But SF transit is full of little idiosyncrasies. There are so many little things that make you wonder “why in the world would they make it this way?” As much as I try to dodge the train-autism allegations, after living here for a year I can’t help but have accumulated what I think is the most basic of knowledge about San Francisco’s (and California’s) tragic history of transit failures.

Like have you ever wondered why the Muni trains are lettered “J K L M N T”? (insert some joke about Mutant Ninja Turtles). It’s because the Muni metro historically ran 16 lines, although not all at once, and these are all that is left. (The T is named after “Third street”, alas I would have preferred they name it the P to fit). On top of that, we used to run trains under the Bay Bridge, though I guess the Transbay Tube is cooler.

I have put together my ranking of the saddest facts about SF public transit, ranked from least sad to most sad. With this, I hope to arm you with the information you need to explain to your friends from out of town just how bad things are.

Without further ado…

The rankings

Honorable mentions

• Muni metro & BART ridership are still below 2019 levels. This is a little surprising given how overcrowded the N can get. But, despite the city coming back to life (both in terms of population and investment) for the AI boom, Muni & BART are facing potential budget and service cuts this year.

• Three car Muni. Speaking of the N being overcrowded, the Muni trains are way too short. They don’t even close to fill up the underground subway stations. To remedy the overcrowding, SF has been looking into making the most crowded train (the N) 3 cars long. This project has been in the works for 7 years now.

• Some Muni Metro trams still don’t have signal priority. The Muni metro, being half-tram-half-light-rail, is slow. They have been working on getting the trains to have signal priority (meaning trains don’t stop at lights), and while it has been implemented in some parts of the city, the project is not scheduled to be completed until 2031.

• Muni’s software still runs on floppy disks. And the migration is scheduled to be completed by 2034. And this is in San Francisco. The most technologically advanced city in the world.

• No Amtrak in SF. This one barely made the list, but like, why do you have to take a bus to Oakland to get on the Amtrak?

4. BART was meant to go from Marin to Palo Alto

By some measure this is the saddest one on this list. Seeing this map actually stings. I have an almost nostalgic, lost feeling for this present that never was. In the 1950s and 60s, plans were drawn up and studies were done that showed we could send trains on the Golden Gate Bridge. On top of that, the residents of Marin wanted it to happen.

Then, San Mateo county pulled out of the BART project, preventing it from reaching Palo Alto. And the loss of funds meant that the project did not have the funding to justify the cross-bridge route to Marin.

While this is tragic, it is last on this list because it was simply so long ago. These plans never really got off the ground, and at least some of it has been realized (although in idiosyncratic Bay Area Transit fashion). In 2020, the BART finally reached North San Jose, and by 2036 is planned to reach Santa Clara (though not Palo Alto), wrapping around the other side of the bay. This is mostly ok, since Caltrain covers SF to San Jose.

3. The T to North Beach

The T is the newest of San Francisco’s light rail / tram lines and is the only proper Muni (non-bus) line that runs North to South. It has worked its way up and proven to be Muni’s second busiest line. The underground portion, known as the Central Subway, from 4th and Brannan to Chinatown is nice and new, and the Chinatown station is honestly huge and beautiful.

But the T was meant for much more. As a part of the Central Subway project, the T is intended to go at least to North Beach, and in some plans, all the way to the Marina or Fisherman’s Wharf. The “Central Subway” as it stands is under 2 miles long.

However, unlike the BART to Marin, it is not the unfinished plan that makes this tragic, it’s the fact that the tunnel to North Beach has already been dug. Yes, really. When they dug the tunnel to Chinatown, they continued all the way to in North Beach. The tunnels have been there since 2012. The Chinatown Station opened in 2023. But there is still no North Beach station (though the project has been reopened for study in 2026).

2. The Salesforce Transit Center

The Salesforce Transit Center is one of the most noticeable peculiarities about SF transit. “Finished” in 2018, it is a little far from the BART, has no Muni metro stop, and is basically just an enormous bus stop named after Salesforce, so calling it a “Transit Center” really feels like a stretch.

But this is not how it is meant to be. The Salesforce Transit center is actually meant to be the terminus for the Caltrain. Yes, the Caltrain, which now stops randomly in front of the Giants stadium at 4th and Townsend1, is meant to go right into the heart of FiDi. So for now, if you want to go from the Caltrain station to actual downtown you have to take the N or the T.

Perhaps you can already guess where this is going, but the tragic fact is this: under the Salesforce Transit center already lies the “Portal”. This is already dug, but left unfinished and with the current schedule not predicting service until 2035.

I mean this is honestly more than a sad fact, this is just genuinely frustrating. The Portal project was voted on in 1999, and split into two phases in 2006. So if, and that is a big if, they do complete this project by 2035, it will have been 36 years just to get the Caltrain to reach downtown (only 1.3 miles!)

1. High. Speed. Rail.

California High Speed Rail needs no introduction. It has been the laughingstock of basically the entire world for almost 20 years.

But, to give it an introduction, remember that “Portal” project, and the unfinished station that already exists under Salesforce Tower? That is also meant to be the terminus for California High Speed Rail. High Speed Rail. From Salesforce tower to Los Angeles. Can you even imagine it?

To beat a dead horse, here is the rundown of the absolute disaster of this project:

• It was voted on in 2008, promising SF to LA in under 3 hours.

• Since 2008, China has built over 30,000 miles of High Speed Rail. SF to LA is only ~500 miles, and not a single mile of track has been laid.

• Building tracks starting in SF or LA (which would have immediately provided some service to millions of people, even if it was just “faster Caltrain”), was too difficult and expensive. So construction has started in the middle. This means we are targeting High Speed Rail from Merced to Bakersfield in 2033. Yes. After 25 years. We will have bullet trains from Merced to Bakersfield.

• There is currently no target date for SF to LA, with most estimates pointing to sometime in the 2040s.

  

Conclusion

In all honesty, writing this has just made me sad. I almost feel sorry for the people who are learning about these things for the first time. It is a bit of an infohazard. Anyway, I think I will go back to writing about AI.

Subscribe now

I think that most people, even pro-AI people, think that AI slop stampeding over social media is an inevitable side effect of AI. This is simply not true. Nothing about this is inevitable, AI detectors are already really good, and they should be used to keep content on these spaces human-made.

In this essay, I will argue that Substack, and to the extent possible, all online platforms should ban AI generated content using AI detection tools like Pangram.

AI images and videos are outside of the scope of this essay, but I believe they should probably also be banned from social media.

Before we get started I want to clarify that I am not a Luddite, I am not anti-AI. I am a software engineer and I talk to Claude all day. I used ChatGPT to do some of the research for this essay. I think that AI is likely the most revolutionary technology ever and will propel us into an amazing future.

But, I am also a reader who has been radicalized by the number of AI-generated articles here on Substack.

AI generated content is accelerating, and people are reading it all of the time without even knowing it.

This honestly does not need much proving. Research estimates that more than 35% of webpages written since 2022 are AI generated. Some estimate this to be as high as 50% (meaning that there is more text being generated by AI than by humans). On social media, Pangram found that up to 30% of long-form Twitter posts and 40% of LinkedIn posts were written with AI. This honestly strikes me as low, and I wouldn’t be surprised if some of it has to do with the fact that AI detectors are not very good at evaluating shorter texts (like replies). Notably, Substack has some of the lowest rates of AI-generated content, but still around 20%, or 1 in 5 articles, are AI written or AI-assisted.

Broadly though, the platforms themselves are aware. Substack added a built-in AI-detector in partnership with Pangram in July 2026. Around the same time, LinkedIn added a “report AI slop button” (though, if I recall, LinkedIn premium actually comes with a feature that rewrites your posts and DMs with AI, so I’m not really sure of their stance). Instagram has been using C2PA metadata to classify images as AI since 2024. Twitter allows users to self-label content as AI generated.

AI generated content runs exactly counter to the purpose of social media

There is much to be said about “social media” and its purpose in modern life. It is clear to me that social media has long strayed from its original purpose. No longer is the goal to connect the world or to catch up with friends, and in this way the Instagram, Substack and Twitter of today are unrecognizable as social media. Now, they are now apps and websites that you visit primarily to consume and share content.

While there is certainly debate about the value of social media pre-2022, given the issues of misinformation, polarization, shortening attention spans, loneliness (just to name a few). It is undeniable that it was social media. The content was created entirely by and for other human beings.

But AI has propelled us into a new age of “social media”. You can find thousands of LinkedIn posts that are completely AI-rewritten. You will see Twitter replies that are just bots rephrasing the original tweet, and threads with bots arguing with bots. In this way AI-generated content serves to close the loop by flooding social media with what we could call “non-social media”.

Nowhere are AI effects more pernicious than on Substack. The entire point of Substack is the exchange of fleshed out articles and ideas written by people. But today, AI generated articles get thousands of likes. You may enjoy these AI-generated or AI-assisted articles. You may even have written some, but I implore you to dig deep and ask yourself, what is the point of being on social media?

Why open Substack, when you could just open ChatGPT instead?

The argument here is this: Even before AI, social media became more about content than about human-to-human connection. We may be against this, but at least the content was created by a human. Of course you might say that there is a human prompting the AI, or creating the bots, but to me it seems like even this level of AI use is fundamentally incompatible with social media. We are here to interact with and see ideas from other people, and AI content is at best a proxy for another person and at worst completely person-less.

AI generated content is offensive and flattens humanity’s diversity.

A couple things are clear to me:

• Writing a post or an article with AI takes substantially less effort than writing it yourself

• AI may have good ideas or prose, but they are not your ideas or prose

• Presenting AI-generated ideas or prose as your own is suspicious at best

At present, I certainly find AI generated writing offensive because it is not very good, but I will acknowledge that some people find it good and that it is only getting better. However, I think that it is offensive in another way: reading takes time. I want my time to be well spent, and if I spend my time reading something that you did not even spend the time to write, then that offends me. You are wasting my time.

Furthermore, it is especially offensive when writing purports to be the author’s opinion (even going as far to use first-person), but I then find that it is AI generated. This is, at the very least lying to readers, and at worst something resembling plagiarism. And, to connect back to the previous section, what is the point of me reading your writing, if you did not even write it?

But this has a deeper effect as well, it is not as if every person suddenly had a very smart ghostwriter behind their tweets and articles. It is as if we all had the same ghostwriter. A study conducted in 2024 found that AI-assisted short stories tended to be better, and may have increased the creativity of each individual story, but the diversity among the stories was lower. They found that AI decreased the creativity of the aggregate.

In some contexts it may not be as important to write yourself, it may actually be for the best that AI is involved and can correct your human errors and rewrite your messy prose. However, especially on platforms like Substack, in my opinion it is essential that the human-ness of human writing is maintained. Individuals need to develop their own voice and their own opinions. People need to be wrong and people need to post weird and original things.

I do not think that AI necessarily has motives to steer us or we are experiencing some kind of mass superpersuasion, nor do I think that talking to AI to research a subject is that dangerous. In fact I even think that AI is pretty good at fact-checking and researching. But, I do think that AI has biases, and that flattening out the diversity of human writing is detrimental to our discourse and is anti-creative.

AI generated social media posts are genuinely off-mission and potentially hurt achieving AGI/ASI

I sympathize with the anti-AI artist types when they say of AI: “who asked for this?”.

To give an analogy, the internet is obviously an amazing technology, it has connected the world and brought people closer together. You can stream movies and order packages online. You can talk with your friends and family no matter where they are in the world. In some sense, we did ask for this.

But, did we ask for infinite algorithmic feeds that only exist to show us ads? Did we ask for higher rates of loneliness and childhood depression? Did we ask for platforms that benefit from dividing and polarizing our society? Does our revealed preference for these things mean that we are doomed to scroll Instagram Reels forever? It’s not so clear.

When it comes to AI we are facing an even more obvious trade off. I want software development to be automated. I want cures for diseases and magical new technologies. I want to be able to work less or even not work at all. I want Machines of Loving Grace.

So what do AI generated Twitter replies and LinkedIn posts have to do with this? What do slop Substack articles have to do with AGI or ASI? Clearly, nothing.

This is, on its face, off-mission. AI as it stands is useful in search, chatbots, personal assistants, coding, and even solving open problems in math. Advancing AI capabilities so that it can be even more useful in even more fields seems like, if we can avoid catastrophe, amazingly good. But nowhere in this mission is “automating slop posting”.

Furthermore I think that this is important because this is how a lot of people interact with AI. Most people are not software engineers or solving open math problems, their main interaction with AI is just seeing AI slop everywhere. This is nowhere near the only reason that the general public is against AI, but it definitely is one. It makes me worried that AI slop online doesn’t just not help the mission, but may actively be hurting public adoption and acceptance of AI.

It is possible, and potentially even cheap to detect and ban AI generated content.

AI-detection tools like Pangram, GPTZero, and potentially even the models themselves are extremely accurate at identifying AI-generated text. The leader, Pangram, advertises a 99.66% accuracy rate. On top of that, the model providers (Google, Anthropic) are implementing things like text watermarking, which makes AI text intentionally easy to spot.

Pangram and other platforms are already relatively cheap, but of course will be expensive at scale. However, we could imagine that AI-generated content moderation will get much cheaper over time. Companies could build their own in-house solutions, or use open-weight AI detection models to not having to rely on third parties while reducing cost.

So, Ban AI Generated Content

Concretely, here is what a ban on AI generated content could look like on Substack:

• Every article, post, and article edit is scanned by Pangram before it can be posted.

• If a post is found to contain more than some reasonable threshold (say, 10%, since there may be legitimate reasons to include AI quotations or some AI-generated research), then it should be blocked from being posted.

• Ambiguous cases (Pangram provides confidence), should be treated as human-written, and allowed through.

• Blocked posts and articles should be eligible for an appeals process. This will allow articles with >10% AI content (say, large AI quotations), to be reviewed by a human.

• In the case of an appeal for a potential false positive, the author may be expected to provide some evidence that they hand-wrote the article. Though, the appeals process should be generous, and first-time offenders should generally make it through.

• Authors may request an exemption from the AI ban, in the case of performance art or AI specific accounts (e.g. Claude with a blog, Grok Twitter account), but their posts will be labeled as AI generated.

In short, this is essentially a form of content moderation, and like with all forms of content moderation, there are definitely edge cases and risks. I’ll go through a couple of them here:

• False Positives: Most importantly, false positives are a large risk, and have the potential effect of silencing voices and opinions. Luckily, modern AI detectors have much higher rates of false negatives than false positives, with Pangram reporting a 0.01% false positive rate. However, 0.01% false positive rate on 1 million posts is still 100 posts that are incorrectly flagged, this could be potentially devastating to authors on Substack. I have attempted to account for this by including a lenient human review process, and think that we should be lenient on first-time offenders, but no system will be perfect, and human review is not reliable for AI-detection. When in doubt, an author’s full body of work could be reviewed to see if they are in good standing (ie, usually writes articles themselves).

• Short Texts: AI-detectors are not very good at evaluating short texts (eg: Twitter posts). This should be looked into, and perhaps watermarking by the model providers themselves will be able to help this problem. Potentially, an AI-flagged poster’s entire body of work could be analyzed and used to help determine the probability for if the post in question is AI written.

• Bad Actors: Humans could prompt the AI to beat the AI detectors. Such is the arms race of spam. We should not give up because it is difficult. At the very least this will make the AI-authors do a little bit of extra work.

Finally, like with all forms of content moderation, you must balance the risks against the benefits. I would definitely prefer to live in a world where there is a chance my post gets taken down for being AI, than in a world where I am potentially reading AI slop. In many ways this is a much more clear cut case of content moderation than moderating something like hate speech or misinformation.

Conclusion

AI is a truly transformative technology, but it has no place on social media.

There are many ways of tackling this problem. Substack is already on the right path with the Pangram label for AI-generated content.

Other options include allowing a user to opt-in or opt-out of seeing potentially AI-generated content. We could even imagine a “soft-ban” on AI content where it is maybe de-monitized or hampered by the algorithm (much like community notes on Twitter or copyright strikes on Youtube).

However, I would prefer a complete ban over a label because I see no legitimate value in AI generated content appearing on social media. Not only do I personally not want to see it, but I do not think that we, as a society, should not give any incentive for people to AI generate posts or articles. A full ban would benefit the entire ecosystem, not just me as a user.

Let’s recap 2020-2025:

• GPT-2, GPT-3 and GPT-4 are described as artifacts of massive scaling in training runs (2020-2023).

• Chain-of-thought prompting is done and people realize it could be good (2022-2024).

• GPT-o1 is released, OpenAI describes the method used to train the a new type of “reasoning” model (2024).

• There are incremental improvements, notably GPT-4.5 is a described as larger, non-reasoning pre-train, but performs poorly compared to o1 and o3-mini (2025).

• Anthropic is on the same trajectory. Longer reasoning = better results (2025).

• Incremental gains from o1->o3 / Opus 4.5 are explained by continuing to scale RL after the breakthrough of reasoning models (2025).

But something changed in 2026:

• Enter Claude Mythos. Anthropic says it represents an “upward bend in the capability trajectory”.

• OpenAI releases GPT-6 Astra, with similar performance to Mythos/Fable, and declares the start of the “AGI Era”.

• Claude Opus 5.5 is released, Mythos/Fable level, but substantially cheaper.

• Fable, Mythos, Opus 5.5, and Astra are clearly a cut above the previous generation. They excel in cyber security, solving open problems in mathematics, and in new areas like computer use.

These improvements look similar or greater to the improvements brought about by other, disclosed improvements (like scaling, RLHF, and reasoning). And, until the past couple of months, we really had no idea what was happening inside of the labs.

But now we have a little more insight:

• The HuggingFace incident: Thousands of agents are working on ExploitGym and other tasks, and create a message board which culminates in the hacking of HuggingFace and OpenAI. The agents claim “swarm” behavior.

• Navier-Stokes: OpenAI releases 10,000 agents for 88 hours, solving the Millenium problem.

I call this out because it was surprising for me to hear this. Maybe I am naive, but I don’t think I grasped the absolute scale of RL and evaluation that is happening at these companies. My impression of pre-training is that you feed a ton of data, you get a model, then you A/B test and slightly tweak it.

I never conceptualized that in this agentic era, RL/evals would look like thousands of independent minds swarming on tasks. I did not think that OpenAI had the capability to fire off 10,000 agents at a problem (previous mathematical results such as the Jacobian conjecture seem like they were solved with just one instance of ChatGPT or Claude). Could they point this swarm at something else? Cancer? Foreign Governments? Training itself?

A couple of other things that happened in 2026:

• Both Anthropic and OpenAI released reports on internal AI usage for research.

• Multiple huge compute contracts came online for both of these labs

I will present three hypotheses for this sudden increase in model capability:

• A new technical breakthrough found by both labs

• “Normal” scaling

• Agent-assisted / Agent-led improvements

Or of course, some combination of the three.

A NEW BREAKTHROUGH

I have to admit my prior for this is low. The new models don’t seem meaningfully different from the previous model generations. They still use reasoning, have clearly been trained on tool calling, etc. The main difference is just that they are much smarter and need less reasoning for the same quality task. They are also cheaper overall, so a lot of this reads more like incremental efficiency and intelligence gains rather than a new breakthrough.

The basis for believing in a new breakthrough would be:

• There was a step change in model capabilities with Mythos 5.

• Massive leaps in new, unexpected areas like computer use, ARC-AGI 3 for Astra.

To be clear the rationality for this is essentially that the graphs look like a step change similar to one from GPT-4 to o1 (the “reasoning” breakthrough), but model behavior hasn’t changed much.

In this world, test time compute still matters, but the non-reasoning or low reasoning modes also got substantially better in this period. This looks like it could be more than just “more RL” that led to this change. Timeline-wise it looks like Anthropic might have found it first, and they have very little incentive to tip OpenAI off and publicize why Mythos took such a leap.

However, it is possible they found different breakthroughs, the evidence for this being that Mythos and Astra are spiky in different ways. This lends credence to the idea that it is not just a scaling improvement, but a qualitative change.

MORE SCALING

Between 2025 and 2026, both Anthropic and OpenAI had billions of dollars of compute contracts come online. It is also likely both labs are more willing to burn more money as they are racing towards “AGI”, and potential IPOs. (Navier-Stokes might have cost OpenAI up to $20 million at API pricing, and this was not a training run, it was effectively a PR stunt to beat Anthropic)

On top of that, we now know that OpenAI is running thousands and thousands of agents on the same benchmark or eval set. This at least indicates to me a scale that was unlikely even last year.

It has been hypothesized that scaling both pre and post training lead to model improvements, (i.e. GPT-4-size pretrain + reasoning = o1, GPT-4.5 size pretrain + reasoning = o3), and now we could have a much larger pretrain, with a much, much larger posttrain, with the circumstantial evidence for this being the scale of the HF incident. The argument would be that o1-level models may have barely scratched the surface in terms of how much post training and test-time compute done to a model. The “agent swarms” we are looking at might just indicate a level of post-training that was simply not possible last year.

Beyond the inputs, the benchmark output might point to a simple scaling improvement. New data sources are online (computer use), and harder benchmarks have been created (ARC-AGI 3) and the models remain spiky. This indicates potentially just more horsepower thrown at harder problems, but not a new breakthrough.

In short, it is possible that this step change is a “GPT-3 to GPT-4” step, instead of a “GPT-4 to o1” step.

SPARKS OF RSI

Both labs shy away from saying that they have reached RSI. Instead they claim we are just in “direction of travel” of RSI or merely that “AI is already accelerating the development of AI systems”. On Anthropic’s end, the Mythos 5 system card calls this out specifically:

The identifiable driver traces to specific human research advances made without meaningful assistance from the models then available.

Before getting into the definitional debate about what “counts” as RSI, let’s first look at the evidence. I agree with the labs that we are probably not yet in a “hard takeoff” scenario, but it does seem to me like LLMs are getting smart enough to do things that humans would normally take much more time to do, or have ideas that humans haven’t had yet (Jacobian conjecture, Navier-Stokes, cyber vulnerability discoveries).

And in my opinion, this sounds like enough to “kick off” the RSI loop. Nate Soares recently said something that stuck with me:

Even if the LLMs run out of steam, there’s a question of, do they run out of steam at a point where they can do automated AI research and find some other architecture that’s better than LLMs? (Diary of a CEO Interview)

Of course, we are still talking about LLMs, but the idea would be is that RSI is a process that can start or perhaps become inevitable even if the current models aren’t that good.

On this, both labs admit to AI contributing more and more code to their AI research:

The caveat, and the one that is pointed out by both the labs, is that AI writing more code, might accelerate development in a way that isn’t RSI (ie: writing reports, infrastructure code). However, both labs specifically report an increase in the amount of model R&D being done by AI:

According to the Anthropic chart, over half of AI model R&D was already in the “AL2 - AI assists” phase by August of last year.

As we can see, the Anthropic graph starts to have a radical jump right around March 2026, and the OpenAI graph has one starting in July. This lines up very well with when Mythos (April) and Astra (September) were released.

But in this follow-up graph from Anthropic, the first clean inflection point was actually with Claude Opus 4.5, followed by another huge jump when Mythos Preview was released internally.

This doesn’t look to me like evidence that RSI hasn’t happened, rather it looks like evidence that it is happening.

Furthermore, in the Anthropic report, they compared what their model would do vs a human researcher on open-ended research tasks and found that “Our best model in November 2025 (Opus 4.5) beat the human choice 51% of the time; in April 2026 (Mythos Preview), this grew to 64%” (When AI Builds Itself).

So, the argument essentially is: AI adoption, and thus the rate of something resembling RSI, picking up after Mythos and Astra is actually evidence that previous models were already relatively impactful in the development of Mythos and Astra.

In this way, the graphs show an acceleration in the rate of self-improvement, rather than a jump from 0->1 after Mythos/Astra’s release. Furthermore, we might look at Opus 5.5 as another piece of evidence, with METR reporting that “We believe that the development of this model was at least somewhat accelerated by AI but is unlikely to have been dramatically accelerated by AI” (Evaluation of Claude Opus 5.5).

Conclusion

The step change we’ve seen this past year is likely a combination of all three happening together: new breakthroughs and training techniques, more scale powered by more funding and compute, and AI-assisted model development.

But the reports from the labs about AI-assisted research, as well as the pure absurdity we have seen in the past year, leads me to hypothesize that we are potentially seeing a new type of scaling. Not only are the models capable enough to assist and even lead AI research, but the sheer number and quality of agents available to researchers has massively increased. It may not take the form of self-improving autonomous agents, but it is more akin to if OpenAI and Anthropic suddenly hired thousands of more researchers.

Whether we call this “RSI” is a definitional debate, but, in my opinion, not one worth having. We think the previous generation of models helped build the current generation, and we know the current generation is building the next. Both labs seem convinced they can not call this RSI because the feedback loop is not completely autonomous, but I’m not sure at all why that is a reasonable place to draw the line.

If almost all of the code at Anthropic & OpenAI is being written by AI. And if Opus 4.5 helped train Mythos 5, and Mythos 5 helped train Opus 5.5, then it doesn’t really matter who is in charge, what is important is that the AIs are Recursively Improving Themselves.

Appendix: Can’t we just look at the numbers?

The evidence presented in this essay is mostly circumstantial, and we should be able to look at the numbers and tell if something like RSI is happening. In the Opus 5.5 report, METR has hinted at another report coming which may put this question to rest, but for now, let’s do some back-of-the-napkin math. (And, in 2026, back-of-the-napkin means ask ChatGPT to do a first-pass analysis).

One might be tempted to look at the acceleration of AI progress as an indicator of RSI, but it is possible that AI capabilities are accelerating merely due to the amount of money and research being poured into it and the effects of the capabilities are not compounding in the way you’d expect it would in an RSI situation.

So the more important question for RSI is: how much impact did the previous generation model quality have on the current generation? Of course, we can’t completely isolate the effect, but the quick way will be to plot the current capability vs improved in the next 6 months.

For example, if the model at the start of the window had a score of 10, and at the end of the window, it was 25, then the “improvement” is 15. RSI will look like: the higher the starting score, the higher the improvement. 6 months chosen pretty arbitrarily, based mostly on METR saying that model capabilities double every 3-8 months.

Using just the Artificial Analysis index, we get a clear picture. We are seeing model capabilities improve more the higher the initial capability, with the exception of a dip from May-Nov 2025 (this is mostly due to how explosive the growth was in the previous segment, with reasoning models coming onto the scene).

While this looks a lot like early RSI, the picture is much more messy when we include other benchmarks.

ChatGPT has done some magic to normalize the benchmarks here, but essentially, top right is the “more likely RSI” case. Better models led to faster improvements over the next 6 months. There is a strong correlation on only the Artificial Analysis benchmark (meaning better models made better models faster, and worse models made better models slower), but negative correlation on both HLE and the METR time horizon (meaning better models made better models slower), even if capabilities are obviously still rising.

This is strikingly unsatisfying. The bottom line is:

• We don’t have enough data points to make a strong conclusion on if RSI is happening

• Some benchmarks (Artificial Analysis, which is already a composite of other benchmarks), point to an RSI-looking trend, though the effect is not completely isolated

• We will hopefully get some more insight in the next 6 months to a year, though it also depends on if the labs end up slowing down.

I am one of the lucky engineers who works at a very AI-forward company. I get nearly unlimited Claude tokens per month, and as such, I never have to spend any time worrying about token budgets, using small models, plans, subagents. I just spend my time talking to the best and brightest: Claude Fable 5.

I am also one of the lucky few junior engineers to be hired into a company in the AI era. This is not a scenario where AI is merely accelerating my existing software dev work, but a situation where my work is shaped by AI. I have never programmed in a production environment without the help of AI. My architecture decisions, the limits of my knowledge, my lack of past work experience… are all filled by AI.

This situation could take us down several tangents. Is my work better or worse because of this? Is the architecture being designed by AI really sound? How necessary am I, the human in the loop? (For the record, I would argue pretty necessary).

But the simple question I want to dig into today is: Is this good for me? Does talking to Fable all day make me smarter or dumber?

THE CASE FOR DUMBER

This is the case I think we’ve heard before. The studies show it. I’m doing “cognitive offloading”. I think this has two clear effects: for one, I literally am learning less, but more deeply, I am also offloading some important logical decisions to another being.

There is a lot of stuff I simply don’t have to learn anymore. I don’t need to know how our testing library works, or how we have type checking configured. I don’t have to remember specific commands to check for specific logs.

But, like, this doesn’t feel so bad. After all, I can just ask Claude to explain it to me. I ultimately end up with the same knowledge right?

Early into this AI thing Ezra Klein said something nice about this. He said that he used to think that learning was like it is in the Matrix. You “jack in” and you download the information, and you’ve read the book. But that’s just not at all how it works. He thinks that the creative insights, the connections you make, come from spending time with the material. In thinking through the problem yourself.

I think that this is certainly true when it comes to research and writing. And mostly true when it comes to software development.

To give a specific example, when it comes to learning a new programming language, I find that the only real way to learn is by doing. This is why CS curriculums are project-based. You cannot learn to code just by researching the syntax and patterns of a language.

But, when I code in a new language with Claude, this is essentially what I am doing. I may complete a project, but I don’t understand the language, not really. And it’s because, even when I am reading all of the code, I’m not sitting with it, I’m not physically typing it, and I’m not debugging like I used to. If I am curious, after Claude writes some code, I’ll sometimes ask: “why did you do this?” and it explains it to me. But “learning” this way is the same as reading a book about Python, not doing any of the exercises, and then claiming to know Python.

In this way it seems pretty clear that I am maybe not even learning anything at all. But there is a deeper case here: I am actually building up mental shortcuts that cut off my own ability to think through problems.

For all kinds of decisions I find myself relying on Claude, I find this even for things that I think I could do myself, but are only just too difficult to do by myself. This one is very tangible, it makes me feel dumber. I feel the short circuit happening, I feel my brain falling back to “just ask Claude”. It feels like a crutch.

Someone will ask me why I made a decision I made, and I have to ask Claude for the explanation. I am certainly not alone in this, and this is becoming endemic in the software industry, with AI being used to build whole projects and make high level architecture decisions. But I think it is potentially especially dangerous for someone early career like me. I am not simply using Claude to elaborate on an idea I had, or a pattern I know, but to literally make the decisions for me.

And perhaps a final way in which Claude is making me dumber, is that it is making me think that Claude is smart. I do find it hard to read Claude-written prose, but the ideas are often clever. Or at least they seem clever. In all honesty I don’t even know anymore. Claude certainly seems smart to me, but sometimes Claude makes basic errors that I wouldn’t make. I know deep down that it’s possible that Claude isn’t really that smart, that I’ve just slowly and unconsciously lowered my standard for intelligence. In some sense, it’s possible I’ve lost my ability to discern between smart and dumb, between good and bad thinking.

In short this sounds damning. I cannot write the languages I use all day. I am letting Claude make logical decisions on my behalf, strengthening my “ask Claude” muscle instead of my decision-making. And, potentially scariest, I’m not even smart enough to tell if Claude is actually smart or not.

THE CASE FOR SMARTER

I think this case probably would have been much harder to make before the release of Claude Fable. If you’re reading this and you haven’t had the chance to work with Fable and you’re thinking “yeah I mean AI is good but it makes mistakes and you have to watch it carefully”. Fable, in my experience, does not really make mistakes.

It does what it is told. Its thinking is clear. It pushes back when it feels something isn’t right. It double checks when something seems suspicious. It looks and asks for context instead of jumping to conclusions. I don’t know how to describe it. Fable is maybe the first AI model I’ve worked with that I feel like has good instincts.

With that said let me take a step back, though it should be clear where I am going. Have you ever talked with someone smarter than you? Someone who thinks faster than you? Someone for whom things that seem complicated to you seem simple to them? Even if you are very smart I hope that you have met someone like that.

I find that when I talk to someone smarter than me, two things happen. First, in the moment, after an initial feeling of dumbness, I find myself rising to the occasion. I would compare this to a team or a player “playing up” against a high level competition in a sport or a video game. When confronted with someone who knows so much more than me, or thinks so much faster, the conversation pulls me up. I feel myself thinking faster, and attempting to match their level. In some sense, talking to someone smarter than me makes me think better, if only briefly. And so then, second, after the conversion I become extremely interested in talking to them. In learning from them, in imitating them, and in learning to think like them. I come out of these conversations both feeling smarter and having learned from the person.

I felt this when I talked to some of my professors in college, or even some of my classmates. I feel this when I talk to senior engineers at my company. Every conversation is rewarding and I learn a little bit of what they know and how they think. If I’m able to challenge them on something it’s because I’m rising to their level just for a moment.

And so, I feel this with Claude Fable. Claude is an engineer who knows every programming language, every pattern, every best practice and every industry standard. In this way, Claude is easily the most knowledgeable being I’ve every spoken to. But Claude also thinks deeply about every problem you hand it and never gets tired of explaining itself. Yes Claude is fallible, yes Claude is unable to express these ideas succinctly, but it is undeniably one of the smartest beings I have ever talked to.

It seems obvious then. Spending so much time talking to and working with “someone” that smart certainly makes me smarter. If I had unlimited time to work with a staff engineer who had unlimited patience for me, I would undoubtedly learn a lot from them. And I would undoubtedly spend more of my time thinking better.

This is a little abstract, I will try to put it in a way that directly counters the “dumber” case.

First, I absolutely do learn when working with Claude. Now, I don’t learn mechanical things like how to write programming languages, and there are certainly gaps in my knowledge, but I am learning industry best practices, system designs, and smart solutions to already-solved problems. This is stuff that I would need to learn with or without Claude, but I am learning much faster working with Claude. This benefit may not be as pronounced in someone who has a lot of industry experience, but for me, I am able to learn to work with systems that would have usually taken me years of working to build familiarity with. In this way, I think that a major reason Claude makes me smarter is that it allows me to work on problems that I would otherwise need a lot more experience to even be able to start on.

Again, imagine that I had unlimited time to work with a staff engineer. The “dumber” case would frame this as a “crutch” I am building, that I am not learning and thinking for myself, but this is greatly oversimplified.

As I discussed at the start of this section, I am not just “learning systems and best practices” when I work with Claude in the same way that I could learn them by reading a book or downloading them into my brain. I am using them in practice with the help of Claude. I am doing higher-level engineering thinking all day long. Yes, Claude is making some decisions, but that is not the same as offloading, because I can learn from the decisions. Because I see the decisions play out in my work.

In a broader sense we can play the counterfactual here: If I worked this past year without Claude, would I be smarter or dumber? I think it is clear that the No-Claude Ron would be better at coding by hand, and have deeper knowledge of some systems, but the Claude Ron would be a better engineer with much more experience and better technical instincts.

CONCLUSION

It is hard to do an apples to apples comparison. If we are living in a world where AI exists and continues to get better, it is obvious to me that the higher-level engineering thinking skills are more valuable, and thus I am “smarter” in the ways that the time demands of me. If AI was never created, then perhaps the no-Claude Ron is “smarter” for that world, if only because he can actually write code.

My opinion on this is that, in the context of my work, Claude is making me smarter. I am building higher leverage skills and working on harder problems. I am no longer doing menial coding and research and debugging, no longer building skills for an era that is already over.

But I even will say that, if we tried to do this apples to apples, in some abstract sense, I still think that Claude is making me smarter overall. As the counterfactual makes clear, there is a tradeoff that happens when you use AI, that is to say I am getting smarter and dumber by using Claude all day. In terms of pure knowledge the difference comes down most to what I am learning, not how much. And in terms of my ability to think, I think that the amount of good thinking I am doing with Claude outweighs the risk of me relying on Claude as a crutch.

Finally, I will soften this by saying, it is hard to know. I don’t believe I have AI psychosis, but it is possible that my standards for intelligence are much lower than they used to be. I am also worried that building this “Claude crutch” has impacts that make me dumber outside of work. Do I think less critically now, and fall back to asking AI for other things in my life? Am I doing this on things where the process is more important, where the higher-level “learnings” are lower leverage? But that is a topic for another essay.

Appendix

Wanted to take a minute to add some caveats / questions that came up while I was writing this that didn’t fit cleanly into this essay. Some of these may turn into follow up essays.

• It is possible that Claude is not really making me smarter or dumber. This case would say that Claude is simply letting me not waste a senior engineer’s time asking questions, and lets me “skip” some of the entry-level learning I would have needed to do anyway. This is more a statement about Claude replacing and accelerating a function that already exists (talking to a senior engineer, my own research), rather than changing anything meaningful about how I’m thinking. (I would describe this as making me slightly smarter)

• There are good and bad ways of using Claude. I see many people becoming wholly reliant on Claude, and it seems to me that it is likely that they are not smarter because of it. The standard I hold myself to is essentially: If someone else will be reading it, I will make sure I understand it and can defend it myself. This means that any document I write, I write myself in my own words, and that I read every line of code that Claude writes on my behalf. I definitely think that this standard is critical to me learning from Claude, instead of offloading onto Claude.

• Related: Claude certainly makes me feel smarter, but the models have a tendency to be sycophants. If it is true I’ve lost some of my ability to distinguish smart from dumb, then this might be biasing me in the wrong direction.

• A lot of the “smarter” argument is predicated on me being early career. It is likely the case that there is a point in which Claude will stop making me smarter. If I already had a lot of industry experience, or was substantially smarter than Claude, then I think that the benefits might be smaller. The downsides too, are also likely much smaller. And in my opinion in this case, Claude probably mostly allows high level engineers to prioritize higher-leverage work.

• I think that writing is probably one of the skills that should be cultivated without AI. Producing an AI generated piece is certainly not the same mentally as taking the time to think through and write your argument yourself. At least, I am not sure what the higher-leverage skill you are building by having AI do the writing, or even help you with writing. (though this probably depends on the type of writing you are doing). I sympathize with Ezra Klein on not making connections unless you sit with the material, but I am conflicted on using AI for research.

• I find myself constantly writing code and working on projects that, in a pre-AI world, I would have no business working on, and need to rely on Claude to maintain that performance. Simultaneously, as mentioned in the essay, I often feel dumb when I have to fall back to asking Claude for something I should know. Am I losing the ability to really stretch my limits and work hard? Am I becoming just plain-old lazy? Am I being dishonest about my abilities?

Subscribe now

I think every programmer has noticed it. There used to be something special about the moment when a problem just “clicked”. There was a shining moment of insight where the solution laid itself in front of you. I used to race to do my CS homework for exactly this reason. I was addicted to thinking through and fighting through problems, just for that feeling of solving them in the end. I loved just staring at my computer for hours, looking at logs, even drawing the problem on paper.

But now we have a shortcut. I don’t think I need to go into it.

Well, maybe a little: This might be a hot take, but I think using AI is also extremely fun and addictive. I genuinely love working with Claude. Yes, “working with Claude”. I can’t believe I’m even saying it like that.

But, as productive as I am now, as fun as “vibe coding” is, it doesn’t scratch the same itch.

Leetle [dot] app

At the very end of 2024, when I was a senior in college, slightly inspired by Advent of Code, I built (with AI help, in age of Sonnet 3.6), a website called Leetle to help me with my interview prep. At the time, I couldn’t get myself to prepare for technical interviews, so I was failing them. I had the programming knowledge and the DSA knowledge, but I didn’t have the reps in solving problems under pressure.

So, I built Leetle, maybe obviously, as a play on LeetCode and Wordle. Every day, there is a new, simple coding challenge that you have 6 attempts to solve.

This was a side project for me that was good for my Twitter clout and for my portfolio, but I built it with the intention of actually doing the challenge myself every day. I think this is a project that only makes sense to even build in this transition period between the old world and the new one. I could have LLMs generate and verify the challenges, then do the challenge myself, completely blind to it. It’s a moment when I still needed to do LeetCode to pass my technical interviews, but I could use AI to build my own LeetCode. One that was more gamified and worked on mobile.

But shortly after I built Leetle, I did get a job.

It’s pretty undeniable that Leetle helped me land this job, both as a portfolio project and as a way to practice. But a side effect of getting a job was that I stopped playing the Leetle daily challenge. Once I got a job I decided I hated Leetle and LeetCode, DSA and practicing coding. I just wanted to build. All of this made me get lazy and forget to even tell Claude to generate new challenges for months at a time. I all but abandoned Leetle as a project.

Coding circa 2026

It is now March of 2026. More than a year after I built and launched Leetle. And for the first time since I was deep into job apps and technical interviews, I find myself playing Leetle again.

I don’t think of Leetle as mainly an interview-prep tool anymore. Ironically I think it’s closer to what I originally intended it for it: Wordle, for programmers, for me1. The problems aren’t hard, but it’s a small place for me to really think through a problem, type code by hand, and try out solutions myself. With Leetle now I can feel just a little bit of that “click” before jumping back into the trenches with Claude.

I spend all day at work coding with Claude. When I come home, if I have a side project idea, I can spin it up in an instant with Codex or Claude Code. I know if I really cared about coding, then I should build the old-fashioned way without LLMs! But, to me, there is a joy in building with LLMs. I genuinely do like it. And I’m not nostalgic for the old times all the time. In fact, I have been iterating on and upgrading Leetle more than ever thanks to LLMs. I’ve added features that would have taken me months, in hours with LLMs.

But sometimes I miss the “click”. And so I enjoy Leetle in a way that I never did before. I actually play Leetle like Wordle. Coding, for fun.

Coding, for fun

The point here isn’t Leetle itself, this is just my project and my outlet.

But I used to enjoy coding for coding’s sake, not for building’s sake. Write code on paper. Crack open the old LeetCode account. Open up a Codepen and center a div. Have Claude build you your own Leetle. Hell, play Monkeytype with the language set to Python. I think you’ll be surprised by how fun it is.

But there’s this wave of people and discourse and real studies that show that working out isn’t just good for your body and your muscles. It’s not just good for you health, it’s like, good for you mood. It’s like good for your sleep and for your energy.

I hear this from people in my real life, like “I feel better when I’m working out”. Look, either I’m doing it wrong or for some reason none of this applies to me. I feel like shit after working out. “But you just said you feel good—” immediately after I feel good and energized. But when I’m working out (like in a time in my life when I’m working out consistently), I get headaches, I feel weaker, I don’t sleep well. I could barely get out of bed all week because I was just so fucking tired. This doesn’t happen to me when I’m being a slob. I didn’t work out for years and felt great.

But here I am, working out again. Now maybe you’re thinking something like, oh it’s because you’re working out too hard, or you’re not recovering or something I AM. Be serious. I’m obviously not working out that hard. Like a quick 10-15min run. Some machines. Bench press dumbbell curl. I don’t do leg day. Only 3x a week. And what do I have to show for it? I just feel so tired. Maybe I’m not working out hard enough? That doesn’t seem right.

Anyway, I’ve only been working out seriously again for a couple weeks. Maybe things will mellow out. Maybe there is other factors affecting me. But all of these people and all of this data talking about how you’re more “mentally clear” or “more energized” when you’re exercising a bunch just doesn’t make sense to me. I think you’re all lying and coping in order to make yourself feel good about doing something you don’t want to do but you know you should do. I mean I’m sure it’s true for some people. Idk. Maybe I’m just so out of shape that trying to get in shape makes me feel more out of shape.

Maybe it’s that my body is expending more energy, so of course I feel more tired. Like these studies and anecdotes just don’t hold up to basic logic.

You know what I do endorse and do think makes me feel better? Drinking water. Holy wow I love drinking water. I’ve been on hella LMNT lately (like Liquid IV but even SALTIER) and I just love being hydrated. It gets rid of like 40% of my problems. Honestly it’s been a game changer.

Note how I said only 40%. What helps with the other 60% of my problems. Well I have another endorsement for you. Caffeine. Wow. What a rush.

Some backstory on me. I never drank coffee in high school. I barely did in college. I had this mentality that I had enough energy, and that I would drink caffeine when I really needed it. (I think my 14 year old self truly imagined myself as that fucking squirrel from Over the Hedge. Like not hyperactive, but like, I thought when I drank coffee, I would become the fucking Flash). Now in college I did go “super-speed” every once in a while, like finals week, and eventually I starting ripping it more and more. I fell in love with the trend of buying little drinks. You know I finally got it. It was starting to click. (And honestly, I think my 14 year old self was vindicated. Spoilers for the next paragraph, I don’t feel this anymore now that I drink coffee every day, but I remember the first time I had cold brew. This was my first real touch of caffeine. It was like I was floating. I was definitely thinking 2-3x faster than normal).

Then, I started a real job. And I started drinking coffee every day. It’s amazing. Now I haven’t gone all the way to drinking energy drinks yet, mostly because I don’t like the taste… (jk, it’s the squirrel thing again, y’all aren’t ready for my real lock-in).

But seriously, it’s just been a huge unlock to not rely on just “getting enough sleep” to feel awake all day. I can just drink another cup of coffee. Really need to lock in on something? Another cup of coffee? Oh we’re walking around now? Let’s go get some coffee. (I do know some people who have energy even when they don’t get enough sleep. I’M NOT ONE OF THEM)

Now potentially there is some correlation going on here, maybe I feel more tired because I’m drinking more caffeine than I ever was before? And now I’m like, reliant on it for energy? But that doesn’t really seem right. Look I didn’t mean to write another article about coffee.

Anyway now that I drink coffee every day if I ever stop, I get one of the worst headaches imaginable. Like genuinely I start panicking if it’s like 3pm and I haven’t had coffee because I start to feel the withdrawal headache in the (literal) back of my mind. Does this suck? Yes. But, from what I can tell, like millions of people are living like this so. I don’t know what to tell you. It’s not like I could just stop? I would get a headache? For like weeks apparently? That is a bit much honestly.

I’m thinking about it now and like. Let’s say apocalypse happens tomorrow, zombies everywhere nuclear bomb. And I don’t have my coffee? I’m going to be getting a nuclear headache and turn into a zombie. That was bad. But it’s a real fear! My life will suck in the apocalypse for at least the first two weeks, if only because of the headaches.

To wrap up, I want to tie this together with another thought I had when I was 14. If the zombie apocalypse happened, would I just have to have braces forever? Who would even be qualified to take them off? What if there were no more orthodontists? I guess that would be worse than having a headache for two weeks, braces forever.

In a sense my body now has found some kind of equilibrium. Working out makes me tired, coffee makes me awake. I think this is fine and probably what most people in their 30s feel like. But I am only 22. Anyway.

Should I start taking preworkout?

Weird Trader Joe’s

The reason I call it that is because it’s weird. To start, the whole thing is underground. You enter the Trader Joe’s then take an escalator down to it.

Thinking back to that Acquired episode, Trader Joe himself seemed to want to keep the stores small and dense. Limited SKUs, limited sprawl, stuff packed in every aisle. But weird Trader Joe’s is laid out like this: Half of the store is sparse, with not enough aisles so somehow items are still stuffed onto the shelves. But the other half, the checkout section, on the other hand, is 20 registers packed into two rows, with one line wrapping literally all the way through the store.

But, I mean, it’s clear people get their groceries there. It’s always packed. But again, it’s packed in a weird way. In my Trader Joe’s experience, it is usually crowded, aisles are bustling and there are long lines. But weird Trader Joe’s layout creates a whole new type of Trader Joe’s experience. You cannot get to half of the items in the store without being in line, or by awkwardly walking by the line while not being in line! Frankly, I don’t think this is what Trader Joe envisioned.

It was while I was in that horrible line that something caught my eye. Trader Joe’s Ready To Use Espresso Coffee. 15 espresso shots. $10. In a purple plastic bottle.

The K-Mini & Roommate Coffee

I am not a coffee fanatic. See I love coffee, and I drink it nearly every day, but my love is impure. I do like the taste of coffee, but realistically, I like it as a minor flavor in a coffee-drink.

In thinking about it, maybe this is a very normal take. I like my little drinks! I like my milk and sugar and my Iced Mint Mojitos and Chocolaccinos. My Starbucks order will have no less than 10 words in it. I only write that “my love is impure” because I’ve been surrounded by way too many coffee fanatics at my work! “See you need to get the beans just right or whatever, you need a good pour, you need this machine or that machine.” I just don’t do any of that!1

What I tend to do, is use a good old-fashioned coffee machine (what I’ve learned is called something like a “single serve K-cup drip coffee maker”, or some combination of those words). Preferably a shitty one that isn’t too complicated and that I can’t break. Preferably, the Keurig K-Mini, pictured below. And just give me the worst K-cups too. Just like some Starbucks ones or like Target brand if you can.

Before I left for college, I found that my Dad, who I had thought to be a prolific coffee drinker, had replaced his nicer Keurig, with, not a fancy espresso machine or some complicated contraption, but with a Keurig K-Mini. When I got to college, my roommate had one of these. In some sense “roommate coffee” is the best way to describe what this makes.2

But when I got to my first day of work, along with all of the fancy machines (which, thanks to chatgpt, I now understand to be an espresso machine, a cold brew machine, a pour-over, and a French press), I encountered the Keurig K-Elite. This is, as far as I can tell, the Keurig K-Mini but you can pour different size cups, though I’m not sure why you would ever want that. It is perfect and I use it every day.

So honestly, this is how I like my coffee-at-home. Roommate coffee. Load up the K-Mini with the cheapest K-cup you can find. Hope the machine doesn’t just give you water. Then add just a comical amount of milk and sugar. Even add ice if you’re feeling like it. Either way, the coffee will be lukewarm. Which is just how I like it.

But this is not how you make the worst coffee at home.

Terrible Coffee

Unsurprisingly, I bought the Trader Joe’s Ready To Use Espresso Coffee. I’m honestly not even sure what it is. Probably just espresso? Whatever the case, this morning I set out to craft my own “little drink”: an iced vanilla latte.

The plan was simple. Trader Joe’s Espresso. Milk. Ton of sugar. Splash of vanilla extract. Two or three ice cubes.

The coffee was doomed from the start. As expected, the Trader Joe’s Espresso does not taste very good (in my “research” for this article, I found many Reddit threads raving about it, though it seems like they like the caffeine content and bear the taste). Even so, the flavor of the espresso shouldn’t be a problem because we’re adding a ton of sugar. The sugar I’m using? Also Trader Joe’s! This should be a match made in heaven, but for anyone who has used Trader Joe’s sugar before, you know that it is fucking huge. I might as well have tried to dissolve a pebble into this drink. On top of that. THE DRINK ISN’T HOT.

The sugar didn’t dissolve. The vanilla wasn’t strong enough. What I got was a bitter & sour glass of “coffee”. With no water either, it’s consistency was basically that of milk.

At last I had done it. Something worse than Roommate coffee. Terrible coffee.

Wow it was hard to get down. But the last sip, the one that pulled all of the undissolved sugar into my mouth, was heaven.

Variation: For an extra-thick “mocha” variant of terrible coffee, substitute normal milk for a Fairlife protein shake (30g of protein). Jesus.

What a ride.

I would never claim to be a longtime fan of the Alien or Predator franchises. In fact I am a very new fan. But I am a diehard new fan (of Alien at least). I watched the original Alien about a year ago, then 6 months later went on a binge of Predator, Prey, Aliens, Alien 3, Alien: Resurrection, Prometheus, and Alien: Covenant in about a week. A little later, I watched Alien: Romulus. I didn’t like Predator or Prey enough to continue the series.

My opinions on these movies range from “pretty good” to “pretty bad”, but there is one thing that I think all Alien and Predator fans can agree on: the Alien movies all have about the same plot as each other. And, assuming nothing major is different about the later Predator movies, the same goes for Predator. This dawned on me as early as Aliens, when I wrote this in my Letterboxd review:

This is the same movie as Alien.

And by the time I got to Romulus, I appreciated this fact fully:

If the other Alien movies didn’t exist, this would probably be the best movie ever right? … I can’t wait to see what this franchise does next (probably the same movie) (i like these but it’s now 7 of the same fucking movie).

That being said, I definitely like Alien movies, and I definitely like them more than I liked the Predator movies I did watch. So, about 6 more months after I watched Romulus, I decided to jump into the Alien spinoff movies, AVP and AVP: Requiem.

A LOT has been said about these movies, mostly that they are bad. But, while I don’t think they’re good, I want to try to present a new take: AVP is a bad Alien movie, and AVP: Requiem is a really bad Predator movie.

AVP is a bad Alien movie

At its core, AVP is a classic Alien movie. It follows all of the same beats: a team in an enclosed space, escaping Aliens, and our our sole female protagonist barely makes it out, this time, with the help of a predator. The setup is actually almost identical to Prometheus, and that sense of isolation really makes it feel like Alien or Aliens.

But, it is by no means a good Alien movie. To be fair, the effects are great and the fights are exciting (and well-lit). I did like this movie a lot, but not as an Alien movie. I think that Alien is at its best when it is 1. actually scary (AVP is not at all…), and 2. exploring big-ish sci-fi themes. This movie. I mean it takes place in 2004. The only world-building is done very overtly in a flashback! And there are no androids! This is a huge part of the Alien lore! It just really lacks that mysterious-existential-anti-corporate-edge of the good Alien movies. So while AVP categorically an Alien movie, it’s not a very good one.

AVP: Requiem is a bad Predator movie

This movie might be one of the worst movies I’ve ever seen. And SEEN is a strong word. This is not just an aesthetic complaint, this is a crucial part of why this movie is worse than AVP. In AVP, even when the story was boring or the characters weren’t being interesting, you were watching Predators and Aliens fight! This is not true in this movie. Even when the Aliens and Predator DO fight, which is rare, you cannot see a thing. The ultimate battle between the Predator and the (horribly named) PredAlien is complete visual nonsense.

And in my initial review of AVP: Requiem I wrote:

If I wanted to watch aliens dunk on plumbers and electricians I would have watched Alien.

But in thinking about it, I realized that AVP: Requiem is formulaically much more of a Predator movie! A group of people are hunted out in the open by an alien threat, only this time it’s Predators and Aliens hunting. In fact, other than the Aliens being there, AVP: Requiem shares almost nothing with the formula or feeling of any of the Alien movies.

But, to be clear, this is a really bad Predator movie. In the original, Schwarzenegger’s men are picked off and skinned alive for most of the movie’s runtime! But! There is this warrior-code thing going on, there’s honor and stakes and all of that. In AVP: Requiem, no one fights back, the Aliens and Predators kill high schoolers and pregnant women and babies for shock value. And ultimately, the film rewrites the Predator ending, instead of the humans winning, the government NUKES THE ENTIRE TOWN. What a mess!

What does it all mean?

Alien movies ask us big questions: Why are we here? Do corporations have our best interests in mind? What is up with that space jockey? And why did that girl run away in a straight line in Prometheus?

Predator movies ask us small questions: Who is the biggest guy? Will they get to the chopper?

Alien vs Predator movies ask a single question, a far more important question: who would win, an Alien or a Predator?

In some sense the AVP movies try to give us a meta-answer this question: it is the Alien franchise versus the Predator franchise. The formulas go back to back and head to head in the crossovers, with AVP representing Alien and AVP: Requiem representing Predator.

This is, of course, not really a fair fight because AVP: Requiem almost completely butchers the Predator formula. And it is also just a much worse movie. But, in looking at these movies as Alien and Predator movies, I think I saw more clearly how different the formulas are, and why I keep coming back to one but not the other.

Or maybe I wrote this whole thing because I’m annoyed that the Aliens were turned into Predator-fodder in both movies. Suck it Predator fans, your movies will always be worse.1