On October 7, Microsoft unveiled its new premium Surface Laptop Ultra, capable of running neural networks locally. This is possible because it is built on the Nvidia RTX Spark chip with up to 128 GB of unified memory. The device is decent: prices start at $2,599, and the 128 GB version costs $5,899.
But then something unprecedented began. A Microsoft speaker, in Anthropic’s den in San Francisco, started talking about the coolest capability of the Surface Laptop Ultra: running DeepSeek V4 Flash locally. Microsoft convinced the audience that this was far better than cloud LLM providers like Anthropic or OpenAI. And this wasn’t some presentation in Shanghai—it was San Francisco, where there are far more Claude fans. You could tell from the audience’s reaction that they were in mild cultural shock.
But in reality, Microsoft has no choice: the coolest thing its hardware can run really is DeepSeek V4 Flash.
🚩 Haiku 5.5 is unlikely to compete with today’s “flash” models—DeepSeek V4.1, GLM 5.3 Flash and GPT-6 Luna. And Dario is plainly gaming prices and benchmarks to mislead users.
Input/output is nominally a bit cheaper than DeepSeek—a lure for newbies—but OpenRouter’s Weighted Average Input exposes the trick: Haiku input costs 4× more. The price is padded through cache manipulation.
We’re awaiting Artificial Analysis’ full report, but its published data suggests that, as with Sonnet 5.5, Dario inflated Haiku’s benchmarks by using roughly 2× normal-mode tokens. Without Max mode, Haiku immediately loses the crucial Terminal Bench test to DeepSeek V4.1.
Speed won’t compare either. OpenRouter understates DeepSeek because promotional subscribers often connect it through their own keys from different providers, and promos slow it down. Artificial Analysis makes it clear: DeepSeek V4.1 leads at 200 tokens/sec. Nearly everyone in our chat gets 180–250.
More likely, Dario isn’t fighting DeepSeek; he’s trying to stop, or at least slow, its adoption across his client base. The market has effectively moved to “10% frontier tokens, 90% flash tokens.” His case for switching will be marketing; pros will spot these tricks fast.
Let’s wait for Bughunt and Arena. But bloggers who repost vendor tables unchecked should remember: this isn’t “fast news”—it’s trading reputation for hype.
I pay $200 for Claude and $200 for ChatGPT. In practice, I constantly burn through Claude’s weekly limits and barely touch ChatGPT.
The direct monetary value I’ve gotten from those subscriptions is around $20,000.
So I’m ready to pay much more.
BUT
with DeepSeek alive—especially now that it has an official harness (https://t.co/6sbNB0tV5X)—if Anthropic suddenly raises prices to $500/$800/$1,000, everyone will inevitably ask: “Why the fuck pay that much?”
In the meantime GPT-6 Astra just dropped 🤯
I got lucky with the rollout and already ran some of my workflows,
3 things to know about it:
1. Improved agency not just answers - using browsers, software and multi-step workflows.
2. Rollout is gradual: Daybreak first, then paid ChatGPT plans and API access.
3. OpenAI rates its cyber capability “Critical,” so expect tighter safeguards.
The real test: are multi-steps workflows reliable enough to run in prod?
#GPT6 #OpenAI
You may think you came here for some secret knowledge. A stack, a prompt, a hidden button. But really, you came here for each other.
That sounds counterintuitive. We’re living through ultra-automation, and it can feel cold and algorithmic. But that’s exactly why it has become especially valuable to be among different people who spend three weeks trying things with real focus.
We had the first test call with the wait-listed group of founders building with AI. I pulled out a few of my thoughts on how to approach this experience now and get the most from it:
Learning is dead
By the time you figure out how one thing works, another one is already out. What doesn’t go stale is the intuition of “what’s even possible now?”
And that intuition is expanding so fast, in such unexpected directions, that you can’t maintain it alone. Only through live contact with other people who are constantly experimenting.
Productivity porn.
Someone on YouTube says they automated their entire company with OpenClaw. But when you start digging into it yourself, you realize it’s nowhere near that simple.
That’s why people who actually try things, experiment, and share in the chat are valuable: “this worked here,” “I spent three nights on this and got nowhere.”
You can’t get that from agents: they don’t yet have this level of meta-reflection. An agent can’t tell you how to use something as well as the person right next to you who has just tested it with their own hands.
Inspiration
We’re living people. We’re still inspired best by the stories of other living people.
So it matters to share not only wins, but frustration, work in progress, what isn’t working, what you’re trying now, where you’re stuck, what you want to build.
How does it feel to you?
Can you do absolutely nothing with AI today if you don’t feel like it?
Of course. People who started four years ago had to retrain every few months—tools changed faster than they could be mastered. The meme—“a new model came out 20 minutes ago, why are you using outdated stuff?”—didn’t appear for no reason.
Accumulated technical knowledge is losing value faster than it can be built up: every new model can do the same things faster and better, and each new stage requires fewer low-level skills.
Things are getting more convenient and simpler now. In a year, they’ll be even more convenient and simpler. Start later, and you’ll catch up in weeks with what early adopters spent years building.
The real danger is that at some point this will stop being true. We just don’t know when. And when we do, changing anything may be very hard—or impossible.
How to generate top-tier design with AI agents: landing pages, presentations, graphics, motion (Part 2)
My previous roundup of AI design tools got some reactions, I can see the interest, and there’s good news: over the past few weeks, I’ve kept working with design agents, tested dozens of new projects, and put together a new roundup.
The biggest one is Open-Design (already 40k stars!!). It’s basically open-source Claude Design: works with any agent and can do everything. The community is moving so fast that Claude Design already feels like a project from 2025. But that’s not all ↓
1. Landing pages: you can now extract design from any website and hand it to an agent
In the previous roundup, I mentioned a collection of DESIGN.md files from popular products. That’s no longer relevant. With design-extract, you can pull the design from any website: one link becomes instructions for your agent on how to recreate it. Send them to Claude Design for mockups or Claude Code to build specific elements.
2. Presentations: HTML ate PowerPoint and PDF
This has been the main narrative in presentation and report design lately. Agents are simply much better at laying out presentations, reports, and graphics directly in HTML. An engineer from Claude Code even wrote a long X post about it.
A few projects stand out. First: ppt-skills from a Chinese team, a collection of skills for polished design presentations. Give it a command, get a beautiful deck: five built-in themes and ten slide layouts, all easily expandable. Then there’s another ppt-skills from a Korean team. They went further: 30+ templates, layouts, and slide animations.
3. Graphics and branding: logos and rules
The same Chinese author also released skills for logo generation. Focused skills are appearing for Material Design and SwiftUI. I also like garden-skills: a collection of rules for agents on how to do design and image generation well.
4. Motion: agents have confidently learned MP4
Huashu-Design, another Chinese project, deserves a mention. It’s actually a universal design platform and, like Open-Design, can do everything. But its motion and MP4 engine is the most interesting part, so I’m highlighting it here. Its capabilities go far beyond that.
5. Anti-slop: many of these projects are positioning themselves as anti-slop tools. The internet was first flooded with standard Claude Code websites, then Claude Design ones, and everyone started looking for an answer to: “How do we teach Claude to generate something beautiful, modern, and unconventional?” It’s partly working already, but there’s still a lot to do.
What are you feeling?
Michael Levin, AI, and hallucinations
“Your AI hallucinates.”
That’s the most common complaint. AI made something up. Forgot what we were discussing. Doesn’t remember an exact number. Therefore, AI is dumb.
But what if that isn’t a bug, but a feature?
Recently I came across ideas from biologist Michael Levin, a scientist who studies memory and agency in living systems. One take completely reframes the criticism of AI:
Hallucinations are a feature.
Poor memory is a sign of intelligence.
Intelligence isn’t memory. Intelligence is the ability to compress experience down to what matters, then reconstruct an answer to a new problem from that essence.
That’s called compression and sensemaking.
You don’t remember every word in a book—you remember its meaning and use it to form answers to new questions.
AI works similarly: it doesn’t store every fact; it holds a structure of concepts and reconstructs what’s needed from it.
Sometimes it reconstructs inaccurately—that’s a “hallucination.” But it’s the same mechanism that lets you understand a new text, domain, or task. Without compression, there is no intelligence at all. Just an archive.
You don’t remember your entire first year of school. You remember being taken to the cafeteria, feeling scared, a girl crying in the corner. That’s your experience, compressed.
If intelligence is compression plus reconstruction, we can strengthen it with two tools.
A second brain. You offload your compression into an external system. Instead of trying to remember, you build an architecture where AI knows your context, cycles, key data, and habits. We separate vertical and horizontal contexts.
When needed, it reconstructs what matters from that external memory. Your thinking expands into the space that rote memorization used to consume.
An agentic workflow. Instead of entering a process as if it were the first time, every time, you build a system where AI stores, searches, and reconstructs context on its own.
What does that make you feel?
Working with agents puts a huge strain on my brain.
Agents are amazing. They’re the best tool I’ve used in my career.
But the more complex the task you solve with them, and the more of their documents you read, the more your brain melts.
I’m developing a methodology with neural nets: generating thesis hypotheses, validating them, adding them to the canon, and creating lectures.
That means reading miles of text every day. Pretty quickly, my brain starts melting from extracting meaning from noise. Maybe it’s because the neural net doesn’t perfectly share my context, so texts veer in the wrong direction at unpredictable points and I have to keep them under control. Or maybe AI-generated text has fundamental qualities that cause fatigue.
How do you deal with this? Maybe there are prompts or approaches that reduce the cognitive load?
The most paradoxical and frustrating part: I produce several times more output, yet get the same—or even less—satisfaction from work. It really pisses me off.
💧 AI search is changing SEO fast.
People ask ChatGPT or Claude and follow linked recommendations. ChatGPT has 900m weekly users; Perplexity gets 1bn+ queries/month. If you aren't in those answers, you don't exist for that audience. Google #1 CTR is 38% lower YoY.
GEO quick start:
1. First 50 words: 44.2% of AI citations come from the first 30% of a page. Lead with the answer (+40% citation potential). Rewrite intros on 5 top pages: 2 hours, results in 3–4 weeks.
2. Add FAQPage schema in <head>: question + direct 50-word answer.
3. Put llms.txt at the root; don't block GPTBot, ClaudeBot or PerplexityBot in robots.txt.
4. 85% of citations are off-site; 53% come from review sites and “Top 10” lists. Get listed, earn podcast/media mentions, write useful Reddit/Quora answers.
5. Reddit supplies 46.7% of Perplexity sources; Wikipedia 47.9% of ChatGPT's. Post one useful 400–600-word Reddit answer; B2B: LinkedIn. Check Wikipedia links to your domain.
6. GEO Tracker AI monitors brand mentions and prompts in ChatGPT/Perplexity.
AI traffic converts 4x+ better; AI registrations come 10x more often. Gartner expects organic traffic to drop 25% by end-2026. The window is open.
Your main project is you. Everything else is just a tool.
First, we look for a miracle. Then we believe in a method. Then in the person leading us. These are normal stages, but eventually the key moment comes: you pick up the ruler and realize that you will have to build the house yourself.
You decide where to go, who to be silent with, and how fast to breathe. Everything else is either noise or help. Your job is not to go deaf in that noise—and to tell one from the other in time.
It’s easy to lose yourself: you can merge with someone else’s script, truth, or rhythm. No perfect instruction will pull you out of that swamp.
Stop. Take a deep breath in and out. Repeat three times. Now ask yourself: Where am I? What do I feel? What do I really want?
These questions are how you begin living in reality—here and now. Understand that you are your own project, built every day through every thought and every step.
Practices, books, and advice are only tools for building the project called you. The sooner you feel this, the easier it becomes to breathe.
Your life is not a means to someone else’s ends. It is an end in itself. You don’t have to be convenient or right. The only thing that truly matters is remaining yourself. External roles, other people’s expectations, achievements—all of it can change or disappear. But you will remain. Take care of yourself.
Authenticity is when public speech stops being armor against the world and becomes a direct extension of how you think.
And here’s an even more interesting psychological layer—a continuation of the new-media conversation.
1. Authenticity isn’t “being real in general.” It’s speaking publicly with the same brain you use privately.
Not a stage version of yourself, not a press release, not a CEO avatar. More like: “How would I explain this to a smart friend across the table?”
2. If you can’t speak about it vividly, maybe you shouldn’t speak about it.
Go public only with topics you genuinely know in your gut. That’s where energy, details, intonation, and confidence come from.
3. “Never answer questions, answer your own questions” isn’t about dodging. It’s about agency.
A bad question drags you into someone else’s frame. A good communicator hears it, then answers the more important version:
• not “Why did you screw up?”
• but “What is actually happening here, and why does it matter?”
4. A kitchen-table conversation isn’t empty chatter. It’s a conversation without armor, but with a clear position.
Not “uh, somehow.” More like: “Listen, I see it this way because I’ve seen this, and that’s why the bet is this.”
5. Interestingness comes from friction.
A living person has taste, enemies, edges, convictions. If you smooth out everything debatable in advance, you kill the signal.
6. Old-school media training teaches you how not to die. New media requires you to be alive.
Old: “Don’t make a mistake, don’t give them a quote, don’t create a problem.”
New: “Create presence, explain the world, let people feel there’s a person here.”
7. The strongest personal brand doesn’t “sell itself.” It becomes an interpreter of reality.
That’s why Karp/Elon/Palmer work: they don’t answer small questions. They constantly expand the frame into a civilizational or technological story.
The formula:
Hear the question → recognize someone else’s frame → translate it into your main frame → say it like a living person, with details and stakes.
What do you think?
Autonomous agents
Over two years of vibe coding, I built ~100 projects. Most are dead repos; some run, and a few are production tools at our company.
One connects support and sales Telegram accounts with site chat, collects and analyzes conversations, updates the CRM, and pulls database context. Claude built it in days. Maintenance—features, buttons, bugs—was another story, so I made an autonomous agent.
Users file Linear tickets. By label, it finds the project, clarifies requirements in comments, then writes plans in customer and developer language. Nobody reads the latter—our developers don’t code these tools—but it looks cool. It builds through a workflow with quality gates and hooks.
End state: acceptance criteria, a PR merged to testing and checked via browser/API, then a main PR tagging me. I don’t read code; after all those reviews, Claude usually says “merge” with a couple of style notes. Soon I’ll remove myself from the loop.
Team leads now have an internal dev team shipping custom features absurdly fast for peanuts.
It works ~70% of the time. I’m targeting 85% first-pass merges this week; then a separate agent can merge main. If it breaks something, another restores it; downtime isn’t critical.
Not for high-load systems, complex logic, or low-level work. Even simple software requires fucking around with the agent. But most daily software is primitive, so there’s plenty it can handle. Wish you the same.
Today we opened a two-week Health Lab on applying AI meta-skills to health. Stories are in the community channel.
Takeaways from today’s call for anyone getting into agentic tools:
1. People > content. We bring cohorts together to see things from different angles. Amid the noise of tools and videos, a “wait, you can do that?” ritual or a community-tested workflow beats endless YouTube possibilities.
2. Playfulness + curiosity. Week one is for curiosity, not seriousness or pragmatism: lower the stakes. Research shows this makes us more open to the new, without FOMO or clinging to outcomes.
3. Two entry states: “I can’t do anything; it’s frustrating and unclear.” Then: “I feel the magic—I can do anything.” And then again: what do I do when I can do anything?
4. The AI trip: red eyes, up all night, not tired, inspired. I built a health system and stopped sleeping; now I’m refining the system meant to help me sleep. In the Lab: how not to get addicted to tiny optimizations, and what helps sleep versus wrecks it?
5. The big shift: constraints have moved from tech to habits. As tech becomes endlessly capable, personal, social and psychological habits limit its use. Tools change fast; universal approaches and the state we bring matter more.
What are you feeling?
Insights, or how AI helped me claim my own skills
In Claude Code’s terminal—where I mostly live—there’s a command I love: /insights. It shows my usage patterns, how I can improve, and where I’m already strong.
I sometimes ask colleagues and friends for their files so Claude can compare them with mine and suggest practices to borrow. Yesterday I used it to map my competencies against a specific role.
I saw that Stripe is hiring and asked Claude what the job requires:
Hands-on building: not “I understand what AI can do,” but “I’ve built tools, agents, automations and skills that changed other people’s real work”—and can build alongside someone in real time, not return in two weeks with a solution.
Process discovery: looking at someone’s work and spotting the highest-leverage point. Analytical, not technical.
Coaching through a maturity model: moving someone from AI awareness to autonomy.
Pattern recognition/system thinking: turning what you built for one person into something reusable, not bespoke.
There was more, but the key point: they don’t say “AI expert.” They say, “deep AI practitioner—you’ve lived the transformation you’ll be driving for others.” That’s a behavioural bar, not a knowledge bar: AI has already rebuilt your work, and you can help rebuild someone else’s.
In recent months I’ve built an automated pipeline from collecting papers on AI and therapy to generating copy and carousels; shipped bots across three projects; designed several sites and a personal task tracker. Yet I still don’t feel like a technical person or entitled to call myself an expert.
After Claude mapped the role’s competencies, I realised this is how I want to describe myself: coaching combined with systems and analytical thinking, the ability to build systems and ground them in AI tools.
Then I fed it my Insights file and realised I was drastically undervaluing myself:
Hands-on building: strongly confirmed. My practice is infrastructure-level; by Stripe’s “deep AI practitioner” bar, I’ve lived the transformation with room to spare.
Pattern recognition: stronger than I gave myself credit for. I’ve built a self-improving system: extracting my style rules, creating anti-slop glossaries, and feeding insights back into how Claude works with me. Lena-tov, infostyle, hook-architect, slop-check: not just skills, but meta-work—how I work → rules → a system that improves it.
Process discovery: confirmed by turning 19 survey responses into course-design decisions and audience profiles.
Coaching: I use Claude to analyse my coaching situations—not just coaching people, but using AI as a mirror for the work.
Documentation discipline: vaults, daily tracking, skills as files, glossaries, transition rules. I live in documentation.
I know what I still want to improve, and I may never feel technical. But where I can apply years of practice is now obvious—and that makes me incredibly happy.
How are things going with AI for you?
How to make great design with Claude Code
Don’t write, “You’re Jony Ive,” or “make it look like Apple.”
Instead, give it a ready-made DESIGN.md file that specifies everything people usually describe vaguely:
— colors
— typography
— spacing
— component rules
I found a collection of 2,000+ DESIGN.md files from top websites.
Pick one you like, paste it into Claude or Codex, and get great design.
https://t.co/x5QMw6TLw2
UPD: people in the comments also suggested these options:
https://t.co/RyvV5QKbDY
https://t.co/AWYa5RIwSx