In partnership with

😶‍🌫️ The Illusion Of Thinking: Still Useful?

The latest wave of LLMs like Claude 3.7 Sonnet Thinking, OpenAI’s o-series, and DeepSeek-R1 often seem to “think.” They produce detailed, reflective responses, evaluate their own work, and appear to reason step by step. But a recent Apple research paper, “The Illusion of Thinking,” suggests that what we’re witnessing isn’t genuine human-style reasoning. Instead, it’s a sophisticated act of memorization, pattern-matching, and probabilistic recall masquerading as thought.

Yet, even this illusion can be remarkably powerful, at least until its limits become clear.

What Did the Study Reveal?

Apple’s researchers put these AI models through their paces using classic logic puzzles such as Tower of Hanoi, River Crossing, and Blocks World. These puzzles allowed the team to control the complexity of the task and closely examine both the final answers and the “thought process” the models used to get there.

Here’s what they found:

Three Distinct Performance Zones

• Low Complexity: Simpler, non-reasoning LLMs actually outperform their “reasoning” counterparts delivering faster, cheaper, and more accurate results.

• Medium Complexity: Reasoning models pull ahead, with more elaborate “thoughts” leading to better answers.

• High Complexity: Both types of models falter accuracy plummets, and what looks like reasoning devolves into educated guessing.

Surprising Patterns

• Less Effort as Tasks Get Tougher: As problems become more complex, models actually “think” less, even when they’re allowed more time or space to reason. This unexpected reduction in effort before failure highlights a fundamental scaling limitation.

• Overthinking and Inconsistency: On easy tasks, models sometimes find the right answer quickly and then keep going and talk themselves into a wrong answer. On harder puzzles, they often start with incorrect solutions and only stumble onto the right one after much trial and error, if they get there at all.

One of the study’s most striking findings: Even when given the correct solution method (like the algorithm for Tower of Hanoi), models frequently fail to execute it correctly. Rather than logically working through the steps, they’re echoing familiar patterns from their training data.

This means today’s LLMs don’t truly reason by building and checking logical steps from scratch. They simulate reasoning by recalling and remixing what they’ve seen before. It works until it doesn’t.

Why Is This Still a Big Deal?

Because even simulated reasoning is incredibly useful. This mimicry can power knowledge retrieval, math tutoring, basic planning, and many low-to-medium complexity logic tasks that are still useful for a wide range of real-world applications.

Such as:

1. Software engineering

2. Ticket routing

3. Customer support agent

4. Content moderation

5. Legal summarization

In fact, this stage, where models appear smart but are mostly remembering may be a crucial milestone in AI’s evolution.

What’s Next for AI Reasoning?

The real opportunity isn’t to polish the illusion, but to build models that can:

• Generalize reasoning patterns, not just recall them

• Scale their reasoning effort with task complexity, rather than giving up

• Combine memory with true symbolic logic, instead of just stringing together tokens

This means we must develop new architectures and training methods grounded in what current models can (and can’t) do. If we want AI that truly reasons, not just remembers, we need to move beyond today’s tricks.

The Bottom Line

Understanding the illusion of AI reasoning is the first step toward building systems that can genuinely think. The journey from recall to real reasoning is just beginning and it’s an exciting one.

INFRASOCIAL 
July 1 | Los Gatos  https://lu.ma/infrasocial_jul1

Join us July 1 at Hapa’s Brewing Co. in Los Gatos for INFRA SOCIAL, an evening of networking and conversations at the intersection of AI infrastructure, cloud architecture, and data innovation.

STEP SF 2025
August 20–21 | San Francisco https://lu.ma/StepSF25

Step SF returns to Silicon Valley with a global lens on innovation. Bringing together 2,000+ attendees and 150+ companies, this event spotlights AI agents, GenAI, and cross-border scaling strategies. Founders, operators, and investors from across emerging ecosystems will convene to share insight and forge new partnerships.

Bay Area Startups Collectively Secured $3.69B this week

$3.69B went to Bay Area startups this week, on top of the $16.6B from the first two weeks of June. Thinking Machines Lab, led by Mira Murati (ex-OpenAI CTO) closed a $2B seed round at a reported $10B valuation today, preceded by Applied Intuition's $600M Series F earlier in the week.

Meta not done: Last week's $14.3B investment in Scale AI by Meta (and Scale's CEO Alexandr Wang moving to Meta to lead a new AI group) turns out to be more “acqui-hire” and exit than not, as a substantial portion of the investment will be paid out to investors. Now CEO Mark Zuckerberg is recruiting Daniel Gross, CEO of Safe Superintelligence and ex-Gitlab CEO Nat Friedman – both prolific investors in AI companies - to join that new AI group. This after a recent offer by Meta to acquire Safe Superintelligence failed. And OpenAI's CEO Sam Altman reports that Meta has offered $100M compensation packages to several OpenAI researchers. More here.

M&A continuing to fill in for IPOs: In the face of a continuing drought of IPOs, M&A has become a critical path for exits. Founders and investors are moving towards M&A as a quicker, more viable path to liquidity. Per Pitchbook, large M&A continues to be rare because of elevated interest rates, uncertain economic growth, increased regulatory scrutiny, and stock market volatility. But smaller acquisitions are increasing, due to a combination of lower startup valuations, the need for liquidity, and a lack of new funding options. More...

Follow us on LinkedIn to stay on top of what's happening in 2025 in startup fundings, M&A and IPOs, VC fundraising plus new executive hires & investor moves.

Early Stage:

• Thinking Machines Lab closed a $2B seed round, an artificial intelligence research and product company that is building a future where everyone has access to the knowledge and tools to make AI work for their unique needs and goals.

• Lumana closed a $37M Series A, provides modern, AI-powered video security to enhance security and safety, streamline operations, and enable immediate incident response

• OneSkin closed a $20.4M Series A, a longevity company on a mission to transform the way we think about skin and aging.

• Ember Flash Aerospace closed a $9.8M seed round, building the future of disaster response, from our off-grid wildfire monitoring system, Vigilant, to our ultra-long endurance UAS sensor platform, Vigilant Sky.

• RIIICO closed a $5M seed round, pioneering the synchronization of physical factories with their digital counterparts, in a time of great transformation processes.

Growth Stage:

• Applied Intuition closed a $600M Series F, Applied Intuition is a Tier 1 vehicle software supplier that accelerates the adoption of safe and intelligent machines worldwide. 

• Juniper Square closed a $130M Series D, transforming the private markets investing experience with a full range of modern, connected fund software and services.

• Collective closed a $71M Series B, an all-in-one financial platform designed for self-employed entrepreneurs, it handles LLC and S Corp formation, payroll, taxes, monthly bookkeeping, compliance and more all in one place.

• Pano AI closed a $44M Series B, AI-powered wildfire detection that provides advanced early detection and situational awareness solutions to fire agencies, utilities, governments, and private landholders.

• Browserbase closed a $40M Series B, the browser infrastructure platform for agentic software, with powerful APIs, a growing ecosystem, and now no-code tools.

Follow us on LinkedIn to stay on top of what's happening in 2025 in startup fundings, M&A and IPOs, VC fundraising plus new executive hires & investor moves.

“Chain-of-thought and agentic models are breaking the old token economics.”

- RK Anand, Co-founder & CPO, Recogni

At AI INFRA SUMMIT, RK Anand didn’t just join a fireside, he set one. In a sharp exchange with Keith Newman, he laid out why the next frontier in GenAI isn’t more scale. It’s efficiency.

🚨 Compute volatility is spiking—10x to 100x swings are already happening
⚙️ Agentic models aren’t just powerful—they’re unpredictable and expensive
🔋 Energy is the new bottleneck, especially in U.S. deployments
⚡ The fix? Inference infrastructure that delivers 10x efficiency gains

RK’s call to action was clear: GenAI’s growth won’t be saved by bigger clusters—it’ll be saved by smarter silicon.

Your Feedback Matters!

Our mission is to provide an insider's view of Silicon Valley's undercurrents – insights often overlooked by mainstream sources. While many newsletters offer broad market overviews, we focus on delivering a unique, in-depth understanding of the local ecosystem. We share behind-the-scenes conversations, introduce key players we meet at events, and offer exclusive insights.

Your feedback is crucial in helping us refine our content and maintain the newsletter's value for you and your fellow readers. We welcome your suggestions on how we can improve our offering. [email protected] 

Logan Lemery
Head of Content // Team Ignite

Find out why 1M+ professionals read Superhuman AI daily.

In 2 years you will be working for AI

Or an AI will be working for you

Here's how you can future-proof yourself:

1. Join the Superhuman AI newsletter – read by 1M+ people at top companies

2. Master AI tools, tutorials, and news in just 3 minutes a day

3. Become 10X more productive using AI

Join 1,000,000+ pros at companies like Google, Meta, and Amazon that are using AI to get ahead.

Sign up and start learning AI