Peter T.

Systems engineer × AI builder | Asking the questions Big Tech hopes you won't.

[Peter T. @piotrtam]

https://t.co/6Zc9xXkwgk

[Peter T. @piotrtam]

@ronisarkar_exe https://t.co/8TKtmNC1K6 - AI Native complete digital business cards platform

[Peter T. @piotrtam]

Anthropic just passed OpenAI in business adoption for the first time.

Ramp AI Index (April 2026):
- Anthropic: 34.4%
- OpenAI: 32.3%

One year ago Anthropic had under 8%. They quadrupled.

The scarier number: only 50.6% of businesses use ANY AI.

Half the market hasn't even started.

[Peter T. @piotrtam]

ChatGPT vs Claude vs Gemini vs Perplexity.

The 30-second framework:

Need to create or automate? → ChatGPT
Need to think or build? → Claude
Need to search or verify? → Perplexity
Need to work across Google? → Gemini

Most people need two. Pick the two that match your actual workflow.

#AI #BusinessTools #Productivity

[Peter T. @piotrtam]

AI Is Fast. The Cleanup Is Slow. Nobody Tells You This Part.

Every week I talk to a business owner who says: "We're using AI now. We built in two weeks what used to take three months."

Give it six months. Different tone.

"Nothing works properly. The team we hired to fix it says they need to rebuild from scratch. It's going to cost more than if we'd done it properly the first time."

I've been working with AI since 2021 - before ChatGPT existed, before the hype. Over 8,000 hours across five years. That's four full years of someone's working life. Eight hours a day, five days a week, for four straight years.

The pattern I see repeating across businesses right now is alarming.

The miracle phase

AI is incredible at the start. You describe what you want. It builds it. The demo looks perfect. The speed is intoxicating.

You start thinking: maybe we don't need that senior developer. Maybe AI really can handle it.

This phase lasts about 8-12 weeks. You're ahead of schedule, under budget. You start telling other founders about your AI strategy at dinners.

The fire phase

Then things start breaking. Not in obvious ways. In subtle, expensive ways.

A customer's payment goes through but their account isn't updated. Your team can't figure out why - the code looks correct. Because AI is brilliant at producing things that look correct.

Your team tries to add a feature. It breaks three others. They fix those. Two more break. They're playing whack-a-mole with a codebase nobody fully understands - because AI generated most of it and nobody questioned whether it was actually right.

Why this happens

AI doesn't think. It predicts.

It produces the most plausible-looking output. Not the most correct. Not the most robust. The most plausible.

The difference between "looks right" and "is right" is where businesses lose hundreds of thousands in cleanup costs. "Looks right" didn't handle the edge case where a customer cancels mid-checkout. "Looks right" assumed the database would always respond instantly. When it doesn't - everything downstream fails silently.

A human developer might introduce one or two blind spots per week. AI introduces dozens per day. Each one invisible. Each one a future production incident.

The real cost

I've seen businesses spend $30,000 building something with AI in eight weeks, then spend $120,000 over the next six months trying to fix it.

The AI didn't fail. The process failed. Nobody was controlling the quality. Nobody was stress-testing the output. Nobody was asking "what happens when this goes wrong?" because everyone was celebrating how fast things were going right.

The terms and conditions problem

You know what happens when a website shows you terms and conditions? You scroll to the bottom and click Accept. You don't read it.

That's exactly how most teams use AI.

AI proposes a plan. It sounds confident. It sounds reasonable. So the team clicks Accept. Go ahead. Build it.

Nobody asks: did you consider what happens when two customers book the same slot? Do you even know who our competitors are? Do you understand our business - or are you producing the most generic version of what I asked for?

I push back constantly. In one month alone, I rejected AI's proposed approach 72% of the time. Not small things - fundamental decisions that would have caused serious problems three months later.

Every single one of those rejections felt like slowing down. Every single one saved weeks of cleanup later.

Who on your team is doing this? Who treats AI's output like a first draft from a talented but uninformed intern - not like gospel from an expert?

If nobody is pushing back, everyone is clicking Accept. And you already know what happens when you accept terms you haven't read.

The bottom line

AI is the most powerful multiplier in the history of building software. But a multiplier works in both directions. It multiplies good decisions into extraordinary speed. And it multiplies bad decisions into extraordinary mess.

You wouldn't hand a $2 million machine to an untrained operator and hope for the best. AI is the same. The output quality comes from whoever is directing it. Not from the model. Not from the prompt. From the person who knows what "right" looks like and refuses to accept anything less.

The question isn't whether to use AI. The question is whether you have someone who knows how to control it.

I've been building production software with AI since 2021. 8,000+ hours across five years. 300,000+ lines of production code shipped as a solo founder. Not because AI is magic - because I learned where it breaks and how to stop it before it does.

[Peter T. @piotrtam]

I'm really disappointed @AnthropicAI :( #ai #claude) and #googlegemini did it right!

[Peter T. @piotrtam]

@Prathkum @withneo They only fail as the process of product requirements as well as lack of mudular structure.

[Peter T. @piotrtam]

Hey @AnthropicAI What's happening with your models? I've been running the same prompt template for three months to generate many reports, and they were always perfect. Now, it seems like Sonnet 4 can't handle it and is mixing up data. Yet, I'm running the exact same prompt with Gemini 2.5 Pro, and it handles it just fine. I'm a Max user, and I'm wondering if this is a temporary problem or if I should switch to Google, as Sonnet 4 is no longer reliable. Anyone experiencing similar issues?

[Peter T. @piotrtam]

@manusai Looking good! Great feature! This one saves my day too: https://t.co/gz6fRyQaCy . What was bothering me is that there were no API first markdown to google docs conversion with support for MCP servers, until i found this above. Techically u can connect to anyting , anytime.

[Peter T. @piotrtam]

@RobertLicau Congrats!

[Peter T. @piotrtam]

If any SEOs here I just wanted to share a plugin for for wordpress that allows to edit SEO meta of multiple posts / pages within one screen and Google SERP preview ,something I was missing for many years,so I finally developed it. https://t.co/9fy5ZhYhtr That helps to save a lot of time.

[Peter T. @piotrtam]

Finally! Editing multiple SEO page metas on a one screen - wordpress. https://t.co/Z0yom7i2NO. #seo

[Peter T. @piotrtam]

@georgesttock AI

[Matthew Berman @MatthewBerman]

This blew my mind.

PhD student Jiayi Pan reproduced the emergent “thinking” behavior in a 1.5b model using the DeepSeek R1 technique for just $30.

This means we can give “thinking” to pretty much any model!!

I broke down the findings in my YouTube video below 👇

[Peter T. @piotrtam]

#Bangkok 😂

[Peter T. @piotrtam]

@MatthewBerman Better python environments than next js dependencies! :p

[Peter T. @piotrtam]

@MatthewBerman Yup, basically it's an industry standard.

[Peter T. @piotrtam]

@MatthewBerman VRAM is mainly decided for graphical processing, so if we talk about running normal text models in ram which has speeds as VRAM I'm not sure if there would be any difference.

[Peter T. @piotrtam]

@sengineland @jtabeling Answer Engine Optimisation is on the rise!