ICN | B.IA.seed Edge

AI video creator cheat sheet: compare the cost of 10 seconds across model makers and aggregators, find free credits, contests & creator programs with prizes and deadlines. Sources linked. Updated monthly. Free for followers. Enter your handle: yst.media/en/x

Images 2.6: Flare for drafts, Sunburst for finals In the OpenAI API, Images 2.6 has two variants: Flare is fast, Sunburst is detailed. We run the storyboard on the first and take final frames from the second. Plus GIF animation and precise edits. What OpenAI claims The news is from November 4. Generation is now 50% faster, and Flare at quality=low can produce an image in six seconds. That is claimed by the developer; I did not measure the speed myself. Sunburst at medium or high gives higher quality and keeps consistency between edits better. For characters this is the main point: the hero should not drift between iterations. What about the price There is no single price per image. Both models cost $30 per 1M output tokens, and input is the same too: $5 for text and $8 for images. The same rate does not mean the same bill: token usage can differ between models and quality settings. In OpenAI's calculator example, 823 output tokens come to about 2.5 cents, but the page does not show for what size and quality. So Flare's gain on drafts is first of all in time, not in rate. What else is in the release GIFs work for mascots, characters and storyboards. The API now has edits by visual marks and SVG export. The complexity parameter controls scene planning, quality controls detail. For most edits OpenAI recommends low complexity, for scenes with many elements medium or high. My setup is simple: Flare for drafts, Sunburst for the final. We run storyboard variants fast and cheap in terms of time, and once a frame is chosen, we make the final in a separate pass on the detailed model and refine it with precise edits without rebuilding the frame.

OpenAI raised the limits. The ChatGPT bonus can go toward images OpenAI added an API bonus to eligible Pro and Business plans, and on October 29 raised the limits for Work and Codex. What's changing with the limits On the $100 and $500 Pro plans and Business Premium, Work and Codex usage limits have doubled. There's a detail with the $200 Pro plan: the lower limit was raised to the previous higher level, while the higher limit stayed the same. The planned reduction was canceled. Plus is unchanged. The calculation "I pay $200 and get $4 000 worth of usage at API prices" hasn't been confirmed yet. That needs a calculation based on specific models and tasks. The actual API bonus on the $200 Pro plan is $200 for a full month. What's cool here The $200 itself doesn't impress me that much. What's much more interesting is that you can use it to pay for images through the API. Image generation in chat was already included in the subscription. Now you can add infographics, post covers and other social media visuals to your own automation. Instead of going into chat every time, you can run it from your own workflow. That's super cool. Voice generation and audio transcription through OpenAI's API are also covered by this bonus. Who gets it and how to claim it On Pro, you get $100, $200 or $500 a month, depending on your plan. On Business: $20 per Standard seat or $100 per Premium seat, up to $500 a month per workspace in total. Plus, Go, Free and Enterprise aren't eligible. You need a paid subscription that's at least seven days old. On Pro, open Settings, then Billing and API credits; on Business, the owner activates the bonus in the workspace's billing settings. Link an API organization you own. Your ChatGPT and API email addresses must match. Adding a card for the API is optional. Availability depends on the country and user group. The first credit is prorated for the days remaining in the period. With monthly billing, any unused balance expires at the end of the billing period.

Weekends are for resting

My friends in poker say a car should cost no more than 1-2% of your capital. youtu.be/axeNSZBf8sE?si=ojz9… The new Porsche mission S concept is out. Estimated at 500к$. I like it. Coming by 2029. Capital needs to be 25-50кк$ by then. Building it up.

How we improve our AI salesperson's script and check conversion We compare three responses to the same customer message and choose which one to use as the basis for changing the sales script. Then we check the statistics to see how the change performs against the previous version. Continuing the "Let's Build a Business with AI" series. At Animora, we built a system that talks to customers on its own, analyzes conversations and suggests changes. But what is its analysis actually based on? What to measure mistakes against You can feed an agent a conversation history and ask it to find poor moves. But what will it use to decide that a move was poor? So we built the foundation around a reference script and a data-driven approach. The script describes our version of the optimal customer path. Branches extend from it for different situations. On the map, these are step cards and the connections between them: what to find out, what to offer and where to go next. Both the sales reps and the agent follow this path. We look at which stages people drop off at and why. When a sales rep departs from the script successfully, the agent flags it as a strong move. Unsuccessful departures go into execution errors, which we then review with the team. How observations turn into changes There's also blind analysis: the agent reads the conversation without knowing our reference script. It looks at the conversation as an outside observer and proposes its own hypotheses. Observations from different conversations are then grouped into broader hypotheses. We discuss them with the sales reps and decide what is worth adding to the system. This lets us improve the existing path and spot things it doesn't account for at all. But another problem came up: it's hard to keep the whole map in your head all the time. To assess a single idea, we had to imagine how it would change the conversation. Choosing based on the response to the customer Now we take one customer message and compare how the sales rep responded, how the agent would respond without the hypothesis and how it would respond with the new hypothesis. A fictional example for a personalized film. Customer: "I'll think about it. I'm not sure what story to tell yet." Sales rep: "All right, message us when you've decided." Agent following the script: "The screenwriters will help put the story together. Would it be more convenient to continue this evening or tomorrow morning?" Agent with the hypothesis: "You could start with a couple of your memories, and the screenwriters will help put them together into a story. Would it be more convenient to continue this evening or tomorrow morning?" The hypothesis here is to give the person a clear first step. We can use the sales rep's response, keep the agent's current version or take the hypothesis forward. If we choose the third option, the editor prepares a change to the script card: suggest starting with a couple of memories. We review the before and after, then apply the change in a separate step. The agent keeps track of the map, and we look at the result of its work. This has made reviewing conversations much easier for us. For each change, we collect statistics on how customers move through the stages and the resulting sales conversion rate. We compare the results with the previous version. Just because we liked a response doesn't mean it performs better. For me, the main takeaway here is about the interface. You can build a very complex system, but people need to understand how to use it. So when I'm building, I picture a simple scene: I open it, see a specific situation and make a decision. Then I keep refining it.

Agents are still far from universal When you work with AI every day, it feels like everyone is already in on it: everyone has Claude Code or Codex, agents, their own systems. But our social circle seriously skews our sense of scale. Less than one percent In the "The AI Funnel" infographic, around 6 billion people have barely interacted with LLMs. Another 1 to 2 billion use them occasionally. And only around 20 million work with advanced models professionally, mostly in coding and math. If you're already using agents in Claude Code or Codex to tackle work tasks, on this chart you're roughly in that group of 20 million. About 0.2% of humanity. These are rough estimates from an image, not precise statistics. There's much further to go In its narrowest segment, the chart shows people with internal access to advanced systems: swarms of agents working on large projects, from cybersecurity to science. The range of possibilities there is even greater, while most people are still a long way from even ordinary work with agents. I don't think we've ever had systems that gave a person this much of an advantage across so many kinds of activity at once. So the feeling that "everyone has already figured it out" is misleading. If you want to gain that advantage, go deeper: take on more complex tasks, learn to work with agents and put them together into systems.

Claude and Codex: two websites for Animora In the previous post, I showed our studio. Now for the promised comparison: how Claude and Codex built a website for it. The shared starting brief: video, 3D and interesting transitions, so the page itself would feel like a film. Codex: GPT-6 Astra In reasoning max mode, it built a short prototype: video on the first screen, a photo turning into a character, and a 3D gallery. The movement is tied to scrolling: you scroll through the page and move through frames of a film. animora-concept.domainfortes… Claude Opus 5.5 It built a full page with a single 3D world: a screen with clips from our films, a projector beam and camera movement through seven chapters of the story. It kept all 12 sections and the text in five languages. This session used both reasoning max and medium modes. animora-v2.domainfortestsfor… I liked Claude's version more, and after some edits, we used it for the main website. Both links lead to the English versions, so you can open them side by side and compare for yourself.

Building a business with AI. A look at Animora Video Hi everyone. It's been a while since I last wrote to you: I've been busy working on our studio, Animora. I'm continuing the series on how we're building a business with AI. We're already fully up and running, growing and picking up the pace. A studio grew out of our experiments We started with AI video: we spent several months building our own generation systems. We ended up with a working product that became the foundation of Animora. We also have our own sales system, which the company is now using. So AI is involved both in making films and in running the business itself. What we do We turn personal stories into animated films about loved ones. A first meeting, a wedding, adventures with friends, parents' stories. People already have the memories, and we help turn them into a film. They share stories and photos. We write the script, create characters and animated scenes, add voice and music, and edit the film. The photos are references for the characters. Before the final handover, they watch the film, and we make the agreed revisions. Italy and the finished website There are days when our conversion rate jumps to 10%. We recently launched in Italy. And I'm finally ready to show you our Animora Video website. I built it with Claude. You can watch a compilation of clips from our films there and see how we turn a personal story into a film. Here's the English version: animora.video/stories/ I'll write a separate post about building the website and how Claude and Codex handled the task.

Just sealed my whitelist application. Encrypted NFTs feel like the natural next step for confidential tech: seen by all, known by one. After months of building on Zama with C4* Dev Hub, I'm glad to see a collection put it to real use. @fungolabs

Opus 5.5 is now publicly available. Goodbye, limits 😄

Chapter 4. Business. We solved payments and multilingual packaging early. The sales system now fits our operation and has its own developer agents, so we ship 5-10 updates every day. Here's some of what we've built. 1) A multi-user system with access for every department: sales, production, head of sales, head of production, and owner. 2) Automatic translation to and from the client. Russian-speaking managers can now sell worldwide while every customer gets natural language, without feeling like they're talking through Google Translate. 3) Statistics run through every layer: finance, managers, production, marketing, plus end-to-end conversion tracking for ad campaigns. We can see almost every data point. 4) The system handles part of the production work itself and keeps customers warm while it does it. That helped lift conversion. 5) And of course, the sales agent. For the business, the payoff is simple: one team can sell across languages, every department works in the same system, and I can trace an ad campaign all the way to revenue and production. The agent doesn't sit beside the process. It does part of the work and knows when to hand control back to a person. That last part deserves more detail. We wrote a sales script and put it into the agent's operating logic. It understands the conversation context, knows which stage of the sale it's in, and adapts when a customer pulls the conversation sideways. If it doesn't know what to write, it calls a human. It follows up based on what was actually discussed, can pause when someone asks to continue later, and can send payment links. But none of that matters if the agent can't sell. Ours does. We're 4 weeks in. We built the sales script without specialists, and every improvement came from live work. During the first week of testing, we obviously lost more leads while tuning the agent and the surrounding systems. Even with that included, we reached 5% in the first month and hit 12% on individual days. My read: that's pretty good. And this still isn't the most interesting part. The agent training system is still in development. I'll share some of the work behind it next.

Launching a business with AI. Chapter 3: sales. We got the leads, then hit the sales problem. Tried phone calls first and got 80% no-answers from day one. My read: calls can still work if your audience is 45+, Generation X. Ours is millennials and younger. They don't want to pick up, especially when an unknown number calls. We spent about 1.5 weeks testing it. Result: 0 sales. So we moved to chats. Lesson 3: chats and social media are where b2c sales happen. I went deep on sales systems here. While we were still testing calls, I researched how AI could fit into the process. I built an agent that could talk to people with a fairly natural voice. Technically, it worked. Commercially, there were too many problems. Even the best voice agents still don't sound fully human. Once people realize they're talking to AI, many stop engaging because the conversation feels devalued. Regulators also require companies to disclose that the caller is AI, with large fines across the EU for getting this wrong. Calls weren't the answer. Fine. Chats were different. I could move fast, control the experience, and build around the way our customers already preferred to communicate. I spent a couple of days researching the market and found around 80 systems for managing chat sales. None had the full set of features I wanted. More importantly, none let me connect my own agent and customize the system around it. Which makes sense. Giving customers that level of control would mean opening up a large part of the product. But I already had 2 things: a feature list assembled from the entire market and configured agents with full limits. After 3 days, our system had 50% of the key features from the combined scope of those products. After a week, it passed 80%. That's the product level of companies valued at tens of millions of dollars, including Moneychat and similar systems. I've written this before: if you know how something should work, you can now build it very quickly. AI has compressed the distance between understanding a system and shipping one. We're already living in that reality. The business result came fast too. We got our first sale around day 4 of testing. New market, new niche, new sales system. The hypothesis was confirmed. Now we had to turn all of it into a BUSINESS.

Launching a business in the age of AI Chapter 2: Traffic About 1 month ago, my partner Sergey ran our first ads for a product that didn't exist yet. He also writes about launching traffic with AI on Instagram. The job was simple: test the niche before we built around it. We went straight to European markets and got leads at a good price. Then we tested several countries. The obvious assumption was that more expensive countries would produce more expensive leads. They didn't. The relationship wasn't linear. In practice, a country with a higher average order value could still give us a cheaper lead. For a founder, this means purchasing power and lead cost don't move together. Test the market before committing the team and budget. Lesson 2: Test before concentrating your effort.

Yesterday we got the first payouts from our new project. Important milestone. Here are my lessons from launching a company in the AI era. Chapter 1: R&D Around March, we went deep into AI video generation. We built several systems of our own for making videos. This was the R&D stage: we were trying to build a product that could produce content at an acceptable quality. Our systems got better as the models got better. Eventually we had a working product. It's not perfect, but it works. That's what matters. I wrote about the danger of this stage before. R&D keeps expanding, and you can get stuck there for a long time. My read: 95% of projects people are building right now will never be used. We entered a completely new market. Nobody on the team had experience with generation. We didn't even have experience making content in the first place. So 5 months of learning doesn't sound that long. But looking back, we could and should have gone to market much earlier. Lesson 1: get to your first revenue as fast as possible. A usable product with customers teaches you more than another month of polishing. Tomorrow I'll share what happened with our first traffic and first customers.