#029 • 15 July 2025

Contents of this Newsletter

🎭

AI Tools: Perception ≠ Reality

⏰ : 6 min

📓

Blog Picks

⏰ : 11 min

🌊

Windsurf Acquisitions

⏰ : 3 min

🗞️

Weekly AI News

⏰ : 6 min

MAIN STORY

AI Got a Reality Check This Week

You’ve probably seen the headlines. Elon Musk claims that Grok 4 can solve PhD-level problems and answer engineering questions that aren’t even documented. Reports say it scored 26.9% on “Humanity’s Last Exam” beating Gemini 2.5 Pro and GPT-4, without using external tools.

Alibaba backed startup Moonshot dropped Kimi K2, an open-source coding model. It outperformed Claude Opus and beat GPT-4.1 on several benchmarks. It’s also cheaper. The same day, OpenAI delayed its own open-source release again. If you just looked at the benchmarks, you’d think AI is leaping ahead.

But not everything moved forward this week.

The METR Study: AI Made Developers Slower

A new study from METR took a more grounded approach. They wanted to measure how AI actually affects experienced developers working on real projects and actual repo issues.

They asked 16 experienced open-source devs to work on tasks from projects they already knew. Half the issues allowed AI. The other half didn’t.

Here’s the surprising part: Developers using AI were 19% slower.

They expected AI to make them faster. They felt faster. But the numbers told a different story.

Why It Slowed Them Down

The study found five reasons AI held them back:

Developers trusted suggestions from AI that weren’t helpful.
They already knew the codebase but AI just added extra steps.
In large codebases, AI made changes that didn’t fit.
Less than half the suggestions were usable. The rest had to be fixed.
AI missed context that wasn’t documented like decisions, patterns, constraints.

So instead of building, devs ended up prompting, reviewing, waiting, and cleaning up. This talks a lot about how their perception of their performance was far from the reality. And this part matters.

You can’t measure real progress by how many pull requests went up or how many lines of code changed. Output doesn’t always mean impact. Sometimes, it just means rework. And AI, when misused, can turn productivity into noise.

Where AI Actually Helps

At Quash, we’ve always been pro-AI. We believe that when it's built with the right guardrails and applied in the right places, it can absolutely move the needle. And we’re already seeing that happen.

That’s why tools like Perplexity’s Comet browser stand out. It doesn’t try to do everything, instead it focuses on structured tasks, handles them fast, and doesn’t require constant back and forth. We’ve been using it this week for internal ops, and it’s been genuinely useful.

Prakhar

@pr_kharshakya

tried @PerplexityComet by @perplexity_ai for the day, the first browser since Arc that actually feels new.
Not just search. It “does” things.
Thanks to @eschadiol & @jeremiahjw for an invite.

9:2 PM • Jul 11, 2025

Retweets

Likes

Read 1 replies

For us, this is a great example of what it looks like when AI is designed to support real workflows. It’s a reminder that the goal isn’t to make AI look smart. It’s to make it useful. And that starts by being clear about what it should and shouldn’t do.

This is the kind of thinking we’re betting on.

What We’re Building at Quash

We realized early on that this is where AI shines: execution. When the task is clear, and the stakes are precision and speed, AI becomes an engine. A good one.

That’s why we’ve had scriptless test execution in the works for a while now.

It’s simple. You select the test cases, and Quash runs it. No scripting or setup. No YAML files, test runners, or device configs to deal with. Just pick your test cases, upload your build, choose devices and go.

It’s especially useful for fast-moving teams who already know what matters. Whether you're verifying a fix, checking a critical flow, or running a regression suite, Quash lets you skip the grunt work and get straight to results which are complete with logs, screenshots, and insights.

We’re not building AI to do everything but to take over the parts that slow you down.

Check out our Recent Blogs

SOCIAL

How Windsurf Went from OpenAI to Google to Cognition

This week, Windsurf became the center of a whirlwind power struggle in the AI coding space. OpenAI had been in talks to acquire the fast-growing IDE startup for a reported $3 billion, but when the offer expired, Google made a surprise move by hiring Windsurf’s CEO, cofounder, and key researchers in a $2.4 billion reverse-acquihire. While headlines swirled about Google’s strategic play, the remaining team at Windsurf was left without a clear future. That changed almost overnight. Cognition, the company behind Devin, announced it had acquired the rest of Windsurf including the full product, IP, and 200+ employees in a deal that came together over the weekend.

One key piece of the story is Claude. Back in June, Windsurf lost access to Anthropic’s Claude models which was a major setback, reportedly triggered by rumors of OpenAI’s acquisition talks. The move hurt Windsurf’s momentum and pushed some customers toward competitors like Cursor. With Cognition stepping in, Claude access is now back giving the Windsurf IDE a second wind, and putting it back in the hands of thousands of developers who rely on it daily.

Zooming out, the real takeaway isn’t just about who owns Windsurf. It’s signaling the evolution of developer tools. The line between AI-powered IDEs and autonomous agents is fading fast. Developers want one tightly integrated workspace that’s fast, flexible, and context-aware. Cognition now owns both ends of that spectrum: the agent (Devin) and the environment (Windsurf). In a market where giants like Google and OpenAI are watching every move, this deal just raised the bar for what a modern AI dev tool needs to deliver.

news

This Week in AI

Meta Bets Big on AI Superintelligence with Massive Data Center Push
Mark Zuckerberg says Meta will spend hundreds of billions building multi-gigawatt AI data centers as part of its superintelligence strategy. Projects like Prometheus (launching 2026) and Hyperion (scaling up to 5 GW) are already in motion. Despite recent Llama 4 setbacks and top departures, Meta is doubling down—reorganizing efforts under a new “Superintelligence Labs” division and betting on future cash flows from Meta AI, smart glasses, and next-gen ad tools.

Harmonic raises $100M to push “Mathematical Superintelligence”
AI research startup Harmonic announced a $100 million Series B funding round to accelerate the development of their Aristotle system, which aims to advance AI’s ability to solve open problems in mathematics, physics, and science. The goal is real-world “mathematical superintelligence,” expanding the impact of AI in scientific discovery

Apple eyes Mistral acquisition to boost its AI ecosystem
Apple is reportedly considering acquiring French AI startup Mistral. Mistral, Europe’s most valuable AI startup at $6.2B, is known for its fast-growing model family, strong OCR capabilities, and its chatbot, Le Chat. Apple’s interest signals a serious move to strengthen its AI stack, especially after lukewarm reception to its in-house efforts. Talks follow past rumors around Perplexity, though those were denied. An acquisition like this could mark a turning point in Apple’s AI narrative.

All Newsletter Issues | Security | Contact Us | Blogs | Test Case Generator

Unsubscribe · Preferences

AI Feels Fast. But Is It? 🏃🏻💨