stuck on a multi-step browser task (computer use), don't just re-run it — add a 'checkpoint' instruction: 'After each major step, describe what you see on screen before proceeding.' This forces the model to verbalize its state, dramatically reducing silent failure loops and making errors easier to catch and correct.

❓ FAQ

What is OpenAI's Jalapeño chip and what does it do?

Jalapeño is OpenAI's first custom AI inference chip, co-developed with Broadcom. Unlike GPUs used for training, Jalapeño is optimized specifically for running (inferencing) large language models at scale, targeting lower latency and better energy efficiency for serving ChatGPT and API requests.

Does Gemini 3.5 Flash computer use work like Claude's computer use?

Yes, conceptually. Gemini 3.5 Flash can now control a computer — clicking, typing, scrolling, and navigating UIs — similar to Anthropic's Claude computer-use feature. Both operate via screenshot-and-action loops. Gemini 3.5 Flash is Google's faster, cheaper model tier, making this capability more accessible for high-volume automation.

What did Anthropic accuse Alibaba of doing with Claude?

Anthropic publicly accused Alibaba of illicitly extracting Claude AI model capabilities — essentially reverse-engineering or distilling Claude's behavior into Alibaba's own models without authorization. This is one of the first major public model-theft accusations between top-tier AI labs and may have significant IP and legal consequences.

Why did the NSA lose access to Anthropic's Mythos tool?

According to the New York Times, the NSA lost access to Anthropic's Mythos AI intelligence tool amid a dispute between the two organizations. The specific nature of the dispute was not fully disclosed, but it highlights the risk of relying on third-party AI API access for critical government or enterprise workflows — access can be revoked abruptly.

---

Explore AI for Anything to learn and get certified in the tools that matter.