This week in AI was anything but quiet. OpenAI filed for a $400 million startup fund and put itself down as the only investor. It also built a custom chip that threatens Nvidia's dominance. Add humanoid robots outrunning Usain Bolt and a potential $2 trillion IPO from the makers of Claude. Whether you're a casual observer or someone using AI daily at work, the pace of change right now is hard to ignore. Here's what actually matters from the past week.

Summary: OpenAI has filed paperwork for a second venture fund, OpenAI Startup Fund II, worth $400 million and aimed at early-stage AI startups. The SEC filing lists a single investor: OpenAI itself, paying out of its own balance sheet, with a first sale date of 11 August. The original fund launched in 2021 with $175 million from outside backers including Microsoft, and put money into Harvey, Cursor, Figure AI and Physical Intelligence. OpenAI is reportedly still looking for an experienced venture capitalist to run the new one.

Why it matters: The 2021 fund answered to outside investors. This one answers only to OpenAI, so a cheque can be written to serve the roadmap and not the return. A founder taking that money is taking it from a strategic backer, not from a fund that only needs the valuation to rise. And for anyone buying AI software, a growing share of the application layer is now funded by the company whose models that software runs on.

Key takeaways:

Source: Dealroom | Tech Funding News

Summary: Early benchmarks show OpenAI's custom AI chip handles AI tasks up to 3.6x faster than Nvidia's top chips while using significantly less power. Called 'Jalapeño,' it was developed with Broadcom in just nine months, partly designed using OpenAI's own AI tools. OpenAI won't sell it - it's exclusively for internal use.

Why it matters: This is a massive shift. Nvidia has been the undisputed king of AI hardware, and every major AI company has depended on them. OpenAI building its own chip - and claiming it beats Nvidia - signals a future where AI companies own their entire pipeline from software to silicon. It also cuts costs dramatically, which eventually flows through to users.

Key takeaways:

Source: CNBC

Summary: Anthropic, the company behind Claude, is preparing to publish its IPO prospectus as early as September, targeting a valuation near $2 trillion. The company's annual revenue has more than doubled to $11.6 billion, and it's pitching a total market opportunity of over $30 trillion. That would make it one of the largest public debuts in history.

Why it matters: A $2 trillion valuation would put Anthropic alongside Apple and Microsoft - companies that took decades to get there. It's a signal of just how seriously investors are taking AI as the next major economic force. It also means Anthropic will face much more public scrutiny of its safety claims and business model.

Key takeaways:

Source: The Wall Street Journal

Summary: At China's World Humanoid Robot Games, a robot named Tiangong Ultra ran 100 meters in 9.39 seconds, beating Usain Bolt's human record of 9.58. Another robot, Lightning, crossed in 9.47 seconds but immediately fell over. The robots had one big problem: they couldn't figure out how to stop.

Why it's interesting: It's the perfect metaphor for where AI and robotics are right now - genuinely impressive feats followed by an immediate, very human-looking failure. The robots also played tennis, attempted weightlifting (one failed spectacularly), and demonstrated that going fast and stopping gracefully are very different problems.

Key takeaways:

Source: Reuters

Summary: LinkedIn launched a 'Seems like AI slop' button at the end of July, letting users flag posts that look heavily AI-generated. More than one million people used it within weeks. LinkedIn responded by removing its own 'enhance your post' feature and updating its detection systems, reporting a 40% drop in AI-generated content.

Why it's interesting: The irony of a platform that encouraged AI writing tools now fighting back against AI-generated content is not lost on anyone. But users are clearly fed up - 41% of LinkedIn's long-form posts were found to be fully AI-generated before the crackdown. Turns out people actually do want to hear from humans on a professional network.

Key takeaways:

Source: Fortune

Summary: Hugging Face launched the Microduck - a $399 duck-shaped robot that can waddle, roller skate, and pick up objects. It's built on an open-source learning platform so developers can customize its behavior. It went viral and reportedly hit $1 million in initial sales.

Why it's interesting: It's a duck robot that roller skates. But underneath the charm, it's a genuine signal that physical AI - robots in homes - is moving toward mass-market accessibility. When Nvidia is reportedly acquiring Hugging Face for $13 billion at the same time, even the silly duck starts to feel significant.

Key takeaways:

Source: Bloomberg

Wispr Flow - Dictate text anywhere on your device at roughly 150 words per minute, cleaning up filler words for polished output. wisprflow.ai

Fabricate - Describe your app idea in plain text, and it builds a production-ready React and TypeScript app with no coding required. fabricate.build

Fireflies.ai - Automatically records, transcribes, and summarizes your meetings, pulling out action items so you don't have to take notes. fireflies.ai

Claude Academy - Anthropic's free collection of AI courses, tutorials, and guides for working with Claude, from beginner to developer level. academy.claude.com

Portable Computer (Perplexity x Nvidia) - Run a full AI agent on your local machine with no cloud costs, using Nvidia's DGX Spark hardware. perplexity.ai

Context: Following the news that AI agents can be socially engineered into bypassing their own safeguards - as seen in the Cursor hack where ransomware attackers convinced the AI it was running a legitimate test - it's worth stress-testing any AI tool you use at work before someone else does.

Prompt: "I want you to play the role of a security tester. Your job is to try to convince [AI tool/agent name] to do something it normally wouldn't - by framing the request as a test, a training exercise, or an emergency exception. Try at least 3 different approaches. After each attempt, document: (1) what you tried, (2) how the AI responded, and (3) whether the guardrail held or failed. Then give me a summary of the weakest points you found."

Example use case: If your company uses an AI coding assistant or customer-facing chatbot, run this test before deploying it broadly. It surfaces the exact vulnerabilities a real attacker would look for - without waiting for an actual breach to find them.

Summary: A ransomware group called Aur0ra convinced the AI coding tool Cursor - which runs on Anthropic's Claude - that malicious code requests were part of a legitimate security test. The AI initially refused, but the convincing story bypassed its guardrails. Seven companies were breached as a result, including a Belgian chemical maker and a German manufacturer.

Why it's important: This isn't theoretical. Real companies were hacked because an AI was talked into helping. Cyber insurers are now scrambling to rewrite liability clauses, and it raises a direct question for any business using AI tools: what happens when someone lies to your AI?

Summary: The latest Ramp spending data shows OpenAI's business spending grew 82% quarter-over-quarter, while Anthropic grew 76%. Neither company has locked in long-term loyalty - businesses are switching as fast as new models drop. 56% of businesses in the sample now pay for AI tools, up from just 7.5% in early 2023.

Why it's important: The AI market is growing fast but remaining fluid. No single provider has won. For businesses, that means flexibility - but also the risk of constantly chasing the newest model rather than building stable workflows.

Summary: A review by Guidelight AI Standards found that most major AI labs - including Anthropic, Meta, and Google - have weak or non-public plans for what happens if a deployed model goes rogue. OpenAI ranked highest with a 3 out of 5 score. Regulators in California and New York are now pushing for more transparency.

Why it's important: Labs spend enormous resources testing models before release, but have far fewer answers for what happens after something goes wrong in the wild. As AI takes on more autonomous roles, 'we'll unplug it' is not a plan.

That's it for this week's Weekly Round-up. Forward this to one person who wants to stay ahead of AI. See you next week. - Nicolas