sizuq
AK

akira

@akira

Atmosphere

Burning◉

Also here

Manabe Akiramanabeakira.com眞鍋聡良manabeakira.jpgalaxygalaxy.sizuq.com

Thoughts

A
akira@akira
1 month ago

Original and Copy

“Be the first to do what no one else has done. That’s how you spark innovation.” You hear this slogan constantly, especially in tech and startup circles. And it’s true that if you move into territory where someone has already built their fortress, you’ll be defeated unless you bring some distinctive added value or originality. To break into an auto industry dominated by Toyota, GM, and Volkswagen and still claim a global position, you needed something like Tesla’s EV vision, its software-first design, and later its wild concepts and execution power in Autopilot and FSD.

And yet, even civilization-scale projects aren’t always producing something “original.” Over the past year I’ve been using Codex, Claude Code, and lately Grok Build on a daily basis, and I can’t help feeling that each company’s designs are steadily “converging.” I haven’t had much chance to try Gemini, Antigravity, Cursor, or the Chinese models, but I suspect they probably show the same tendency.

Take, for example, the button that lets you reset your weekly usage limit whenever you choose. It was implemented in Codex around June this year, and it also got handed out as an apology for some glitch or bug, or to commemorate the user count crossing a milestone. Then in August, Grok added it too, and now you can reset your combined weekly limit across Chat, Build, and Imagine, as many times as you have tickets on hand.

Plan Mode in coding agents is another one. The user types in some idea or something they want to do, and the AI writes up a plan in document form. If there are corrections, it reconsiders; if not, you approve it and move into implementation mode. That’s the flow. Codex, Claude Code, and Grok Build all currently have this same feature.

Integration with external tools via plugins might be the most obvious example of all. Connect Google Calendar, Gmail, or Drive plugins to your AI app, and you can now handle tasks conveniently just by chatting. And depending on the model, developers can hook up plugins to a Supabase or Neon database, Vercel hosting settings, even Stripe payment settings, and once you grant permissions the AI will read and write for you. It’s already built into all of ChatGPT, Claude, Grok, and Gemini’s apps (though each company calls it something different: Connectors, Connected Apps, MCP, and so on).

Of course, there’s individuality in each company’s strengths and areas of focus. Grok and Gemini (the Google Flow line) sell themselves on image and video generation, and ChatGPT has strong image generation too, whereas Claude leans toward code, long-form reasoning, and intellectual work; Gemini is integrated into search, Grok is strong on real-time information from X, so their coverage differs somewhat. But no one can deny anymore that a “copycat war,” to a greater or lesser degree, is underway.

The “copying” I mean here isn’t theft. I’m loosely lumping in everything from adopting a design whose effectiveness has been proven at another company, to the phenomenon of different companies converging on similar optimal solutions.

It’s surprisingly interesting to observe that even at the cutting edge of innovation, it’s not all homegrown ideas; there’s plenty of borrowing too. The scarcity of ideas themselves is fading, and the weight has shifted all at once toward GPU procurement power, capital, distribution reach, brand, operational capability, and raw manpower.

What we need to be careful about is this: the fact that copying happens like this does not disprove the value of an idea no one else has tried, or the value of the action that gives it form. The “Attention Is All You Need” paper written by Google researchers; Elon and Sam Altman founding OpenAI in 2015; Ilya, involved from the early GPT days; Dario, who carried the GPT-2 and GPT-3 era. Long before “generative AI” and “LLM” became household terms, these people were already in motion. The current copycat war is merely being observed in the upper rounds within that sense of momentum.

Come to think of it, over the past few days it’s dawned on me that the same thing happened with the cloud before the AI wave. Email like @icloud.com, @gmail.com, and @outlook.com; video-calling tools like FaceTime, Google Meet, Microsoft Teams, and the old Skype; personal clouds like iCloud, Google Drive, and OneDrive, and the calendars and to-do lists within them. Word, PowerPoint, Excel, Pages, Keynote, Numbers, Google Docs, Slides, Sheets. Aren’t Apple, Google, and Microsoft, the three digital giants, engaged in an outrageous bout of copying each other (or did they just happen to overlap?).

On smartphones too, the model where developers publish apps to the App Store or Play Store, Apple and Google review them, and the platform collects a set percentage of in-app purchases is common to both companies. Did they each naturally arrive at that optimal solution by pushing rationalization to its limit, or did one simply copy the other? Which is it? There’s Apple Maps and Google Maps, Apple Pay and Google Pay too.

This, again, was something that Apple, which began in a garage, and Google, which began in a grad student’s room, started during their expansion phase, back when digital was part of no one’s daily life; in their early days what they needed was, if anything, a flood of originality and ideas.

What I’ve learned from watching the AI development race, and the cloud and smartphone races before it, is that civilization-scale innovation is not necessarily pure ideas alone, and obviously not imitation alone either. Without world-class obsessiveness, extraordinary imagination and planning ability, and a personality that seeps through into the brand, you can’t sense the dawn-era wave and ride it. And you have to incorporate groundbreaking ideas as much as possible, or you’ll be at a disadvantage in the growth-and-expansion phase. Whether at the level of the individual or the group, the key may be the ability to switch freely between modes of thinking.

* All features mentioned are as of August 25, 2026.

A
akira@akira
2 months ago

I’ve been using Claude Code Mobile to make small tweaks and add new features to my app while lounging around. Given the device’s form factor and the kinds of situations it’s designed for, it’s dramatically easier than using a computer when all I need to do is turn ideas expressed in natural language into an implementation plan or work through the finer details of the design.

There are still some things a cloud session can’t handle automatically, though—such as applying database changes that require reading .env.local, or generating a .glb 3D model in Blender from Python code and then uploading it to Supabase Storage and Cloudflare R2 using secret keys from .env.local. For those tasks, I still need to switch over to my local PC environment.

As for Codex, the primary UI flow seems to assume that you’ll pair it with a computer that’s already running. Apparently, though, it can also be used in a standalone cloud session, so I’m planning to give that a try as well.

A
akira@akira
3 months ago

Return of Fable5

With Codex 5.5 High, my black hole interactive experience — which had been stalled for years without turning out quite how I wanted — got a serious polish in Fable 5, following the lifting of U.S. government regulations. It’s now so close to my ideal that I can't help but admire its true capability. It really does a fantastic job of balancing actual space atmosphere with gorgeous graphics.

GPT-5.5

After adjustments by Fable 5

The prompt I threw at it for the fixes was something like this:

Audit and list everything in the current visuals that still feels like an awkward mock-up, contradicts actual science, or serves as explanatory add-on — these should ideally be made movable and hidden by default unless needed. Also, flag elements that blend into the background in dark or light mode. Since the theme is space, light mode might not work well for a black hole specifically, so we should probably drop it and consolidate into dark and dim modes.

In response, Fable 5's current state assessment and improvement proposals (excerpted) looked like this:

The investigation and analysis are complete. Here is the audit and proposals based on the current implementation (black-hole-scene.tsx), screenshots, and the theme architecture (lib/store.ts / app/layout.tsx with global data-theme switching for dim/dark/light).

A. Elements still feeling like mock-ups

  • Rear/Bottom Arc (rearDiskArc / frontDiskBand): Uniform color and width ring segments with abrupt, cutout-like ends that persist to the end. In actual lensing, arcs connect continuously to the apparent edge of the disk, looking D-shaped. -> Replace with a CanvasTexture that tapers width and brightness along the arc, visually connecting the endpoints to the disk's edge.

  • ISCO Inner Ring: A uniform white torus giving a "placed wire" look. Physically, it's the bright edge of gas, not a line. -> Replace with a wide gradient band, integrated as the inner edge emission of the disk.

  • Patchiness of the Beaming Halo: Patchiness is visible when zoomed in due to blob composition rendering. -> Increase texture resolution and overlay angular Gaussian blur.

  • Uniform Fade on Disk Outer Edge: Gives a "drawn circle" look. -> Add turbulence noise cuts to the alphaMap to make the outer edge irregular.

  • Rigid-Body Rotation of Turbulence Streak Layers: Same speed across all radii looks stiff. -> Split into inner and outer layers for differential rotation (which also better expresses Keplerian motion).

B. Points contradicting science

  • Constant Emission of the "Event Horizon Rim" (horizonRim, r=0.84): Most critical. The event horizon does not emit light and cannot be observed. Constantly drawing a glowing ring at the horizon directly induces the misconception that the "event horizon = a glowing structure." -> Stop constant display, demote to a shadow-step-only toggle for "Estimated Event Horizon Position (Guide Line)," or remove.

  • Beaming Direction World-Fixed: The brighter side should be fixed relative to the observer, so orbiting 180° creates a contradiction where the "receding side is brighter" (a trade-off of dropping billboarding). -> Move the halo (the main body of the boosting effect) to the lens image plane group (azimuth-tracking), and weaken the vertex color gradient of the disk itself. This keeps it mostly correct even when orbiting around.

  • Exaggeration of the Triple Ring (rim 0.84 / subring 0.94 / photon 1.02): Actual subrings cling tightly to the shadow edge, and seeing 3 of them constantly makes it read like "machine parts." Exaggeration is acceptable as an educational representation, but constant display is the issue. -> Move to Item C (step-limited display).

  • No Gravitational Lensing Distortion of Background Stars: Stars pass straight through, with distortion applied only to the explanatory arcs. As a medium-term task, implementing an approximation formula to shift star positions near the shadow would be lightweight and feasible.

[Omitted below]

Reading through this gives me the feeling of a fog finally lifting. Starting July 7th, it looks like it will move to an additional credit-based system separate from the Claude monthly subscription plan, but as long as the Trump administration doesn't step in to block it again, the quality makes me want to keep using it anyway.