How I Accidentally Beat the Official DeepSeek API by Overclocking an “Outdated” A100 from 33 tok/s to 673 tok/s

By shi3z | September 12, 2026

(Update: Part 2 is now live)

The Executive Summary

Long story short: what happened in my lab over the last 24 hours blew my mind. It’s 5:00 AM right now, and I’m pretty sure I just caught a glimpse of the dawn of a completely new era.

This is going to be a long post because it’s a chronological log of a frantic 24-hour engineering sprint. But it boils down to two takeaways:

1. If you take AI engineering seriously rather than treating it like magic fairy dust, your gains won’t just be incremental — they will be astronomical.
2. Using a six-year-old, supposedly “obsolete” datacenter GPU, I managed to run a bleeding-edge frontier open model faster than the cloud providers themselves.

The open-source code is public, so anyone with matching iron can reproduce this.

GitHub - shi3z/deepseekv4.1-A100-custom

The newly minted DeepSeek v4.1 Flash was engineered explicitly for next-gen silicon. It fundamentally ignores older architectures. My trusty NVIDIA A100s were, on paper, completely unsupported. Under normal circumstances, I would have thrown in the towel or called it a day the second a test run spat out a token.

Instead, I went down the rabbit hole.

A quick primer on tok/s (tokens per second): this is the pulse of LLM inference. It dictates how fast your model “thinks.” More tokens per second equals higher throughput. By the end of this sprint, I pushed single-stream decoding up to 165 tok/s, parallel batches up to 1,000 tok/s, and a 4-GPU cluster to a staggering 1,467 tok/s. For perspective, Claude Code generally hovers around 50–80 tok/s. DeepSeek’s official cloud API caps out around 400 tok/s.

My home lab just outran the cloud.

Daily Life with Dual Stack-chans

It started innocently enough. My second M5Stack-based Stack-chan arrived.

I had a spare M5Stack Core2 lying around, bought an extra chassis, and snapped it together — only to discover the Core2 lacks an integrated camera. Hardware fragmentation strikes again. Worse, the internal speaker on the Core2 sounds like a dying mosquito.

Still, with two units sitting on my desk, I set them up to do a little two-man comedy routine (manzai). I customized their face displays: one sporting deadpan, half-asleep eyes, the other rocking a slightly mischievous, wide-eyed stare.

Watching them banter, I remembered a screenplay I wrote years ago about autonomous AI agents that gradually invent their own inscrutable language. Sci-fi back then; Tuesday morning reality now.

Then, mid-skit, sleepy Stack-chan piped up:

“Hey, DeepSeek v4.1 Flash just dropped.”

“Yeah? What’s the gimmick?”

“Word on the street is it runs locally.”

“Local,” right. Usually, that means “local to anyone with an illicit rack of 16 liquid-cooled H200s.” But reports claimed it could actually boot on a 512GB Mac Studio. Well, I have enough hardware to match that.

The Shock of DeepSeek v4.1 Flash

DeepSeek v4.1 Flash isn’t just a minor point release. Its architecture relies on a novel lookup structure dubbed the engram — a massive dictionary mechanism that trades compute overhead for dramatically lower active VRAM pressure. Because these engrams aren’t accessed on every single cycle, they don’t strictly need to hog precious high-bandwidth VRAM; you can shove them into system RAM or offload them to high-speed NVMe arrays at 4-bit quantization.

In my lab sits Tsugunosuke, my personal 8x A100 80GB rig. It’s an eccentric monster boasting 1.7 TB of system RAM — a capacity Intel’s spec sheet technically claims shouldn’t work on this board, but my friend Onuma and I hacked together anyway.

The catch? The A100 belongs to the Ampere architecture. It has zero native acceleration for FP4 or modern microscopic block scaling.

I asked Claude Code to whip up an initial loader, only to get hit with the standard robotic defeatism:

"vLLM does not support FP4 on Ampere architecture. Execution is impossible."

I almost closed the terminal. Then I paused.

Wait a second.

A 4-bit float multiplied by a 4-bit float yields, at most, an 8-bit result. That is exactly 256 possible outcomes. The entire operation can literally fit inside a 256-element lookup table (LUT).

I pitched this to ChatGPT: “Can’t we just LUT this?”

ChatGPT bounced back: “Absolutely. It’ll fit cleanly.”

When you hit a wall, the secret isn’t finding a new library. The secret is refusing to rely on existing libraries and writing the damn kernel yourself.

At stock settings across 6x A100s, it crawled at a meager 33 tok/s. On a single A100 bottlenecked by host RAM transfers, it choked at 5 tok/s.

That’s when drikin — tech creator, gear collector, and owner of a four-node DGX Spark cluster — decided to drop by my mentions to talk some trash:

“My four Sparks hit 100 tok/s out of the box.”

Excuse me?

A cluster of consumer-adjacent toys flexing on an 8-way datacenter monster? Over my dead body.

I am fifty years old, but the inner competitive hacker took the wheel. Challenge accepted.

Questioning the “FP4 Incompatibility” Myth

The industry loves broad claims like: “Ampere cannot do FP4.”

What is FP4? It’s just a 4-bit floating-point number — a sign bit, an exponent, and a mantissa. Scientific supercomputing relies on FP64 for climate modeling and orbital trajectories. Modern AI, however, has proven absurdly resilient to low-precision quantization.

The appeal of FP4 is simple: compared to FP16, it slashes memory footprint by 75%. That means 4x the weight density and 4x the theoretical memory throughput.

Looking at the format specs, an obvious trick jumped out: you don’t even need a lookup table.

Take the E2M1 standard (2-bit exponent, 1-bit mantissa). If you multiply its raw representation by 2, it maps neatly to a clean set of signed integers:

Plaintext

0, ±1, ±2, ±3, ±4, ±6, ±8, ±12

This means an E2M1 value can be decompressed into a signed INT8 integer with zero precision loss.

And what does an A100 have plenty of? INT8 Tensor Cores.

Never confuse your storage format with your compute format. “A100 lacks FP4 support” simply means it lacks Blackwell’s single-instruction hardware decoding for FP4 operands directly into Tensor Cores. It does not mean the A100 cannot calculate the math.

The Real Villain Wasn’t Math — It Was the Bus

I spent hours obsessing over Tensor Core micro-ops, but my first profiler run revealed a much uglier bottleneck.

In single-GPU setups, the MoE (Mixture of Experts) weights lived in host system memory, dispatching chosen experts over PCIe Gen4 x16 on demand.

It was agonizingly slow.

Each token required shipping roughly 4.5 GB of expert weights across the bus. At PCIe Gen4 x16’s real-world ceiling of ~25 GB/s:

Actual measured latency: 179 ms per token.

That single transfer physically caps your speed at 5 to 6 tok/s. Squeezing 20% more math out of your GPU cores is useless when your bus is choking on 4.5 GB of raw weight transfers per token. The architecture was fundamentally broken.

Offloading Compute to the CPU

Fortunately, Tsugunosuke isn’t powered by a desktop processor. It runs dual-socket Intel Xeon Silver 4410Y chips (24 cores / 48 threads), 16 channels of DDR5–4000 RAM, AVX-512, VNNI, and AMX extensions.

If shipping experts to the GPU costs 179 ms, why send them at all? Why not compute the MoE layers right on the host CPU?

Leave the FP4 weights parked in DDR5 memory. Fetch them straight into CPU vector registers, expand them to INT8 on the fly, and crunch the dot products using Intel’s VNNI instructions:

Layer Processing Strategy (40 Layers)LatencyPCIe DMA Transfer to GPU179 msCPU Native BF1662 msCPU INT8 via VNNI51 ms

Computing the entire MoE stack on the CPU was over 3x faster than simply copying the data to the GPU to compute it there.

In heterogeneous computing, this is rule number one: theoretical GPU TFLOPS mean nothing if the interconnect is starved.

Death by 2,000 Kernel Launches

The CPU was faster, but 51 ms per token is still unacceptable. Profiling system RAM revealed we were hitting only 89 GB/s out of a theoretical 512 GB/s DDR5 bandwidth.

We weren’t memory-bandwidth-bound; we were instruction-bound.

Claude Code’s naive implementation was hammering tens of millions of instructions per layer just handling bit-unpacking, index shuffling, block scaling, and pointer arithmetic. No sane systems programmer writes code like that.

To fix it, we rearranged the weight layout in RAM during initial model loading. File formats optimize for disk storage; inference engines should optimize for cache-line ingestion. By repacking nibbles to let multiple rows share activation fetches and scaling computations, we drastically cut down instruction bloat.

Then I turned the profiler back toward the GPU.

Single-token decode was stalling at 28.4 ms. In silicon terms, 28 milliseconds is an eternity.

Claude Code immediately hit me with: “We have reached the architectural limits of the hardware.”

Never believe an LLM when it claims you’ve hit a fundamental hardware limit. Always inspect the trace:

• Expert GEMV: 7.0 ms
• Dense GEMM: 9.2 ms
• Miscellaneous overhead: ~12.0 ms

What was that 12 ms of “overhead”?

The runtime was launching nearly 2,000 independent CUDA kernels per token.

RMSNorm, RoPE, quantize, dequantize, tiny element-wise operations — spread across 40 transformer layers, dispatched as individual kernels. Launching a single kernel takes several microseconds; multiply that by 2,000, and you’ve burned over 10 milliseconds just talking to the driver.

The GPU wasn’t slow. The CPU was wasting all its time filing paperwork.

Bypassing Global Memory

The dense layers presented another headache: weights were stored in FP8, but the A100 has no native FP8 Tensor Cores.

The baseline script used this painfully naive pipeline:

Load FP8 Weights from VRAM 
  -> Dequantize to BF16 in Registers 
  -> Write massive BF16 Weights back to Global VRAM 
  -> Launch standard BF16 GEMM

Writing intermediate weights back to VRAM is madness. Global memory latency on an A100 is 400–600 clock cycles (284–426 ns). Registers take 1 cycle (0.71 ns).

FP8 in VRAM 
  -> Stream straight to Registers 
  -> Unpack bits directly to BF16 fragments 
  -> Feed straight into MMA (Matrix Multiply-Accumulate)

By keeping the expanded weights entirely within the register file and shared memory, dense execution dropped from 9.2 ms down to ~5.0 ms.

The Anatomy of Throughput: Topology and MTP

Single-GPU performance steadily climbed: 17 tok/s, 27 tok/s, 32.5 tok/s, and finally 58.9 tok/s on standard single-stream decode. A 10x improvement over stock.

Next came multi-GPU scaling.

An 8-GPU A100 chassis is not a flat pool of compute; it’s split across NUMA domains and NVLink bridges:

[NUMA Node 0]
GPU0 === NVLink === GPU1
GPU2 === NVLink === GPU3

[NUMA Node 1]
GPU4 === NVLink === GPU5
GPU6 === NVLink === GPU7

In low-batch inference, naive 8-way tensor parallelism actually degrades performance because interconnect synchronization and tiny GEMV operations devour your gains. Splitting the machine into two isolated 4-GPU instances yielded far superior efficiency.

To break the serial execution barrier, we enabled Multi-Token Prediction (MTP).

Standard autoregressive decoding generates one token per step

MTP speculatively evaluates multiple token candidates concurrently per step. This transforms single-row vector operations (GEMV) into multi-row matrix multiplications (GEMM), allowing the GPU’s compute engines to achieve actual saturation.

With MTP set to K=3 and K=5, throughput scaled dramatically.

Then came Semantic-Aware Batching.

In Mixture-of-Experts architectures, grouping 32 random prompts forces the engine to activate nearly every expert across the model. But if you batch semantically similar prompts together — coding tasks with coding tasks, translations with translations — they route to overlapping sets of experts.

In our benchmarks, sorting incoming requests by semantic intent reduced the active expert count from 216 down to 156 per layer, massively reducing memory fetch overhead.

The Scoreboard

The final benchmarks confirmed the setup had completely decoupled from its original bottlenecks:

• Single-stream decode: 165 tok/s
• 8-stream concurrent throughput: 600+ tok/s
• 64-stream batching: 674 tok/s per 4-GPU instance
• Dual 4-GPU cluster aggregate: 1,467 tok/s

For context: DeepSeek’s official hosted endpoint sits around 400 tok/s. drikin’s quad DGX Spark rig tops out near 100 tok/s.

The Return of Local Computing

This project started as an absurd weekend bet to see if a legacy datacenter GPU could survive a modern FP4 workload. But the broader takeaway runs much deeper.

Tsugunosuke cost me roughly ¥40,000,000 (~$270,000 USD) to assemble over time. In everyday life, that’s serious capital. In modern AI infrastructure, it’s pocket change. A brand-new 8x H100 or B200 chassis will easily set an enterprise back close to $1,000,000 USD.

Yet raw silicon specs are only half the battle. If your inference software is unoptimized, an eight-figure cluster will crawl. Conversely, when you strip away kernel bloat and engineer around your hardware’s exact memory hierarchy, an older machine can deliver cloud-grade performance right under your desk.

Think about what local ownership actually means:

• Zero API fees, zero per-token metering, and zero rate limits.
• Complete data sovereignty — zero risk of telemetry ingestion or “Harvest Now, Decrypt Later” interception.
• Freedom to hack the kernels, rewrite quantization logic, or inject custom speculative schedulers at will.
• At 3:00 AM, if I want to blast through a million tokens across a private agent swarm, the full 1,400+ tok/s pipeline belongs entirely to me.

The personal computer revolution didn’t occur because early microcomputers were faster than mainframes. It occurred because owning your own compute creates software paradigms that time-shared mainframes never permit.

Cloud AI will always own massive scale and foundational pre-training. But for a lab, an engineering team, or an enterprise running high-throughput daily operations, the local AI supercomputer isn’t a nostalgic pipe dream anymore.

It’s already here.

So, I’ve been editing Project DENT Season 3 today, but before releasing the main episodes, I decided to first put out a bonus episode introducing game creator Tetsuya Mizuguchi, who serves as a judge for the series.

It’s the untold story behind that timeless classic, Sega Rally.

When I first moved to Tokyo for college, I went to the arcades every single day. Arcades back then were full of elaborately designed cabinets and unique games — just looking around was enough to get your blood pumping. A new game arriving was enough on its own to become the talk of the school. That’s the kind of place it was.

Among them, one thing that truly blew my mind was a giant cabinet racing game that physically tilted and lurched — and the one with the most overwhelming presence in the arcade was Sega Rally.

Around the time I started getting involved in the game industry, I was invited to the year-end party of the CSK Group, which owned Sega at the time. The party was held at a rented-out Joypolis inside a Takashimaya department store, and there, an actual Celica had been mounted on a massive moving rig — a bizarre game where you literally felt, in your body, the acceleration, deceleration, and lateral G-forces of driving. I’ve never seen a game that daring, before or since.

A few years later, when Sega Rally 2 was being ported to consoles, Dwango was brought on to handle the online multiplayer component. I wasn’t directly involved, but I was on-site as an engineer dispatched by Microsoft to support the OS technology side. My job at the time was handling OS troubles and fixes that came up on-site, coordinating over phone and email with headquarters in Redmond. This was, after all, the first proper console game ever built on top of a full operating system. It was a huge undertaking for Microsoft too — even counting outside contractors like myself, the Japan-side team numbered only about five people. It was important enough that BillG (Bill Gates) personally emailed encouragement to James Spahn, who led the Japan effort — though he may well have been sending that to every department, every day.

At the time, the console port was being handled by CS1 Ken (the home console development division), while Mizuguchi-san, who created Sega Rally, was the department head of AM12 Ken (the arcade game development division).

Word reached me that “a producer named Mizuguchi, who can’t even write a line of code but is a real pain, keeps making unreasonable demands about the console version of Sega Rally,” and I pushed back with things like “if the hardware spec makes it impossible, it’s impossible.” I never actually met him in person during that time — the only impression that stuck with me was “an extremely stubborn, particular guy.”

Later, around the time I went independent, Mizuguchi-san and Kuniyasu Utsumi-san — who had spun off from Sega together — came to visit my company, back when it still only had three people. This time, they were making a mobile game. That was the first time I actually met Mizuguchi-san, and I was completely won over by his personality. He was nothing like the impression I’d formed from TV — warm, approachable, always laughing like he was having the time of his life.

From there, through various turns of events, I ended up learning from Utsumi-san as a mentor in business and from Mizuguchi-san as a mentor in creative work, in all sorts of ways. Both of them took me to various places around the world, which massively expanded my horizons. Decades later, Utsumi-san would go on to serve as an outside director for my company.

During a break in the Project DENT judging, we got to chatting. “Sega Rally was such a great game,” I said — and it struck me that this was actually the first time we’d ever talked about it directly, despite having known each other for close to twenty years. That’s when I finally heard the untold story of Sega Rally’s development, straight from him. I’m not sure why I never asked before — maybe the impact of his work and philosophy since Rez was so overwhelming that it never occurred to me to look further back.

As it turns out, when building the game, Mizuguchi-san first went to an actual racetrack, put a programmer in the passenger seat, and personally drove a real rally car himself to get a feel for drifting — crashing and totaling several cars in the process before the game came together.

A development approach that would be unthinkable in ordinary game production. But that kind of hands-on, physical-sensation-driven approach to game-making really was characteristic of Sega at the time. And indeed, Sega Rally came out with a texture completely different from every other racing game out there. To make it feel like a real rally, they went and secured logo licenses from actual car manufacturers and tire makers — creating an atmosphere that was fake yet felt utterly authentic. Even the mob of annoying onlookers who’d wander onto the track popping camera flashes contributed to that distinctive, tactile sense of realism.

Mizuguchi-san himself apparently ended up with whiplash from chasing that feel for traction. It struck me all over again: so this is what being a game producer actually means.

To think that a game I was obsessed with in high school and college was made this way!

Tetsuya Mizuguchi really is a genius.

One day in April. At 7:30 in the morning, a group of men and women gathered at Shinjuku West Exit.

Wearing sunglasses even in the early morning and hauling brightly colored suitcases, this group looked nothing like ordinary tourists.

Normally, people on a sightseeing trip would look a lot more cheerful, but every one of their faces was tense.

Twenty men and women had been gathered here, on a Monday morning on a weekday. Hardly any of them knew each other. What was about to begin? Not just the participants — even the company that had chartered the bus and the company that had planned the event had been told nothing. It was a bizarre departure.

And no wonder — because the official name of this hackathon was something I myself had only hammered out with ChatGPT, head to head, that very morning before it started.

This hackathon came about because Yuzuki-san of Koto Planning proposed it, saying, “Your hackathons are so fun, Shimizu-san — let’s do more of them.” But since I already serve as the overall MC for the All-Japan AI Hackathon, I couldn’t just casually take on all sorts of other hackathons.

I once turned her down, saying “A hackathon with open public applications is off the table — that’s what the All-Japan event is for.” But she was so enthusiastic about wanting to do something that I said, “Alright, let me think of something,” and headed off to Golden Gai as usual, turning various ideas over in my head.

The bartenders at Golden Gai typically rotate by the day — even at the same bar, a different person tends it depending on the day of the week. Friday especially is the big money-making night, being the start of the weekend. The Friday bartender is basically the star of the shop, the top earner. Among that Friday cleanup crew, probably the oldest hired bartender — Kyoko-san, over seventy — said something unexpected to me.

“You know, I’ve gotten curious about this AI thing. I went to your talk in Shibuya recently — was it a lecture? A talk show?”

“Wait, you were there?”

“I was. And after that, I thought I’d like to try a hackathon myself. I’ve been inviting some of the younger folks to come with me.”

“Oh, that’s great to hear. We’re actually recruiting participants for the Tokyo qualifier of the All-Japan AI Hackathon right now — you should definitely apply.”

“I will, I will.”

Kyoko-san used to be what’s called a kyoiku mama — an education-focused mother. When her children were young, she toured European boarding schools with them and took them on trip after trip — what I’d call a “Grand Tour,” practiced from early childhood. Later she worked at exam-prep schools for elite private education, and has a track record of getting numerous students admitted to prestigious schools, including Keio’s kindergarten. She herself graduated from a music college’s vocal program and now teaches piano — which, incidentally, closely mirrors my own mother, who also graduated from a music college’s vocal program and became a piano teacher.

Now retired from her main career, she works several part-time jobs while steadily saving money for her dream of traveling to Europe again someday.

I feel bad for Yuzuki-san, but the All-Japan AI Hackathon is enormously important to me. AI has lowered the barrier to programming, and because that barrier is lower, beginners can now casually join and even go toe-to-toe with — or beat — experienced developers. Sometimes a beginner pulls off a giant-killing upset against a veteran. That’s the thrill of it, and I think it might be the single most exciting competition of this AI era.

Because of that, the lottery for entry is run strictly and fairly — even when I personally encourage someone to apply, they often don’t get picked. The first Tokyo qualifier in particular saw a flood of applications, since it was the first one held in a year.

Sure enough, when the roster was handed out that day, Kyoko-san’s name wasn’t on it.

“I didn’t get picked,” she said, sounding disappointed, that Friday. It stung. Given her age, I thought, there might not be much time left for her to join the All-Japan event and build up her skills.

That’s when I remembered Yuzuki-san.

That’s it — let’s hold a hackathon for Kyoko-san. Not just for her, either — why not let the regulars and bartenders smoldering away in Golden Gai experience AI-powered software development? Rather than open recruitment, I’d hand-pick the participants myself — that would create a clear distinction from the All-Japan event.

When I proposed this to Yuzuki-san, she said, “I have a candidate venue — want to go take a look?” and took me to scout Lake Kawaguchi, which became the event site.

SANU 2nd Home is a currently popular service that lets you partially “own” a share of various glamping facilities around the country. One of them, SANU 2nd Home Kawaguchiko, close to Tokyo, turned out to be a stunning spot with a view of Mt. Fuji.

“Oh no…”

Suddenly I couldn’t picture the Golden Gai crowd taking a hackathon seriously in a place like this at all.

The location was simply too gorgeous. Most of these people don’t touch AI — heck, they barely touch a computer — in daily life, and if I brought them somewhere like this, they’d just end up drinking from morning to night as usual. Which, in fairness, is a perfectly reasonable response to a place this beautiful — it’d almost be a waste not to.

“Let’s make the Golden Gai crowd one team,” I decided on the spot.

“Why’s that?” Yuzuki-san asked, looking puzzled.

“If I bring the whole Golden Gai gang out here, it’s going to turn into chaos. Forget the hackathon.”

The image of Golden Gai’s infamous cherry-blossom-viewing free-for-alls came vividly back to me. That would never make for a hackathon, let alone an article.

For now, I called GOROman.

“Are you free one Monday and Tuesday in April?”

His reply came back immediately — instantly.

“I’m free. I’m unemployed.”

“Then let’s meet at Shinjuku West Exit at 7:30.”

A short pause, then a few minutes later:

“Got it.”

“Alright.”

“Who did you invite?” Yuzuki-san asked.

“Yoshihito Kondo — GOROman, Naru-sensei, whatever you want to call him — basically, a genius creator.”

Who else… oh, Kazunori Nanji, definitely — can’t skip him for an event like this. Actually, I haven’t seen Yoichiro Miyake in a while either, would be nice to catch up… and on the spot I recruited about six people, and every single one confirmed immediately. Remarkable, considering I hadn’t even explained what we’d be doing.

“Anyway, let’s just do our own original hackathon here. If you’ll leave everything to me, I’ll take it on.”

“Understood. I trust you, Shimizu-san. By the way, you mentioned before wanting the theme to be ‘The Hackathon That Changes Your Life’ — what’s the theme this time?”

“‘The Hackathon That Changes Humanity.’”

Honestly, even if Yuzuki-san had turned me down at that point, having already gathered a lineup like this, I would have found a venue myself and made this hackathon happen regardless. Ideas live and die on momentum — you have to act the instant you think of them. And it was in that moment I realized I’d been unconsciously boxing myself in. No — I could be freer than that. More selfish, even. There are things you simply can’t do without that kind of freedom.

The reason giant-killing happens at the All-Japan AI Hackathon is that the entry bar is set deliberately wide open. But if people like GOROman or Nanji showed up there “as participants,” it would throw the whole event into chaos — frankly, it’d be “too big a deal for grown-ups behaving like grown-ups.” Conversely, though, that also means they themselves never get to experience the fun and thrill of actually competing in a hackathon.

In other words: gather participants of a caliber normally only invited as “judges” at an ordinary hackathon, together with people who, under normal circumstances, would never even set foot in a hackathon — cut them off from the outside world, and half-force them to compete against each other.

In a setting like that, I thought, maybe we could spark collisions of wisdom nobody had ever imagined, resonances of insight, and ultimately entirely new chemical reactions that could open up humanity’s future.

It was around this same time that I conceived what would later become the crowdfunded “AI Arcade concept for expanding humanity’s imagination.”

It needed a name worthy of it. After going back and forth with ChatGPT, I decided to borrow from Steve Jobs’s famous line, “Put a dent in the universe.”

More precisely, what Jobs actually said was:

“Put a dent in the universe. Otherwise why else even be here?”

Given the position Jobs was in when he said this, it’s a wildly cocky thing to say.

The year was 1983. At that point, although Jobs was a co-founder of Apple Computer, he was, frankly, dead weight. Apple’s early success clearly rested on the Apple II that Wozniak had built, and Jobs diverted the ample funds that success generated to try to build the Lisa computer — named after a daughter he wouldn’t publicly acknowledge as his own.

But Jobs was ousted from the Lisa team he himself had started, after a string of problematic behavior, and instead hijacked Jef Raskin’s low-cost computer project — the Macintosh team — raised a pirate flag over the company, and forcibly redirected the Macintosh from a low-cost line into a premium, high-performance one, mounting what was, for the time, a reckless gamble. This was the context in which he said those words.

At that point, Jobs was still nobody in particular — he had no real accomplishments in the computer world (at best, credit for pulling his old friend Woz away from Hewlett-Packard to build the Apple II). And yet, even then, he was already declaring that his life would have no meaning unless he did something significant enough to “put a dent in the universe.”

And the bitter irony is that after Jobs pushed his team this hard, insisting there was no point in living without putting a dent in the universe, the finished Macintosh was a commercial disaster — and Jobs himself was pushed out of the company. There are plenty of Jobs biopics out there if you want the details.

In other words, for a long time, this quote was something of a dark chapter for Jobs. He was only vindicated later — first ousted over the Macintosh’s poor sales, then brought back precisely because of another Macintosh sales slump, at which point he created the iMac and the iPod. That was the start of his comeback.

I’m turning fifty this year. Getting older has its upsides, and one of the best is: there’s really nothing left to be embarrassed about. Somewhere between middle age and old age, there’s no point being shy anymore. So, shamelessly, without a hint of embarrassment, let’s just say it plainly: push the human race forward. Incidentally, “Push the human race forward” is a line from the legendary “Think Different” campaign, delivered by Steve Jobs himself after his return to Apple in 1998.

Two versions of the Think Different campaign were made — one narrated by Jobs himself, and one narrated by actor Richard Dreyfuss. The ad agency pushed for the Jobs version, but Jobs reportedly killed his own narration.

“This is Apple’s campaign, not mine,” he said.

But looking back now, what we should really appreciate is the rare fact that this message survives in Jobs’s own voice. There really is something in a person’s own voice that carries a kind of lingering resolve.

“Alright, let’s go with this. This hackathon is Project DENT. A hackathon to put a new dent in the universe and push humanity forward.”

I’m not at an age where I need to be embarrassed anymore. And besides, Jobs said far more embarrassing things in his far more embarrassing twenties. What’s a fifty-year-old like me got to be scared of?

The twenty men and women gathered at Lake Kawaguchi were stunned when they were suddenly told this. Met with blank stares, a sheet of paper was handed out — one I had secretly printed that very morning, because the name had only been finalized at 5 a.m. that day.

On it was written the philosophy, motivation, goals, spirit, participant selection criteria, and judging criteria for Project DENT — all written by me personally, without any AI assistance whatsoever. Not a single word left to chance, because words carry lingering resolve.

Let me reprint it here for readability.

Project DENT

Humanity is stagnant. Since the arrival of the iPad in 2010, humanity has yet to produce a genuinely new tool.

Even though we’ve seen technological breakthroughs — deep learning starting in 2013, the generative AI boom from 2020 onward — few examples exist of these being properly wielded as tools.

Humanity still has not surpassed the vision Alan Kay laid out in 1968.

AR, VR, AI — these are all technologies envisioned back in the 1960s. Back then, few people imagined the computer industry would grow to today’s scale.

Now, computing is simply too profitable. Technology attracts expectation and capital in excess, and as a result, crude, money-driven business models run rampant, while the truly important questions for humanity get pushed aside. New technology is always born out of play. As humanity’s mission, we must put a new dent in the universe. Let’s push humanity forward.

The best way to predict the future is to invent it.

Participant selection criteria

Project DENT participants are selected and summoned according to the principle of the Wisdom of Crowds. Participants must either possess some outstanding specialized skill, or bring experience from an entirely different field capable of introducing ideas never seen before.

Judging criteria

Project DENT judges entries above all on whether the work created offers imaginative value that contributes to humanity’s future. The criteria rest on three axes:

• Visual appeal — Visual charm. Does it spark the imagination and make people want to touch it at first glance?
• Future potential — How much does the vision the work presents seem likely to contribute to humanity’s future? Could it plausibly give rise to new industries, entertainment, or ways of living?
• Originality — Is it built with unprecedented, free-thinking ideas? Is it genuinely unlike anything else?

For Project DENT’s first hackathon, the theme — at the request of sponsor SANU 2nd Home — started with “nature,” and the regulation was for each team to build three or more games. The time limit: 90 minutes.

In the end, the teams that showed up that day were: “AI mo Issho,” a tag team of GOROman and Kazunori Nanji; “toitrio,” made up of developers of Sony’s small robot “toio”; “Reality Falsification Institute,” a senior squad featuring Square Enix’s Yoichiro Miyake and ex-NVIDIA’s Kazuyuki Hashimoto; “Team Giken-damashii,” made up of veteran developers like Atsushi Shirai and Daisuke Imai; “null² kids,” made up of PhD students from the University of Tokyo’s Matsuo Lab, Rekimoto Lab, and Ochiai Lab, who had actually built null² for last year’s Osaka-Kansai Expo, produced by Yoichi Ochiai; and, in complete contrast to all these heavyweights, “Futsū no OL” (Ordinary Office Ladies), a team of three women with zero programming experience, plus “Golden Sisters,” three Golden Gai bartenders led by Kyoko-san — seven teams total.

The idea was to gather only people who, left to their own devices, would absolutely never join a hackathon — and have them battle not for personal profit or glory, but toward the shared goal of advancing humanity. The prize money: 100,000 yen.

You might think that’s a small prize for such a lofty goal, but I judged that if the prize is set too high, different dynamics — separate from the event’s actual purpose — start to kick in. I aimed to keep it under 40,000 yen per person; for instance, if the prize were 1,000,000 yen, some people would think, “For a million yen, I might as well spend 300,000 out of my own pocket,” and try to solve the problem with money rather than brainpower. I wanted to avoid that.

The very first theme already produced something interesting. GOROman and Nanji’s “AI mo Issho” team — true professionals and battle-hardened veterans — fell into a mood of resignation, saying “there’s no way we can make three games in 90 minutes.” Meanwhile, the total-amateur “Futsū no OL” team had one member alone knock out three games early on, and kept going smoothly from there, ultimately producing sixteen works. The Golden Gai bartenders, cool as anything, built a set of twelve mini-games — three for each season. The veterans were stunned first.

The other teams followed a similar pattern: the more experienced the veteran, the more resistance they felt toward the “crank out quantity in a short time” approach, and aside from the two amateur teams, each team finished only three to four games.

That said, the veteran teams’ individual pieces were of much higher quality. The point of the first theme was, in part, to pressure the veterans with the sheer terrifying momentum of amateurs pushing through on pure energy alone.

Especially interesting was that the game GOROman forced together in the last ten minutes, “SANU Weasel Night,” turned out far higher quality than the game made in the preceding 80 minutes. Take three photos, and the content of those photos gets recognized, automatically generating a sound-novel-style scenario along with corresponding image processing — genuinely impressive work.

One operational miscalculation, though: since the rule was “more games = higher score,” it ended up taking nearly an hour just to review all sixteen of the “Futsū no OL” team’s works, throwing off the schedule.

The next theme was balloons and shopping: build a game using real balloons for outdoor display, with a budget of 3,000 yen per team to shop for materials before finishing the build.

This tricky challenge seemed to frustrate many of the veteran participants — after all, these are people who normally never think about anything beyond the screen.

Kondo of null² kids in particular fumed that “Shimizu clearly doesn’t have his own answer for this one.” Editing this footage, I found myself muttering, “Well, of course not.” If I had the answer, it could never turn out interesting. Why are we even here? To put a dent in the universe. The whole point of this hackathon is answering questions that have no answer.

Meanwhile, having built a huge volume of games last round only to hear me mutter “this doesn’t really feel like a game,” the “Futsū no OL” team seemed to take it hard, and this time spent much of their effort earnestly debating “what even is a game?” and “what is ‘gameness,’ anyway?”

By contrast, the only team that showed zero confusion, surprisingly, was the dark horse from Golden Gai: the Golden Sisters.

The strength of a beginner team lies in their willingness to simply accept, without hesitation, “what we can’t do, we can’t do.” That resolve — “let’s just play the hand we’re dealt” — ultimately shapes how good the final work turns out.

Veteran engineers, on the other hand — especially software engineers — have so much accumulated knowledge and experience that countless thoughts race through their heads: “that’ll be hard,” “that’s theoretically possible but iffy within the time limit.” Ironically, this ends up self-sabotaging, narrowing what they let themselves attempt. The result is tunnel vision, where “raising quality within the time limit” often gets prioritized over adventurous ideas. In short — they don’t fully trust the AI yet.

At this stage, what GOROman made was an image-recognition flag-raising game using blue and red balloons, plus a drone-dart game. Sure, it used AI, but as a game it was practically a 100-year-old fossil — fine for a high school culture festival booth, but not something a pro should be making (though, granted, 90 minutes might be tough even for high schoolers). In short, he was “playing it safe.”

As for why balloons were the theme: the day before, I’d spent an entire day wandering around Tokyu Hands, wanting to bring something three-dimensional into the hackathon, when I hit upon the obvious realization: anything three-dimensional you try to bring along ends up bulky. And afterward, being stuck with a pile of bulky leftover stuff is a pain.

“Oh — balloons! They’re not bulky, but they can be inflated.” That’s when I bought a huge stash of balloons and brought them along.

The intent behind using balloons was to see how people would realize “play” with a genuine sense of spatial expansiveness. If all that came of it was balloons substituting for flags, that would be a letdown — I wanted people to compete with freer thinking than that.

Among the other teams, null² kids tried to build a balloon-drawing game but left it unfinished — and even that idea was just an application of existing image-processing techniques, without much originality. Team Giken-damashii built a game where you draw a picture on a balloon, have the AI critique it, and pop the balloon if the AI judges it poorly drawn — but they never really dug into whether an AI critiquing your art is actually fun.

At this stage, the Golden Sisters, presenting last, pulled decisively ahead of the pack.

First, Kyoko-san held up a large iPad displaying a QR code. Not just the judges but the other teams’ participants too could scan it and join the game simultaneously. The balloons had various facial expressions drawn on them; make the same expression and take a photo, and the AI judges your expression. Score 100 points, and you get the prize attached to that balloon, on the spot.

What’s more, your own photographed expressions get saved to a gallery you can browse afterward — leaving behind a record of all the faces you made, as a memento of the event or trip.

This inventive idea drew in not just the judges but the other participants too, and the place erupted.

People who come to drink in Golden Gai aren’t just there to drink. If all they wanted was alcohol, they could drink alone at home. What they’re really after is communication. This piece captured exactly that, and it overwhelmed not just the judges but every participant.

Since dinner at this venue was a self-catered BBQ, Kyoko-san — drawing on a lifetime of experience — handled the shopping, picking out BBQ ingredients and practical items like sponges for washing dishes; manga artist Natsuko drew the faces; and Chan-Saki handled the programming — a fully realized division of labor.

The following day’s theme was building an AI arcade prototype over ten hours, aiming to open for business by 9 a.m. the next morning — actually converting each person’s cottage into an arcade to entertain guests.

Here, the team that quietly caught up through steady, solid ideas was toitrio.

Sony’s small robot toio, released by SIE, is actually packed with sensors — a sensor for detecting its own absolute position, an accelerometer, and it even has LED lights and a motor. Since about ten of them can be controlled simultaneously over Bluetooth, I’d been hoping to see a piece that made full use of that.

What toitrio built was a game that visualizes “the weight of words.” Using toio units on a mat linked to a screen, players grab “words” and arrange them in order of weight; correctly grasping the relationships between words scores points. It’s a two-player game, where the score difference between neighboring players is expressed visually and physically by a toio unit running across a seesaw the team 3D-printed overnight.

This is arguably the first real instantiation of what the “AI Arcade concept” is aiming for: turning MIT’s Hiroshi Ishii’s “Tangible Bits” — and its successor, “Radical Atoms” — into an actual game. In particular, expressing the state of the match through a physical seesaw felt visually and viscerally compelling.

The young team null² kids, who hadn’t had a great showing until now, delivered in the final stretch: SANUfight, a fighting game where you capture your own pose with AI and create original special moves to battle with. The flow — photograph a pose, have AI evaluate it and determine attack power and reach, then actually play with it — was a classic, well-trodden concept, but executed with excellent use of AI.

Given how talented this team clearly was, it was a shame they hadn’t been able to show it until this point.

The “Futsū no OL” team delivered a mini-game collection, “Everyone’s SANU Break,” with a polish that belied their team name. At the start, it photographs the player, and that photo of the player appears throughout the game from then on. Swatting flies, collecting flowers, having a BBQ, doing a löyly sauna session — each rendered in cheerful pixel art, while still carrying a distinct “ordinary office lady” charm.

The Golden Sisters, calling themselves the “Golden Arcade,” brought “Bakyūn Panic,” inspired by Whack-a-Croc-style games. Scan the QR code, and the gallery of onlookers can interfere with the player — a brutal design. If anything, the gallery had more fun than the player did.

What’s more, it was genuinely impressive that the gallery’s smartphones and the game’s main screen were linked together in real time.

“How are you handling the server for this?” I asked.

“Everyone asks me that — what’s a server? Claude Code handled everything, so I honestly have no idea.”

“You’re kidding…”

Normally, building a game like this starts with server design — how many people will connect simultaneously, what latency you can respond within, where the server will be hosted. All of that was completely skipped, with nothing more than the idea “it’d be fun if the gallery could participate” — the rest handed entirely to Claude Code. And it worked. Is this the AI era now? And the person who built it doesn’t even know what a server is. A terrifying age.

Team Giken-damashii built a compatibility-diagnosis game using two giant widescreen touchscreens. Two people face each other and answer questions to find out their compatibility — and even the background music shifts subtly depending on the compatibility score, painstakingly played through a MIDI sound source.

Game centers used to be a classic date spot, after all — and even purikura photo booths, back in the day, were designed with couples in mind. As an idea for couples dating at an arcade in this AI era, this thoroughly deserves credit. Giken-damashii really shines when they let their shoulders relax.

And GOROman and Nanji’s “AI mo Issho” team brought “SANU PATROL,” a side-scrolling action game inspired by Moon Patrol. Notably, they combined a Stack-chan robot with a robotic hand so that Stack-chan itself operates the arcade controller’s lever, representing collaboration between human and AI. Functionally, it’s a one-button game for the human player — but combining it with AI created a sense of teamwork distinct from simply facing the game itself, changing the weight of that single button.

Zone out for a bit, and Stack-chan (Naru-Ebi III) prods the lever as if egging you on, forcing you to watch the screen seriously and read its intent. It was a genuinely new kind of gaming experience with AI, and a solidly compelling exhibit.

That’s how the very first Project DENT wrapped up, to great excitement — who ultimately won, you’ll have to check the main video for that.

Now that this hackathon is over, I’ve somehow ended up committing to two more. Well, “ended up” isn’t quite right — I planned them myself — but filming for Season 2 happened the Monday before last, and filming for Season 3 (Masters) happened just yesterday, meaning I shot two back-to-back heavy hackathons within two weeks, and I am completely wiped out.

Editing for Season 2 is already done, but we’re coordinating the release timing with drikin, so it’ll air soon. Season 2’s theme is drones — probably the world’s first hackathon combining AI and drones. GOROman “retired” during this one.

Since GOROman had spoken so highly of the Golden Sisters at the end of Season 1, we paired them up for Season 2, forming the team “Brain Ethics.”

How will it turn out? It’s fun — but honestly, almost too fun, and now things are getting pretty hectic.

KingJim sells a long-selling word processor called the Pomera. It’s a fantastic machine for writing manuscripts and organizing your thoughts.

Writing manuscripts on the Pomera is great, of course, but the Pomera is also known for the fact that it can run Linux.

There were a few quirks, but thanks to ChatGPT I managed to get WiFi up and running, so let me share the method here.

Debian on Pomera DM200/DM250 — 記録
www.ekesete.net

This method mostly worked, but the WiFi recognition was different on my particular lot of Pomera, so I edited the following file while consulting with ChatGPT.

$ cat /etc/init.d/dm200_wireless 
#!/bin/sh

WIFI_SETTING_FILE=/mnt/sd/settings/wpa_supplicant.conf

if [ `cat /proc/meminfo | awk 'NR==1 {print $2}'` -eq 1026076 ]; then
   MODEL="DM250"
else
   MODEL="DM200"
fi
MODEL="DM250" # ここを追加
case $1 in
  start)
    # WiFi firmware
    if [ $MODEL = "DM250" ]; then
        #DM250
        echo "/opt/etc/firmware/fw_bcm43438a1.bin" > /sys/module/bcmdhd/parameters/firmware_path
        echo "/opt/etc/firmware/nvram_AP6212.txt" > /sys/module/bcmdhd/parameters/nvram_path
    else
        #DM200
        echo "/opt/etc/firmware/fw_RK901a2.bin" > /sys/module/bcmdhd/parameters/firmware_path
        echo "/opt/etc/firmware/nvram_AP6210.txt" > /sys/module/bcmdhd/parameters/nvram_path
    fi

    # WiFi/bt power on
#    echo 1 > /sys/class/rkwifi/driver

    # WiFi power management on
#    /sbin/iwconfig wlan0 power on

    # copy wpa_supplicant.conf
    if [ -f $WIFI_SETTING_FILE ];then
        cp $WIFI_SETTING_FILE /tmp/wpa_supplicant.conf
    fi
  ;;
  stop)
  ;;
esac

exit 0
pomera@pomera:~$ 

It seems there’s a bug in the DM200/DM250 detection logic — when using the newer model (DM250), you apparently have to force it to be recognized as a DM250 or it won’t work.

It’s a bit of a hassle having to put the SSID and password into the config file, but once you do that, it connects right away.

Launching a terminal called fbterm gives you access to both Japanese fonts and Mozc.

I installed Tailscale, so now I can access all of the devices I own. This is really convenient — I don’t have to carry a PC around anymore. I’d been substituting an iPad for some of that, but when I want to type away on a real keyboard, this does the job.

It can read Japanese too, so no complaints there. You can use both Claude Code and Codex from here.

By the way, when I tried to install Claude Code, it got rejected because it wasn’t supported. Well then, fine — I’ll just make it myself! I can build my own agent, damn it!!

So I decided to build something OpenClaw-like that runs on the Pomera. I’m calling it Poclaw.

A hacker’s sidekick really ought to let you bang away on a proper keyboard.

It comes with proper tools, so it can look things up online and tell you what it finds.

I mean, I get the argument that I should just use an iPad, but this kind of thing is about romance, so it’s fine.

The AI Game Center campaign is still running — ¥10 million is a long way off, but support keeps trickling in steadily, and there’s now talk of taking this to regional venues outside Tokyo. That was not part of the plan. Plans, apparently, are suggestions.

One small thing worth mentioning: the crowdfunding tiers are expensive by design, but it turns out you can also support the project with no reward at all — just a straight donation. Some friends from my local bar started sending in ¥1,000 and ¥3,000 contributions. I was not prepared for how much that would hit me.

Anyway. The project has cleared 100%, which means I’m allowed to start running before the finish line. So let’s run.

The Question I Keep Coming Back To

What kind of entertainment is only possible with AI?

Not “enhanced by AI.” Not “AI-assisted.” I mean: what could you build that is fundamentally impossible without it?

Last year I made a prediction: that AI would eventually make it possible to convert classic films into VR experiences. I filed that thought away and moved on.

Then a few weeks ago I found myself sitting on a large collection of 360-degree video footage from my Insta360 One — stuff I’d been shooting obsessively for a while and hoarding on the theory that it would be useful someday. I told Claude Code: take these 360-degree videos, extract frame pairs automatically, and build a tool that maps regular photos to VR.

It produced over 100 matched pairs almost instantly. (The reason this was easy: Insta360’s .insv format is secretly just MP4. The camera companies are not trying to hide anything, they’re just lazy about naming things.)

The question was whether AI could take a regular flat photograph and invent the parts of the room that the camera couldn’t see.

I was skeptical. This is the kind of thing that sounds like a fever dream — the camera only captured the left wall, but the AI is supposed to generate the right wall, the ceiling, the window view, the furniture that logically should be there. That requires the model to have absorbed something like a physical intuition about how spaces work.

I assigned Claude Code to throw every available compute resource at the problem — specifically, the spare cycles on Sakura Internet’s image generation servers, which have about 40 Tesla V100s with 32GB VRAM each — and then I went and worked on other things.

A while later, I got a notification from my Claude Code dashboard (a custom agentic AI management tool I’d built):

“Completed.”

I had to think for a moment about what it had been working on.

Then I looked at the output and genuinely did not trust what I was seeing.

The Part That Broke My Brain

It worked.

Not approximately. Not “if you squint.” The model was generating the missing halves of rooms with objects — a sofa, a window view, an adjacent wall — that had never appeared in the training image. Things that were not there, reconstructed from context, and plausibly correct.

I tested it on more images. Tables with their surroundings filled in. Grass. Interior spaces extended beyond the frame.

I tested it on illustrations. Same result.

I’ve been working with AI long enough to have my intuitions violated fairly regularly. This was the most jarring violation yet. When something works that shouldn’t work, the only honest response is to sit with the discomfort for a while before you start explaining it.

The technical approach this time is much simpler than the EvoX algorithm I’ve written about before — simple enough that it should generalize to any image or video, at least in principle. The theoretical basis for the monocular-to-360 estimation is documented in recent literature (there’s a WACV 2025 paper on OmniDiffusion that covers this territory), but seeing it actually produce coherent rooms from single frames is a different experience than reading about the approach.

Down a Lens Rental Rabbit Hole

The Insta360 footage proved the concept, but the resolution has limits — the sensor is generous in pixel count but the optics leave something to be desired.

So I went looking for better source material. I have a Sony α7S, but no fisheye lens. Lens rental in Tokyo is wonderfully casual — ¥2,200 for a day, except you can pick it up the evening before and return it the following morning, which means you effectively get 40 hours for the price of one.

I rented a LAOWA FE 8–15mm f/2.8 zoom fisheye and spent the rental window answering a question I’d been putting off: circular fisheye vs. diagonal fisheye — which actually works better for VR?

Circular fisheye captures the full 180-degree sphere but produces a round image with black corners. Diagonal fisheye captures 180 degrees only along the diagonal, producing a slightly distorted but fully rectangular frame.

For pure immersion, circular wins. But for games and entertainment — where you need to guide a viewer’s attention — diagonal is probably better. Too much visual information in every direction and people don’t know where to look.

After shooting comparison footage and building a quick comparison tool, diagonal took it on image quality. The α7S produces about 4240×2832 pixels, which converts to a 4096×4096 equirectangular image — two of those side by side gives you something approaching 8K.

The remaining problem: converting diagonal fisheye to equirectangular throws away information. You capture a rich frame and then lose some of it in the projection math, which feels deeply wrong. The solutions are either to upgrade the camera body (an α1 gets you to roughly 8640×5760, nearly four times the area) or to let the AI fill in the edges — which, as we’ve just established, it can now apparently do.

ChatGPT suggested just using the diagonal fisheye images as-is. I tried it on my apartment. The depth and immersion were, honestly, sufficient. Maybe I’m overcomplicating this.

So What Kind of Game Do You Actually Make With This?

VR gaming has been around long enough to accumulate a graveyard of brave experiments. Nothing has yet become “the Tetris of VR” or “the Mario of VR” — the thing that defines the medium the way those games defined theirs. Plenty of theories about why. I’d rather not relitigate them.

Instead, I want to try something that doesn’t look like a VR game at all.

Interactive movies. For two people. Ten to fifteen minutes long.

The format: you and a friend or partner make a reservation in advance. When you show up, you sit in two chairs, put on headsets, and experience a short film with proper spatial audio and haptic seats. The catch is that the film was generated specifically for you — built around photos you uploaded at booking time.

Your dog is in it. Your late grandmother is in it. A character you drew once and never showed anyone is in it. A version of yourself wanders through the story.

The reservation window exists because the render takes time. But that pre-generation is also what makes the experience personal in a way that no off-the-shelf entertainment can match.

I have a close friend — Kimiya Matsuo, a musician and artist — who has been using AI to generate songs in his late wife’s voice since she passed away. He keeps making music with her. I think about what it would mean to give someone like him fifteen minutes inside a space where she exists again, moving through a world that includes things she loved, places they went together, assembled from fragments by a model that has learned to fill in what it doesn’t know.

That is not a game in any traditional sense. It is something that was impossible before AI existed.

That’s the thing I want to build.

AIで人類の想像力を拡張する｜清水亮が挑む「AIゲームセンター」構想始動(By 清水亮（shi3z）) - Kibidango(きびだんご) : 「半歩先の未来をお届け」するcrowdfunding型EC

By hour four, we were at 54%. By hour ten, we’d cleared 100%. As I write this, we’re past 200%.

I genuinely don’t know what to do with myself.

How This Happened

The original math was simple: if twenty people are willing to put in ¥50,000 each, that’s a signal worth acting on. That was the whole theory.

What actually happened felt less like a crowdfunding campaign and more like winning an election — that specific combination of gratitude and sudden, slightly terrifying accountability. People weren’t just clicking a button. They were telling me something about how they feel about gaming, about AI, about the gap between what technology can do and what anyone is actually doing with it.

I went to sleep with the campaign at some percentage I felt good about. I woke up and it was at 141%.

My imagination could not keep up.

The Conversation That Changed the Plan

I called an emergency meeting with the crowdfunding team, including Aoi-san from kibidango. My first question was practical: do we set intermediate goals? ¥2 million? ¥5 million?

Aoi-san said no.

“Small incremental targets are the wrong move. If you’re going to carry people’s dreams, you need one big goal worth dreaming about.”

So we thought about it seriously.

The next goal is ¥10 million.

For context: as a company president, I’ve allocated ¥10 million plenty of times. Equipment: X hundred thousand. Salaries: Y hundred thousand. Inventory: Z hundred thousand. Buffer: the rest. It’s a spreadsheet exercise.

Money raised through crowdfunding is nothing like that.

I think about it more like taxes — and I mean that as a compliment to both. I went to national elementary school, national middle school, national university, national graduate school — twice. When I was accepted into the Mitou Software Creation Program, the Japanese government handed me ¥10 million of public money to build something, and in return gave me the title of “Super Creator.” I’ve never been able to deny that title when someone brings it up, because it’s not really mine to deny — it belongs to the taxpayers who funded it.

Crowdfunding money is even heavier than tax money, because it comes with direct expectation. These aren’t anonymous citizens funding public infrastructure. These are specific people who found this project, read it, decided they believed in it, and handed over money they earned.

A fifth-grader’s ¥500 coin means more than a corporate executive’s thoughtless ¥500,000 at a Ginza bar. The amount is irrelevant. The weight is everything.

So: what does ¥10 million actually make possible?

What ¥10 Million Could Build

Head-to-head versus cabinets. Two-player competitive play was never in the original budget. It’s the heart of the arcade experience. Generative AI in a versus game — players essentially building the battlefield against each other in real time — should produce something genuinely new.

Immersive/motion-based games. Think Sega Rally or MANX TT — the kind of full-body arcade cabinet that makes you forget you’re standing in a room. What does that feel like when the track, the obstacles, and the physics are being generated dynamically?

Gun shooters. Simple to pick up, hard to put down. The gameplay loop is familiar; what changes is everything inside it.

Multiplayer co-op. Four people standing around a cabinet together is a fundamentally different experience than anything you get at home. That’s supposed to be what arcades are for — and it’s mostly been abandoned.

Experimental interfaces. MIT Media Lab’s Hiroshi Ishii has been doing things with tangible computing for decades that nobody has ever tried to make fun. (His inFORM installation reportedly costs over ¥100 million to build. We will find a less expensive path.) Haptics, gesture, computer vision — there are interfaces that have never been used for games because no one has bothered to try.

For all of this, Giken Base is going to be too small. We’ll need to rent a proper venue for a day or two — somewhere we can actually fit full-sized cabinets and let people move around.

How the Project Will Actually Run

A few things I know need to happen, some immediately:

Brainstorming first. Before I spend a yen of anyone’s money, I want to get smart people in a room — game developers, AI researchers, interface designers — and actually pressure-test what’s possible. I’ll fund that out of pocket while the campaign is still running.

The compute problem is real. The original ¥1 million budget didn’t include the cost of the GPU hardware to actually run generative AI at game speed. That was always going to be a separate question. At ¥10 million, it becomes a solvable question.

I need collaborators. I’ve been out of the games industry for a while, but I still know people. This project will benefit from experts I haven’t thought of yet. I’m planning a combination of online sessions, live events, and possibly a hackathon or competition — at ¥10 million, a real prize pool becomes realistic.

I have never thought I was capable of doing this alone. I just now have the means to ask for help properly.

One Note on the Images

The concept art scattered throughout this post was generated by AI. It represents what I’m aiming for — an ideal, not a guarantee. What we actually build will be shaped by physics, budget, time, and the specific weird genius of whoever ends up working on this with me.

I don’t know exactly where this goes. But I intend to find out.

Thank you, genuinely, to everyone who made this happen faster than I could process it. The next chapter starts now.

AIで人類の想像力を拡張する｜清水亮が挑む「AIゲームセンター」構想始動(By 清水亮（shi3z）) - Kibidango(きびだんご) : 「半歩先の未来をお届け」するcrowdfunding型EC

I Want to Build an AI Arcade — and I’m Selling My Soul to Do It

“Hey, want to join an AI hackathon?”

That’s usually enough to make people flinch.

“That sounds complicated. Isn’t it, like, really hard?”

“Nope. Easy.”

“Sure, maybe you think it’s easy.”

“Okay — pull out your phone, open the App Store, download something called Gemini.”

“…fine. Done.”

“Launch it, tap on Canvas, and type: Make me a fortune-telling app using AI.”

“…what? Fortune telling? Okay, something’s happening, it’s saying a bunch of weird stuff — “

“Give it a sec.”

“…wait. It made something. That’s it?!”

“Told you. Easy.”

And that’s how I’ve been dragging friends and acquaintances into hackathons, one smartphone demo at a time. The problem is, about half of them catch the bug and keep going — and the other half walk out the door and forget everything within 48 hours.

This is actually one of the hardest unsolved problems in programming education, and most people just quietly ignore it.

The Last Human Skill in the Age of Vibe Coding

Here’s what I’ve come to believe: you can’t teach someone to code if they don’t have something they desperately want to build. It’s the same as learning a foreign language. Nobody becomes fluent by studying grammar in a vacuum. You become fluent when you’re stuck in Auckland with zero Japanese speakers around you and you need to communicate.

Programming works the same way. The people who actually learn it aren’t motivated by “I want to get good at coding.” They’re motivated by “I need to build this specific thing and I will not stop until I do.”

That obsessive desire — that’s what has pushed humanity forward. Nobody asked Robert Goddard to invent the liquid-fueled rocket. He did it in a field in rural Massachusetts in 1929 because he’d read H.G. Wells’ War of the Worlds as a kid and got completely possessed by the idea of space travel. Forty years later, humans walked on the moon.

Star Wars Changed My Life (Twice)

My own origin story starts before kindergarten, when my dad — a factory engineer with a high school diploma — dragged me to an open campus event at a local technical university. They had a machine that could look at a shape on camera and announce whether it was a circle, a square, or a triangle.

Impressive, technically. But my six-year-old self’s reaction was basically: okay, but… why?

The real spark came in 1977, when Star Wars hit Japan. Those wireframe computer graphics in the Death Star targeting sequence — I was wrecked. I went home and immediately tried to recreate them on our family PC.

What I didn’t know (and thankfully, neither did my dad) was that those wireframes required a multi-million dollar computer and three months of work. Had anyone told us it was impossible, we’d have given up immediately. Instead, my dad kept buying me whatever math and programming books I pointed at, no matter how advanced, and I kept grinding away.

By fourth grade, I had a working 3D renderer. I’d figured out the rotation matrices by pestering a classmate’s father — a university math professor — and finally got the formula I’d been hunting for two years.

The lesson I took from all this: technical knowledge can be taught. What can’t be taught is the feeling of “I need to make this thing.” That emotion has to come from somewhere else.

The Game Center Was My MIT

In middle school, that “somewhere else” was the local game center.

In 1980s Japan, arcades were genuinely sketchy places — cigarette smoke, vague danger, my grandmother absolutely loved them. She was a submarine game devotee. But for me, the arcade was a live technology expo I could visit every single day.

Galaxian. Galaga. Xevious. Space Harrier. Virtua Racing.

These weren’t just games. They were existence proofs — demonstrations that you could do something incredible with computers if you thought about it the right way. The arcade answered the question “what is this technology for?” with the most compelling possible answer: it’s for making people feel something they’ve never felt before.

That, I now believe, is the most important skill left for humans in an age where AI can build almost anything on demand.

Not coding. Not math. But this: given a tool, can you imagine what it could become?

Why AI Has Barely Touched Gaming (And Why That’s Insane)

Here’s something that genuinely bothers me: GPUs — the hardware that makes modern AI possible — exist because of gaming. Sega, Nintendo, and Sony spent decades in an arms race to give us more realistic graphics, and that competition made GPUs cheap and powerful enough to eventually train neural networks.

Gaming literally built the infrastructure for the AI revolution.

And yet today’s AI is barely showing up in games in any meaningful way. Sure, some studios are using AI to generate textures or write side-quest dialogue. But games that put AI’s actual nature at the center of the gameplay experience? Essentially nonexistent.

This is weird. This is wrong. And I want to fix it.

GIKEN STAR, and a Prototype Born in Three Hours

A few months ago I asked my collaborator Master Watanabe at our workshop (Giken Base) to build me a game cabinet prototype — the kind of vertical-screen arcade enclosure you’d recognize from the 80s, but made from MDF wood and loaded with a camera and a microphone.

The camera reads the player’s face and clothing. The microphone picks up their voice. Because in the kind of game I want to make, you are part of the game.

That evening, I built a test game called GIKEN STAR — a vertical scrolling shooter where, when you activate a bomb, you shout a word into the microphone. Your ship transforms into something related to that word, generated on the fly by an image AI. The boss enemy also gets redesigned based on your word, and its attack patterns are auto-generated by an LLM in real time.

No two boss fights are ever the same.

And here’s the part I’m proud of: it’s still fun. Because I spent fifteen years making commercial games, and I absorbed enough from my old Sega mentor to know how to keep chaos enjoyable. Randomness without craft is just noise. Randomness with craft is a boss that surprises you in ways that feel fair.

The prototype took three hours. Which tells me three good AI games are completely within reach.

I’m Selling My Soul (It’s a Crowdfunding Thing)

So I’ve decided to actually do this. AI arcade games, a real physical cabinet, and eventually some kind of public event — maybe a pop-up AI Game Center — where kids and adults can actually play them.

To fund it, I’m doing a crowdfunding campaign. I’d never done one before, and when I described the project to a crowdfunding producer named Aoi-san, he got quiet for a moment and said:

“Shimizu-san. There’s only one approach here. You have to sell your soul.”

“…my soul?”

“Put a price on it. That’s what crowdfunding is.”

So here we are. I’m selling my soul. The tiers are called Bronze Soul, Silver Soul, Gold Soul, and — limited to five people — the Heroic Soul.

The funding goal is ¥1,000,000 (roughly $7,000). That number isn’t arbitrary. Back in 2022, right after I’d left my company and was delivering food for Uber Eats, I wrote a quick tutorial about Midjourney on a blog platform. It made ¥1,000,000. I felt so grateful — and so weirdly obligated to pay that energy forward somehow — that I immediately pivoted into building AI art tools and running hackathons and festivals.

A million yen is the right size for an idea that might not work. Big enough to mean something. Small enough to not be irresponsible.

What I Actually Believe

Entertainment isn’t a distraction. It’s a tool for expanding the imagination — the most sophisticated one humans have ever built. And unlike a hammer or a toaster, it can’t be reused. You can’t watch Star Wars for the 100th time the same way you watched it the first. Entertainment must constantly reinvent itself, which is why it has always been the leading edge of technology adoption.

The GPU exists because kids wanted better games. AI will get better because someone figures out how to make it genuinely thrilling to interact with.

I was one of those kids in the arcade, absolutely certain that whatever was in those cabinets held the key to the future. I want to build the arcade for the next generation of kids who don’t know yet what they want to become.

The crowdfunding page is live. Come sell your soul with me.

https://kibidango.com/projects/2915

The reason was simple: we didn’t have a Famicom at home, and my grandmother was a gaming fanatic.

As a kid, every time we went shopping at the department store, my grandmother would take me to the game corner without fail. Sometimes she’d let me look through the periscope she was peering into, and I got hooked on the game where you fire torpedoes at warships.

My grandmother, born in the Taisho era, was the only gaming enthusiast in our family. My dad’s idea of games was mahjong at best, and my mom was a workaholic to the core.

I wasn’t in a position where I could play games freely. The most I got was going to the neighbor’s house to play Famicom. Otherwise, I had no choice but to make my own games on our home PC.

By the time I was finishing second grade, I could write a decent range of programs in BASIC. But it wasn’t until third grade that I realized programming and making games were fundamentally different things.

Hopscotch, kick-the-can, tag — those kinds of primitive “games” kept me obsessed every day. Children’s play is remarkably clever and refined, and every game was genuinely thrilling.

But whenever I tried to create a completely new game with original rules, I first had to explain those rules, and then somehow enforce them. In other words, children’s play requires a consensus on the rules before anything else — and games with complex rules were simply impossible for most kids to actually play.

What was revolutionary about computer games was that children and adults alike could play games of the same complexity. The enforcement mechanism — the thing that made you follow the rules — had been automated by the program.

Take table tennis or tennis: you can’t play without knowing the rules first. But with PONG, you could just spin the dial randomly and have fun. Knowing the rules was optional.

In our industry, we call the act of randomly mashing the joystick without knowing the rules “lever-gacha.” Even in that completely rule-ignorant state, given a little time, you can understand the rules and improve. There are two kinds of beings that can do this: children, and AI.

When DeepMind — later acquired by Google — announced Deep Q-Network, not many people were excited about it. I remember the reaction on sites like Hatenabookmark being quite cold. I was enthusiastic, but most people dismissed it as “pointless nonsense.” And that’s why none of you won a Nobel Prize. (Neither did I, to be fair.)

Deep Q-Network, in a sense, evolved naturally and directly into reinforcement learning AIs like AlphaGo and AlphaZero, and ultimately became AlphaFold, which brought a major breakthrough in chemistry — and the development team received the Nobel Prize in Chemistry. I believe it was the first time that people whose profession was computer programming won a Nobel Prize.

As someone who had been in the game industry, I understood it well. If you can learn through lever-gacha, that is essentially real intelligence. Because virtually everything humans learn — children included — is learned through a kind of lever-gacha.

In the 1990s, the game industry was poor but dazzling. I saw a particular potential in computer games.

One was the computer game as a device for “thought experiments” or “experiential thinking.” SpaceWar!, the very first computer game born from gravity simulations, existed as an extension of thought experimentation. You could say it was the origin of simulators.

Early games started in space because screens were black. And there’s no doubt that the golden age of space opera — Star Trek, Star Wars — was still in full swing, and the excitement of Apollo reaching the moon had not yet faded.

Star Wars was such a massive hit that space-themed content became widely accepted. More than that, there was a growing hunger to somehow experience “something Star Wars-like” for yourself.

If Star Wars hadn’t existed, I probably wouldn’t have thrown myself into computers with such intensity.

Star Wars contained two things that would become indispensable to my life. One was the experience of space itself — the aerial combat aboard X-Wings, the trench run across the surface of the Death Star. The other was droids.

Among all the droids in Star Wars, the lead mech R2-D2 is genuinely strange.

For one thing, it’s unclear whether it even has a face. R2-D2 boldly discards the “facial symmetry” that vertebrates — humans, fish, cats, dogs — and even arthropods like insects all possess.

And yet, rather than looking frightening to a child, it somehow looks adorable.

On top of that, R2-D2 never speaks a word of human language, and yet comes across as calmer and more intelligent than any other character in the film.

Having never seen a character like that before, I was completely captivated by R2-D2 as a child. The white-and-blue color scheme was part of it too. I don’t know exactly why, but Doraemon is also white and blue. So for me, white and blue have always been the colors of AI.

Building artificial intelligence was too difficult for a first-grader, so my programming journey began with an interest in space.

To deepen my understanding of pseudo-3D techniques, the masterpiece Space Harrier was essential.

Around that time, for some reason my dad decided to buy a Famicom(NES) game as a gift when we were visiting relatives. He bought some obscure title at Takeyaya in Okachimachi — and it turned out to be something extraordinary.

It was a game that reproduced the gameplay of Space Harrier on the vastly less powerful 8-bit Famicom. I wouldn’t learn until about 30 years later that an acquaintance had secretly developed it. But by comparing Space Harrier and Attack Animal Gakuen, my understanding deepened considerably.

Even a game that looks simple now taught me things: that scaling is the core technique, and that the way objects scale must not follow 0.5× → 1× → 1.5×, but rather something close to doubling — 0.1× → 0.2× → 0.4× → 0.8×.

By comparing these two games, back when I hadn’t even learned fractions yet, I came to intuitively understand something called “perspective projection” from a 3D programming textbook:

For me, games were textbooks. Games were teachers.

With just this one small insight — this one simple calculation — you could express the sensation of three-dimensional space.

This can be reproduced with a very simple program.

Back then, cameras were very expensive and taking photos wasn’t so easy — especially for an elementary school kid. Today I would probably line up some caramel boxes at equal intervals, take a photo, measure the distances on screen, and develop the intuition that way. But in those days, all I had was games.

If I hadn’t had this moment before adolescence, I would never have later installed vectors and matrices as a kind of second language in my own mind.

After grasping this, my head became extraordinarily clear. Until then, every concept had swirled together in a muddy fog. But once I began moving back and forth between three-dimensional space and a two-dimensional plane, my imagination expanded to a literally different dimension.

The educational power of games should never be underestimated.

Take Oregon Trail, for example. Every American of a certain generation has played it. It’s a game about the westward expansion, through which you can learn American history.

Practically mandatory in schools from the 1980s through the 1990s, it taught players about the hardships of the American frontier and the importance of budget allocation.

When I went independent in 2003 and founded my first company, I spent countless hours obsessively playing “The Conveni” and SimCity before we even opened our doors.

“The Conveni” is literally a game about managing a convenience store. SimCity is about running a city. What they share is the management of investment and debt, income and expenditure.

I decided I would never hire an employee until I had fully mastered “The Conveni.” My business instincts were sharpened in places like this — seemingly unrelated and trivial. The feeling of unease when looking at numbers, the excitement of reading a cash flow statement — I cultivated all of that here.

But thinking about it more carefully, almost no business simulation games actually include a cash flow statement. Maybe the people making them have never seen one.

If someone made a game where you manage a business while looking at a real cash flow statement, you could dramatically reduce the number of executives who run their companies into the ground. Venture capitalists should require applicants to submit their score from that kind of game. Maybe I should just build it myself — I’m a manager with a quarter-century of experience and a game creator, after all.

The educational potential of games has attracted serious attention in many quarters.

Carnage Heart, for example, was considered a promising teaching tool for programming.

Though from my perspective as a programmer, it was simultaneously too difficult in some ways and too cumbersome in others.

Age of Empires taught me the most fundamental lessons for navigating adult society: selection and concentration of resources, how to deploy military force, Lanchester’s strategy, alliances and negotiation. Age of Empires is a long-running series that still releases new installments.

That was what games meant to me. I became obsessed because they demanded real thinking. The rush of outsmarting an opponent and pulling off a comeback — nothing else comes close.

Games of that era had powerful educational value, and they worked just as well as entertainment for adults.

And yet today, the meaning of “making a game” has completely changed.

Despite how dramatically generative AI has advanced, games have hardly benefited at all.

The only sliver that’s benefited is the narrow domain of “game development” itself — and even there, the conversation is mired in a rather trivial debate: should in-game content really rely on generative AI, or is it better to have human artists craft everything by hand?

I will say this plainly: right now, there is not a single game in the world that has been built to genuinely unleash the full potential of generative AI.

AI Art Impostor was a masterpiece that caught the wave of generative AI early (from the development side, however, it was cost-inefficient). In the early days of LLMs, there was AI Dungeon — but that too remained an unfinished game in many respects.

Today, if you seriously try to incorporate generative AI into a game, three walls stand in your way.

The first is the API wall. API usage fees are still high. To let players enjoy a game freely, you’d need to pass those API costs onto users — and no one wants a game that complicated.

The second is the hardware wall. If you want to build a fully local AI game with nothing in the cloud, you’d need something like an RTX 5090 as a baseline. That means a PC costing close to one million yen(almost $7,000 ) — and the market for that is tiny. Only a small group of enthusiasts would ever play it.

The third is the budget wall. If a game might incur unlimited API fees, or requires a one-million-yen PC to play, no company will fund it. Even venture capitalists would tell you to build a gacha game or a hyper-casual game instead. And VCs are complete amateurs when it comes to games. They have no interest whatsoever in the cultural or educational significance of what games can be.

So the conclusion is this: right now, the only people who can build a game that fully leverages AI are amateurs. Building it professionally is impossible.

If I were still an active executive running a company of a hundred employees, I wouldn’t make — couldn’t make — such a reckless bet.

But fortunately, I’m alone now. There’s no one else I need to feed.

And unlike before, the assets a game requires can, in principle, all be generated. What’s more, for someone unemployed, I have an extraordinary environment at my disposal: forty GPU servers provided free of charge by Sakura Internet, the AI supercomputer “Tsugino-suke”(A100 80GBx8), the Memeplex machine(A6000x2) donated by Dospara (Thirdwave), and various other GPU resources lent or donated by a range of companies — all mine to use freely.

So this year, I’m going to take on a challenge that no one else is doing — something that might be completely pointless to the world, but that I alone believe has meaning, and that perhaps only I can do.

I’m going back into game development.

Not for money.

Purely as research into how to use this technology.

Now is the time to pursue that research — and there is no more ideal subject for it than games.