Reflections on 2026, and another year of my life

My birthday, which falls within the last four months of the year, is coming up or has already passed by the time I post this. I’ll be turning either 21, 22, 23, 24, or 25 years old. I do not wish to reveal my exact date of birth, as online privacy is very important to me. Nevertheless, knowing one’s approximate age helps in viewing their statements and actions within better context. LLM-enabled de-anonymization will likely make it easier to guesstimate age. Better to just get it out of the way. I selected the publication date for this text using a random number generator.

I don’t intend to write about my private life here or anywhere in my blog. One reason is the need for privacy; another is that it is likely not particularly interesting to people who do not know me personally (I am, after all, not a “build in public” Instagram Reels influencer who promotes doomed-to-fail SaaS products via undisclosed ad posts). I try to be mindful of other people’s time and circumstances, and I don’t believe that reading about my life is the best way for others to spend their time. Ironically, I actually quite enjoy reading about other people’s lives whenever they choose to share.

Moreover people are quickly forgotten while their work or impact endures. There’s many European scientists who have passed away long ago and whose names are forgotten, yet but what they invented or discovered remains of crucial importance to our modern society. Therefore, it is better to direct my energy in my writing toward the work itself and the insights I have reached, rather than toward myself.

I really resonated with Brennan’s Brown article on “I want to be unremarkable”. While I agree and disagree with him on many other statements in the article and blog, the main point of this one particularly struck a chord with me.

All I will say is I am grateful for my friends and family that surround me. I’m lucky to be here writing this and as I deal with short-term personal problems, such as finding a job, and “long” term societal changes, such as the arrival of recursively self improving machine intelligence, to confront, I take solace in enjoying the here and now. Days like these are good day to reflect on my life so far and receive gifts from friends and family willing to send.

However, I use my birthday (and the New Year’s holiday) to reflect on the world at large, resulting in insights that’s more relevant to the average reader.

State of my projects

I’m currently focused on Yugo. The limitations I encountered with C++ during Aburi development were getting on my nerves. I felt my time and tokens were better spent elsewhere. I’m still committed to C and C++ (still the world’s most consequential languages and the languages in which the Yugo bootstrap compiler are written) but the next time I work on Aburi will be during the Yugo rewrite/remake.

In my announcement post for Yugo I said that one good and useful program in a new PL is better than 100 poorly made vaporware programs. To this extent, I’m building a PC emulator/virtualizer in Yugo. I am well versed in low level programming and it is something I’d use (I currently use Parallels to isolate my computing workflows for privacy/security and 86Box to play old retro games). Progress is excellent (odlično!) and the process is exposing multiple language gaps in Yugo.

The virtualizer development process shows why humans are still needed (for now). For example, LLM came up with the genius idea of stopping all concurrent CPUs (HVF backend) for a device access, which is not how real hardware or other emulators/virtualizers work. This solved a bug the agent was facing right away, but it resulted in serious performance problems later, and I had to tell LLM that this is a bad idea. The Yugo concurrency primitives weren’t as polished at the time, complicating matters.¹

Another instance of humans still needed is when I was working on bringing up the mouse/display for higher resolutions, the mouse on the host suddenly stopped aligning with the mouse on the guest despite absolute positioning enabled. The cause was a mismatch between the Windows resolution and the resolution reported by the virtual GPU. If I prompted and left agent unattended it probably wouldn’t have caught this.

Building performance-sensitive software in newer programming language is harder than it seems because performance issues can be in the generated code or in the Yugolang itself: a distinction that might be hard for models to determine.

State of the industry

Earlier this year, coding agents became good enough for serious work, although it wasn’t immediately widely recognized. In my observation there’s a time lag between when AI is capable of performing certain tasks and when people en mass realize. I expect this gap to narrow over time as more people use AI.

It seemed like a matter of time before events like the frontier labs closing off models and AI risk/safety becoming a major topic of discussion, and lo and behold, both happened. I’ll share my predictions for the future later, but this has undoubtedly been a pivotal year for language modeling.

It seems that entry-level positions in the software industry have practically ceased to exist. The story doesn’t look much better for mid-senior level engineers: talented people such as Jacky Alcine are looking for a job and don’t have one. I’m currently looking for a software job: time will tell how that goes. I had a whole section dedicated to my search but deleted it to maintain the “no personal sob stories” rule.

State of the world

I really don’t see the point of higher education for anything other than social connections (and research with professors) at this point. Teachers use AI to generate presentations, students use AI to do assignments and so forth. I sometimes pity the high schoolers that frequent sites like the ApplyingToCollege subreddit or College Confidential because they get so excited about a world that will no longer exist. The signaling value of top schools like Harvard might count for something: I was visiting a West African town outside the capital (no point in guessing which one it might be) and I was talking to someone there and the topic of American colleges came up. He had only ever heard of Harvard.

If models continue to evolve and take over almost all white-collar jobs, let alone blue-collar jobs, then AI will become the most important political issue, crowding out virtually everything else. AI Alignment/safety and economic impact of AI are going to be front and center. Already economic issues (e.g: no jobs for young people) are the number 1 concern among the older people in my life who don’t believe in agentic AI or don’t understand it.

In my opinion, AI diffusion among certain non-tech industries will be slower than what some people think. For example, telecom companies in US have large amount of unionized workers that I’ll call “pale blue” workers, a mix between white and blue collar workers. These would include fiber and landline technicians that would do anything from programming a fiber/cell tower switch to setting up the wires within and between them. AI could help tune the machines, but there could be resistance from unions. For similar reasons I predict police/fireman/security services will be the last jobs to be replaced with AI, if ever, and healthcare too (but for different reasons). As a reminder most Americans have either not touched AI or use it for basic productivity tasks and can’t fathom the above possibility.

I understand the annoyance some people have with certain AI enthusiasts - someone I know boasts about how great LLMs are and said agents have complete access to everything on his computer including his government ID and other private information, as a privacy person I could never imagine doing such a thing (I use multiple Parallels, soon to be replaced with my virtualizer, VMs to isolate my computing and agentic activities). It’s easy to forget but ultimately, “agents” serve the labs and their creators, not the end user. That being said, people need to internalize that this is the least we will ever hear about AI. There isn’t bubble in the way someone like Ed Zitron would describe it. Each model unlocks new tasks never thought possible in the pre-LLM days.

Predictions

I am no Nostradamus: this is just what I think will happen 2027 and the immediate months before/after. The big caveat is many of these are already in the early/mid stages of happening and my predictions are that it will continue to intensify. With shorter timelines then all this happens and more, with longer timelines then some of this stuff happens but not all.

• AI biology takes off and we begin to see interesting results. We will see cures for diseases drop like solutions to math problems did this year.
  • There’s tremendous financial incentive here because AI labs can use this repair their reputation. In fact, if I remember correctly, some labs said something similar to the investors.
  • It’d be personally welcome, having dealt with chronic conditions in the past. Assuming the cures are readily accessible anti-AI people² will be forced “hand it to them” or “broken clock right twice a day” this one, lest they look extremely stupid.
  • While the concern is real, I don’t see a bioweapon being released by some “vibe kiddie”. It’s more likely that the tools needed to create bioweapons will be much more tightly controlled. Ultimately a solution that doesn’t involve model guardrails has to be found. Otherwise you’re kicking the bucket down the road instead of preventing it entirely.
• AI in programming gets more advanced. What AI can accomplish with one prompt/minimal work increases.
  • We will see ports of large-scale multi-faceted major softwares like Chrome and Linux rewritten in languages like Rust, Go, Zig (or Yugo :) with few or even one prompt.
  • Hardware agentic harnesses will improve and models will get better with hardware/chip design.
• AI safety becomes a mainstream issue. People understand “AI could kill us all” mindset, even if they don’t agree.
  • Many people I know personally had never entertained idea AI could exterminate humanity until the Jacob Coxon situation happened. However, many IRL heard about it weeks after the initial tweet. It’s a good reminder that not everyone’s social media profiles are the same.
  • An “agent swarm” or some other AI mechanism will cause real world damage in a way that will be of significance to the average person (e.g: stole from peoples bank accounts, mislead a human in major ways, cause disruption to infrastructure). This is different from “almost avoided”, humans tend to be bad at taking warnings seriously.
    • Hugging Face drama is still too high level for the average American to understand let alone care.³ However, you don’t need average American to understand, just the people with the most power/money. The important people will still be more moved by an event like this compared to the Hugging Face incident.
    • I personally think a pause or a slowdown to align models and improve alignment techniques like mechanistic interpretability is very much warranted and necessary. In fact that seems to be the prevailing view among people that understand the potential of future AI rather than just seeing it as “just another technology”
• The “ChatGPT/Claude Code moment” that happened to programmers and mathematicians will expand to other fields (obvious), but as AI advances faster, non-technical fields may face a compression of adaptation time compared to software engineering (not so obvious). This is cause for concern as some programmers have cognitive dissonance even 2 years in, let alone non-tech workers who have less time to adapt. The societal and industrial reactions to this shift will be critical to observe.
• China begins to close source their models and be more protectionist. They will still be more open than American labs but less open than before.
  • US and China will remain the only two countries to be relevant in AI. Individual contributions by people from many differing countries will continue.
• Explicit extreme anti-AI sentiment in programming circles wanes exponentially (the inverse curve of AI progress). Those that absolutely can’t handle it just quit, and I’m told some already have. Others will retain vocal opposition, but will otherwise use or tolerate it in their own development life in varying ways. Good-faith criticism and practical complaints (which do exist) will remain: every technology has flaws.
  • When I first released Aburi (the name was formerly Aburiscript), John Regehr of Csmith fame shared the project on Fediverse/Mastodon and almost no comments were about the compiler (Aside from the ast2llvm code which I later deleted for cir2llvm in 0.2, a good portion of the code was hand-written: I started the project before December 2025). In Dec 2025-Jan 2026 I was using the Gemini Code Assist plugin in CLion which didn’t always generate working code (meaning you still had to manually modify and change code) and the plugin interface made you interact with code in ways modern coding harnesses including Antigravity don’t.
  • One of the commenters was commenting on the code being “public domain” instead of Apache/MIT (which in my view is effectively public domain, its just that some jurdisticions dont accept public domain as a license IIRC). I don’t know if he/she/they intended to “sting” me but it doesn’t bother me because what was missed is that now code is cheap, pennies on the dollar (or pesawas on the Cedi) per line quite literally.⁴ AI is capable of doing clean room implementations of closed source software, let alone rewriting a codebase to use a different license if one was inclined to do so. This doesn’t matter because a fully functional C++ compiler can be obtained by downloading MSVC, GCC or Clang (or EDG, which was released as open source recently). Aburi was not and still isn’t better than any of those projects.
  • Most of the commenters were expressing their opinion: that’s fine, and I don’t intend to use this blog as snark for other people or projects. I will directly call out Per Vognsen’s comment about “you didn’t make it” as it’s a mild personal attack, and say he didn’t make Bitwise or any of those games. Clang/MSVC/GCC did and he’s just taking credit for the hard optimization work those compilers did 😂 😂 😂. His perception of the project might partially be my fault: at the time, the documentation in the repo was AI generated and I didn’t have as much material on my website as I do now (and he might have been unfamiliar with how agentic coding actually works), therefore his comment isn’t that big of a deal. For the record I did spent a lot of time working on and learning from Aburi. Yugo is the biggest evidence.
  • Generally I don’t understand why people spend their time working on bringing others down instead of working on their own stuff. There’s a programming language dev who is supposedly a popular figure in the free software community that has a defamatory list of people he doesn’t like and calls them “weird little guys”, and all while having some major skeletons in his own closet (skeletons that are significantly bigger than the ones of any of the people he’s listed). To me it seems only reason he isn’t on his own list is because he made it. But hey, what do I know, I’m just an observer.
  • I also saw some compiler folks criticize “EAs” and “rats” when Jacob Coxon situation happened. In my opinion that’s not a good use of time either.
    • Adapting a quote from The Darkness 2 (one of the best games of the 7th generation, the best video gaming generation, but that’s a topic for a different post): “Go go go! Miscompiles to fix, optimizations to add, ICE (internal compiler errors) to melt!”. No time to intensely focus on things that don’t affect you or me much.
• AI, while trying to complete one task, will solve some other difficult open problem/task along the way. For example, a model may be asked to find a solution to some extremely difficult problem X and somehow manage to solve difficult Problem Y and Z along the way.
  • I already see this in my experience when I ask LLM to solve or add a feature and it fixes bugs along the way.
• Detecting AI output will become more important as agents take on human personas. There will be movement from both the LLM companies (model watermarking) and external actors (AI detectors).
  • On a blog meta note, I do use AI and language translators to clean up writing slightly and remove identifying writing tics that might result in deanonymization. That means pangram or whatever might give some percentage of AI or mixed authorship (My Aburi 0.2 blog post was 30% AI on Pangram IIRC, funnily enough it marked a sentence I did not have AI reword as an AI sentence) Rest assured all the writing originally came from me.
  • In my opinion, LLM watermarking is good as long as it doesn’t trace back to a specific account. I presume there will be/is a way to watermark open models.
• The era of “super-persuasion” and “personal slop” intensifies. People will use AI to create their own entertainment/cinematic/gaming experiences. Loosely related to what i’m doing with my software and the Balkans/Ghana/Turkey theming but it will be in a more accessible format and more people will be doing it. This is already happening to some extent but better models will allow more immersive experience.
  • I am not a fan of AI imagegen/videogen models and if it were up to me I’d send it back to the stable diffusion era or it just doesn’t exist.
  • No human will be immune to this, there will be slop for everyone and everything. And you’ll enjoy it :)
    • For example I thought this “Africa Arise” AI generated-song,this AI generated Instagram reel of Tupac standing with the Serbian studenti, and this Albanian nationalist AI slop Instagram account, to be funny.

Conclusion

I recently came across this lovely Madonna song: “This Used to Be My Playground”. On the surface, the lyrics speak about the past in general terms — but I was struck by how perfectly they capture what the old guard of programming feels about the field’s current direction. Programming (and other fields like math) was never just about results; for many people, it was about the community it gave us. Countless people grew up in harsh real-world environments where computers, the internet, and code were their only source of joy. The early and mid-era internet, and arguably the earliest days of social media, represented a level of unbounded human creativity we may never see again. And now, with AI increasingly dominating the digital world, both directly and indirectly (devouring ever more of the world’s compute), it feels as though the playground that was our pride and joy is being demolished. Yet it’s happening, and we humans must adapt.

2027 will be an interesting year both for me and for the world at large. If you scrolled all the way down: thanks for reading.

Here’s another nice retro-anime song to close out..

Happy early Republic Day to Turkey! (you thought I’d forget?)