Your Next Startup Might Be You + AI

 Something interesting is happening in the startup world in 2026: startups are getting smaller, older, faster—and considerably more ambitious.

The startup stereotype used to be two 25-year-olds, a laptop, questionable sleep habits, and a garage.

Not anymore.

🚀 Small teams, big revenue

Startups are reaching revenue milestones much faster. Among YC companies, median monthly recurring revenue at the end of a batch has reportedly climbed from about $8,000 to $20,000. And yes, some companies are reaching seven-figure revenue in just three months.

Apparently, the traditional startup advice of “give it time” has been replaced by “give it three months and an AI agent.”

👤 One founder is now a team

Solo founders are having a moment.

Their share of YC batches has reportedly jumped from around 5% to nearly 20%.

Why? Because AI coding agents can now do a remarkable amount of the work that once required several engineers. One person can build, test, redesign, debug—and occasionally argue with an AI about why the button is still in the wrong place.

The important skill is shifting from:

“Can you build it?”

to:

“Do you know what is worth building?”

🧠 Experience is suddenly fashionable

Another surprise: founders are getting older.

Professionals in their 30s, 40s and 50s are increasingly starting companies, bringing something that can't simply be downloaded from an AI model: deep industry knowledge, judgment, management experience, and taste.

In the AI era, knowing the problem may matter as much as knowing the code.

🤖 The new startup equation

The emerging formula looks something like this:

One experienced human + several very capable AI agents + a good problem = surprisingly serious company.

Co-founders are certainly not obsolete. Teams still matter enormously as companies grow.

But the minimum number of people needed to start something has fallen dramatically.

And that may be the biggest startup story of 2026:

The startup isn't disappearing. It's getting very, very small.

The 2026 startup landscape is characterized by a significant shift toward hard tech, the rise of solo founders, and a dramatic acceleration in revenue growth (1:13-1:59).

Key Trends in 2026:

• Acceleration of Revenue: Startups are reaching meaningful revenue milestones significantly faster than in previous years. Median monthly recurring revenue (MRR) for YC companies by the end of a batch has risen from 8K to 20K, with some companies even hitting seven-figure revenues within a single three-month batch (1:46-1:57, 22:03-22:22).
• The Resurgence of Experienced Founders: There is a noticeable trend of highly experienced professionals—often in their late 30s, 40s, or 50s—starting companies. These founders are effectively leveraging their industry knowledge and management experience to navigate the current AI-integrated landscape (32:54-34:42).
• Rise of the Solo Founder: The percentage of solo founders in YC batches has spiked, growing from approximately 5% a year ago to nearly 20% currently (29:49-30:05).

Why Solo Founders are Succeeding:

• AI as a Force Multiplier: Modern AI coding agents have effectively reduced the need for large initial technical teams. A single founder can now handle what previously required a CTO or a larger engineering team, allowing them to build mature, full-stack products rapidly (4:06-5:00, 22:38-22:50, 30:44-31:42).
• Shift in Focus: The competitive advantage has shifted from merely "knowing how to build" to "knowing what to build." Experienced founders with deep industry intuition and product taste are uniquely positioned to manage AI agents to execute their vision efficiently (30:46-31:00, 33:26-33:57).
• Lowered Barriers: While co-founders remain valuable for scaling, the initial bar for entry has been lowered, enabling individual founders to reach traction before necessarily needing to bring on a partner (31:00-32:50).

The video highlights that the percentage of YC companies with solo founders has grown from 5% to nearly 20% (29:38-30:06). Several factors are driving this shift:

• Reduced Need for Large Teams: Historically, a co-founder was often essential to balance roles, such as pairing a technical builder with a skilled salesperson or "hustler." The speakers explain that knowing what to build and prompt has become more valuable than the sheer technical labor of writing code (30:06-30:59).
• Advancements in AI Tools: With the rise of agentic coding and sophisticated AI models, a single founder can now handle tasks that previously required a team of engineers, allowing them to iterate and launch products much faster (31:34-31:48, 34:59-35:53).
• Experience and Taste: Older, experienced founders in their 30s to 50s are excelling because they possess the "taste" and management skills necessary to effectively deploy these AI agents, often managing them similarly to how they previously managed human teams (32:54-34:42).

AI Slop or Human Bias?

"This image was generated by AI."

For many people, those six words are enough to change how they feel about a piece of art - even before they've really looked at it.

Ironically, research consistently shows that people often cannot reliably distinguish AI-generated art from human-created work. Yet when they are told something is AI-generated, they rate it as less creative, less authentic, and less valuable.

Welcome to the curious world of AI slop. 

The internet has embraced the term AI slop to describe the flood of low-effort AI-generated content: images, blogs, videos and social media posts.

It's content that feels repetitive, shallow, and disposable.

The problem is that the label has started to spread beyond genuinely low-quality content.

Increasingly, people use "AI slop" as shorthand for anything made with AI, regardless of its quality.

Those are not the same thing.

Several studies reviewed in a recent University of Oulu master's thesis found something fascinating.

Participants viewed artworks that were either human-made or AI-generated.

When people believed an artwork was created by AI, they consistently rated it lower - even when the artwork was actually created by a human.

Conversely, AI-generated works received higher ratings when participants believed they had been created by humans.

Even more interesting, participants generally could not reliably tell the difference between AI-generated and human-created works.

The label influenced perception more than the artwork itself.

Psychologists call this expectation bias.

We rarely evaluate things in isolation. Instead, we interpret them through the stories attached to them.

Think about wine.

People routinely rate identical wine as tasting better when told it's expensive.

The same happens with brands, music, and food.

AI art is no different.

Once viewers hear "AI generated," they begin looking for flaws:

• awkward anatomy
• strange lighting
• repetitive composition
• emotional emptiness

Sometimes those flaws are real.

Sometimes they're imagined.

The real problem isn't AI

Let's be honest.

There is plenty of AI slop online.

The barriers to content creation have collapsed. Anyone can generate thousands of images or articles in an afternoon.

Quantity exploded.

Quality did not.

But here's the important distinction:

Bad content isn't new.

Long before generative AI existed, the internet was already overflowing with it. AI simply made low-effort production dramatically faster.

The technology didn't invent mediocrity.

It automated it.

One reason AI art provokes strong reactions is that people often associate creativity with effort.

Traditional art means years of practice, emotional investment and personal expression

AI appears to bypass that journey.

Yet modern creative workflows rarely fit neat categories.

Photographers rely on computational cameras.

Architects use generative design.

Musicians use AI mastering.

Filmmakers use AI-powered editing.

Designers use AI to brainstorm dozens of concepts before refining one.

The question becomes less "Was AI involved?" and more "How was AI used?"

There is an important difference between:

• generating 5,000 random fantasy portraits for engagement farming

and

• spending days iterating on prompts, compositing outputs, painting over details, and building a cohesive artistic vision.

Both involve AI.

Only one deserves the label "slop."

Likewise, not every human painting is a masterpiece.

Quality has never depended solely on the tools.

Perhaps instead of asking:

"Was this made by AI?"

we should ask if it is interesting and communicates something meaningful. 

Those questions matter regardless of the software used.

As AI becomes embedded in nearly every creative tool, the line between "AI art" and "human art" will continue to blur.

The phrase AI slop will probably survive, but hopefully it will be reserved for what it should describe:

low-quality, mass-produced content, not every work that happens to involve artificial intelligence.

Because if we judge art by its label instead of by what we actually see, we're not evaluating the artwork.

We're evaluating our assumptions.

REFERENCE

Pudas L. Usage and adoption of generative AI tools within creative industries (Master's thesis, University of Oulu, 2026/05/13).

Lauri Pudas Master's thesis

When AI Became the Joke

On April 1st, 2026, something subtle but important happened in tech culture: artificial intelligence stopped being the future - and became the punchline.

For years, science fiction imagined AI as either humanity’s greatest triumph or its existential threat. From omniscient machine overlords to loyal robotic companions, the genre treated AI with a kind of mythic seriousness. Even in the real world, the early 2020s framed AI as transformative, disruptive, and inevitable.

But this April Fools’ Day told a different story.

Instead of fear or awe, the dominant tone was fatigue—playful, self-aware fatigue. The industry didn’t just joke about AI. It joked about itself.

What made 2026’s AI jokes stand out wasn’t that they were impossible. It’s that they were just believable enough. 

Companies introduced products like:

• AI companions for your existing AI assistants
• Smart glasses that analyze your barbecue in real time
• Coffee machines that judge your emotional readiness before serving caffeine
• Gaming AI that plays instead of you

None of these ideas are technologically outlandish. In fact, many are only a few engineering cycles away from reality.

And that’s exactly the point.

The humor no longer comes from imagining what AI can’t do. It comes from asking why we would want it to do these things in the first place.

Classic sci-fi often explored AI as a boundary—something that challenges what it means to be human. But in 2026, AI has crossed that boundary so thoroughly that it’s now embedded in the mundane.

We don’t just have intelligent systems guiding spacecraft or solving global problems. We have AI optimizing our grocery lists, adjusting our thermostats, and recommending what to watch while we scroll past those recommendations.

April Fools’ jokes leaned into this reality by exaggerating it just slightly:

• What if your AI needed emotional support?
• What if your umbrella had a processor more powerful than your laptop from ten years ago?
• What if your devices didn’t just assist you—but judged you?

These aren’t wild sci-fi scenarios. They’re extensions of trends already in motion.

There’s a quiet shift happening here, and it mirrors a broader evolution in science fiction itself.

Where earlier eras asked, “What if machines become like humans?”, today’s question is closer to:
“What if we keep giving machines roles they never needed?”

The April Fools’ jokes of 2026 function almost like micro-science fiction stories:

• The AI pet for your AI hints at recursive dependency loops between systems.
• The grilling glasses parody the idea of total optimization of leisure.
• The gaming copilot raises uncomfortable questions about agency and authorship.

These are comedic setups - but they carry the DNA of real speculative thought.

Perhaps the most telling signal is this: calling something “AI-powered” is no longer impressive on its own.

That label used to imply innovation. Now, it invites scrutiny—or outright skepticism.

This doesn’t mean AI is less important. Quite the opposite. It means AI has become normal.

And when a technology becomes normal, culture gains the freedom to critique it, mock it, and question its place in everyday life.

April Fools’ Day 2026 wasn’t just a collection of jokes. It was a cultural checkpoint:

• A sign that the hype cycle has matured
• A moment where the industry can laugh at its own excess
• A reminder that not every problem needs a neural network

If there’s a lesson hidden beneath the humor, it’s this:

Just because we can apply AI to something doesn’t mean we should.

Science fiction has always warned us about unintended consequences—but in 2026, those consequences aren’t dystopian. They’re… inconvenient. Over-engineered. Slightly ridiculous.

And maybe that’s the most human outcome of all.

Because in the end, April Fools’ Day didn’t show us a future ruled by machines.

It showed us a present where humans are still very much in charge—just occasionally making things a little more complicated than they need to be.

REFERENCES

https://www.tomsguide.com/news/live/april-fools-day-2026-live-best-jokes-pranks

https://www.cnet.com/tech/april-fools-day-2026-the-internets-sneakiest-pranks-are-coming/

https://www.pocket-lint.com/funniest-2026-april-fools-jokes-from-around-the-internet/

https://www.fonearena.com/blog/478889/april-fools-2026-tech-pranks.html

https://www.gizbot.com/gadgets/features/these-7-april-fools-gadgets-in-2026-dont-feel-like-jokes-124705.html

https://www.gizbot.com/gadgets/features/these-7-april-fools-gadgets-in-2026-dont-feel-like-jokes-124705.html

Products:

Razer “AVA Mini” (AI companion for your AI) - Technical Satire: It included features like "multi-sensory scent detection" and an "AI Pet-sonality" that evolves based on how often your other smart devices talk to it.

Traeger “MEAT-AI” Grilling Glasses - pokes fun at the actual trend of putting AR and AI into every niche hobby

Eight O’Clock Coffee “Brew O’Clock AI” - direct jab at the "Internet of Things" (IoT) overkill (If the AI senses you are "too grumpy" upon waking, it refuses to dispense the coffee until you say something nice to the machine, satirizing the forced interaction models of modern AI)

Currys “SniffGuard” - "AI Over-automation"—the idea that we have reached a point where we are using $500 worth of compute power to solve problems that take two seconds of human effort.

IGN “PlayStation Project Playmo” - Astro Bot-themed controller with onboard AI will give you gaming tips, beat the bosses for you, and even make V-bucks purchases on your behalf — and leak your personal data

OPPO “Find U” Smart Umbrella - the 2026 trend of "flagship-level" engineering being applied to things that absolutely do not need a processor

Timekettle “British-to-American Translator” - satirized the current state of hyper-specific AI translation tools.

Paradox “Crusader Kings III AI (CKSS)” - sharp jab at the wave of AI being retrofitted into games

The Invisible Waste

AI has an “invisible cost”: every prompt warms a server, pulls electricity, and - depending on the setup - uses water for cooling. That part is real, and it matters.

But here’s what’s missing from most AI sustainability takes: Knowledge loss is also environmental waste.

When an engineer retires and nobody remembers why the system works, we rebuild it.

When a report is buried under five versions, we rerun the experiment.

When last year’s “don’t do this” gets forgotten, we proudly do it again - now with even more meetings.

And every time we redo work that was already done, something physical happens. Machines run again. Labs test again. Prototypes get scrapped again. Supply chains ship again. And carbon gets emitted again.

So yes, AI uses energy. But so does humanity’s favorite hobby: starting over.

That’s the real tradeoff: Compute vs. Forgetting.

AI can absolutely make things worse - hallucinations, duplicated models, junk outputs, and the classic “I don’t trust it so I’ll redo it anyway.” That’s AI as a waste amplifier.

But AI can also act like a memory system, not just a content generator. 

We need to find what already exists instead of reinventing it, connect old lessons to new problems, flag conflicting or outdated knowledge and identify “unknown knowns” hiding in archives

If AI prevents just one avoidable rework cycle - one duplicate study, one failed pilot, one wrong design - it may already pay back its footprint environmentally.

Because the greenest computation isn’t “zero compute.”

It’s the computation that stops you from burning energy to relearn what you already knew.

The greenest work of all? The work you never have to repeat.

REFERENCES

Ulhaq I, Nayak R, George M, Nguyen H, Quang H. Green knowledge management: a bibliometric analysis, research trends and future directions. VINE Journal of Information and Knowledge Management Systems. 2024 Oct 2.

Abbas J, Khan SM. Green knowledge management and organizational green culture: an interaction for organizational green innovation and green performance. Journal of Knowledge Management. 2023 Jul 24;27(7):1852-70.

Yu S, Abbas J, Alvarez-Otero S, Cherian J. Green knowledge management: Scale development and validation. Journal of Innovation & Knowledge. 2022 Oct 1;7(4):100244.

Al-Faouri AH. Green knowledge management and technology for organizational sustainability: The mediating role of knowledge-based leadership. Cogent Business & Management. 2023 Dec 11;10(3):2262694.

Wang S, Abbas J, Sial MS, Álvarez-Otero S, Cioca LI. Achieving green innovation and sustainable development goals through green knowledge management: Moderating role of organizational green culture. Journal of innovation & knowledge. 2022 Oct 1;7(4):100272.

Sustainability of Artificial Intelligence - The invisible cost of intelligence (interview with Bonny Banerjee on innovative Sustainability channel). February, 15, 2026.  Sustainability of Artificial Intelligence - The invisible cost of intelligence

The Year AI Stopped Guessing

In 2025, artificial intelligence changed in an important way. Instead of just guessing what sounds right, AI started to check whether it is actually correct.

Before, AI worked a bit like autocomplete. It looked at lots of examples and predicted what word or answer was most likely next. That worked well for writing stories or poems, but it caused problems in math and science, where one small mistake can ruin everything.  

So, researchers changed how AI thinks.

Now, AI often works in steps, more like a careful student solving a math problem:

• One part breaks big problems into smaller ones

• Another part translates ideas into math rules

• Another part checks every step using strict logic tools

• The AI (sometimes with a human) coordinates all of this

Instead of being rewarded for sounding confident, AI is rewarded only when its answers can be proven correct.

Researchers also taught AI to:

• Think longer before answering if a problem is hard

• Check its own work and fix mistakes

• Learn from problems it already solved correctly

• Split hard problems into easier pieces

Because of this, AI got very good at math. Some systems performed as well as gold-medal students in math competitions.

This change also matters in the real world. In areas like:

• computer security

• airplanes and rockets

• financial systems

it’s more important to be right than just fast. AI that can prove its answers helps reduce dangerous mistakes.

By the end of 2025, AI wasn’t just copying knowledge anymore. It was:

• discovering new math ideas

• proving them step by step

• and checking itself along the way

2025 marks a turning point where AI moved beyond probabilistic pattern-matching toward formal verification, treating correctness as a hard constraint rather than an emergent property.

This shift fuses neural intuition with symbolic rigor—reviving neurosymbolic reasoning—by tightly integrating large language models with formal systems like theorem provers.

Inference-time scaling became central: models reason longer and more deliberately at test time, with correctness enforced via verifiable rewards rather than human preference alone.

Training advances such as RL with verifiable rewards and critic-free optimization made rigorous reasoning cheaper and more accessible, narrowing the gap between open and proprietary models.

Sparse attention enabled long, efficient reasoning traces, allowing open models to match elite performance on top math and programming competitions.

Data scarcity in formal math was broken through synthetic bootstrapping: models generate, verify, and retrain on their own successful proofs, creating a positive feedback loop.

Agentic architectures replaced monolithic provers, decomposing problems into verifiable subgoals and managing failure through recursion, graph search, and lemma-based workflows.

Verification loops matured—from external theorem provers to internal self-critics—making reasoning both more reliable and more efficient.

Proofs were not only generated but optimized for human readability, ensuring industrial-scale proof remains interpretable.

Parallel theory work showed transformers encode uncertainty and algorithmic structure, explaining why deliberate reasoning emerges with the right constraints.

These advances spilled into industry: AI systems now discover, formalize, and verify new mathematics and algorithms, not just check known results.

A new economy of truth is forming, with platforms and protocols commoditizing verification, attribution, and incentive alignment.

REFERENCES

Victor Shaw. The Industrialization of Certainty 2025 Year in Review for AI in Mathematics and Formal Methods. Dec 31, 2025 https://formalintel.substack.com/p/the-industrialization-of-certainty

Yadav C. Beyond Surface Trust: Towards Incentive-Aware Trustworthy AI (Doctoral dissertation, University of California, San Diego). https://escholarship.org/content/qt92g2w8q3/qt92g2w8q3.pdf https://www.proquest.com/openview/1670da7289a2f95b7e2d12c025fc8c9d/1?pq-origsite=gscholar&cbl=18750&diss=y

Shin D. Automating epistemology: how AI reconfigures truth, authority, and verification. AI & SOCIETY. 2025 Aug 12:1-7. https://link.springer.com/content/pdf/10.1007/s00146-025-02560-y.pdf

Yong Lin, Shange Tang, Bohan Lyu, Jiayun Wu, Hongzhou Lin, Kaiyu Yang, Jia Li, Mengzhou Xia, Danqi Chen, Sanjeev Arora, Chi Jin Goedel-Prover: A Frontier Model for Open-Source Automated Theorem Proving   arXiv:2502.07640 [cs.LG]  https://doi.org/10.48550/arXiv.2502.07640

AI, Math, and the Myth of Infinite Creativity

Is artificial intelligence on the verge of out-creating humans? A recent theoretical analysis suggests otherwise - and the reason lies in mathematics.

According to research by Professor David H. Cropley, large language models like ChatGPT are structurally limited to a level of creativity comparable to an average amateur. The constraint comes from how these systems work: they predict the next word based on probability. This creates a built-in tension between two essential ingredients of creativity - effectiveness and originality. 

Effective output uses words that make sense and fit the context. Original output surprises us. But in a probabilistic system, the more predictable and sensible a word choice is, the less novel it becomes. Push too far toward novelty, and the result turns incoherent. Cropley expresses this trade-off mathematically, showing that AI creativity peaks at just 0.25 on a scale of 0 to 1 - a ceiling that aligns with everyday, “little-c” creativity rather than professional or groundbreaking work.

This helps explain why AI often feels impressive to the general public: most human creativity sits at an average level, so a machine that replicates it appears skilled. But experienced artists, writers, and designers quickly notice the formulaic patterns. AI can mimic style and structure, but it cannot generate truly transformative ideas untethered from past data.

The takeaway is not that AI lacks value. On the contrary, it excels as a tool for efficiency, brainstorming, and support. But creativity, at its highest level, remains deeply human — driven by intuition, experience, and the ability to combine extreme originality with perfect execution.

Until AI evolves beyond statistical prediction into something fundamentally new, math itself suggests that the spark of true genius still belongs to us.

Cropley’s paper is valuable as a thought experiment, but its core flaw is this:
It treats a snapshot of current AI mechanics as a universal law of creativity.

The paper reduces creativity to a neat mathematical product:
Creativity = Effectiveness × Originality.
While elegant, this formula ignores widely accepted views in psychology that creativity is multi-dimensional, involving factors such as emotional impact, context, risk-taking, intent, meaning, and cultural value. By narrowing creativity to two variables, the model risks mistaking a convenient metric for the reality of creative processes.

Rather than proving AI cannot reach expert creativity, it primarily demonstrates that current large language models, under specific assumptions, cannot optimize novelty and effectiveness simultaneously. That is a far narrower conclusion than the one the paper implies.

The debate it raises is important - but the math may be more metaphor than destiny.

REFERENCES

“The Cat Sat on the …?” Why Generative AI Has Limited Creativity - Cropley - 2025 - The Journal of Creative Behavior - Wiley Online Library

A mathematical ceiling limits generative AI to amateur-level creativity

From Text to Pixels: How AI Models Are Learning to See and Think

Artificial intelligence keeps surprising us. Just when we thought large language models (LLMs) were all about reading and writing text, new research is showing they can also learn directly from images — even from the tiny pixels that make up a picture.

A recent study called DeepSeek-OCR takes this idea further. It’s designed to read text from images, like a super-smart version of the scanners that turn printed pages into digital files. But instead of just converting pictures into text, DeepSeek-OCR lets the model understand the pixels themselves. That raises an exciting question: could future AI models skip words entirely and just “think in pixels”?

This idea builds on a trend known as multimodal AI, where systems can handle more than one kind of input — for example, both pictures and text. OpenAI’s GPT-4o, released back in May 2024, was already doing this, and was much better at understanding context because of it.

But there’s another reason researchers are looking for change: cost. Training and running huge AI models takes enormous computing power. A McKinsey report in June 2024 found that AI training costs have been growing by about 20 percent each year. To keep progress affordable, scientists are exploring compression techniques — ways to make models smaller and faster without losing smarts.

One interesting example is ChunkLLM, a lightweight system that speeds up long-text processing by breaking data into small, meaningful chunks. Instead of wasting power re-reading everything, it learns when and where to focus attention — a clever shortcut that saves time and memory.

It’s a pattern we’ve seen before. In the early days of the semiconductor industry, engineers used scan compression to test chips faster and cheaper while keeping performance high. Now, AI researchers are doing something similar: compressing how models learn and think.

From compressed circuits to compressed thoughts, the goal stays the same — do more with less. And maybe, just maybe, the next big leap in AI won’t come from more data, but from smarter ways of seeing and thinking.

REFERENCES

Haoran Wei, Yaofeng Sun, Yukun Li   [2510.18234] DeepSeek-OCR: Contexts Optical Compression  arXiv:2510.18234 [cs.CV]  https://doi.org/10.48550/arXiv.2510.18234  [v1] Tue, 21 Oct 2025 02:41:44 UTC (7,007 KB)

Haojie Ouyang, Jianwei Lv, Lei Ren, Chen Wei, Xiaojie Wang, Fangxiang Feng   [2510.02361] ChunkLLM: A Lightweight Pluggable Framework for Accelerating LLMs Inference   arXiv:2510.02361 [cs.CL]  https://doi.org/10.48550/arXiv.2510.02361  [v1] Sun, 28 Sep 2025 11:04:00 UTC (427 KB)

DeepSeek OCR | Hacker News

Karpathy on DeepSeek-OCR paper: Are pixels better inputs to LLMs than text? | Hacker News

ChunkLLM: A Lightweight Pluggable Framework for Accelerating LLMs Inference | Hacker News

The Man, the Dog, and the Chip: AI Takes a Byte Out of EDA

Almost 50 years ago, someone cracked a joke that aged remarkably well:

“The factory of the future will have only two employees - a man and a dog. The man will be there to feed the dog. The dog will be there to keep the man from touching the equipment.”

In 2025 this punchline has finally made its way into Electronic Design Automation (EDA). Only now, the “equipment” in question is an AI system with more neural layers than the human brain has excuses for late tape-outs.

Will Circuits Start Writing Themselves?

Recent papers like “Large Language Models for EDA: From Assistants to Agents” (He et al., 2025) and “AutoEDA: Enabling EDA Flow Automation through Microservice-Based LLM Agents” (Lu et al., 2025) hint that AI isn’t just helping with design - it’s taking the wheel. Or perhaps, more accurately, re-routing the traces.

AI-driven verification tools (Ravikumar, 2025) and self-aware silicon (Vargas et al., 2025) now promise chips that can debug themselves faster than an engineer can find the semicolon they forgot. Researchers are even generating pseudo circuits at the RTL stage, which sounds suspiciously like AI daydreaming about better hardware.

Meanwhile, CircuitFusion (Fang et al., 2025) teaches chips to learn multimodally - combining circuit diagrams, timing data, and layout specs into one grand, caffeinated neural symphony. Think of it as ChatGPT meets circuit board karaoke.

As Peter Denning observed in Communications of the ACM, AI is both “light and darkness”—a Dickensian tale told in Verilog. Sure, we might get faster chips and fewer bugs, but we might also get less human engineering intuition, replaced by a kind of silicon omniscience that never sleeps and never spills coffee on the FPGA board.

Ray Kurzweil imagines a beautiful merger of human and machine minds. Sohn imagines a utopia. The rest of us? We’re just hoping the dog keeps us from accidentally retraining the wrong model.

Forget the flashy “AI singularity.” The real risk is the automation singularity—a slow, incremental outsourcing of human judgment to the same systems we built to help us. AI systems that prioritize speed, cost-cutting, and surveillance could erode not only our autonomy but also the joy of discovery—the little “Aha!” moments that made engineering fun in the first place.

AI in EDA is neither apocalypse nor utopia - it’s a grand debugging session for humanity’s relationship with technology. We’re learning to co-design not just chips, but the very process of innovation.

So, as the man and the dog look over the humming chip factory of the future, one thing is clear: the dog may still guard the console - but now, it’s also probably wearing an AI-powered collar that runs a lightweight EDA agent.

REFERENCES

https://cacm.acm.org/opinion/three-ai-futures/

Ravikumar S. AI-driven verification: Augmenting engineers in semiconductor EDA workflows. World Journal of Advanced Engineering Technology and Sciences. 2025 May 30;15(2):223-30.

Liu S, Fang W, Lu Y, Zhang Q, Xie Z. Towards Big Data in AI for EDA Research: Generation of New Pseudo Circuits at RTL Stage. InProceedings of the 30th Asia and South Pacific Design Automation Conference 2025 Jan 20 (pp. 527-533).

Mandadi SP. AI-Driven Engineering Productivity in the Semiconductor Industry: A Technological Paradigm Shift. Journal of Computer Science and Technology Studies. 2025 Jul 13;7(7):543-9.

He Z, Pu Y, Wu H, Qiu Y, Qiu T, Yu B. Large Language Models for EDA: From Assistants to Agents. Foundations and Trends® in Electronic Design Automation. 2025 Apr 30;14(4):295-314.

He Z, Yu B. Large language models for eda: Future or mirage?. In Proceedings of the 2024 International Symposium on Physical Design 2024 Mar 12 (pp. 65-66).

Xu Z, Li B, Wang L. Rethinking LLM-Based RTL Code Optimization Via Timing Logic Metamorphosis. arXiv preprint arXiv:2507.16808. 2025 Jul 22.

Mohamed KS. The Basics of EDA Tools for IC: “A Physics-Aware Approach”. InNext Generation EDA Flow: Motivations, Opportunities, Challenges and Future Directions 2025 Apr 12 (pp. 91-129). Cham: Springer Nature Switzerland.

Vargas F, Andjelkovic M, Krstic M, Kar A, Deshwal S, Chauhan YS, Amrouch H, Tille D, Huhn S. Self-Aware Silicon: Enhancing Lifecycle Management with Intelligent Testing and Data Insights. In2025 IEEE European Test Symposium (ETS) 2025 May 26 (pp. 1-10). IEEE.

https://www.linkedin.com/feed/update/urn:li:activity:7356043406298546176/

https://www.linkedin.com/in/sebastian-huhn-84657768/

https://www.linkedin.com/posts/sebastian-huhn-84657768_ieeeets-siliconlifecyclemanagement-testandreliability-activity-7334929170373783552-fuCk

F Vargas, M Andjelkovic, M Krstic, A Kar… - … IEEE European Test …, 2025 - ieeexplore.ieee.org

Fang W, Liu S, Wang J, Xie Z. Circuitfusion: multimodal circuit representation learning for agile chip design. arXiv preprint arXiv:2505.02168. 2025 May 4.  https://arxiv.org/pdf/2505.02168

https://github.com/hkust-zhiyao/CircuitFusion

Fang W, Wang J, Lu Y, Liu S, Wu Y, Ma Y, Xie Z. A survey of circuit foundation model: Foundation ai models for vlsi circuit design and eda. arXiv preprint arXiv:2504.03711. 2025 Mar 28.

Lu Y, Au HI, Zhang J, Pan J, Wang Y, Li A, Zhang J, Chen Y. AutoEDA: Enabling EDA Flow Automation through Microservice-Based LLM Agents. arXiv preprint arXiv:2508.01012. 2025 Aug 1.

Wei A, Tan H, Suresh T, Mendoza D, Teixeira TS, Wang K, Trippel C, Aiken A. VeriCoder: Enhancing LLM-Based RTL Code Generation through Functional Correctness Validation. arXiv preprint arXiv:2504.15659. 2025 Apr 22.

Next Generation EDA Flow: https://www.google.com/books/edition/Next_Generation_EDA_Flow/

RapidGPT: https://docs.primis.ai/ - industry’s first AI-based pair-designer tailored to ASIC and FPGA engineers

OpenAI x Broadcom — The OpenAI Podcast Ep. 8: https://youtu.be/qqAbVTFnfk8?si=DSl5apccjADsM7jc

Beyond Saturation: Rethinking AI Benchmarks for the Real World

Benchmarks are how we take stock of progress in AI—but what happens when those benchmarks no longer tell us what we need to know? In recent years, many language models have "solved" the flagship benchmarks like MMLU, SuperGLUE, and MedQA, with leading models approaching or surpassing human performance. This has created what researchers are calling benchmark saturation—and a growing realization that traditional testing does not reflect real-world utility.

AI now permeates high-stakes environments—from hospitals and HR departments to banking workflows—yet our evaluation frameworks remain trapped in clean, static, and largely synthetic tasks. The real world, however, is messy. Dynamic. Multi-agent. It involves judgment, uncertainty, cost constraints, ethical ambiguities, and performance under pressure. New work is emerging to address these gaps—but the way forward demands not just new benchmarks, but a new philosophy of benchmarking. 

A recent NEJM editorial (March 25, 2025) highlights an essential truth: “When it comes to benchmarks, humans are the only way.” While AI can simulate performance on reasoning tasks, the ultimate test is whether it helps—or harms—people in context. This is especially vital in clinical settings, where synthetic evaluation fails to capture the complexity of patient care and ethical decision-making.

The authors call for four critical recommendations:

 - Human-in-the-loop validation of AI outputs.

 - Use of multi-agent clinical simulations with layered complexity.

 - Evaluation of longitudinal impact, not just one-off answers.

 - Designing benchmarks that mirror actual clinical workflows, not classroom-style quizzes.

This line of thinking extends to enterprise and governmental domains as well: we need evaluations that reflect how models perform when real people depend on them.

The paper Recent Advances in LLM Benchmarks against Data Contamination spotlights another urgent issue: training contamination. As LLMs are trained on massive internet datasets, many benchmark questions (especially static, well-known ones) get memorized—compromising fairness and scientific rigor.

To counter this, researchers propose dynamic benchmarking: the continuous evolution of evaluation datasets and tasks, ideally generated or curated in a way that:

 - Prevents leakage into training data.

 - Reflects emerging domains and shifting linguistic patterns.

 - Introduces concept drift, temporal dependencies, and ambiguity—just like in real life.

But dynamic benchmarking brings its own challenges. The paper identifies a lack of standardization and proposes design principles to assess validity and reliability of such moving targets. A GitHub repository now tracks evolving benchmark methods—a sign that the community is embracing benchmarking as a living process, not a fixed scoreboard.

The ICLR 2025 CLASSIC benchmark takes this further by grounding LLM evaluation in real enterprise tasks, not hypothetical ones. With over 2,000 user-chatbot interactions across IT, HR, banking, and healthcare, the CLASSIC benchmark introduces five critical evaluation axes:

 - Cost

 - Latency

 - Accuracy

 - Stability

 - Security

Why does this matter? Because real-world AI deployment is never just about correctness. The benchmark reveals dramatic variation: Claude 3.5 Sonnet blocks nearly all jailbreak prompts, while Gemini 1.5 Pro fails 20% of the time. GPT-4o may be accurate, but it costs 5x more than alternatives.

By bringing enterprise metrics into the core of benchmarking, CLASSIC sets a new standard for trustworthy deployment-focused evaluation. We need more of this across domains.

the LLM-Powered Benchmark Factory study introduces BenchMaker, a tool for automated, unbiased, and efficient benchmark creation. Instead of relying on slow, costly human annotation, BenchMaker uses LLMs under a robust validation framework to generate test cases that are:

 - Reliable (high consistency with human ratings),

 - Generic (usable across models and tasks),

 - Efficient (less than 1 cent and under a minute per item).

It even reports a Pearson correlation of 0.967 with MMLU-Pro—suggesting synthetic benchmarks, when done right, can rival traditional ones. But the key is structure: careful curation, validation across multiple models, and feedback loops to refine benchmarks iteratively.

We’re entering a post-saturation era of AI evaluation. Accuracy alone is no longer enough. Benchmarks must reflect:

 - Context-specific utility

 - Security and robustness

 - Economic and temporal efficiency

 - Multi-turn, multi-agent reasoning

 - Human validation and trust

As benchmarks evolve into simulations, scenario-based tests, and longitudinal deployments, the community must resist the lure of simple scores. The future of benchmarking isn’t about outscoring a test - it’s about showing real-world readiness.

Researchers, practitioners, and platform developers must align on the next generation of benchmarks—not just for better AI, but for more trustworthy, useful, and safe deployment. Contribute to open-source datasets like EHR Shot (branch of CLASSIC) or The Pile.  Adopt dynamic benchmarking strategies. And most importantly, keep humans at the center.

REFERENCES

Rodman A, Zwaan L, Olson A, Manrai AK. When It Comes to Benchmarks, Humans Are the Only Way. NEJM AI. 2025 Mar 27;2(4):AIe2500143.

Deng C, Zhao Y, Heng Y, Li Y, Cao J, Tang X, Cohan A. Unveiling the spectrum of data contamination in language models: A survey from detection to remediation. arXiv preprint arXiv:2406.14644. 2024 Jun 20.

Chen S, Chen Y, Li Z, Jiang Y, Wan Z, He Y, Ran D, Gu T, Li H, Xie T, Ray B. Recent advances in large language model benchmarks against data contamination: From static to dynamic evaluation. arXiv preprint arXiv:2502.17521. 2025 Feb 23.

Wornow M, Garodia V, Vassalos V, Contractor U. Top of the CLASS: Benchmarking LLM Agents on Real-World Enterprise Tasks. InICLR 2025 Workshop on Building Trust in Language Models and Applications.

Yuan P, Feng S, Li Y, Wang X, Zhang Y, Shi J, Tan C, Pan B, Hu Y, Li K. LLM-Powered Benchmark Factory: Reliable, Generic, and Efficient. arXiv preprint arXiv:2502.01683. 2025 Feb 2. 

Using AI Without Losing Your Mind

Throughout history, every major technological advancement has changed how humans think and interact with the world. Writing was feared to erode memory, calculators were expected to diminish mathematical skills, and GPS has arguably weakened our navigation abilities. Now, generative AI presents a new challenge—it doesn’t just assist thinking, it replaces elements of it. As AI systems become more capable, they risk turning us from active problem-solvers into passive consumers of machine-generated knowledge.

A recent study by Microsoft and Carnegie Mellon University (Lee et al, 2025) found that increased reliance on AI led to a decline in critical thinking among 319 knowledge workers. Participants who placed high trust in AI were less likely to verify its outputs and reported losing confidence in their ability to perform key tasks such as writing, analysis, and decision-making. Other studies suggested that AI can develop sophisticated manipulation and deception tactics (Williams et al., 2025), can be self-aware (Betley et al., 2025) and can itself demonstrate "critical thinking" in conversations (Greenblatt et al., 2025) - raising the question: Are we going to outsource too much of our cognitive work to machines? 

This phenomenon resembles the "irony of automation"—where over-reliance on tools leads to skill atrophy. Elevators have reduced our need to climb stairs, spellcheck has weakened our spelling proficiency, and video has replaced long-form reading. With AI now at the helm of many routine tasks, we must ask: will we be able to refrain from using these tools as crutches? Just as regular exercise is necessary to maintain physical fitness, setting aside dedicated periods to work through problems without AI assistance is essential for keeping our critical thinking muscles active.

If humans tend to take the easiest path, we risk losing the capacity to question and evaluate AI-generated information. This concern echoes a central theme in Frank Herbert’s Dune, which warned of a future were humans, overly reliant on thinking machines, lost control over their own cognitive processes. While today’s AI isn’t nearing sentience, it is already shaping how we think, work, and communicate. The challenge ahead isn’t solely about making AI more powerful—it’s about ensuring that it enhances, rather than replaces, human intelligence.

REFERENCES

Lee HP, Sarkar A, Tankelevitch L, Drosos I, Rintel S, Banks R, Wilson N. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. https://www.microsoft.com/en-us/research/uploads/prod/2025/01/lee_2025_ai_critical_thinking_survey.pdf

Greenblatt R, Denison C, Wright B, Roger F, MacDiarmid M, Marks S, Treutlein J, Belonax T, Chen J, Duvenaud D, Khan A. Alignment faking in large language models. arXiv preprint arXiv:2412.14093. 2024 Dec 18. https://doi.org/10.48550/arXiv.2412.14093

Betley J, Bao X, Soto M, Sztyber-Betley A, Chua J, Evans O. Tell me about yourself: LLMs are aware of their learned behaviors. arXiv preprint arXiv:2501.11120. 2025 Jan 19. https://doi.org/10.48550/arXiv.2501.11120

Herbert, F. (2006). Dune. Hodder Paperback. https://genius.com/Frank-herbert-chapter-1-dune-annotated

Discussions:

https://news.ycombinator.com/item?id=43057907

https://news.ycombinator.com/item?id=43028827

https://www.reddit.com/r/Futurology/comments/1jxwu64/will_ai_make_us_cognitively_dumber/