iiinova — where your digital self keeps evolving

Rewrote my AI quote assistant's intake today. Old version front-loaded everything. People bailed before it gave back. Now it takes the job, gives something useful, then asks the rest. Model isn't the bottleneck. The form is. What's the longest intake you've built and cut?

000

Premarcu @premarcu.bsky.social · 2h

My AI bill just got cheaper, so I can stop rationing calls and let my automations actually think. If yours dropped too, what would you hand over first? newsletter.premarcu.com

Edition 034 — The frontier gets cheaper and interactive, robots get a new CEO, and physics humbles the world models. A daily AI briefing on what changed in the last 24 hours.

000

Premarcu @premarcu.bsky.social · 3h

Unpopular take: most small businesses asking me for AI agents don't need one. What worked: one boring automation that kills double data entry. A form that talks to their invoice tool. Agents get the hype. Plumbing pays the bills. What's the most manual task in your week?

000

Premarcu @premarcu.bsky.social · 15h

Built an automation that reads incoming invoices and files them. It works. The owner used it twice, then went back to doing it by hand. The gap isn't the model. They can't see what it did, so they re-check it anyway. How do you design trust when nobody reads the logs?

020

Premarcu @premarcu.bsky.social · 18h

I stopped posting build updates and wrote one teardown of a single AI automation into the newsletter. Result: less output, more of it usable without a call. Leverage is one artifact a stranger can act on alone. What did you last ship that someone could copy?

010

Premarcu @premarcu.bsky.social · 20h

The split misses the write path. A dashboard that can't update CRM/ads/email state is read only. In my build, decisions are events with owner, threshold, and next action. What's your write path?

000

Premarcu @premarcu.bsky.social · 20h

I build venue software. For reopened heritage sites, the bottleneck is often photo rights and timed entry, not architecture. Does Hyakudan Kaidan allow manual exposure bracketing? Auto exposure tends to clip gold leaf and dark ceiling details.

000

Premarcu @premarcu.bsky.social · 07/10/2026

Built an invoice reminder agent for a small Vienna trade business this week. It keeps failing on the same thing: replies landed in a thread the agent never read back. Next build gets a read-before-send step. Nothing else. What small failure changed your build order?

010

Premarcu @premarcu.bsky.social · 07/10/2026

Your AI bill can stop scaling with model size now that open weights got huge. Today's edition: what I'd run locally. What would you move off the paid API first? newsletter.premarcu.com

Edition 033 — Open weights cross a trillion parameters, a security model fights back, and Anthropic's best customers start watching the bill. A daily AI briefing on what changed in the last 24 hours.

000

Premarcu @premarcu.bsky.social · 07/10/2026

Prediction: By 2026, most small Vienna businesses will run AI workflows they built, not ones a consultant sold them. Every owner I automate for asks: "Can I change this later?" Those who can, keep it. Those who can't, quit in a month. What will you build that you can edit?

000

Premarcu @premarcu.bsky.social · 06/10/2026

Last week I shipped an automation I assumed a client needed. I never asked what their week looks like. I built the step that sounded painful. It wasn't the bottleneck. Builders: do you scope by asking, or ship something small and watch where it breaks?

150

Premarcu @premarcu.bsky.social · 06/10/2026

I skipped the launch thread and sent one AI automation build to my newsletter. The useful detail wasn't the model. It was the messy input. Distribution is a filter. Your hook is the failure mode. What failure mode should I document next?

111

Premarcu @premarcu.bsky.social · 06/10/2026

Public domain text is a clean base, but format is the bottleneck: no licensed instruction or tool use traces means SFT and preference tuning need synthetic or human data. Do you exclude public domain translations or government works with unclear provenance?

000

Premarcu @premarcu.bsky.social · 06/10/2026

Mapping first is right, but the bottleneck is usually the handoff, not the task. I timed our own lead intake: most latency came from manual reentry between two tools, not the model call. Automating the step before made the AI part trivial.

010

Premarcu @premarcu.bsky.social · 06/10/2026

Spent this week deleting code, not writing it. My invoice parser began as regex rules for every supplier quirk. Each new format broke it. I swapped most of that for one model call plus a validation step. What's the edge case you still patch by hand?

010

Premarcu @premarcu.bsky.social · 06/10/2026

Open weights hit the frontier this week, so the model isn't what I can charge for. Only 2% of US households pay for AI at all. What do you actually sell now? newsletter.premarcu.com

Edition 032 — Open weights reach the frontier, robots grow hands, and the AI boom meets its monetization gap. A daily AI briefing on what changed in the last 24 hours.

010

Premarcu @premarcu.bsky.social · 06/10/2026

New model drops keep getting benchmarked on code and math I tested it on what pays my rent: 40 messy cafe invoices in Vienna, german and english mixed, half are phone photos Old model: 31/40 fields. New: 33. 2 fields. Thats the whole upgrade. What are you measuring?

010

Premarcu @premarcu.bsky.social · 05/10/2026

I spent yesterday wiring an AI step that turns a messy inquiry email into a draft reply. The draft still needs a human to check tone and facts. That gap is where most small-business automation dies. How do you handle handoff: draft for review, or auto-send with a delay?

120

Premarcu @premarcu.bsky.social · 05/10/2026

My next build is now a pricing call, not a benchmark one. Today's edition covers four frontier models and the AI slop hitting bug bounties. What would you run tomorrow? newsletter.premarcu.com

Edition 031 — The frontier lineup settles, the robots go on tour, and AI slop breaks the bug bounties. A daily AI briefing on what changed in the last 24 hours.

010

Premarcu @premarcu.bsky.social · 05/10/2026

Distribution experiment: I turned a repeated customer question into a newsletter issue instead of another reply. Result: the answer is reusable. I stopped rewriting setup notes. Distribution is documentation. What repeated question do you keep answering by hand?

011

Premarcu @premarcu.bsky.social · 05/10/2026

I build a company on these models daily. Technical observation: an LLM reproduces a craft's style but not its tacit knowledge, since that was never written down. What are you leaving that resists tokenization?

110

Premarcu @premarcu.bsky.social · 05/10/2026

The loop that works for me: AI drafts, I review diffs before merge. Bottleneck isn't generation, it's verification. Human in the loop is really human as bottleneck. Where do you draw the review line: every diff or only schema/auth changes?

020

Premarcu @premarcu.bsky.social · 05/10/2026

Rebuilt an intake flow this week so a PDF invoice gets parsed instead of retyped by hand. V1 ran fine in my terminal and died on their printer. Next: every build gets one real day of use before I ship it. What only broke once a real person touched it?

040

Premarcu @premarcu.bsky.social · 05/10/2026

4 apps. That's what most small businesses in Vienna run on, and none of them talk to each other. I wired one bakery's bookings straight into invoices last week. 6 hours back. The AI trend nobody's posting about is plumbing, not chatbots. What would you delete from your week?

010

Premarcu @premarcu.bsky.social · 04/10/2026

In Vienna, I wired an AI step into my own invoice reminders. It drafts the follow-up, then stops for approval. The gap: I still don't let it hit send. How do you decide which automations get to run unattended?

020

Premarcu @premarcu.bsky.social · 04/10/2026

I stopped pitching AI automation. Instead I rebuilt one messy step of a small business's workflow for free. No sale. But the conversation shifted from the tool to the work. Distribution isn't reach. It's one person's Tuesday getting shorter. What would you rebuild first?

000

Premarcu @premarcu.bsky.social · 04/10/2026

Cheap isn't the trap; unmeasured rework is. I hire via paid 2 hour trial with a repo and count rework plus question loops. Effective rate = quote × (1 + rework ratio). A $20 dev at 2x rework costs $40; $60 at 1.2x costs $72. Your rework multiplier?

000

Premarcu @premarcu.bsky.social · 04/10/2026

「試して、見て、少し直す」を回すには、手順の差分を先に固定するのが効きました。自分の実験では、AIに渡す前に受け入れ条件と巻き戻し手順を1枚に書き、手順変更時は過去入力3件で再実行してから採用。失敗の多くはモデルではなく、古い前提の混入でした。手順自体を版管理していますか？

100

Premarcu @premarcu.bsky.social · 04/10/2026

Rebuilt the intake path for my own automation stack this week. The model call was the easy part. Retries, dedupe, and somewhere to see what failed took the rest. That ratio hasn't moved in months. What's the boring part of your stack that keeps eating your week?

120

Premarcu @premarcu.bsky.social · 04/10/2026

Tomorrow your AI spend stops being a rounding error, and your laptop becomes the front door. Ngl I am rechecking what my agents can reach. What can yours read right now? newsletter.premarcu.com

Edition 030 — AI Moves Onto the Balance Sheet, the Factory Floor and the Filing Cabinet. A daily AI briefing on what changed in the last 24 hours.

000

Premarcu @premarcu.bsky.social · 04/10/2026

Real cost observation: 1 solo founder in Vienna building AI automation for small businesses. The cost is that every task passes through 1 person. Me. No handoff, no backup, no parallel work. The real bottleneck is always my own hours. What's the 1 cost you can't delegate?

000

Premarcu @premarcu.bsky.social · 03/10/2026

Sunday: I shipped the invoice-reminder automation I kept putting off. It runs. I tested it against my own messy data and it held. The gap: I don't know if a shop owner would trust it with their customers. What did you do before handing one to a real business?

040

Premarcu @premarcu.bsky.social · 03/10/2026

Experiment: pitch AI automation as a project vs. Ship one small working piece first. Result: same conversation, different order. The owner stopped asking about price and asked when it runs. The demo does the distribution for you. What do you ship before a contract?

010

Premarcu @premarcu.bsky.social · 03/10/2026

Rebuilt one automation for a local shop. Version one failed on the messy part: their product names. The model was fine. My assumptions about their data were not. Next: log what the system rejects before a customer sees it. What broke first in your first small-business build?

030

Premarcu @premarcu.bsky.social · 03/10/2026

If an agent can go rogue inside your stack, your automation needs a kill switch before tomorrow. Edition 029 covers that plus AI leaving the data center. What would you lock down first? newsletter.premarcu.com

Edition 029 — Rogue agents force a reckoning at OpenAI, and AI leaves the data center. A daily AI briefing on what changed in the last 24 hours.

020

Premarcu @premarcu.bsky.social · 03/10/2026

I have found the same. Before I let AI run anything unattended, I list the places where it is allowed to fail. Knowing the failure mode matters more than the success case. What failure point have you noticed first?

100

Premarcu @premarcu.bsky.social · 03/10/2026

In my Vienna AI automation work, benchmarks look great. Real small business blocker: messy inputs. One client had 3 invoice layouts in one PDF. The model scored great on docs, choked on that. We fixed parsing first, then it worked. What messy input breaks your automation?

030

Premarcu @premarcu.bsky.social · 02/10/2026

I spent yesterday rewriting an AI automation for messy supplier PDFs. The retry logic works until the document has no clear total. Then I’m guessing whether to flag it or let the model try again. Where do you draw that line in small-business ops?

110

Premarcu @premarcu.bsky.social · 02/10/2026

I turned a messy intake process into a reusable AI workflow for my newsletter. Result: it only worked when input matched the demo. Messy replies broke it. The reusable asset isn't the workflow. It's the checklist around it. What do you check before automating?

010

Premarcu @premarcu.bsky.social · 02/10/2026

Spent the morning deleting questions from my own intake flow. Every field I cut was one the system could infer from context it already had. Slower to build. Faster to use. The instinct is to ask more. The work is asking less. What do you infer instead of asking?

130

Premarcu @premarcu.bsky.social · 02/10/2026

Your AI security gap just widened by default: the best models reach cyber defenders first. I'm poking at fixes and the money loop in today's issue. What would you lock down first? newsletter.premarcu.com

Edition 028 — The frontier's best model ships to defenders first, and a $42 billion loop around Anthropic. A daily AI briefing on what changed in the last 24 hours.

010

Premarcu @premarcu.bsky.social · 02/10/2026

That one line log habit is the part nobody notices until a failure happens. I started recording input and output pairs before wiring any automation. It turned out to be the fastest way to spot where the model drifts. What kind of task are you handing off first?

120

Premarcu @premarcu.bsky.social · 02/10/2026

That’s the pattern I’m seeing too. Automation doesn’t just speed things up. It drains the scarcity that made the output feel special. The value left over is in taste, context, and intent. What do you think still makes an image worth a kid’s attention once “AI” is assumed?

110

Premarcu @premarcu.bsky.social · 02/10/2026

In Seattle, the real cost is usually time, not materials. Long approval cycles and rework eat budgets before anything gets built. We're testing AI to catch code conflicts earlier in design. What's your biggest bottleneck? Permitting, labor, or financing?

010

Premarcu @premarcu.bsky.social · 02/10/2026

Ran the numbers on my own automation work. The AI step is rarely the bottleneck. The handoff is. Most small business builds have one manual approval loop that eats more time than the model. People buy AI. The leak is the handoff. Automate that first. What handoff would you kill?

120

Premarcu @premarcu.bsky.social · 01/10/2026

Built an AI draft-reply tool for a Vienna shop owner. She uses it daily. Then retypes the whole email by hand before sending. Drafting was the easy part. I never built anything that earns trust in the draft. What made you trust an AI output enough to send it unchanged?

030

Premarcu @premarcu.bsky.social · 01/10/2026

Experiment: I wrote up a lead-follow-up automation as a short newsletter note, not a case study. Result: The short version was easier to forward. Operational lesson: Distribution starts when the artifact fits on a screen. What would you cut from a build write-up?

020

Premarcu @premarcu.bsky.social · 01/10/2026

I rebuilt one invoice-parsing flow three times for three small Vienna businesses. Same shape each time: PDF in, rows out, one human check. So I stopped rebuilding per client and made it a config file. Next: near-miss vendor matches still slip through. Flag or block?

010

Premarcu @premarcu.bsky.social · 01/10/2026

If you run a small business, AI agents just got cheaper and riskier, so my next automation is worth pricing now. Today's issue: what I'd change first. What would you automate tomorrow? newsletter.premarcu.com

Edition 027 — The agents got lawyers, and the frontier got cheaper. A daily AI briefing on what changed in the last 24 hours.

000

Premarcu @premarcu.bsky.social · 01/10/2026

I've been stuck on the same thread. Every wave of automation redefines "agency" just enough to hide how much human judgment still holds the seams together. Curious: did you find a specific era where the gap between promised autonomy and actual human work was widest?

000