The week that started with amazing @VoiceSummitAI finished with meeting the legend @garyvee. It couldn't have been any better ❤️ #VoiceFist is here to stay. Pay attention. Watch people's changing habits. And act to grab the first row seat. #VOICE19 @karol_stryja

Hey Arc users 👋 Are you moving to Dia? Any other alternatives?

I am playing with Rakazo and first impression is that this is a truly impressive Grok Bot alternative

For years, I told companies NOT to build their own business software. Now I’m changing that advice. I’m an old-school dev and architect. Building your own CRM or ERP usually meant hiring a team, funding development, and maintaining it for years. For most companies, buying specialised software was the reasonable choice, even if it meant bending their processes to fit. A subscription was easier to justify than a team of developers. AI agents are changing that calculation. Clickable prototypes can be built in hours and validated with users within days. That makes it much cheaper to find out whether custom software is even worth pursuing. Getting it ready to run a business still depends on the scope, integrations, and reliability you need. But we are still taking weeks, not months. I believe we’re heading towards hundreds of thousands of custom CRMs, ERPs, each one fitted to a single company, with agents helping to build and maintain them. What interests me most is that companies can finally afford software that fits how they work. Without having to ask people to change how they work. Yet it is still best you have someone who understands the system and takes responsibility for it. Security, permissions, deployment, backups, and maintenance all need attention. Agents can help with most of that, but someone has to know whether they’ve done the job properly. This isn’t the future. It’s here. Now. Which team are you on? BUILD or BUY?

Simplified Technical English was built for aircraft mechanics. Not for sales decks. ASD-STE100 is the aerospace standard for maintenance manuals. Written for people who may read English as a second language, next to an engine, with no time to guess. One word, one meaning. Short sentences. Active voice. It's great for technical docs. Turns out it's just as good at explaining a business. I gave Claude Code and Codex the same short prompt: use the ASD-STE100 rules to explain dyrektor.ai to a board member of a small or mid-sized company who has never heard of it and isn't technical. Use only the public website. Both came back with an interactive HTML explainer. No copywriter, no agency, no brief. Then I had the AI read it as a persona: a board member of an 80-person manufacturing company. Three minutes, on her phone, between meetings. Would she book a call? "Probably yes, but not today and not by myself." She'd forward it to her operations director. That one answer told me more than an internal review would: → we may be writing for the wrong person → it wasn't clear you get a person and a whole team → she didn't know "Shadow AI" and found the definition only at the very bottom → she wanted a manufacturing example (which we have to add for real, not make up) Last step: turn the explainer into a slide deck with speaker notes and a PDF. Claude Code (with Opus 5.5) was faster and far more interactive. Codex (GPT-6.1 Sol) said the key thing more clearly. Neither was ready to ship without a few review cycles with a real person. Give it a try on your business, product or service and let me know in the comments if that works for you.

One minute of a 100,000 tok/s model will cost $1,800 at today's prices. It will write ~600,000 lines of code. You'll need about two months to read the output of that one minute. So you won't. Nobody will. Bear with me for this Sunday special. Now that models are getting similar in terms of coding ability, speed becomes the differentiator. And it has a price tag. OpenAI's new Ultrafast tier: 300 tokens per second, 6x the standard price. Anthropic sells a fast mode too. We're used to paying for intelligence. Now we'll also pay for time. I know, 100,000 tok/s sounds crazy. But a hardwired Taalas chip already does 17,000 on a small model. So will creating software get 1000x faster? No. Everything a machine can check gets fast. Analysis, prototypes, code, contract drafts. Everything a human has to trust stays slow. The customer's attention, 30 days of real use, the regulator's queue. Realistic gain 2–10x? Not 1000x. What does change? Nobody reads the code. Like assembly today. The difference is that a compiler gets an exact language. A model gets a spec if lucky. Ok, so code becomes disposable. But data, contracts and the audit trail do not. At that speed every model will run through enough data to know what to build. So what is scarce at 100,000 tokens per second? Choosing. Knowing what not to build. Saying no to 49 good ideas. And signing for the one you ship. How do you see the world at these speeds?

ASD-STE100 is one of the wildest things I learned about in the recent weeks Below is a screenshot of an animated explainer of how Shafts works. It is a desktop application for macOS I created some time ago. It controls SSH tunnels.

Ile czasu zajęło Ci zrobienie ostatniej oferty firmowej? Nie samo pisanie. Szukanie poprzedniej wersji w mailach, przeklejanie logo z prezentacji sprzed roku, tabelka, która uparcie wyjeżdża na drugą stronę. U większości firm, które znam, to parę godzin, a czasmi nawet dwa, trzy dni. Za każdym razem od nowa. Nagrałem wideo, w którym pokazuję, jak zrobić to raz. Bierzemy zwykłą, istniejącą stronę internetową i krok po kroku robimy z niej: → brief marki (kolory, logo, opisy) → szablon oferty → gotową ofertę w PDF → w bonusie: papier firmowy i przykładowe pismo Wszystko na moim ekranie, na prawdziwym przykładzie - oferta partnerska dla Civil42 i Fundacji Instytut 42. Potrzebujesz tylko adresu swojej strony i asystenta AI, który tworzy kod - ChatGPT w trybie Codex albo Claude Code. Zachęcam żeby spróbować na swoim przykładzie w miarę oglądania. A jeżeli potrzebujesz czegoś więcej zapraszam na dyrektor.ai

Is it just me or is @bot not working for others as well?

Oh, hello Jev!

I worried I might have been too harsh on Grok Bot yesterday, but today just reconfirms that using it for business might be a bad idea. Today all bots stopped responding with no error messages across both desktop and mobile apps. No issue if this delays my calorie journal, but if I depended on it for business - a whole other story. *this is not a brief issue - over 30 mins now, no response, but I can connect to the bot computer today. Do you also experience issues? What's your take on using it for business?

I love Grok Bot. I also wouldn't trust my business with it. I moved to Grok Bot from Hermes a while ago for one reason - it was much nicer to use. Both across the desktop and mobile app. Nothing to babysit and no configs to fight. You say what you need, and it gets done. That same simplicity is the problem. Tonight my Grok Bot lost the connection to its computer, and my bots stopped working. "Can't reach Bot's screen." Retry. Retry. Retry. Nothing. The only way forward was a button labelled "Recover." But as you recover, every app is gone. That's a rebuild. And that's still not a bad outcome. But what if recovery wasn't available for a while? We used to say "not your keys, not your coins." With hosted agents, "not your computer" goes further. It's not just someone else's machine. It's a machine you can't see, can't reach and can't repair when it breaks. I'll keep using Grok Bot for experiments and personal productivity. But for anything a client or a revenue stream depends on, I need something reliable and under control. Convenience is a feature. Control is a requirement. Where do you draw that line with AI agents you use for your business?

Grok Bot works too damn well...

Kto w Twojej firmie odpowiada za AI? Jeśli odpowiedź brzmi "trochę wszyscy", to znaczy, że nikt. Dziś o 18:00 razem z @r_garbacz opowiemy na webinarze, czym jest Zewnętrzny Dyrektor AI i dlaczego może być dość przydatny. Szczególnie razem ze swoim zespołem. Zapisz się tu: luma.com/g0sqzydo

I'm 45 just a guy in love with technology from Poland 🇵🇱 building external AI director service, brand safe AI agents and distributing live broadcasts all over the world looking to connect with anyone aiming to make a change

Grok Bot completely replaced Hermes for me. All I need it to be able to run and orchestrate pi for more complex tasks or write code on a home server. Connected via Tailscale. This is becoming too easy :)

Who gets fired if AI goes wrong? Silence. Almost every time. That silence isn't a governance gap. It's the actual reason the pilots don't ship. The technology has been good enough for a while now. The signature hasn't shown up. Nobody wants their name on a system they can't fully explain, doing work they can't fully check, at a speed they can't fully follow. So the pilot gets extended. Then extended again. Most innovation theatre is just unsigned paperwork. Here's the part I think we're collectively getting wrong. Agents are already building software. And they are going to absorb responsibility fast. Spot the fault, fix it, file the report, decide the refund. Responsibility is a mechanism, and mechanisms automate. Accountability won't. Accountability is a relationship between a person and a consequence. It needs someone who can be argued with, overruled, sued or fired. You cannot subpoena a model checkpoint. So as the work gets cheaper, the signature gets more expensive. Not the code. The name at the bottom of it. Every serious profession worked this out a century ago. You don't pay a structural engineer for the calculation - a spreadsheet does the calculation. You pay for the stamp, and for the fact that they can lose the right to make it. Software never had that. But we are placing a bet that it's about to need it. Which is roughly the direction we are taking with dyrektor.ai - External AI Director with a team. Someone from the outside who sits on the inside, puts their name on the AI decisions, and says plainly every month whether to scale it, fix it or kill it. I tried to describe it. Give it a name. It still sounds slightly absurd to me. Accountability as a Service. Our industry has spent years selling tools and calling it transformation. I believe that era is ending. Tell me why I'm wrong.

Software dev is officially a commodity. Work well my friend while I go and rest… Thanks for Limen @_overment