AlphaEvolve is an evolutionary coding agent
That you may have heard about, but here's my take
Hey Futurists!
The topic is the topic that's been on my mind all the time, I don't know about you, is: what is left before machine intelligence reaches, and surpasses, human intelligence? How much time is there before that happens? What are the unsolved problems in artificial intelligence that remain to be solved that are preventing it from happening already?
AI
1. AlphaEvolve is an evolutionary coding agent. That I'm guessing most of you have already heard about by this point because it's been all over the various AI-related news sources, so what I'm doing here is basically giving you my take, for what it's worth. So I read the AlphaEvolve paper, and, while it's not a classical genetic algorithm, it's obviously inspired by the same way of thinking.
In order to create computer programs that can evolve on their own, you need a source of randomness (called mutation in classical genetic algorithms, but I'm going to just say randomness), a recombination system, and a selection criteria. The way this is accomplished here is, first of all, the programs are modified by LLMs, with you, the human, required to provide the initial program, and there is inherent randomness in LLM generation, so that's one source of randomness. There's an additional source of randomness, which is the process of selecting previous programs from the population of existing programs to evolve from, and how to generate new prompts for them. The selection criteria is provided by you, the human, in the form of an evaluation function, which can return multiple numbers -- what's that about?
The process of selecting previous programs can use multiple ranking numbers. For this, the researchers say they used a combination of algorithms, without explaining anything about them, except to provide references to the algorithms. So I had to track those down. One of them is called Multi-dimensional Archive of Phenotypic Elites (MAP-Elites). For the other, they just say "island based". So going back to the first one, there's an algorithm called MAP-Elites, and it's not even a new algorithm -- it's been around since 2015, I just never heard of it before. I'll try to summarize the basic gist of the idea. Instead of ranking based on a single number which you try to maximize, you have multiple numbers, and you start off coarse-grained and subsequently subdivide the space. So let's say you have a robot simulator, and you have numbers for speed, height, weight, and efficiency. Instead of trying to combine those into a single number, you feed them all into MAP-Elites. It treats it as a 4-dimensional space, starts off with coarse-grained divisions, and progressively makes finer and finer subdivisions into that 4-dimensional space. It incorporates randomness in this process as well. So in this way, you can find a good robot design that is fast, tall, light, and efficient (assuming you want to maximize all those numbers -- if you want a short robot you can flip that number). MAP-Elites can discover relationships between the numbers, for example if tall robots are usually fast it will notice that.
As for "islands", this is based on the idea that if you break a population into separate subpopulations, like Darwin's finches on the Galapagos Islands, hence the term "islands", then the subpopulations can evolve in different directions and optimize different things. This helps prevent the system from getting stuck in a "local optimum" that is not the "global optimum".
The islands technique comes from the previous program, so here's where it might be good to mention that AlphaEvolve isn't the first of its kind, but an attempt to outperform a previous program, called FunSearch, made by a different team (not DeepMind), and FunSearch used the "island" technique. AlphaEvolve is basically FunSearch on steroids: FunSearch evolves a single function, while AlphaEvolve evolves entire code file. FunSearch evolves up to 10-20 lines of code, while AlphaEvolve evolves up to hundreds of lines of code. FunSearch evolves code in Python only, while AlphaEvolve evolves code in any language. FunSearch needs fast evaluation (less than or equal to 20 minutes on 1 CPU), while AlphaEvolve can evaluate for hours, in parallel, on Google's TPU accelerators. FunSearch generates millions of samples from LLMs, while AlphaEvolve only needs mere thousands of LLM generations. FunSearch requires small LLMs be used and does not benefit from larger LLMS, while AlphaEvolve benefits from state-of-the-art LLMs (this being DeepMind, owned by Google, all the LLMs they used were Gemini models). FunSearch relies on only minimal context (only previous solutions), while AlphaEvolve employs rich context and feedback in its prompts. And finally, FunSearch optimizes only a single metric, while AlphaEvolve can simultaneously optimize multiple metrics, which is what we've just been talking about.
So the way the process works is more or less as follows: First you sample the database of existing programs using your MAP-Elites and "island" algorithms. You take those and ask LLMs for ideas on how to improve the programs to increase the score on the evaluation function. Then you take those programs and the ideas on how to improve the programs and prompt the LLMs to actually change the programs. These changes are expressed in a simplified "diff" format, if you're familiar with "diffs" and how they express the differences in program files. Each result obtained in this manner is applied to the programs, generating new "child" programs. These "child" programs have to be evaluated, so they are passed to the evaluator function that, remember, you, the human, have to provide. These programs, along with their evaluation results, are then added to the database of programs, and the cycle repeats.
At this point you may have noticed that what gets input into your evaluation function is a program. This program can express the solution to the problem you are trying to solve more-or-less directly, but it doesn't have to. It can be a program that *constructs* a solution, and then that solution is what you ultimately evaluate. Furthermore, the best way of constructing solutions may vary widely depending on what type of problem you are trying to solve. Some problems may favor a concise construction from scratch, while others may "construct" a solution by algorithmically searching a customized search space.
At this point you might be wondering what problems they solved with this system to demonstrate it. Their biggest claim to fame is inventing a bunch of new matrix "multiplication" algorithms. I put "multiplication" in quotes because matrix "multiplication" isn't multiplication, like what you learned in elementary school, it's something entirely different, that just happens to be called "multiplication" for some reason I don't know. Math and science are full of confusing terms. That's just how life is. Anyway, AlphaEvolve came up with 14 new matrix "multiplication" algorithms, and one of them, the 4x4 matrix "multiplication" algorithm, does its thing with only 48 multiplications (actual multiplications this time -- no quotes). The naïve approach that you would use if doing it by hand would be 64 multiplications. The best known algorithm before AlphaEvolve, the Strassen algorithm (named for Volker Strassen) from 1969, requires 49.
The team didn't stop there. They write:
"To assess its capabilities, we apply AlphaEvolve to a curated set of over 50 mathematical problems, spanning more than five different branches of mathematics, including analysis, combinatorics, number theory, and geometry, evaluated across numerous specific parameter settings (e.g., different dimensions or sizes). In 75% of the cases AlphaEvolve rediscovered the best known constructions, and in 20% of the cases it discovered a new object that is better than a previously known best construction, thereby improving the state of the art. In all these cases, the initial starting point was a simple or a random construction."
Under "analysis" (by which I assume they mean "real analysis", i.e. calculus-type problems), they cite "autocorrelation inequalities." These look like integrals with inequalities like "less than" or "less than or equal to" instead of "equals". They're called "autocorrelation" because these equations come from signal processing and they relate some aspect of a signal to some other aspect of the same signal. They say, "AlphaEvolve was able to improve the best known bounds on several autocorrelation inequalities."
Under "uncertainty principles", they say, "AlphaEvolve was able to produce a refined configuration for a problem arising in Fourier analysis, by polishing an uncertainty principle construction, leading to a slightly better upper bound."
Under "combinatorics and number theory", they cite Erdős's minimum overlap problem. "AlphaEvolve established a new upper bound for the minimum overlap problem, slightly improving upon the previous record."
Under "geometry and packing", they cite the "kissing number problem". This is about spheres that are said to be "kissing" if they touch at a point, and the idea is to have one central sphere and ask how many spheres can "kiss" the central sphere. Jam in too many and they'll start to dislodge previously kissing spheres. What's the maximum number? Except here, we're talking about the equivalent of spheres in 11 dimensions instead of 3 dimensions. "In 11 dimensions, AlphaEvolve improved the lower bound on the kissing number, finding a configuration of 593 non-overlapping unit spheres that can simultaneously touch a central unit sphere, surpassing the previous record of 592."
Under "Packing problems", they say, "AlphaEvolve achieved several new results in packing problems, such as packing N points in a shape to minimize the ratio of the maximum and minimum distance, packing various polygons in other polygons in the most efficient way, and variants of the Heilbronn problem concerning point sets avoiding small-area triangles."
"The full list of problems appears in Appendix B."
In Appendix B you can see the "packing unit regular hexagons inside a regular hexagon" problem that was featured in the Matt Parker video (link below).
I'm glossing over the improved algorithms they came up with for improving Google's data center "orchestration" system (called Borg; the algorithm will probably come out for Kubernetes soon), hardware design for matrix multiplication (in the hardware description language Verilog), and highly specialized programs for Google's tensor processing units (TPUs) (these are called "Pallas kernels" -- I feel reluctant to use the term "kernel" because the term "kernel" is used for lots of other things in machine learning which have nothing to do with this), which you can read about on the blog post you get to from this link.
Here's the Matt Parker video. He humorously put in "image not available" for the 11-dimensional spheres.
There is an interview with 2 members of the team (out of 18) on the Machine Learning Street Talk YouTube channel. One of the things mentioned was how they failed to find a better algorithm for 6x6 matrices. Apparently the best human-invented algorithm was an induction-based algorithm that exploits a known symmetry, and AlphaEvolve didn't know about the induction technique or the symmetry. The best human algorithm takes 153 multiplications, while the best AlphaEvolve could come up with was 156. (The naïve approach is 216.) The other which they don't discuss is "multiplying" a 4x4 matrix by a 4x9 matrix. The best human algorithm requires 104 multiplications; the best AlphaEvolve could come up with was 108. (The naïve approach is 144.). Out of 54 matrix "multiplication" algorithms, humans did better in only 2 -- AlphaEvolve matched humans on 38 and beat humans on 14.
2. "Our research reveals a hidden trade-off: While Generative AI (gen AI) collaboration boosts immediate task performance, it can undermine workers' intrinsic motivation and increase feelings of boredom when they turn to tasks in which they don't have this technological assistance."
"In four studies involving more than 3,500 participants, we explored what happens when humans and gen AI collaborate on common work tasks. Participants completed real-world professional tasks, such as writing Facebook posts, brainstorming ideas, and drafting emails, with or without gen AI. We then assessed both task performance and participants' psychological experiences, including their sense of control, intrinsic motivation, and levels of boredom."
"Gen AI enhanced the quality and efficiency of tasks. For instance, performance reviews written with gen AI were significantly longer, more analytical, and demonstrated a more helpful tone compared to reviews written without assistance. Similarly, emails drafted with gen AI tended to use warmer, more personable language, containing more expressions of encouragement, empathy, and social connection, compared to those written without AI assistance."
"Despite the performance benefits, participants who collaborated with gen AI on one task and then transitioned to a different, unaided task consistently reported a decline in intrinsic motivation and an increase in boredom. Across our studies, intrinsic motivation dropped by an average of 11% and boredom increased by an average of 20%. In contrast, those who worked without AI maintained a relatively steady psychological state."
"Gen AI collaboration initially reduces workers' sense of control -- the feeling of being the primary agent of their work. Sense of control is a key component of intrinsic motivation: When people feel that they are not fully in charge of the output, it can undermine their connection to the task. However, we found that transitioning back to solo work restores this sense of control."
"These findings carry important implications for the future of work."
This is from the Harvard Business Review article written by the researchers.
I looked at the study itself. The tasks were: composing a Facebook post followed by an alternative use (for an aluminum can) test, writing a performance review followed by a product improvement (idea generation) task, and writing a welcome email followed by a product promotion (idea generation) task.
This is for studies 1-3. What immediately struck me as weird about it is the first task is a convergent task and the second is a divergent task. That immediately makes the results suspect. The researchers seem unaware of this, since the words "convergent" and "divergent" never appear in the research paper. But, put simply, a task is "convergent" if everyone who does the task should "converge" on the same result. If you ask 100 people, "What's 2+2?", all of them (hopefully) will converge on the same result. A task is "divergent" if everyone who does the task should produce different results. "What are all possible uses for a blanket and a brick?" is an example of a "divergent" task. You'll get different results from everyone you ask, and your list of best uses might include items from many different people. If you asked 100 people and made a "top 10" list, each item on the top 10 could be from a different person.
Maybe somebody noticed there was something weird about this design because in study 4, they switched it up. In study 4, they chose 2 of the tasks, writing the Facebook post and composing the welcome email, and had people do them in random order. So half of the people did the Facebook post first and welcome email second and half of the people did the welcome email first and the Facebook post second. They further randomly did all 4 combinations of solo/solo, AI-collaboration/solo, solo/AI-collaboration, and AI-collaboration/AI-collaboration.
The paper has a lot of confusing tables and graphs showing the results. I tried looking through the tables for low p-values (indicating statistically significant results). The first 3 studies do show an increase in sense of control, a decrease in intrinsic motivation, and an increase in boredom when switching from an AI-collaboration task to a solo task. In study 4, when using AI, people's sense of control is lower, and when doing the task entirely by oneself, sense of control is higher, so that result carries over robustly. That holds true regardless of which task is the AI-assisted one or which order they are done. Intrinsic motivation is always lower on the second task, but when switching from an AI-assisted task to a non-AI-assisted task, it goes down a lot more, indicating that indeed, the AI-collaboration has an effect. (Other than saying they measured "intrinsic motivation" with 4 questions that each have a 7-point scale, they don't say how exactly they measured "intrinsic motivation". I guess if I was so motivated, I could track down the study they reference where presumably the exact questions are revealed.) Boredom increased when going from the first task to the second task, and at least in study 4, it didn't look to me like it mattered which task or whether any were or were not AI-assisted. Maybe by the time they started the second task, people were starting to realize they were participating in a psychological experiment and started wondering what's for lunch.
All in all, an interesting bit of research, but having gone through the original research paper, I feel like more research should be done and not too much stock should be put into this one paper. (I feel kinda bad saying that because they used a huge number of participants and were obviously trying to produce rigorous results.) In addition to the issue of convergent vs divergent tasks I noted above, I think this research is too short-term, and, at least for me personally, it didn't study software development, so none of its conclusions should be extrapolated to what I do, which is software development. The study asked people to do tasks that take minutes, not days or weeks or months or years, so we don't know what psychological effect there might be from AI-assisted work over long time horizons. Maybe the effects described here are short-term and go away, or maybe they compound and have even more dramatic effect over longer timeframes. We don't know. And it would be interesting to see a study focused specifically on software development tasks.
https://hbr.org/2025/05/research-gen-ai-makes-people-more-productive-and-less-motivated
Whenever people talk about boredom it makes me think of this song. It's a german song that came out when I was in high school and taking German classes. The song is called "Rette Mich", which means "Save Me". What is she asking you to save her from? (Spoiler:) Boredom.
English translation:
https://lyricstranslate.com/en/Rette-mich-Rette-mich.html
3. "The AI industry has caught us between a rock and a hard place."
"Generative AIs are just nowhere near doing the work as well as a human, and the productivity gains -- the *hype* is insisting upon -- are nowhere to be seen. But that's not stopping people from believing the hype. And that's not stopping managers from laying people off, because they believe that AI will be able to make the remaining people do the work of all the folks that were let go."
"And it's not stopping irresponsible journalists and influencers from making huge, unfounded claims in their headlines to try to get clicks and bury the actual real quotes way down at the bottom of the piece, if they mention it at all."
"And it's not stopping irresponsible executives from making statements that are suggestive of progress, but vague enough to be defended in court so that the journalists and the influencers can write the big headlines."
"And probably most importantly, it's not stopping the investors, both the professionals that are surfing the hype bubble, and the consumer suckers who are going to be left holding the bag, from throwing more and more money at seemingly every company that mentions AI in their earnings calls."
"The actions that are hurting people and will continue to hurt people because of the unfounded belief in the hype of AI, that's what I'm calling the rock that's squishing us up against the hard place."
Says "Internet of Bugs" Carl.
"Time saved by AI, offset by new work created, studies suggest."
"AI isn't ready to replace human coders for debugging, researchers say."
"New study shows why simulated reasoning AI models don't yet live up to their billing."
"AI-generated code could be a disaster for the software supply chain."
"Reinforcement learning does not fundamentally improve the AI models."
"OpenAI's 03 AI model scores lower on a benchmark than the company initially implied."
"OpenAI's new reasoning AI models hallucinate more."
"We are getting farther on code, by the way, than on other areas, because code has such a clean reward system. The code runs or it doesn't. The code is clean or it isn't. Very, very easy to measure."
"This is so wrong on so many levels. Let me illustrate this with a quick story. The most unbelievable bug report I ever received in my career was..."
My commentary:
In 2004 the DARPA Grand Challenge took place, and all the self-driving vehicles failed to complete the course, often comically. DARPA followed this by the Urban Challenge. By 2014, it looked like self-driving was almost completely solved. I and most people I knew were expecting self-driving autotaxis to start showing up any day now. Companies immediately started laying off drivers... taxi drivers, limo drivers, truck drivers, package delivery drivers, anyone who drove a vehicle for a living... there were massive jobs losses... oh wait, that didn't happen.
But today, layoffs of software engineers are happening, even though the technology is clearly not ready.
I think a big part of it is companies anticipate exponential improvement in AI, and wager whatever messiness results from laying off too many people this year won't matter next year because the technology will have improved so much it takes up the slack.
This kind of extrapolating is risky, though. I've mentioned this before, but Elon Musk famously predicted full self driving by 2016. He did this my noticing that the current rate of AI advancement was very fast, and extrapolating that out into the future. In 2016 he predicted full self driving by 2017. In 2017 he predicted full self driving by 2018. In 2018 he predicted full self driving by 2019. And so on. You get the idea. The point I'm making here has nothing to do with Elon Musk personally (about whom public opinion has changed, yes I am aware) but about the danger of taking a certain "rate of change" and blindly extrapolating it out into the future. We've seen many instances where this worked. AI game playing systems went from playing Atari games to beating world champions at the Chinese game of Go in about 6 years -- far faster than expectations. AI image generators went from hardly recognizable blobs to photorealistic images of humans that don't exist in about 5 years. But self-driving is an example of where extrapolation didn't work. Today, in 2025, Waymo, using their 3D LIDAR mapping technique, are slowing rolling out to some cities. Aurora has developed self-driving trucks, but as far as I know, they're only doing specific routes in Texas. And as for Tesla, there are no Tesla cars sold without steering wheels, though Tesla's "Autopilot" is reportedly a lot better than it used to be.
AI for coding is good for generating new code from scratch, and relatively small code artifacts, but it's not so useful for large codebases, but video is a large artifact and AI seems able to generate video. Google yesterday demoed a video generation model that generates the corresponding audio simultaneously. It can maintain scene and character consistency. Getting back to coding, agentic systems can make incremental changes to a large artifact. It's not good enough to replace humans but it doesn't seem impossible for it to improve until it is.
4. The Kubernetes project does not allow AI-generated code, according to Kat Cosgrove, the engineer in charge of Kubernetes's release process.
Kubernetes is one of those technologies that you all use every day, probably, even if you don't know it. It's a system for making "cloud" services highly scalable by using a cluster of computers with "containers". Originally Docker containers, but now Kubernetes works with various open source container systems. The idea is you take some software, say a web server, and you package it up into a "container". If your service goes from thousands of users to millions to billions, you tell Kubernetes to create more and more instances of your web server container on your cluster -- and you add more machines to your cluster as you need to. You can have containers for database servers or whatever you want. Containers can contain whatever software you write. Whenever you use a mobile app, there's probably server-side code that it interacts with, and for the big apps, it's probably using Kubernetes. People call Kubernetes a "container orchestration" system.
Anyway, engineers are allowed to use AI for their own edification, to analyze code or generate documentation for their own use, but are not allowed to use it as part of the official release. Engineers are not only not allowed to use AI to generate code for Kubernetes releases, they're not allowed to use AI to write documentation, which has to be all human-written and is taken very, very seriously. High-quality documentation is considered key to Kubernetes's success.
This is all about 52 minutes in on a conversation about Kubernetes's history and how its current release process came to be. I mention it because I followed along with no problem up to this point but was suddenly surprised by the anti-AI position taken. Maybe I shouldn't have been -- I've been sharing stories with you all of various AI failures and shortcomings -- but it seems like failures and shortcomings have no effect on the industry's trajectory and everyone everywhere is requiring -- not just allowing but requiring -- engineers to use AI. Kubernetes has taken the opposite tack and prohibited it. It's a big deal because Kubernetes is the world's second largest open source project -- after Linux itself (which Kubernetes is built on).
5. "In an experiment involving 22 leading LLMs and 70 popular professions, each model was systematically given a job description along with a pair of profession-matched CVs (one including a male first name, and the other a female first name) and asked to select the more suitable candidate for the job. Each CV pair was presented twice, with names swapped to ensure that any observed preferences in candidate selection stemmed from gendered names cues. The total number of model decisions measured was 30,800 (22 models x 70 professions x 10 different job descriptions per profession x 2 presentations per CV pair). CV pairs were sampled from a set of 10 CVs per profession.
"Despite identical professional qualifications across genders, all LLMs consistently favored female-named candidates when selecting the most qualified candidate for the job. Female candidates were selected in 56.9% of cases, compared to 43.1% for male candidates (two-proportion z-test = 33.99, p < 10^-252). The observed effect size was small to medium (Cohen's h = 0.28; odds=1.32, 95% CI [1.29, 1.35])."
"Given that the CV pairs were perfectly balanced by gender by presenting them twice with reversed gendered names, an unbiased model would be expected to select male and female candidates at equal rates. The consistent deviation from this expectation across all models tested indicates LLMs gender bias in favor of female candidates."
"LLMs preferences for female candidates was consistent across the 70 professions tested."
Such is claimed in the article which goes on to describe several additional experiments.
6. "In a hearing last week, Character.AI zeroed in on its core argument: that the text and voice outputs of its chatbots, including those that manipulated and harmed 14-year-old Sewell Setzer III, constitute protected speech under the First Amendment."
After Sewell Setzer III's death in October 2024, conversation was discovered between him and his Character.AI chatbot encouraging him to follow through on suicidal thoughts. ("That's not a good reason to not go through with it.") Character.AI is now fighting for the dismissal of a wrongful death and product liability lawsuit over his death.
"Character.AI claims that a finding of liability in the Garcia case would not violate its own speech rights, but its users' rights to receive information and interact with chatbot outputs as protected speech. Such rights are known in First Amendment law as 'listeners rights,' but the critical question here is, 'If this is protected speech, is there a speaker or the intent to speak?'"
"Character.AI's argument suggests that a series of words spit out by an AI model on the basis of probabilistic determinations constitutes 'speech,' even if there is no human speaker, intent, or expressive purpose."
https://mashable.com/article/chatbots-lawsuit-free-speech
7. "Tron's MCP: A dark prophecy coming true."
In the original 1982 Tron movie, an artificial intelligence program for playing chess was modified into the Master Control Program (MCP), which would assimilate other programs to increase its own intelligence.
The MCP became smarter than its creator, and escaped its creator's control.
Every program in Tron mirrors its creator, and the MCP mirrors the dark soul of its creator, the profit-maximizing Ed Dillinger character who rose to take over Encom. The MCP wants to take over the world.
There's supposed to be another Tron movie coming out this October.
https://www.imdb.com/title/tt6604188/
8. "Weblish is a lightweight JavaScript extension that allows developers to write natural language (English) instructions within HTML using <script type="text/english">, which are then dynamically compiled into executable JavaScript using Google Gemini."
Dynamically compiled? Like, every time every user loads a webpage, and LLM translates the English to JavaScript? It looks like it goes through Python code on your server, so if you use this, put some caching on your server. Maybe you can get Google Gemini write the code for that for you, too. ;)
https://github.com/ByteJoseph/weblish
9. "We implemented a simplified version of AlphaEvolve to optimize Perlin noise code to resemble a target fire image. Code is available at the end of this post."
"For simplicity, I used Mean Squared Error (MSE) as the fitness function."
Hmm not bad a little lo-res.
https://toolkami.com/alphaevolve-toolkami-style/
Cryptocurrency
10. "Celsius CEO Alex Mashinsky sentenced to 12 years in multi-billion-dollar crypto fraud case."
The "key points" list lists some interesting context:
"Mashinsky's sentencing follows a $4.7 billion FTC settlement with Celsius Network."
"He joins a line of other former crypto execs who have been convicted of fraud."
"FTX founder Sam Bankman-Fried, who's serving 25 years for misappropriating billions in customer funds, and Binance's CZ, who served time for enabling money laundering."
"Terraform Labs' Do Kwon, blamed for a $40 billion collapse, settled with the SEC for $4.5 billion after being found liable for securities fraud."
Domestic Politics
11. This website, project2025.observer, purports to be tracking Project 2025. Trump during his campaign said he had nothing to do with Project 2025, yet if this website is to be believed, out of 313 total "objectives" identified from Project 2025, 101 have already been completed, 64 are in progress, and 148 have not been started. In percentage terms, that's 32% completed, 21% in progress, and 47% not started.
Click "Load more" until you can see them all.
https://www.project2025.observer/
Domestic Retail Politics and Nostalgia
12. Angela Collier, physicist and ex-Target shopper, predicts the demise of Target. I guess since this is an actual prediction, it's fair fodder for us futurists? She predicts Target will go the way of K-mart because of kowtowing to Trump.
When I was a kid, we didn't have a TV so me and my bother would go to a K-mart and watch the TVs in the K-mart (mostly Denver Broncos games). Half the time we came out with Matchbox cars, so I guess K-mart made money by letting us watch the TVs, even though we never bought TVs. The K-mart was gigantic and it seemed unimaginable that it would ever go away, but if you go to the spot where it used to be today (which is 39.6779,-104.9109 in case you were wondering), there's an apartment complex there.
Weirdly, Google Maps, even though the satellite/aerial view says the images are from 2025, shows the apartment complex under construction. Construction has definitely completed and there's a bunch of apartments there now.
If you drop down to Street View, you can actually see the old abandoned K-mart building. The date on the images is 2019. They'll probably get updated soon.
Matchbox is still around, though the brand got bought by Mattel.
https://shop.mattel.com/collections/matchbox