Florian (@SrileXR) used Cursor, Claude Code + Antigravity to expand a racing prototype into Shed Racer.
Tip: give the agent working code to adapt, then test edge cases.
Credit: Florian Isikci / dmnshd; Kenney assets, mrdoob port.
dmnshd.gg/blog/from-a-tweet-…
ALT Creator's Shed Racer video thumbnail showing the same miniature racing track in a virtual forest on the left and on an office table in mixed reality on the right. Credit: Florian Isikci / @SrileXR / dmnshd; Kenney assets and mrdoob's port.
Codex Mortis was built with Claude Code.
The maker's bug-fixing workflow grew from one chat to parallel agent sessions, then a custom task dashboard. Human QA still drove the fixes.
Game + image credit: GROLAF / CRUNCHFEST.
store.steampowered.com/app/4…
ALT Official CODEX MORTIS gameplay screenshot showing a necromancer, a large skeletal boss and spell effects in an isometric graveyard battlefield. Game and image credit: GROLAF / CRUNCHFEST.
This fishing game was built with Opus agents, according to @dangreenheck.
Tidewater has public source for its WebGPU renderer, ocean simulation and fishing systems.
Game + image credit: Dan Greenheck.
github.com/dgreenheck/tidewa…
ALT Tidewater gameplay screenshot: a first-person fishing rod points over clear turquoise water from a wooden pier, with a tropical island and clouds beyond. Game and image credit: Dan Greenheck / @dangreenheck.
Daniel Petroff is building Imperium, a space 4X, with Claude Code.
His setup: separate agent worktrees, an MCP bridge into the running game, and automated balance tests.
Game + image credit: Daniel Petroff / Petroff Games.
petroffltd.com/imperium/pres…
ALT Imperium species-selection screen, with a crystalline alien in the center and a grid of species symbols on the left. Official press screenshot. Game and image credit: Daniel Petroff / Petroff Games.
BLINDSIDE's team says Cursor + Codex generated all the code.
Humans made the art and levels. Their agents also wired assets and checked Unity errors.
Credit: Team Woll / @BlasinRee; art + levels: Frank.
blasin.itch.io/blindside
ALT Official BLINDSIDE gameplay screenshot: a white character moves through a geometric orange and cream room viewed from above, with directional light beams and a minimal status display. Credit: Team Woll; art and levels by Frank; producer Blasin.
Claude Code helped @leocooout build this capybara delivery game.
His trick: have the agent build a map editor and cutscene tools, then shape the city by hand.
Creator + image credit: @leocooout
Making-of: reddit.com/r/vibecoding/comm…
ALT Gameplay still from A Game About Capybaras Delivering Food: a capybara rides a red delivery scooter along a hillside road overlooking a city, with a phone map visible on the right. Creator and image credit: @leocooout.
LeMi retweeted
FLUX 3 Image Klein is here.
Open Weights with the quality of a large frontier model. Customizable and fast.
5.6B parameters and runs on a single consumer GPU.
Native 4K, bounding boxes, and pixel-perfect editing.
We finally got a speed update of the gpt models
Day 1/
We have optimized the default speed to be ~50% faster across GPT-6 Astra and GPT-6.1 Sol through the subscription across all our products and partners using Sign in With ChatGPT (including OpenCode, Pi, Amp, Devin, ...).
No changes needed on your end and this should be felt within the next two hours.
Damn, my Codex agent has been working for 10 hours, and I'm on the Plus plan. The 6.1 Sol is efficient. Astra, on the other hand was only able to do 1.5 hours of work before reaching its limits.
LeMi retweeted
OpenAI just drew the line between a chatbot and a real agent.
A chatbot waits.
A dot stays awake.
That sounds like product language, but it changes the job completely.
The useful AI is not the one that writes a good answer in a chat window. It is the one that keeps a thread of work alive after you leave: checks the queue, moves the task forward, and comes back only when a decision needs you.
That is why the small details matter much more than the demo.
What can it do alone?
What requires a quick yes?
Which apps can it touch?
Where does the work happen?
How do you stop it?
OpenAI says dots run on their own cloud computer and work across the apps you choose to connect. That makes the interface less like “ask a question” and more like “give someone a job.”
The risk is obvious: an always-on system can make a lot of small mistakes while nobody is watching.
So the product is not autonomy by itself. The product is controlled autonomy.
Good agents should have goals, boundaries, memory, and a clear moment when they hand the work back.
That is the shift hiding inside the announcement.
AI is moving from a tab you open when you need help to a worker you supervise across the week.
The video is free. Almost nobody will watch it.
LeMi retweeted
OpenAI just made its fastest API tier up to 8x faster than Standard processing.
That is not a benchmark story.
It changes what an agent feels like while you are actually using it at work.
Most people treat latency as the dead space between asking and receiving. In an agent loop, it is the dead space between every tiny decision: read the file, call the tool, inspect the result, choose the next move, ask again.
Add enough of those pauses and a capable agent turns into a thing you check later.
Cut them down and the same workflow starts to feel interactive.
That is the real shift.
Fast models do not make bad plans good. They do not make a brittle workflow safe. They do not remove the need for guardrails or judgment.
But speed changes the size of problem you can keep in one live loop.
You can interrupt earlier. Review more often. Run a tool, see the result, and change direction before the context goes stale.
OpenAI’s docs say Ultrafast is meant for cases where speed justifies the higher cost. That is the correct framing.
Not every task needs it.
The ones that do are usually not chat prompts.
They are agents making many small moves where every pause becomes part of the product.
The demo is free. Almost nobody will watch it.
LeMi retweeted
Hemanshu literally explained why AI coding is about to swallow more than engineering work.
It is not autocomplete.
It is not cheap code.
It is taste, turned into execution.
He used Opus 5.5 to build a playable game.
The surprising part was not that it worked.
The colors felt chosen. The motion felt deliberate. The interface responded like somebody had thought about the next ten seconds of your attention.
That coherence is harder to fake than any feature list.
And he says he did not prompt the UI animations one by one.
That is the real shift.
For years, building a good product meant a relay race. Product wrote a spec. Design created the frames. Engineering rebuilt the intent. QA found the holes left between every handoff.
The expensive part was never typing the code.
It was preserving taste across the handoffs.
When a model can carry the visual logic into the implementation, a single person can reach the moment of truth much faster: does someone actually want this?
That is why these demos matter.
Not because AI can make another game.
Because small teams can now test polished product ideas before they spend months assembling the team that would have built the first version.
His full demo is free on X. Most people will scroll past what it means.
LeMi retweeted
FREYA IS BUILDING ITS OWN VOICE TECHNOLOGY
Adam ranks #1 among AI models on
@DesignArena’s AudioRealismBench.
freyavoice.ai
Freya is an Istanbul-based company working on speech synthesis, speech recognition and voice interaction. Its latest release includes two English AI voices, Adam and Eve, which are gradually becoming available through an API.
Adam currently ranks first among AI models in Design Arena’s blind Audio Realism Bench, behind only recorded human speech.
The team says it develops its own speech models rather than simply wrapping another provider. Its smaller Turkish model, FreyaTTS-small, was trained from scratch, with its code, weights and evaluation dataset released publicly.
Freya’s broader plan is to build the voice layer for agents, business calls and automated workflows. The API is still opening in batches, so this is an early product rather than a finished voice platform.