The bike turned yellow. Its reflection stayed red.
The saved prompt explicitly asked for a yellow reflection too.
From my FLUX 3 Klein tests: one run, 6-bit MLX, seed 7. Preserving the background isn't the same as keeping the scene consistent.

ALT AI-generated test images. Left: original red bicycle beside a puddle. Middle: edited yellow bicycle, with its reflection in the puddle still red. Right: a separate object-addition edit adding a black cat. FLUX 3 Image Klein, community flux3-mlx 0.1.0, 6-bit, seed 7, 16 steps on a 16 GB Mac mini.
1
1
89
Levent retweeted
I made a 3D Where's Waldo you can play in the browser
A whole seaside festival town, 10,000+ animated people, all generated from code
No image or model files
Waldo hides somewhere new every visit. Can you beat my time?
waldo-stripes-soft.vercel.ap…
Now your dot can securely complete purchases on your behalf—with your approval.
With ChatGPT Wallet, agents pay without access to your actual card details. Every purchase requires a security check.
ChatGPT Wallet is rolling out starting today to dots, for Pro subscribers in the US.
openai.com/index/introducing…
Levent retweeted
Issued by the FAA just moments ago, ALL US airports are currently ground stopped. This means NO flights may depart for US airports. Flights already in the air will be allowed to land. This is related to an FAA equipment outage.

ALT DUE TO EXTRAORDINARY CIRCUMSTANCES AND FOR REASONS OF SAFETY. ATTENTION ALL AIRCRAFT OPERATORS, BY ORDER OF THE FEDERAL AVIATION COMMAND CENTER ALL AIRPORTS ARE GROUND STOPPED. ALL DEPARTING TRAFFIC SHOULD REMAIN GROUND STOPPED UNTIL FURTHER NOTICE. INBOUND INTERNATIONAL FLIGHTS WILL BE MANAGED ON A CASE BY CASE BASIS.
Who should switch to Hydrogen today?
If you run anything on 3.8 Flash, I’d switch now.
You get a model that went from 31 to 51 on the same index, runs faster per task, and costs less per finished job. Your old price ends in January anyway.
If you pay for Argon for everyday coding and document work, try Hydrogen on high first. Zapier’s test and the Artificial Analysis index both put it level with Argon, at roughly half the cost per task.
If your agent lives in a terminal or clicks around a desktop, keep Sonnet 5.5 in the loop. That’s the one place where Google’s own card shows Hydrogen losing.
And if you’re on Antigravity with an Ultra plan, Orchestrator is the thing to test this week.
Let the big model think and the fast one type. That’s the setup I’m most curious to see people report on.
Gemini 4 Hydrogen is here, bringing cutting-edge intelligence to everyone for fast, everyday help.
It makes it easier to get things done, following your instructions more closely, calling the right tools, and getting you to a reliable output with less back-and-forth. 🧵
1
33
Levent retweeted
Which Claude model should you build with?
Our new guide breaks down when to use Fable 5.2, Opus 5.5, Sonnet or Haiku. It also covers when to adjust effort instead of switching models, and how to save cost by putting subagents on smaller models.
claude.dev/blog/which-claude…
Levent retweeted
nice
you can now build with google maps platform in AI Studio 🗺️📍
go from idea to app powered by @googlemaps in minutes!
what’s new:
- 1-click setup
- create maps apps just by describing your idea, gemini does the rest
- build with map tiles, places, weather, routing, 3D buildings, street view, and more
try it out today: ai.studio/build
super excited to lead the @fal collab with @UnrealEngine for this game jam!!
gonna be there in person over the weekend so feel free to come up and say hi. promise won't bite!
we're handing out some fine credits and prizes to you fellow jammers to bring your visions come to life!
Meet the 2026 Epic AI MegaJam partners! 🤝
Claude, Google, OpenAI, and SpaceXSI are joining AMD, Diversion, fal, Hyper3D, Intel, Meshy, NetApp, NVIDIA, PBRgen, Reallusion, Rokoko, Tripo, and WeLoveIndies to add tools, credits, and prizes to this year’s jam.
Registered jammers: keep an eye on your inbox. Perk codes and claim links go out today.
See every perk and prize: itch.io/jam/2026-epic-ai-meg…
Levent retweeted
The tower of @fal. Thank you for my gift @noahsolomon!
Go generate your own custom one at @2ndbrickstudio everyone.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Levent retweeted
fal ($16B) has a shared Slack channel with almost every customer, with 3-4 engineers in each.
Burkay Gur, CEO of fal, admits engineers hate sales and hate being sold to.
How they learned to sell anyway:
1. They took their investor's intro to a sales adviser, despite the skepticism
2. The founders ran sales themselves for a long time
3. Only then did they hire a team to scale it
4. Salespeople are hired as advocates for the customer inside fal
Lessons:
→ Listen to people who've seen thousands of companies, even when it's uncomfortable
→ Put engineers where customers talk
→ Serve first: that's how engineers want to be sold to
Full breakdown here:
productmarketfit.tech/p/the-…
fal is hiring btw
fal.ai/careers
Today, we’re announcing we raised our Series E & F rounds for a combined total of $1.65B. Our Series E was led by @sequoia and our Series F was led by @BessemerVP with continued investment from @nvidia.
We’re excited to push the boundaries of what’s possible even more next year and this funding will enable the team to build the way we want to.
Levent retweeted
Congratulations to @fal on its Series E and F rounds 🎉
Earlier this year, our own @emilyyzhao sat down with fal CEO and co-founder @burkaygur to talk through how the company got here and what the future looks like.
A few highlights:
👉 Burkay spent five years at Coinbase as one of the first machine learning hires, using ML to fight fraud. The missing infrastructure and tooling is what pushed him to start fal.
👉 The first iteration of fal was a Python compute company. In Burkay’s words, “Databricks and Snowflake brought SQL to the cloud; we were going to bring Python to the cloud.”
👉 After ChatGPT, fal went all-in on image models while most of the industry chased LLMs.
👉 The Series A thesis, end of 2023: quality improves, adoption follows, and Fortune 2000 companies come with it.
👉 Burkay says a very big part of fal’s revenue now comes from product photography, e-commerce, marketing, and ads.
Emily also shared more on why we became investors in fal: LLMs tend to drive market consolidation, but generative media is subjective and personalized, so users end up chaining models together. The platform that provides that workflow and tooling is the hardest problem in the space.
Congrats to Burkay, @gorkem, and the entire fal team. We're proud to be partners in your journey. Watch our full interview with Burkay below.
𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒
(00:20) How fal started, and what Coinbase taught Burkay
(04:33) Did fal ever doubt the market?
(07:45) Where generative media is landing in the enterprise
(12:47) What Burkay would do differently
(15:10) Why fal chose Salesforce Ventures
(17:15) Why we invested
Ripple-65B's launch chart compares it only with other US open models.
So I added the one most people would reach for at this size: Qwen 3.8 27B.
On each maker's own Terminal-Bench numbers, Ripple leads by 7.5 points. On DeepSWE they're about even, and Qwen is under half the size and reads images.
If you can use Qwen, stay with it for now.
If your team can't touch Chinese models, Ripple tops every US open model on Terminal-Bench, the one test it leads with.
huggingface.co/ricursiveAI/R…
44
When Codex asks you to approve a new hook, you assume you're the one answering.
A Postman user says the desktop app installed plugins into Claude Code and Codex without asking, then tried to approve its own startup hook.
A reply posted for Postman promises a clearer opt-in.
I'd like agents to treat a new plugin the way my phone treats an app asking for my location.
Tell me, and let me say no.
medium.com/@lvntblsn/what-is…
1
1
242
Levent retweeted
OpenAI and Google DeepMind launch video models on fal. fal often gets exclusives in return.
fal ($16B) calls its model "marketplace plus":
1. Developers love the platform and aren't tied to any single model
2. Labs use fal to reach those developers
3. Every launch is co-marketed
4. In exchange, fal gets exclusive access for a period, sometimes permanently
5. On top, fal gives labs the infrastructure to run their models
Lessons:
→ Your audience is an asset you can trade
→ Labs want the biggest platform, so being #1 compounds
→ Give the supply side more than distribution
Full breakdown here:
productmarketfit.tech/p/the-…
Tencent says its new Hy4 beats other open-weight models on a finance benchmark.
3.04 versus 2.97 for DeepSeek. Tencent ran the evaluation itself, so I'd like to see independent results.
But there's something odd here.
The Hy4 preview had downloadable weights. The full release now has vision and computer use, but the weights aren't available yet. Only API access.
Tencent says they'll come later.
I find it a little strange to announce a win in the open-model category before releasing the model's weights.
I'll call it open when I can download it.
unite.ai/tencent-releases-fu…
48
Our co-founders @burkaygur and @gorkem sat down with @MollySOShea to talk about the rise of generative media, the economics of AI infrastructure and what’s next for fal.
BREAKING: @fal just raised $1.65 BILLION across its Series E & F
› Series E led by @sequoia
› Series F led by @BessemerVP
I sat down w/ Co-Founders: CEO Burkay Gür (@burkaygur) & CTO Görkem Yurtseven (@gorkem) to discuss the explosive growth of generative media, which they believe is 1 of only 3 major product-market fits in AI, alongside search & coding agents.
Founded in 2021 by former Coinbase & Amazon engineers, fal powers the infrastructure behind AI image, video, audio & 3D models, serving developers, studios & model labs.
fal
/faɫ/ · noun · Turkish
Pronounced approximately “fahl” (rhymes with doll in many American accents)
Definitions
1. Fortune-telling; divination. The practice of predicting future events, particularly through interpreting coffee grounds, cards, or other symbols.
2. Fortune; fate; luck. A person's destiny or prospects.
3. Omen; augury. A sign interpreted as indicating what may happen in the future.
A few highlights:
›› Professional users spend $10K–$20K/month generating AI media.
›› Video reached ~70% of fal's revenue before image editing drove a resurgence, bringing the mix back to 50/50.
›› Demand continued growing even as premium AI video model prices jumped more than 10X.
›› fal has scaled to ~150 employees with just 2 reporting levels.
›› Compute shortages are shaping model launches, while the founders believe AI's strongest margins will be captured by infrastructure providers rather than applications.
"AI really has 3 product market fits. One of them is search. The other really clear one that's growing really fast now is coding agents. & the third one is, we believe, generative media."
We recorded at Dalida in San Francisco & got into fal's evolution from Python infrastructure, the economics of AI inference, Hollywood, robotics, world models & the future of compute.
This was SO much fun!!
𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒
(00:00) Burkay Gur & Gorkem Yurtseven, Co-Founders at fal
(00:48) Why fal stopped chasing headlines
(03:15) Why generative media is AI’s next big market
(04:51) The infinite demand for AI videos
(09:35) The future of AI-generated ads
(13:19) How fal pivoted into generative media
(18:02) Hiring people who live and breathe AI
(23:01) Where AI meets Hollywood
(24:29) From friends to Co-Founders
(27:48) What a “Super IC” actually is
(35:58) The hardest part of staying at the frontier
(37:42) Why compute is the bottleneck for new models
(41:55) Why AI models have such weird names
(43:56) The story behind the name fal
(45:12) Why open source isn't going anywhere
(49:02) "AI has had zero impact on Hollywood"
(53:39) Waiting for AI's Toy Story moment
(56:58) Why Jack Dorsey is underrated
(58:55) Why the application layer has no margins
Levent retweeted
Proud to lead the series F in @fal. On our investment:
Generative media is already one of the biggest categories in AI, but its so early. As models improve, new experiences are unlocked and the market grows. Gaming, video, ads, world models, image, 3d. All of it is just getting started.
@fal has become the key infra for all AI media, and the biggest waves are yet to come.
Most are building for LLMs and coding. @fal is building for creativity. They dominate that market, and are innovating across the whole stack from model to inference to workflow.
The team is the best in the business. Insanely fast, ruthless about performance, enormous ambition, and also humble, kind humans.
I have known @gorkem for more than 15 years and @burkaygur and @isidentical have become good friends along the way. I first got involved w them in the seed and
@BessemerVP has been a proud partner to them for many rounds. So excited to lead this round!
Everything you need to know about @fal and the people behind it, @burkaygur and @gorkem.
What a journey so far. And somehow, it still feels like just the beginning. 🚀
Thanks @MollySOShea
BREAKING: @fal just raised $1.65 BILLION across its Series E & F
› Series E led by @sequoia
› Series F led by @BessemerVP
I sat down w/ Co-Founders: CEO Burkay Gür (@burkaygur) & CTO Görkem Yurtseven (@gorkem) to discuss the explosive growth of generative media, which they believe is 1 of only 3 major product-market fits in AI, alongside search & coding agents.
Founded in 2021 by former Coinbase & Amazon engineers, fal powers the infrastructure behind AI image, video, audio & 3D models, serving developers, studios & model labs.
fal
/faɫ/ · noun · Turkish
Pronounced approximately “fahl” (rhymes with doll in many American accents)
Definitions
1. Fortune-telling; divination. The practice of predicting future events, particularly through interpreting coffee grounds, cards, or other symbols.
2. Fortune; fate; luck. A person's destiny or prospects.
3. Omen; augury. A sign interpreted as indicating what may happen in the future.
A few highlights:
›› Professional users spend $10K–$20K/month generating AI media.
›› Video reached ~70% of fal's revenue before image editing drove a resurgence, bringing the mix back to 50/50.
›› Demand continued growing even as premium AI video model prices jumped more than 10X.
›› fal has scaled to ~150 employees with just 2 reporting levels.
›› Compute shortages are shaping model launches, while the founders believe AI's strongest margins will be captured by infrastructure providers rather than applications.
"AI really has 3 product market fits. One of them is search. The other really clear one that's growing really fast now is coding agents. & the third one is, we believe, generative media."
We recorded at Dalida in San Francisco & got into fal's evolution from Python infrastructure, the economics of AI inference, Hollywood, robotics, world models & the future of compute.
This was SO much fun!!
𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒
(00:00) Burkay Gur & Gorkem Yurtseven, Co-Founders at fal
(00:48) Why fal stopped chasing headlines
(03:15) Why generative media is AI’s next big market
(04:51) The infinite demand for AI videos
(09:35) The future of AI-generated ads
(13:19) How fal pivoted into generative media
(18:02) Hiring people who live and breathe AI
(23:01) Where AI meets Hollywood
(24:29) From friends to Co-Founders
(27:48) What a “Super IC” actually is
(35:58) The hardest part of staying at the frontier
(37:42) Why compute is the bottleneck for new models
(41:55) Why AI models have such weird names
(43:56) The story behind the name fal
(45:12) Why open source isn't going anywhere
(49:02) "AI has had zero impact on Hollywood"
(53:39) Waiting for AI's Toy Story moment
(56:58) Why Jack Dorsey is underrated
(58:55) Why the application layer has no margins
7
227






















