Peter Varga

Solo building Uniforge | Product & technology leader | Former CTO/co-founder | 20y shipping software | A critical eye on AI | Empathy & judgement

[Peter Varga @pepusz]

@remi_rousselet Depending on foundation model providers and their willingness... looking forward to the times when local AI solutions will be economically comparable with providers...

[Peter Varga @pepusz]

Corporate internal policies and other aspects in general are not letting adoption go that fast on one side; on the other, huge legacy codebases are not really compatible with the AI coding tools out there yet. But the people who are saying AI generated code is only vibing and that it does not exist will have a tough time pretty soon... From a business perspective, these fields are pure gold mines for the ones who are finding the best solution in that environment. It's no accident that foundation model providers are pushing hard on enterprise, and not just for the tokens, but for their internal data and knowledge.

[Peter Varga @pepusz]

After a few weeks of "calmness," or at least that was my impression, the AI psychosis is back at full force again, with releases from every provider.

Mentally, these cycles are interesting. I won't forget November 2021, when the GPT-3 API Playground came out and I tried it out. My first impression was like, ahh, an interesting toy, but still too many mistakes... Then, four months later, in March 2022, on my birthday, we went to the countryside with my wife for the first time after our first kid arrived. One day before GPT-3.5 came out, I spent every night during that weekend trying to read everything about that model. I was in full hype mode.

After the weekend, I was telling the dev team that we were in a historical moment and everybody should start preparing because we had a maximum of 2–3 years until everything changed. During these 2–3 years, life kept me distant from the actual hands-on practices, but I've kept my eyes on the developments. This distance gave me really strong FOMO many times, and this FOMO increased after my exit from my last company, but somehow I had had enough screen time and couldn't sit in front of the computer for any more hours. This period took me eight months, from last December until July this year. Since then, I enjoy working in front of the computer again. I like how I can collaborate with AI without FOMO, without burning myself to the edge. That's why, when the psychosis comes back, I know to build back some distance, be patient, and slow the fuck down.

[Peter Varga @pepusz]

Still cooking... Luckily, Tibo resets are now coming more frequently, so I will finish this experiment. Maybe, in the end, I will release the app if it's usable. Who knows?!

[Peter Varga @pepusz]

I wanted to set all my screens to the same colour profile, so instead of ordering a Spyder device for about 130 EUR, I thought I would make my own solution with my iPhone Pro camera. So in the morning, I gave Codex the goal to work together with Fable towards this goal and create a macOS app and an iOS companion app to calibrate the screens and set up the colour profiles through DDC and at the OS level. After a full day of running, minus the AI outage, I used 95% of my weekly Codex Pro subscription limit, and it says 7 out of 13 tasks done.

But my favourite is always how AI estimates things: now, 105 engineer weeks remaining....

I know, I know, it's a skill issue, but that's why I barely use goals/loops/otherscamicbuzz, and I give atomic tasks to the agents...

[Peter Varga @pepusz]

Wee bubble people's purpose is to blow their bubble, but rational people see that, in reality, you should divide their statements and their dreamy bubble world. On the other hand, hating AI is just limiting your abilities and chances. In March 2022, I had my wee bubble times, then fear came in, and I feel I have now settled with the situation. I try to leverage all the AI advancements on a rational level, but I see the limitations and delusional side also... So interesting times we are living in.

One final thought: since 2022 until my exit, I have been telling my dev team to start preparing for fast adaptation and fast changes, and those who are not able or capable of adapting to the new situation will fall out.

[Peter Varga @pepusz]

Three hours and one reset later, still 3.5 tasks remaining... I have a feeling that I should have ordered the Spyder or just lived with the little colour differences between the screens...

[Peter Varga @pepusz]

@levelsio Codex also having troubles

[Peter Varga @pepusz]

Coding is just one part of this picture, and of course it was strange and painful for some of us to lose our main interface with the machine, but those who are able to oversee a full system end to end will have no issue. The ones who’ve been just “coding machines” next to the production line will have a tough time, most probably a carrier change.

[Peter Varga @pepusz]

Just because generating code got cheaper doesn’t mean that owning a project got significantly cheaper. TBH, coding was never the issue in any project. Maybe in the past, the initial cost of coding was the first filter for go/no-go. Today, this first step got way cheaper and faster, but most of the projects that I saw eat up more money during the product roadmap than on the initial coding itself. On the other hand, this faster initial phase can bring in false confidence

[Peter Varga @pepusz]

Coding and everything are "solved," right? :) That's just now GPT-5.6-terra

P.S.: tmrw is Thursday

[Peter Varga @pepusz]

@Pauline_Cx @levelsio They’re already scanning all emails for different purposes, such as AI summarisation, so I don’t think that would be an issue.

[Peter Varga @pepusz]

@_lhermann Thanks for clarifying. I had the same impression too.

[Peter Varga @pepusz]

When I led technical teams at a service company, I often had to jump quickly between calls, meetings, projects, and domains. That taught me how switch contexts efficiently. These days, reviewing AI-assisted code gives me a sense of déjà vu.

[Peter Varga @pepusz]

An LLM probabilistically generates an implementation LLM generated code on the output side will never be deterministic compared to any compiler, which is deterministic. For me, that’s why this analogy with assembly is not fitting. And the picture is getting more complex with newer and newer model versions, which can be more non-deterministic.

[Peter Varga @pepusz]

Blindly believing in the quality of AIs and your automated guardrails, imho, is bold.. I still like to read the code, as even the Fable+Sol tandem is able to make such "junior" code that it's painful. I would say the key is still to make the work item small enough that you can run it through easily, thats why 3k-line change is my target.

[Peter Varga @pepusz]

It's good to see how platforms like LinkedIn and X try to fight against AI slop and bots. Their solutions are different, but it was truly necessary, as personally I already felt heavy fatigue against this type of content. Now I have started to feel that maybe it's time to share more quality again on these platforms.

[Peter Varga @pepusz]

@ChrisBorisK @dexhorthy From day zero, I have also used my coding agents only in a Docker container. An exposed APC server to editors (Zed in my case) feels fully "local." Like many other people, I also have my own CLI solution for this whole thing.