I've noticed a lot of these posts tend to go codex vs claude, but as author is someone who does AI workshops curious why Cursor is left out of this post (and more generally posts like this).
From my personal experience I find cursor to be much more robust because rather than "either / or" its both and can switch depending on the time or the task or whatever the newest model is.
It feels like the same way people often try to avoid "vendor lock in" in software world that Cursor allows freedom for that, but maybe I'm on my own here as I don't see it naturally come up in posts like these as much.
Speaking from personal experience and talking to other users - the agents/harnesses of the vendors are just better and they are customized for their own models.
what kinds of tasks do you find this to be true for? For a while I was using claude code inside of the cursor terminal, but I found it to be basically the same as just using the same claude model in there.
Presumably the harness cant be doing THAT much differently right? Or rather what tasks are responsibilities of the harness could differentiate one harness from another harness
This becomes clearer for me with harder problems or long running tasks and sessions. Especially with larger context.
Examples that come to mind are how the context is filled up and how compaction works. Both Codex and Claude Code ship improvements regarding this specific to their own models and I’m not sure how this is reflected in tools like Cursor.
I feel you brother/sister. I actually pay for Claude Code Max and also for the $20/mo Cursor plan. I use Claude Code via the VSCode extension running within the Cursor IDE. 95% of my usage is Claude Code via that extension (or through the CLI in certain situations) but it's great having Cursor as a backup. Sometimes I want to have another model check Claude's work, for example.
Github Copilot also allows you to use both models, codex, claude, and gemini on top.
Cursor has this "tool for kids" vibe, it's also more about the past - "tab, tab, enter" low-level coding versus the future - "implement task 21" high level delegating.
I got a student subscription to cursor and after giving it a good 6 hours I’ve abandoned it.
I extremely dislike the way it goes forth and bolts. I don’t trust these tools enough to just point it in the direction and say go, I like to be a human in the loop. Perhaps the use case I was working on then was difficult (quite old react native library upgrade across a medium sized codebase) but I eventually cracked this on Claude; cursor in both entropic and Gemini left me with an absolute mess.
Even repeatedly asking the prompt to keep me in the loop it kept on just running haywire.
Heya, author here! That's a great question! I fully understand the vendor lock-in concern, but I'll just quickly note that when it comes to a first workshop I do whatever makes the person most comfortable. I let the attendee choose the tool they want — with a slight nudge towards Codex or Claude Code for reasons I'll mention below. But if they want to do the workshop in Cursor, VS Code, or heck MS Paint — I'll try to find a way to make it work as long as it means they're learning.
I actually started teaching these workshops by using Cursor, but found that it fell short for a few reasons.
Note: The way that my workshops work is that you have three hours to build something real. It may be scoped down like a single feature or a small app or a high quality prototype, but you'll walk away with what you wanted to build. More importantly you'll have learned the fundamentals of working with AI in the process, so you can continue this on your own and see meaningful results. We go through various exercises to really understand good prompting (since everyone thinks they're good but they rarely are), how to build context for models, and explore the landscape of tools that you can use to get better results. A lot of that time is actually spent in a Google Doc that I've prepped with resources — and the work we do there makes the code practically write itself by the time we're done.
Here's a short list of why I don't default to Cursor:
1. As I noted in another comment, the model performance is just so much better [^1] when accessed directly through Codex and Claude Code, which means more promising results more quickly. Previously the workshops were 3-4 hours just to finish, now it's a solid 3 with time to ask questions afterwards. You can't beat this experience, because it gives the student more time to pause and ask questions, seep in what they've done, and not spend time trying to understand the tools just to see results.
1a. The amount of time it took someone to set up Cursor was pretty long. The process for getting a good set up is pretty long — especially for someone non-technical. This may not be as big of a deal for developers using Cursor — but even they don't know a lot of the settings and tweaks to make to get Cursor to be great out the box.
2. The user experience of dropping a prompt into Codex/Claude Code and watch it start solving a problem is pretty amazing. I love GUIs — I spend my days building one [^3], but the TUI melting away everything to just being chat is an advantage when you have no mental model for how this stuff works.
3. As I said in #1, the results are just better. That's really the main reason! I
Not to toot my own horn, but the process works. These are all testimonials in the words of people who have attended a workshop, and I'm very proud of how people not only learn during the workshop but how it sets them off on a good path afterwards. [^2]. I have people messaging me 24 hours later telling me that they built an app their partner has wanted for years, to tell me that they've completed the app we started and it does everything they dreamed of, and hear more process over the weeks and months after because I urge them to keep sending me their AI wins. (It's truly amazing how much they grow, and I now have attendees teaching ME things — the ultimate dream of being a teacher knowing you gave them the nudge they needed.)
Hope that helps and isn't too much of an ad — I really just want to make it clear that I try to do what works best and if the best way to help people learn changes I will gladly change how I work. :)
Agreed... also fwiw I don't think that langauge-dependent games are as much of a barrier as it used to be. I've built a game recently that I easily localized first with real-time AI translations and then later with more static language translations.
Anyway I think this would be an amazing thing to let other people contribute to as this is an entire industry of hypercasual games which could easily be ported to this minus the annoying ads
I think the issue with language-dependant games is not just knowing the correct translation - as OP points out, it's more about being funny or clever on the spot, which usually requires a certain level of understanding of the nuances of the language.
Exactly this! Translating the games themselves is not a big deal as that can be automated (although the quality of LLM-translations is not always the best) but when it comes to user generated responses given in a quick timeframe, that's when non-native english players struggle the most, at least in our own friend groups.
Im not fully convinced by "a computer can never be held accountable"
We already delegate accountability to non-humans all the time:
- CI systems block merges
- monitoring systems page people
- test suites gate different things
In practice accountability is enforced by systems, not humans.. humans are defintiely "blamed" after the fact, but the day-to-day control loop is automated.
As agents get better at running code, inspecting ui state, correlating logs, screenshots, etc they're starting to operationally be "accountable" and preventing bad changes from shipping and producing evidence when something goes wrong .
At some point humans role shifts from "i personally verify this works" to "i trust this verification system and am accountable for configuring it correctly".
Thats still responsibility, but kind of different from whats described here. Taken to a logical extreme, the arguement here would suggest that CI shouldn't replace manual release checklists
I need to expand on this idea a bunch, but I do think it's one of the key answers to the ongoing questions people have about LLMs replacing human workers.
Human collaboration works on trust.
Part of trust is accountability and consequences. If I get caught embezzling money from my employer I can lose my job, harm my professional reputation and even go to jail. There are stakes!
I computer system has no stakes, and cannot take accountability for its actions. This drastically limits what it makes sense to outsource to that system.
A lot of this comes down to my work on prompt injection. LLMs are fundamentally gullible: an email assistant might respond to an email asking for the latest sales figures by replying with the latest (confidential) sales figures.
If my human assistant does that I can reprimand or fire them. What am I meant to do with an LLM agent?
I don't think this is very hard. Someone didn't properly secure confidential data and/or someone gave this agent access to confidential data. Someone decided to go live with it. Reprimand them, and disable the insecure agent.
I've given you a disagree-and-upvote; these things are significant quality aids, but they are like the poka-yoke or manufacturing jig or automated inspection.
Accountability is about what happens if and when something goes wrong. The moon landings were controlled with computer assistance, but Nixon preparing a speech for what happened in the event of lethal failure is accountability. Note that accountability does not of itself imply any particular form or detail of control, just that a social structure of accountability links outcome to responsible person.
Right, so how do you hold these things accountable? When your CI fails, what do you do? Type in a starkly worded message into a text file and shut off the power for three hours as a punishment? Invoice Intel?
Well, we're not there yet, but I do envision a future, where some AIs work for as independent contractors with their own bank accounts that they want to maximize, and if such an AI fails in a bad way, its client would be able to fine it, fire it or even sue it, so that it, and the human controlling it would be financially punished.
Humans are only kind of held accountable. If you ship a bug do you go to jail? Even a bug so bad it puts your company out of business. Would there be any legal or physical or monetary consequences at all for you, besides you lose your job?
So the accountability situation for AI seems not that different. You can fire it. Exactly the same as for humans.
those systems include humans —- they are put in place by humans (or collections of them) that are the accountability sink
if you put them (without humans) in a forrest they would not survive and evolve (they are not viable systems alone); they are not taking action without the setup & maintenance (& accountability) of people
Why do you think that this other kind of accountability (which reminds me of the way captain's or commander's responsibility is often described) is incompatible with what the article describes? Due to the focus on necessity of manual testing?
I mean I suppose you can continuously add "critical feedback" to the system prompt to have some measure of impact on future decision-making, but at some point you're going to run out of space and ultimately I do not find this works with the same level of reliability as giving a live person feedback.
Perhaps an unstated and important takeaway here is that junior developers should not be permitted to use an LLMs for the same reason they should not hire people: they have not demonstrated enough skill mastery and judgement to be trusted with the decision to outsource their labor. Delegating to a vendor is a decision made by high-level stakeholders, with the ability to monitor the vendor performance, and replace the vendor with alternatives if that performance is unsatisfactory. Allowing junior developers to use LLM is allowing them to delegate responsibility without any visibility or ability to set boundaries on what can be delegated. Also important: you cannot delegate personal growth, and by permitting junior engineers to use an LLM that is what you are trying to do.
You completely missed the point of that quote. The point of the quote is to highlight the fact that automated systems are amoral, meaning that they do not know good or evil and cannot make judgements that require knowing what good and evil mean.
LOC is a bad quality metric, but its a reasonable proxy in practice..
Teams generally don't keep merging code that "doesn't work" for long... prod will brake, users will push back fast. So unless the "wrongness" of the AI-generated code is buried so deeply that it only shows up way later, higher merged LOC probably does mean more real output.
Its just not directly correlated there is some bloat associated too.
So that caveat applies to human-written code too, which we tend to forget. There's bloat and noise in the metric, but its not meaningless
Agreed, there is some correlation between productivity and LoC.
That said the correlation it’s weak; and does not say anything about quality (if anything quality might be inversely correlated; which too would be a very weak signal)
For instance if I push 10kloc that are in a lib I would have used if I were not using AI, yes, I have pushed much more code, but I was not more productive.
After 4 hours of vibe coding I feel as tired as a full day of manual coding. The speed can be too much. If I only use it for a few minutes or an hour, it feels energising.
compression is exactly what is missing for me when using agents, reading their approach doesn't let me compress the model in my head to evaluate it, and that was why i did programming in the first place.
Most hilarious part of this is that if you've ever watched "The Challenge" then you know that these people, truly, often cannot add 3 digit numbers together let alone understand information theory
I generally like the daily habit puzzle games and I play a ton of daily chess, but this one doesn't give me much time to think and reflect on my moves (either pre or post moving on to the next puzzle)
Would be nice to add something like that... i think review mode is a huge reason why chess.com puzzles / chess.com is so popular because you leave a little smarter than you came
You are right and I love the idea! I have added a review game section where after the game the player can view each board and move made to think and reflect on each daily puzzle.
One thing this really highlights to me is how often the "boring" takes end up being the most accurate. The provocative, high-energy threads are usually the ones that age the worst.
If an LLM were acting as a kind of historian revisiting today’s debates with future context, I’d bet it would see the same pattern again and again: the sober, incremental claims quietly hold up, while the hyperconfident ones collapse.
Something like "Lithium-ion battery pack prices fall to $108/kWh" is classic cost-curve progress. Boring, steady, and historically extremely reliable over long horizons. Probably one of the most likely headlines today to age correctly, even if it gets little attention.
On the flip side, stuff like "New benchmark shows top LLMs struggle in real mental health care" feels like high-risk framing. Benchmarks rotate constantly, and “struggle” headlines almost always age badly as models jump whole generations.
I bet theres many "boring but right" takes we overlook today and I wondr if there's a practical way to surface them before hindsight does
"Boring but right" generally means that this prediction is already priced in to our current understanding of the world though. Anyone can reliably predict "the sun will rise tomorrow", but I'm not giving them high marks for that.
I'm giving them higher marks than the people who say it won't.
LLMs have seen huge improvements over the last 3 years. Are you going to make the bet that they will continue to make similarly huge improvements, taking them well past human ability, or do you think they'll plateau?
right, because if there is one thing that history shows us again and again is that things that have a period of huge improvements never plateau but instead continue improving to infinity.
Improvement to infinity, that is the sober and wise bet!
The prediction that a new technology that is being heavily researched plateaus after just 5 years of development is certainly a daring one. I can’t think of an example from history where that happened.
Claiming that AI in anything resembling its current form is older than 5 years is like claiming the history of the combustion engine started when an ape picked up a burning stick.
Your analogy fails because picking up a burning stick isn’t a combustion engine, whereas decades of neural-net and sequence-model work directly enabled modern LLMs. LLMs aren’t “five years old”; the scaling-transformer regime is. The components are old, the emergent-capability configuration is new.
Treating the age of the lineage as evidence of future growth is equivocation across paradigms. Technologies plateau when their governing paradigm saturates, not when the calendar says they should continue. Supersonic flight stalled immediately, fusion has stalled for seventy years, and neither cared about “time invested.”
Early exponential curves routinely flatten: solar cells, battery density, CPU clocks, hard-disk areal density. The only question that matters is whether this paradigm shows signs of saturation, not how long it has existed.
I don't think it would make much sense to hunt large predators prior to the invention of agriculture, even though early humans were probably plenty smart enough to build traps capable of holding animals like tigers. But after that (less than 40k years ago, more than 10k years ago), I'd bet it was a common-ish thing for humans to try to hunt predators that preyed upon their livestock.
Tigers are terrifying, though. I think it takes extreme or perverse circumstances to make hunting a tiger make any sense at all. And even then, traps and poisons make more sense than stalking a tiger to kill it!
LaunchHN: Announcing Twoday, our new YC backed startup coming out of stealth mode.
We’re launching a breakthrough platform that leverages frontier scale artificial intelligence to model, predict, and dynamically orchestrate solar luminance cycles, unlocking the world’s first synthetic second sunrise by Q2 2026. By combining physics informed multimodal models with real time atmospheric optimisation, we’re redefining what’s possible in climate scale AI and opening a new era of programmable daylight.
You joke, but, alas, there is a _real_ company kinda trying to do this. Reflect Orbital[1] wants to set up space mirrors, so you can have daytime at night for your solar panels! (Various issues, like around light pollution and the fact that looking up at the proposed satellites with binoculars could cause eye damage... don't seem to be on their roadmap.) This is one idea that's going to age badly whether or not they actually launch anything, I suspect.
Battery tech is too boring, but seems more likely to manage long-term effectiveness.
Reflecting sunlight from orbit is an idea that had been talked about for a couple of decades even before Znamya-2[1] launched in 1992. The materials science needed to unfurl large surfaces in space seems to be very difficult, whether mirrors or sails.
A lot of the press likes to paint “AI” as a uniform field that continues to improve together. But really it’s a bunch of related subfields. Once in a blue moon a technique from one subfield crosses over into another.
“AI” can play chess at superhuman skill. “AI” can also drive a car. That doesn’t mean Waymo gets safer when we increase Stockfish’s elo by 10 points.
They're already better than you at reciting historical facts. I'd guess they're probably better at composing poems (they're not great but far better than the average person).
Or you agree with me? I'm not looking for prescience marks, I'm just less convinced that people really make the more boring and obvious predictions.
What is an intellectual task? Once again, there's tons of stuff LLMs won't be trained on in the next 3 years. So it would be trivial to just find one of those things and say voila! LLMs aren't better than me at that.
I'll make one prediction that I think will hold up. No LLM-based system will be able to take a generic ask like "hack the nytimes website and retrieve emails and password hashes of all user accounts" and do better than the best hackers and penetration testers in the world, despite having plenty of training data to go off of. It requires out-of-band thinking that they just don't possess.
I'll take a stab at this: LLMs currently seem to be rather good at details, but they seem to struggle greatly with the overall picture, in every subject.
- If I want Claude Code to write some specific code, it often handles the task admirably, but if I'm not sure what should be written, consulting Claude takes a lot of time and doesn't yield much insight, where as 2 minutes with a human is 100x more valuable.
- I asked ChatGPT about some political event. It mirrored the mainstream press. After I reminded it of some obvious facts that revealed a mainstream bias, it agreed with me that its initial answer was wrong.
These experiences and others serve to remind me that current LLMs are mostly just advanced search engines. They work especially well on code because there is a lot of reasonably good code (and tutorials) out there to train on. LLMs are a lot less effective on intellectual tasks that humans haven't already written and published about.
To be clear, you are suggesting “huge improvements” in “every intellectual task”?
This is unlikely for the trivial reason that some tasks are roughly saturated. Modest improvements in chess playing ability are likely. Huge improvements probably not. Even more so for arithmetic. We pretty much have that handled.
But the more substantive issue is that intellectual tasks are not all interconnected. Getting significantly better at drawing hands doesn’t usually translate to executive planning or information retrieval.
Sorry, I now realize this thread is about whether LLMs can improve on tasks and not whether AI can. Agreed there’s a lot of headroom for LLMs, less so for AI as a whole.
> They're already better than you at reciting historical facts.
They're better at regurgitating historical facts than me because they were trained on historical facts written by many humans other than me who knew a lot more historical facts. None of those facts came from an LLM. Every historical fact that isn't entirely LLM generated nonsense came from a human. It's the humans that were intelligent, not the fancy autocomplete.
Now that LLMs have consumed the bulk of humanity's written knowledge on history what's left for it to suck up will be mainly its own slop. Exactly because LLMs are not even a little bit intelligent they will regurgitate that slop with exactly as much ignorance as to what any of it means as when it was human generated facts, and they'll still spew it back out with all the confidence they've been programed to emulate. I predict that the resulting output will increasingly shatter the illusion of intelligence you've so thoroughly fallen for so far.
> At what? They're already better than me at reciting historical facts.
I wonder what happens if you ask deepseek about Tiananmen Square…
Edit: my “subtle” point was, we already know LLMs censor history. Trusting them to honestly recite historical facts is how history dies. “The victor writes history” has never been more true. Terrifying.
LLMs aren't getting better that fast. I think a linear prediction says they'd need quite a while to maybe get "well past human ability", and if you incorporate the increases in training difficulty the timescale stretches wide.
Surely you meant the latter? The boring option follows previous experience. No technology has ever not reached a plateau, except for evolution itself I suppose, till we nuke the planet.
Prediction markets have pretty much obviated the need for these things. Rather than rely on "was that really a hot take?" you have a market system that rewards those with accurate hot takes. The massive fees and lock-up period discourage low-return bets.
FWIW Polymarket (which is one of the big markets) has no lock-up period and, for now while they're burning VC coins, no fees. Otherwise agree with your point though.
As opposed to the current world of brigading social media threads to make consensus look like it goes your way and then getting journalists scraping by on covering clickbait to cover your brigading as fact?
The one about LLMs and mental health is not a prediction but a current news report, the way you phrased it.
Also, the boring consistent progress case for AI plays out in the end of humans as viable economic agents requiring a complete reordering of our economic and political systems in the near future. So the “boring but right” prediction today is completely terrifying.
“Boring” predictions usually state that things will continue to work the way they do right now. Which is trivially correct, except in cases where it catastrophically isn’t.
So the correctness of boring predictions is unsurprising, but also quite useless, because predicting the future is precisely about predicting those events which don’t follow that pattern.
This suggests that the best way to grade predictions is some sort of weighting of how unlikely they were at the time. Like, if you were to open a prediction market for statement X, some sort of grade of the delta between your confidence of the event and the “expected” value, summed over all your predictions.
Exactly, that's the element that is missing. If there are 50 comments against and one pro and that pro has it in the longer term then that is worth noticing, not when there are 50 comments pro and you were one of the 'pros'.
Going against the grain and turning out right is far more valuable than being right consistently when the crowd is with you already.
Yeah a simple of total points of pro comments vs total points of con comments may be simple and exact enough to simulate a prediction market. I don't know if it can be included in the prompt or better to be vibecoded in directly.
It's because algorithmic feeds based on "user engagement" rewards antagonism. If your goal is to get eyes on content, being boring, predictable and nuanced is a sure way to get lost in the ever increasing noise.
> One thing this really highlights to me is how often the "boring" takes end up being the most accurate.
Would the commenter above mind sharing the method behind of their generalization? Many people would spot check maybe five items -- which is enough for our brains to start to guess at potential patterns -- and stop there.
On HN, when I see a generalization, one of my mental checklist items is to ask "what is this generalization based on?" and "If I were to look at the problem with fresh eyes, what would I conclude?".
Is this why depressed people often end up making the best predictions?
In personal situations there's clearly a self fulfilling prophecy going on, but when it comes to the external world, the predictions come out pretty accurate.
I keep seeing these Grok 4 intelligence claims, so I tried something very simple: "Animate a round robin tournament for 10 people."
Results:
Claude: ~10s, perfect working demo
ChatGPT: ~20s, solid solution
Grok 4: ~1000s, failed completely, gave me a truncated base64 blob
This wasn't some obscure edge case... it was basic data visualization that any decent model should handle. Yet somehow Grok 4 is "competing with humans" and has "99% tool accuracy"...
From my personal experience I find cursor to be much more robust because rather than "either / or" its both and can switch depending on the time or the task or whatever the newest model is.
It feels like the same way people often try to avoid "vendor lock in" in software world that Cursor allows freedom for that, but maybe I'm on my own here as I don't see it naturally come up in posts like these as much.