#030 • 22 July 2025

Contents of this Newsletter

🧐

Are AI Models Thinking?

⏰ : 4 min

📓

Blog Picks

⏰ : 11 min

🏦

Claude For Financial Services

⏰ : 2 min

🗞️

Weekly AI News

⏰ : 6 min

MAIN STORY

The Illusion Of Thinking Models

Earlier this month, a quiet but powerful research paper from Apple made the rounds: “The Illusion of Thinking.”

It didn’t come with a new demo or model. Just a systematic teardown of how today’s most “advanced” AI models behave when asked to reason through complex puzzles. They claimed that most large language models aren’t reasoning at all. They're just mimicking answers from overfamiliar benchmarks. And when presented with logic puzzles they haven’t seen before, the cracks start to show.

GPT-4 struggled to solve puzzles designed to mirror real reasoning, things like deduction, abstraction, or chaining multiple ideas. They’re not solving puzzles the way we do: breaking down problems, planning, validating, course-correcting.

But here’s what we think:

Contaminated benchmarks are a valid concern. But the idea that today’s eval systems are totally broken is a stretch. Researchers and teams across the AI space are actively addressing these issues: moving from static to dynamic benchmarks, embedding canary strings to detect contamination, and adopting tools like HELM and BetterBench to track robustness, integrity, and real-world utility.
Evaluation frameworks lacking rigor are just far from the reality. There’s more work happening on systematic evals today than ever before. It’s not perfect yet, but what exists already works, and we’re much further along than people think.
Models aren’t reasoning. But that depends on what you mean by reasoning. What we see in most LLMs today isn’t classical or algorithmic logic. What we’re seeing is pattern mimicry at scale. And that’s not a bad thing at all.

In conclusion, these limitations are real and we think it’s important to acknowledge them. But that doesn’t mean we wait around for perfection. In the meantime, we're focused on building systems and agents that are useful now, while keeping a close eye on what’s coming next. LLMs are still learning and they’ll keep learning. They already do a lot when asked the right way, and they’re only going to get better.

Why Reasoning Isn’t Just One Thing

Here’s what’s really happening with today’s LLMs:

They’ve been trained on so much human-generated content like documentation, forum posts, internal monologues, error logs, Stack Overflow rants—that they’ve learned to mimic the way humans think aloud. Not reason, but mimic.

They aren’t breaking problems down. But they’ve seen enough examples of us doing it that they can generate the illusion of problem-solving, often good enough to be useful.

A separate study found that LLMs respond to persuasion cues like authority, reciprocity, commitment, just like humans. Even when they’ve been trained not to comply, the right framing flips the outcome. Because these systems don’t have beliefs. They just absorb patterns in how we talk, think, and behave.

It means we shouldn’t ask: "Can this AI model reason?" We should ask: "Have we framed the problem in a way the model can handle?"

The Power Of Framing

At Quash, we learned early that you don’t get reliable test automation just by throwing a fancy model at it. You need to design for the model you have, not the one you wish existed. That’s why we focus so much on context curation.

Let’s say you’re running a visual test. You want the AI to click the right button. You can’t just dump a raw XML tree and hope it figures it out.

But what if you format the XML? Structure it. Highlight component boundaries. Filter out irrelevant nodes. Now it’s not a complex puzzle anymore. It’s a simple decision. We’ve taken a high-complexity problem and made it trivial.

Same goes for test case generation. We don’t blindly trust a “reasoning” model to handle it. We benchmark models internally across tasks and complexity levels. We know which models do well at “medium-complexity” problems and where simpler models outperform them. It’s mostly about creating environments where intelligence can emerge reliably.

What AI Can Do

AI doesn't need to think like us to be useful, it just needs to perform when the context is right. We’re not trying to build AGI inside your test suite. We’re here to ship tools that make life easier for engineers and mobile teams, today.

By shaping the input, choosing the right models, and knowing how to guide them, we can turn the illusion of reasoning into actual productivity. And while we focus on what works now, we’re always planning for what’s next.

Check out our Recent Blogs

industry spotlight

Wall Street Has A New Analyst

Finance hasn't always been the most welcoming place for AI. It's a high-stakes, heavily regulated industry built on precision and risk management and not exactly a sandbox for experimentation. But that's exactly why Anthropic chose to focus their efforts there with the launch of Claude for Financial Services.

They're sending a clear signal: It's about thriving in low-tolerance environments where every number has real-world consequences. Claude wasn't just dropped into this space, it was trained for it. And it’s not just about offering an assistant for summarising PDFs or surfacing search results. Unlike tools like Perplexity, which focus on rapid information retrieval, Claude goes a step deeper, grounding every insight in verified data, offering audit-ready outputs, and integrating directly with systems like Snowflake, Morningstar, and S&P. This is research at enterprise scale, built for industries that don’t have room for guesswork.

What sets Claude apart is the way it understands the context of finance. It grasps the structure of financial documents, the nuance of risk models, and the importance of traceability. This isn't generic AI squeezed into a new shape; it's verticalized, tuned for the unique needs of the finance world, and deployed with safeguards that make compliance teams breathe a little easier. That's why firms like Bridgewater and NBIM, the world's largest sovereign wealth fund, are already using it to streamline processes and boost efficiency.

It's becoming a critical piece of financial infrastructure. Claude isn't replacing human analysts, but it is changing how they work. The hours usually lost to mundane tasks are now spent on higher-level decisions. And that's the bigger story here: AI's shift from being merely helpful to becoming indispensable, when built with intention rather than just ambition. Anthropic didn't walk into finance with another generic AI assistant. They brought a system that truly understands what's at stake.

update

We're gonna be on Product Hunt soon!

We’re gearing up to launch our test case generator on Product Hunt! If you’ve been following our journey, hit the Notify Me button and support us!

Coming Soon on Product Hunt. Check it out!

news

This Week in AI

OpenAI & Google Win Gold at International Math Olympiad with AI
OpenAI and Google AI systems both achieved gold-medal-level performance at the prestigious International Mathematical Olympiad, solving five out of six problems using advanced natural language reasoning and demonstrating rapid progress toward AI systems able to partner with mathematicians on frontier problems.

AWS Launches AI Agent Marketplace for Enterprise Deployment

AWS has unveiled an AI Agent Marketplace, offering enterprises a central hub to discover, configure, and deploy AI agents at scale. The platform allows companies to integrate prebuilt agents or customize their own into workflows like customer service, HR, and IT. Built-in guardrails for compliance and performance make it enterprise-ready. This move positions AWS as a key infrastructure provider in the emerging AI agent economy.

OpenAI Launches ChatGPT “Agent” Mode
OpenAI unveiled ChatGPT Agent, an autonomous mode that can take multi-step actions on behalf of users, such as browsing the web, filling forms, making reservations, and shopping. This significant leap toward proactive AI agents emphasizes robust safety guardrails. Initially available to subscribers (not in the EU yet due to regulatory issues), this addition was widely praised for UI and autonomy.

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

Unsubscribe · Preferences

Why Framing Beats Reasoning (and Claude’s Doing It With Billions 💸)