Schedule tasks on the web
code.claude.com
code.claude.com
also, someone rightly predicted this rugpull coming in when they announced 2x usage - https://x.com/Pranit/status/2033043924294439147
The same as charging a different toll price on the road depending on the time of day.
Someone spread FUD on the internet, incorrectly, and now others are spreading it without verifying.
Yes, it was FUD, but ended up being correct. With the track record that Anthropic has (e.g. months long denial of dumbed down models last year, just to later confirm it as a "bug"), this just continues to erode trust, and such predictions are the result of that.
I'm not sure it's a rug pull when their stats show 7% and 2% subscription-level impacts. We're back in the ISP days, and they never said unlimited.
I was trying to get The alibaba plan but missed the mark. I'm curious to try out the Minimax coding plan (#10/mo) or Kimi ($20/mo) at some point to see how they stack up.
For Pricing: GLM was $180 for a year of their pro tier during a black friday sale and GHCP was $100/year but they don't have the annual plan any more so it is now $120. Alibaba's only coding plan today is $50/mo, too rich for me.
If you want stability, own the means of inference and buy a Mac Studio or Strix Halo computer.
At least in my company we are close to that flywheel.
Tickets may well not look like they do now, but some semblance of them will exist. I'm sure someone is building that right now.
No. It's not Jira.
We dont have product managers or technical ticket writers of any sort
But us devs are still choosing how to tackle the ticket, we def don't have to as I’m solving the tickets with AI. I could automate my job away if I wanted, but I wouldn't trust the result as I give a degree of input and steering, and there’s bigger picture considerations its not good at juggling, for now
I am already at the point where because it is just the two of us, the limiting factor is his own needs, not my ability to ship features.
Either way… we badly need more innovation in inference price per performance, on both the software and hardware side. It would be great if software innovation unlocked inference on commodity hardware. That’s unlikely to happen, but today’s bleeding edge hardware is tomorrow’s commodity hardware so maybe it will happen in some sense.
If Taalas can pull off burning models into hardware with a two month lead time, that will be huge progress, but still wasteful because then we’ve just shifted the problem to a hardware bottleneck. I expect we’ll see something akin to gameboy cartridges that are cheap to produce and can plug into base models to augment specialization.
But I also wonder if anyone is pursuing some more insanely radical ideas, like reverting back to analog computing and leveraging voltage differentials in clever ways. It’s too big brain for me, but intuitively it feels like wasting entropy to reduce a voltage spike to 0 or 1.
But this is just not true, otherwise companies that can already afford such high prices would have already outpaced their competitors.
And I sense you would have to throw orders of magnitude more tokens to get meaningfully better results (If anyone has access to experiments with GPT 5 class models geared up to use marginally more tokens with good results please call me out though).
Sadly enough I have not seen this happening in a long time.
If this direction holds true, ROI cost is cheaper.
Instead of employing 4 people (Customer Support, PM, Eng, Marketing), you will have 3-5 agents and the whole ticket flow might cost you ~20$
But I hope we won't go this far, because when things fail every customer will be impacted, because there will be no one who understands the system to fix it
But the entire SWE apparatus can be handled.
Automated A/B testing of the feature. Progressive exposure deployment of changes, you name it.
Of course it's in the areas where it doesn't matter as much, like experiments, internal tooling, etc, but the CTOs will get greedy.
“There are 4 Rs in the word “burberrorrly.” Here they are highlighted: burberrorrly (positions 3, 6, 7, 9)”
Obviously not a real word, but perhaps the fundamental concept remains
All Chinese labs have to do to tank the US economy is to release open-weight models that can run on relatively cheap hardware before AI companies see returns.
Maybe that's why AI companies are looking to IPO so soon, gotta cash out and leave retail investors and retirement funds holding the bag.
Regarding the latter, smaller models are really good for what they are (free) now, they'll run on a laptop's iGPU with LPDDR5/DDR5, and NPUs are getting there.
Even models that can fit in unified 64GB+ memory between CPU & iGPU aren't bad. Offloading to a real GPU is faster, but with the iGPU route you can buy cheaper SODIMM memory in larger quantities, still use it as unified memory, eventually use it with NPUs, all without using too much power or buying cards with expensive GDDR.
Qwen-3.5 locally is "good enough" for more than I expected, if that trend continues, I can see small deployable models eventually being viable & worthy competition, or at least being good enough that companies can run their own instead of exfiltrating their trade secrets to the worst people on the planet in real-time.
I suspect, but don't know, some fields of inquiry will be fruitful when it comes to "good enough" small models. Especially when it comes to constrained tasks like software development. Software development models don't have to generalize to anything a chatbot can be asked or tasked with, the space it's required to generalize on is pretty small compared to literally the whole world.
If I was a betting man, I'd put my money where my mouth is, but I'm not. I am betting with my time and focus that smaller local models are worth it, and will be worth it, though.
Maybe for a non-dev it would be nice to submit a ticket and have it auto-fixed by an agent. But in the devs case, it feels like it would be faster to just do it manually.
I still can't get a good mental model for when these things will work well and when they won't. Really does feel like gambling...
Art isn't about being cool. Art is about context.
When I tell people that art cannot be unpolitical, they react strongly, because they think about the left/right divide and how divided people are, where art is supposed to be unifying.
But art is like movement, you need an origin and a destination. Without that context, it will be just another... thing. Context makes it something.
A PR tells me what changed, but not how an AI coding session got there: which prompts changed direction, which files churned repeatedly, where context started bloating, what tools were used, and where the human intervened.
I ended up building a local replay/inspection tool for Claude Code / Cursor sessions mostly because I wanted something more reviewable than screenshots or raw logs.
(That's basically what A/B testing is about.)
It's the "robots will just build/repair themselves" trope but the robots are agents
Oh wait. That's already here and is working fine.
There's a lots of experimentation right now, but one thing that's guaranteed is that the data gatekeepers will slam the door shut[1] - or install a toll-booth when there's less money sloshing about, and the winners and losers are clear. At some point in the future, Atlassian and Github may not grant Anthropic access to your tickets unless you're on the relevant tier with the appropriate "NIH AI" surcharge.
1. AI does not suspend or supplant good old capitalism and the cult of profit maximization.
Don’t you guys have hard business problems where AI just cant solve it or just very slowly and it’s presenting you 17 ideas till it found the right one. I’m using the most expensive models.
I think the nature of AI might block that progress and I think some companies woke up and other will wake up later.
The mistake rate is just too high. And every system you implement to reduce that rate has a mistake rate as well and increases complexity and the necessary exploration time.
I think a big bulk of people is of where the early adaptors where in December. AI can implement functional functionality on a good maintained codebase.
But it can’t write maintable code itself. It actually makes you slower, compared to assisted-writing the code, because assisted you are way more on the loop and you can stop a lot of small issues right away. And you fast iterate everything•
I’ve not opened my idea for 1 months and it became hell at a point. I’ve now deleted 30k lines and the amount of issues I’m seeing has been an eye-opening experience.
Unscalable performance issues, verbosity, straight up bugs, escape hatches against my verification layers, quindrupled types.
Now I could monitor the ai output closer, but then again I’m faster writing it myself. Because it’s one task. Ai-assisted typing isn’t slower than my brain is.
Also thinking more about it FAANG pays 300$ per line in production, so what do we really trying to achieve here, speed was never the issue.A great coder writes 10 production lines per day.
Accuracy, architecture etc is the issue. You do that by building good solid fundamental blocks that make features additions easier over time and not slower
- performance is continuing to increase incredibly quickly, even if you rightfully don’t trust a particular evaluation. Scaling laws like chinchilla and RL scaling laws (both training and test time)
- coding is a verifiable domain
The second one is most important. Agent quality is NOT limited by human code in the training set, this code is simply used for efficiency: it gets you to a good starting point for RL.
Claiming that things will not reach superhuman performance, INCLUDING all end to end tasks: understanding a vague business objective poorly articulated, architecting a system, building it out, testing it, maintaining it, fixing bugs, adding features, refactoring, etc. is what requires the burden of proof because we literally can predict performance (albeit it has a complicated relationship with benchmarks and real world performance).
Yes definitely, error rates are too high so far for this to be totally trusted end to end but the error rates are improving consistently, and this is what explains the METR time horizon benchmark.
The number of devs will reduce but there will still be large activities that can't be farmed out without an overall strategy
The other thing you're missing here is generalizability. Better coding performance (which is verifiable and not limited by human data quality) generalizes performance on other benchmarks. This is a long known phenomenon.
Because it cannot do it?
Every investment has a date where there should be a return on that investment. If there’s no date, it’s a donation of resources (or a waste depending on perspective).
You may be OK with continuing to try to make things work. But others aren’t and have decided to invest their finite resources somewhere else.
Ah ok so you didn't really read my comment, what is your counter argument? Models are just fundamentally incapable of understanding business context? They are demonstrably already capable of this to a large extent.
> Every investment has a date where there should be a return on that investment. If there’s no date, it’s a donation of resources (or a waste depending on perspective).
what are you implying here? This convo now turns into the "AI is not profitable and this is a house of cards" theme? That's ok, we can ignore every other business model like say Uber running at a loss to capture what is ultimately an absolutely insane TAM. Little ol' Uber accumuluated ~33B in losses over 14 years, and you're right they tanked and collapsed like a dying star...oh wait...hmm interesting I just looked at their market cap and it's 141 Billion.
> You may be OK with continuing to try to make things work. But others aren’t and have decided to invest their finite resources somewhere else.
I truly love that. If you want to code as a hobby that is fantastic, and we can go ahead and see in 2 years how your comment ages.
I’d very like to see such demonstration. Where someone hands over a department to an agent and let it makes decisions.
> This convo now turns into the "AI is not profitable and this is a house of cards" theme?
Where did I say that? I didn’t even mention money, just the broader resource term. A lot of business are mostly running experiments if the current set of tooling can match the marketing (or the hype). They’re not building datacenters or running AI labs. Such experiments can’t run forever.
That's your bar for understanding business context? I thought we were talking about what you actually said which is: understanding business context. If I brainstorm about a feature it will be able to pull the compendium of knowledge for the business (reports, previous launches, infrastructure, an understanding of the problem space, industry, company strategy). That's business context.
> Where did I say that? I didn’t even mention money, just the broader resource term. A lot of business are mostly running experiments if the current set of tooling can match the marketing (or the hype). They’re not building datacenters or running AI labs. Such experiments can’t run forever.
I misunderstood you then, I wasn't sure what point you were trying to make. Is your point "companies are trying to cajole Claude to do X and it doesn't work and hasn't for the last year so they are giving up"? If so I think that is a wonderful opportunity for people that understand the nuance of these systems and the concept of timing.
Unless the AI is inserted into every conversation it won't discover this, or how it changes.
Even if it had access to all this documented it wouldn't then be able to account for politics, where Barry who runs analytics is secretly trying to sabotage the project so it ends up run by his team, etc.
You're missing the point though. "1 + 1" vs "one.add(1)" might both be "passable" and correct, but it's missing the forest for the trees, how do you know which one is "long-term the right choice, given what we know?", which is the engineering part of building software, and less about "coding" which tends to be the easy part.
How do you evaluate, score and/or benchmark something like that? Currently, I don't think we have any methodologies for this, probably because it's pretty subjective in the end. That's where the "creative" parts of software engineering becomes more important, and it's also way harder to verify.
You don't need to benchmark this, although it's important. We have clear scaling laws on true statistical performance that is monotonically related to any notion of what performance means.
I do benchmarks for a living and can attest: benchmarks are bad, but it doesn't matter for the point I'm trying to make.
It doesn't because it doesn't learn. Every time you run it, it's a new dawn with no knowledge of your business or your business context
> better reasoning
It doesn't have better reasoning beyond very localized decisions.
> and can ask humans for clarification and take direction.
And yet it doesn't, no matter how many .md file you throw at it, at crucial places in code.
> We have clear scaling laws on true statistical performance that is monotonically related to any notion of what performance means.
This is just a bunch of words stringed together, isn't it?
It does learn in context. And lack of continuous learning is temporary, that is a quirk of the current stack, expect this to change rather quickly. Also still not relevant, consider that agentic systems can be hierarchical and that they have no trouble being able to grok codebases or do internal searches effectively and this will only improve.
> It doesn't have better reasoning beyond very localized decisions.
Do you have any basis for this claim? It contradicts a large amount of direct evidence and measurement and theory.
> This is just a bunch of words stringed together, isn't it?
Maybe to yourself? Chinchilla scaling laws and RL scaling laws are measured very accurately based on next token test loss (Chinchilla). This scales very predictably. It is related to downstream performance, but that relationship is noisy but clearly monotonic
It quite literally doesn't.
It also doesn't help that every new context is a new dawn with no knowledge if things past.
> Also still not relevant, consider that agentic systems can be hierarchical and that they have no trouble being able
A bunch of Memento guys directing a bunch of other Memento guys don't make a robust system, or a system that learns, or a system that maintains and retains things like business context.
> and this will only improve.
We've heard this mantra for quite some time now.
> Do you have any basis for this claim?
Oh. Just the fact that in every single coding session even on a small 20kloc codebase I need to spend time cleaning up large amounts of duplicated code, undo quite a few wrong assumptions, and correct the agent when it goes on wild tangents and goose hunts.
> Maybe to yourself? Chinchilla scaling laws a
yap yap yap. The result is anything but your rosy description of these amazing reasoning learning systems that handle business context.
Awesome you've backed this up with real literature. Let's just include this for now to easily refute your argument which I don't know where it comes from: https://transformer-circuits.pub/2022/in-context-learning-an...
> It also doesn't help that every new context is a new dawn with no knowledge if things past.
Absolutely true that it doesn't help but: agents like Claude have access to older sessions, they can grok impressive amounts of data via tool use, they can compose agents into hierarchical systems that effectively have much larger context lengths at the expense of cost and coordination which needs improvement. Again this is a temporary and already partially solved limitation
> A bunch of Memento guys directing a bunch of other Memento guys don't make a robust system, or a system that learns, or a system that maintains and retains things like business context.
I think you are not understanding: hierarchical agents have long term memory maintained by higher level agents in the hierarchy, it's the whole point. It's annoying to reset model context, but yet you have a knowledge base of the business context persisted and it can grok it...
> We've heard this mantra for quite some time now.
yes you have, and it has held true and will continue to hold true. Have you read the literature on scaling laws? Do you follow benchmark progression? Do you know how RL works? If you do I don't think you will have this opinion.
> yap yap yap. The result is anything but your rosy description of these amazing reasoning learning systems that handle business context.
Well that's fine to call an entire body of literature "yap" but don't pretend like you have some intelligible argument, I don't see you backing up any argument you have here with any evidence, unlike the multitude of sources I have provided to you.
Do you argue things have not improved in the last year with reasoning systems? If so I would really love to hear the evidence for this.
I love it when people include links to papers that refute their words.
So, Antropic (which is heavily reliant on hype and making models appear more than they are) authors a paper which clearly states: "tokens later in context are easier to predict and there's less loss of tokens. For no reason at all we decided to give this a new name, in-context learning".
> agents like Claude have access to older sessions, they can grok impressive amounts of data via tool use
That is they rebuild the world from scratch for every new session, and can't build on what was learned or built in the last one.
Hence continuous repeating failure modes.
10 years ago I worked in a team implementing royalties for a streaming service. I can still give you a bunch of details, including references to multiple national laws, about that. Agents would exhaust their context window just re-"learning" it from scratch, every time. And they would miss a huge amount of important context and business implications.
> Have you read the literature on scaling laws?
You keep referencing this literature as it was Holy Bible. Meanwhile the one you keep referring to, Chinchilla, clearly shows the very hard limits of those laws.
> Do you argue things have not improved in the last year with reasoning systems?
I don't.
Frankly, I find your aggressiveness quite tiring
having to answer for opinions with no basis in the literature is I'm sure very tiring for you. Your aggression being met is I'm sure uncomfortable.
> I love it when people include links to papers that refute their words. > So, Antropic (which is heavily reliant on hype and making models appear more than they are) authors a paper which clearly states: "tokens later in context are easier to predict and there's less loss of tokens. For no reason at all we decided to give this a new name, in-context learning".
well I don't really love it when people just totally misread a paper because they have an agenda to push and can't seem to accept that their opinions are contradicted by real evidence.
in-context learning is not "later tokens easier" it’s task adaptation from examples in the prompt. I'm sure you realize this. Models can learn a mapping (e.g. word --> translation) from a few examples in the prompt, apply inputs within the same forward pass. That is function learning at inference time, not just "predicting later tokens better"
I'm sure also you're happy to chalk up any contradicting evidence to a grand conspiracy of all AI companies just gaming benchmarks and that this gaming somehow completely explains progress.
> That is they rebuild the world from scratch for every new session, and can't build on what was learned or built in the last one.
That they rebuild the world from scratch (wrong, they have priors from pretraining, but I accept your point here) does not mean they can't build on what was learned or built in the last one. They have access to the full transcript, and they have access to the full codebase, the diff history, whatever knowledge base is available. It's just disingenuous to say this, and then it also assumes (1) there is no mitigation for this, which I have presented twice before and you don't seem to understand it, (2) this is a temporary limitation, continual learning is one of the most important and well funded problems right now.
> 10 years ago I worked in a team implementing royalties for a streaming service. I can still give you a bunch of details, including references to multiple national laws, about that. Agents would exhaust their context window just re-"learning" it from scratch, every time. And they would miss a huge amount of important context and business implications.
also not an accurate understanding of how agents and their context work; you can use multiple session to digest and distill information useful in other sessions and in fact Claude does this automatically with subagents. It's a problem we have _already sort of solved today_ and that will continue to improve.
> You keep referencing this literature as it was Holy Bible. Meanwhile the one you keep referring to, Chinchilla, clearly shows the very hard limits of those laws.
You keep dismissing this literature as if you have understood it and that your opinion somehow holds more weight...Can you elaborate on why you think Chinchilla shows the hard limits of the scaling laws? Perhaps you're referring to the term capturing the irreducible loss? Is that what you're saying?
> Do you argue things have not improved in the last year with reasoning systems? I don't
Then are you arguing this progress will stop? I'm just not sure I understand, you seem to contradict yourself
Having only literature on your side must feel nice.
> They have access to the full transcript, and they have access to the full codebase, the diff history, whatever knowledge base is available.
Yes. And it means that they don't learn, and they alway miss important details when rebuilding the world.
That's why even the tiniest codebases are immediately filled with duplications, architecturally unsound decisions, invalid assumptions etc.
> also not an accurate understanding of how agents and their context work; you can use multiple session to digest and distill information useful in other sessions and in fact
I say: agents don't learn and have to rebuild the world from scratch
You: not an accurate understanding of how agents and their context work.... they rebuild the world from scratch every time they run.
> You keep dismissing this literature as if you have understood it
No. I'm dismissing your flawed interpretation of purely theoretical constructs.
Chinchilla doesn't project unlimited amazing scalability. If anything, it shows a very real end of scalability.
Anthropic's paper adopts a nice marketable term for a process that has little to do with learning.
Etc.
Meanwhile you do keep rejecting actual real-world behaviour of these systems.
> Then are you arguing this progress will stop? I'm just not sure I understand, you seem to contradict yourself
I didn't say that either. Your opponents don't contradict themselves if you only stop to pretend they think or say.
Your unsubstantiated belief is that improvements are on a steep linear or even exponensial progression. Because "literature" or something.
Looking past all the marketing bullshit, it could be argued that growth is at best logarithmic, and most improvments come from tooling around (harnesses, subagents etc.). While all the failure modes from a year ago are still there: misunderstanding context, inability to maintain cohesion between sessions, context pollution etc.
And providers are running into the issue of getting non-polluted trainig data.
---
At this point we're going around in circles, and I'm no interested in arguing with theorists.
Adieu
> Like for example a trusted user makes feedback -> feedback gets curated into a ticket by an AI agent, then turned into a PR by an Agent, then reviewed by an Agent, before being deployed by an Agent.
Once you add "humans for clarifications and take direction" then yeah, things can be useful, but that's far away from the non-human-involvment-loop earlier described in this thread, which is what people are pushing back against.
Of course, involving people makes things better, that's the entire point here, and that by removing the human, you won't get as good results. Going back to benchmarks, obviously involving humans aren't possible here, so again we're back to being unable to score these processes at all.
Benchmarks ==> it's absolutely not a given that humans can't be involved in the loop of performance measurement. Why would that be the case?
Code is effectively becoming cheap, which means even bad design decisions can be overturned without prohibitive costs.
I wouldn't be surprised if in a couple of years we see several projects that approach the problem of tech debt like this:
1. Instruct AI to write tens of thousands of tests by using available information, documentation, requirements, meeting transcripts, etc. These tests MUST include performance AND availability related tests (along with other "quality attribute" concerns) 2. Have humans verify (to the best of their ability) that the tests are correct -- step likely optional 3. Ask another AI to re-implement the project while matching the tests
It sounds insane, but...not so insane if you think we will soon have models better than Opus 4.6. And given the things I've personally done with it, I find it less insane as the days go by.
I do agree with the original poster who said that software is moving in this direction, where super fast iteration happens and non-developers can get features to at least be a demo in front of them fast. I think it clearly is and am working internally to make this a reality. You submit a feature request and eventually a live demo is ready for you, deployed in isolation at some internal server, proxied appropriately if you need a URL, and ready for you to give feedback and have the AI iterate on it. Works for the kind of projects we have, and, though I get it might be trickier for much larger systems, I'm sure everyone will find a way.
For now, we still need engineers to help drive many decisions, and I think that'll still be the case.These days all I do when "coding" is talking (via TTS) with Opus 4.6 and iterating on several plans until we get the right one, and I can't wait to see how much better this workflow will be with smarter and faster models.
I'm personally trying to adapt everything in our company to have agents work with our code in the most frictionless way we can think of.
Nonetheless, I do think engineers with a product inclination are better off than those who are mostly all about coding and building systems. To me, it has never felt so magical to build a product, and I'm loving it.
I'm sorry, but only someone who never maintained software long-term would say something like this. The further along you are in development, the magnitude of costs related to changing that increases, maybe even exponentially.
Correct the design before you even wrote code, might be 100x cheaper (or even 1000x) than changing that design 2 years later, after you've stored TBs of data in some format because of that decision, and lots of other parts of the company/product/project depends on those choices you made earlier.
You can't just pile on code on top of code, say "code is cheap" and hope for the best, it's just not feasible to run a project long-term that way, and I think if you had the experience of maintaining something long-term, you'd realize how this sounds.
The easiest part of "software engineering" is "writing code", and today "writing code" is even easier. But the hardest parts, actually designing, thinking and maintaining, remains the same as before, although some parts are easier, others are harder.
Don't get me wrong, I'm on the "agentic coding" train as much as everyone else, probably haven't written/edited a code by myself for a year at this point, but it's important to be realistic about what it actually takes to produce "worthwhile software", not just slop out patchy and hacky code.
I think using agents to prototype code and design will be a big thing. Have the agent write out what you want, come back with what works and what doesn't, write a new spec, toss out the old code and and have a fresh agent start again. Spec-driven development is the new hotness, but we know that the best spec is code, have the agent write the spec in code, rewrite the spec in natural language, then iterate.
Of course it's still valuable. A real app has plenty of mundane code despite our field's best efforts.
How good are these types of algorithms at generalization? Are they learning how to code; or are they learning how to code migrations, then learning how to code caches, then learning how to code a command line arg parser, etc?
Verifiable domains are interesting. It is unquestionably why agents have come first for coding. But if you've played with claude you may have experienced it short-circuiting failing tests, cheating tests with code that does not generalize, writing meaningless tests, and at long last if you turn it away from all of these it may say something like "honest answer - this feature is really difficult and we should consider a compromise."
"There is no sense in which they are mathematically destined to eventually program well"
- Yes there is and this belies and ignorance of the literature and how things work
- Again: RL has been around forever. Scaling laws have held empirically up to the largest scales we've tested. There are known RL scaling laws for both training and test time. It's ludicrous to state there is "no sense" in this, on the contrary, the burden of proof of this is squarely on yourself because this has already been studied and indeed is the primary reason why we're able to secure the eye-popping funding: contrary to popular HN belief, a trillion dollars of CapEx spend is based on rational evidence-based decision making.
> "How good are these types of algorithms at generalization"
There is a tremendously large literature and history of this. ULMFiT, BERT ==> NLP task generalization; https://arxiv.org/abs/2206.07682 ==> emergent capabilities, https://transformer-circuits.pub/2022/in-context-learning-an... ==> demonstrated circuits for in context learning as a mechanism for generalization, https://arxiv.org/abs/2408.10914 + https://arxiv.org/html/2409.04556v1 ==> code training produces downstream performance improvements on other tasks
> Verifiable domains are interesting. It is unquestionably why agents have come first for coding. But if you've played with claude you may have experienced it short-circuiting failing tests, cheating tests with code that does not generalize, writing meaningless tests, and at long last if you turn it away from all of these it may say something like "honest answer - this feature is really difficult and we should consider a compromise."
You say this and ignore my entire argument: you are right about all of your observations, yet
- Opus 4.6 compared to Sonnet 3.x is clearly more generalizable and less prone to these mistakes
- Verifiable domain performance SCALES, we have no reason to expect that this scaling will stop and our recursive improvement loop will die off. Verifiable domains mean that we are in alphago land, we're learning by doing and not by mimicking human data or memorizing a training set.
What difference do I think there is between humans and an agent? They use different heuristics, clearly. Different heuristics are valuable on different search problems. It's really that simple.
To be clear, I'm not calling either superior. I use agents every day. But I have noticed that claude, a SOTA model, makes basic logic errors. Isn't that interesting? It has access to the complete compendium of human knowledge and can code all sorts of things in seconds that require my trawling through endless documentation. But sometimes it forgets that to do dirty tracking on a pure function's output, it needs to dirty-track the function's inputs.
It's interesting that you mention AlphaGo. I was also very fascinated with it. There was recent research that the same algorithm cannot learn Nim: https://arstechnica.com/ai/2026/03/figuring-out-why-ais-get-.... Isn't that food for thought?
I am saying you are absolutely right that Opus 4.6 is both SOTA and also colossally terrible in even surprisingly mundane contexts. But that is just not relevant to the argument you are making which is that there is some fundamental limitation. There is of course always a fundamental limitation to everything, but what we're getting at is where that fundamental limitation is and we are not yet even beginning to see it. Combinatorics here is the wrong lens to look at this, because it's not doing a search over the full combinatoric space, as is the case with us. There are plenty of efficient search "heuristics" as you call them.
> They use different heuristics, clearly.
what is the evidence for this? I don't see that as true, take for instance: https://www.nature.com/articles/s42256-025-01072-0
> It's interesting that you mention AlphaGo. I was also very fascinated with it. There was recent research that the same algorithm cannot learn Nim: https://arstechnica.com/ai/2026/03/figuring-out-why-ais-get-.... Isn't that food for thought?
It's a long known problem with RL in a particular regime and isn't relevant to coding agents. Things like Nim are a small, adversarially structured task family and it's not representative of language / coding / real-world tasks. Nim is almost the worst possible case, the optimal optimal policy is a brittle, discontinuous function.
Alphago is pure RL from scratch, this is quite challenging, inefficient, and unstable, and why we dont do that with LLMs, we pretrain them first. RL is not used to discover invariants (aspects of the problem that don't change when surface details change) from scratch in coding agents as they are in this example. Pretraining takes care of that and RL is used for refinement, so a completely different scenario where RL is well suited.
> “AlphaZero excels at learning through association,” Zhou and Riis argue, “but fails when a problem requires a form of symbolic reasoning that cannot be implicitly learned from the correlation between game states and outcomes.”
Seems relevant.
Humans learn.
Agents regurgitate training data (and quality training data is increasingly hard to come by).
Moreover, humans learn (somewhat) intangible aspects: human expectations, contracts, business requirements, laws, user case studies etc.
> Verifiable domain performance SCALES, we have no reason to expect that this scaling will stop.
Yes, yes we have reasons to expect that. And even if growth continues, a nearly flat logarithmic scale is just as useless as no growth at all.
For a year now all the amazing "breakthrough" models have been showing little progress (comparatively). To the point that all providers have been mercilessly cheating with their graphs and benchmarks.
I'm just going to ask that you read any of my other comments, this is not at all how coding agents work and seems to be the most common misunderstanding of HN users generally. It's tiring to refute it. RL in verifiable domains does not work like this.
> Humans learn.
Sigh, so do LLMs, in context.
> Moreover, humans learn (somewhat) intangible aspects: human expectations, contracts, business requirements, laws, user case studies etc.
Literally benchmarks on this all over the place, I'm sure you follow them.
> Yes, yes we have reasons to expect that. And even if growth continues, a nearly flat logarithmic scale is just as useless as no growth at all.
and yet its not logarithmic? Consider data flywheel, consistent algorithmic improvements, synthetic data [basically: rejection sampling from a teacher model with a lot of test-time compute + high temperature],
> For a year now all the amazing "breakthrough" models have been showing little progress (comparatively). To the point that all providers have been mercilessly cheating with their graphs and benchmarks.
Benchmaxxing is for sure a real thing, not to mention even honest benchmarking is very difficult to do, but considering "all of the AI companies are just faking the performance data" to be the "story" is tremendously wrong. Consider AIME performance on 2025 (uncontaminated data), the fact that companies have a _deep incentive_ to genuinely improve their models (and then of course market it as hard as possible, thats a given). People will experiment with different models, and no benchmaxxing is going to fool people for very long.
If you think Opus 4.6 compared to Sonnet 3.x is "little progress" I think we're beyond the point of logical argument.
We have all the data now.
I don’t see where the huge gap should come from, as one person before they said they still make basic errors.
Models got better for a bunch of soft tuning. Language and abstractness is not really the same thing there are a lot of very good speakers that are terrible in logic and abstractness.
Thinking abstract sometimes makes it necessary to leave language and draw or som people even code in another coding language to get it.
We’ve seen it with the compiler project it’s nice looking but if you would want to make a competitive compiler you would be as far as starting fresh
I think it goes down in two camps. AI is improving on these issues and people countering.
I don’t know for sure, but to me it seems the last 2 years weren’t necessarily 'intelligence' improvements but post-training improvement and tool connections, also reduced censorship.
I’m know using less AI than ever and I’ve been burning 1000USD/month before Claude Code. I have a couple of really fundamental functions built that help me to solve a big chunk of specific problems I can built a lot on that. Adding functionality became easier not more complicated.
I would think for these business problems that I’m facing AI is less than 30% of the time right. For example deciding on how to setup databases for max efficiency how to write efficient queries. Everything that in the end is really moat to you compared to your vibe coded competitors.
From my personal experience I’ve seen a lot of vibe-cded companies stuck and barely adding nec functionality or features and my guess is that they don’t trust changes anymore.
So even if AI would be as good as a really good coder one thing would still be missing a person that is knowing exactly what is happening.
And I mean okay it might be writing a form real quick. But a modern form needs to do a lot of things and if you have established patterns for all kind of inputs, the implementation is mundane.
It’s like when you learn coding, type it yourself to learn. So if you can’t scale the AI only codebase at one point you have to learn it, and I argue right now most efficient way is to write in it.
And I’m also arguing that it’s really tough to get a software so good that it’s actually an asset on the market vibe-coded only. It seems like its more of a drug for wannapreneurs than it is actually building an asset.
Like it builds you a Netflix clone, but what you see is barely the code you need to write a Netflix competitor.
Stripe is apparently pushing gazzaliion prs now from slack but their feature velocity has not changed. so what gives?
how is that number of pr is now the primary metric of productivity and no one cares about what is being shipped or if we are shipping product faster. Its total madness right now. Everyone has lost their collective minds.
I'm not seeing the apps, SaaS, and other tools I use getting better, with either more features or fewer bugs.
Whatever is being shipped, as an end user, I'm just not seeing it.
Its baffling to see these comments on hacknernews though. I guess you have to prove that you are not a luddite by making "ai forward" predictions and show that you "get it"
So one user's experience is relevant to another, so they can learn from one another?
So, we will give these 3 or 4 trusted users access to an on-site chat interface to request updates.
Next, a dev environment is spun up, agent makes the changes, creates PR and sends branch preview link back to user.
Sort of an agent driven CMS for non-technical stakeholders.
Let’s see if it works.
But I do think even now with certain types of crud apps, things can be largely automated. And that's a fairly large part of our profession.
"Generate the following JSON formatted object array representing the interruptions in my daily traffic. If no results, emit []. Send this at 8am every morning. {some schema}. Then run jsonreporter.py"
Then just let jsonreporter.py discriminate however it likes. Keep the LLMs doing what they are good at, and keep hard code doing what it's good at.
You'll define exactly what good looks like.
You need to write a clearer prompt.
Your AI assistant orders an experimental jetpack from a random startup lab. Would you have honestly guessed that the prompt was "ambiguous" before you knew how the AI was going to act on it ?
Yes, you can give the Big Brain Thing a vague task and expect results, some times it'll do it right
But if you want repeatability, give it tools to determine what is a "disruption".
I have this exact system running on a WhateverClaw with a few simple tools: weather, train time tables (I commute via train) and a read-only version of my "will I be at the office or remote today" -calendar via a tool. Oh and a tool to notify me via Telegram.
It gets this information using the given tools and determines if it's worth notifying me.
TBH this doesn't need a "claw", I could just run the tools in cron, construct a prompt with that data and run any LLM on it.
Yesterday, I spent the entire day trying to set up "Claude on the web" for an Elixir project and eventually had to give up. Their network firewall kept killing Hex/rebar3 dependency resolution, even after I selected "full" network access.
The environment setup for "on the web" is just a bash script. And when something goes wrong, you only see the tail of the log. There is currently no way to view the full log for the setup script. It's really a pain to debug.
The Copilot equivalent to "Claude on the web" is "GitHub Copilot Coding Agents," which leverages GitHub Actions infrastructure and conventions (YAML files with defined steps). Despite some of the known flaws of GitHub Actions, it felt significantly more robust.
"Schedule task on the web" is based on the same infrastructure and conventions as "Claude on the web", so I'm afraid I'm gonna have the same troubles if I want to use this.
"Your plan gets 3 daily cloud scheduled sessions. Disable or delete an existing schedule to continue."
But otherwise, this looks really cool. I've tried using local scheduled tasks in both Claude Code Desktop and the Codex desktop app, and very quickly got annoyed with permissions prompts, so it'll be nice to be able to run scheduled tasks in the cloud sandbox.
Here are the three tasks I'll be trying:
Every Monday morning: Run `pnpm audit` and research any security issues to see if they might affect our project. Run `pnpm outdated` and research into any packages with minor or major upgrades available. Also research if packages have been abandoned or haven't been updated in a long time, and see if there are new alternatives that are recommended instead. Put together a brief report highlighting your findings and recommendations.
Every weekday morning: Take at Sentry errors, logs, and metrics for the past few days. See if there's any new issues that have popped up, and investigate them. Take a look at logs and metrics, and see if anything seems out of the ordinary, and investigate as appropriate. Put together a report summarizing any findings.
Every weekday morning: Please look at the commits on the `develop` branch from the previous day, look carefully at each commit, and see if there are any newly introduced bugs, sloppy code, missed functionality, poor security, missing documentation, etc. If a commit references GitHub issues, look up the issue, and review the issue to see if the commit correctly implements the ticket (fully or partially). Also do a sweep through the codebase, looking for low-hanging fruit that might be good tasks to recommend delegating to an AI agent: obvious bugs, poor or incorrect documentation, TODO comments, messy code, small improvements, etc.
I ran all of these as one-off tasks just now, and they put together useful reports; it'll be nice getting these on a daily/weekly basis. Claude Code has a Sentry connector that works in their cloud/web environment. That's cool; it accurately identified an issue I've been working on this week.
I might eventually try having these tasks open issues or even automatically address issues and open PRs, but we'll start with just reports for now.
Seems trivial.
But you can set up a claude -p call via a cronjob without too much hassle and that can use subscriptions.
> Analyzing CI failures overnight and surfacing summaries
Look like on ec2 with python? Because with Claude, it’s that prompt, and with your solution it’s infra + security groups + multiple APIs + whatever code you actually write
So for example the only "analysis" of CI failures are which systems failed and who/what committed the changes to those things. The only way AI would help me here is if the system was so jank that the sole primitive i can use is textual analysis of log files. Which granted is probably real for a lot of software firms, but I really hope I have better build and test infrastructure than that.
I think this shows the value.
> Which granted is probably real for a lot of software firms
Here's the rub though; for many many people it's a huge improvement over what they have right now.
Expectations - the functionality of "do X on a timer" needs to be offered to users as a proper end-user feature[0], not treated as a sysadmin feature (Windows, Linux) or not provided at all (Android). People start seeing it on their own devices, they'll start using it, then expecting it, and the web will adjust too[1].
UI - somehow this escapes every existing solution, from `cron` through Windows timers to any web "on timer" event trigger in any platform ever. There already exists a very powerful UI paradigm for managing recurring tasks, that most normies know how to use, because they're already using it daily at work and privately: a calendar. Yes, that thing where we can set and manage recurring events, and see them at a glance, in context of everything else that's going on in our lives.
--
<rant>
I know those are hard problems, but are hard mostly because everybody wants to be the fucking one platform owning users and the universe. This self-inflicted sickness in computing is precisely why people will jump at AI solutions for this. Why I too will jump on this: because it's easier than dealing with all the systems and platforms that don't want to cooperate.
After all, at this point, the easiest solution to the problems I listed above, and several others in this space, would be to get an AI agent that I can:
1) Run on a cron every 30 minutes or so (events are too complicated);
2) Give it read (at minimum) access to my calendar and todo lists (the ones I use, but I'm willing to compromise here);
3) Give it access to other useful tools
Which I guess brings us to the actual root problem here. "Run tasks on a cron" and "run tasks on trigger" are basically just another way of saying unattended/non-interactive usage. That is what is constantly being denied end users.
This is also the key to enabling most value of AI tools, too, and people understand it very well (see the popularity of that Open Claw thing as the most recent example), but the industry also lives in denial, believing that "lethal trifecta" is a thing that can be solved.
</rant>
--
[0] - This extends to event triggers ("if X happens, then") automation, and end-user automation in all of every-day life. I mean, it's beyond ridiculous that the only things normal people are allowed to run automatically are dishwasher, and a laundry machine (and in the previous era, VCRs).
[1] - As a side effect, it would quickly debullshitify "smart home" / "internet of things" spaces a lot. The whole consumer side of the market revolves around selling people basic automation capabilities - except vendor-locked, and without the most useful parts.
- IFTTT was great when it started; at some point, it became... weird, in a "I don't even know what's going on on my screen, is this a poster or an app" kind of way.
- Zapier is an unpenetrable mess, evidently targets marketers and other business users; discovery is hard, and even though it seems like it has everything, it - like all tools in this space - is always missing the one feature you actually need.
- Yahoo Pipes, I heard they were great, but I only learned about them after they shut down.
- Apple Shortcuts - not sure what you can do with those, but over the years of reading about them in HN comments, I think they may be the exception here, in being both targeting regular users and actually useful.
- Samsung Modes and Routines - only recently becoming remotely useful, so that's nice, even if vendor-restricted.
- Tasker - an Android tool that actually manages to offer useful automation, despite the entire platform/OS and app ecosystem trying its best to prevent it. Which is great, if your main computer is a phone. It sucks in a world of cloud/SaaS, because it creates a silly situation where e.g. I could nicely automate some things involving e-mail and calendars from Tasker + FairEmail, but... well my mailboxes and calendars lives in the cloud so some of that would conflict with use of vendor (Fastmail) webapp or any other tool.
Or, in short: we need Tasker but for web (and without some of the legacy baggage around UI and variable handling).
The sorry state of automation is not entirely, or even mostly, the fault of the automation platforms. I may have issues with some UI and business choices some of these platforms made, but really, the main issue is that integrations are business deals and the integrated sides quickly learned to provide only a limited set of features - never enough to allow users to actually automate use of some product. There's always some features missing. You can read data but not write it. You can read files and create new files but not edit or delete them. You can add new tasks but can't get a list of existing ones. Etc.
It's another reason LLMs are such a great thing to happen - they make it easy (for now) to force interoperability between parties that desperately want to prevent it. After all, worst case, I can have the LLM operate the vendor site through a browser, pretending to be a human. Not very reliable, but much better than nothing at all.
n8n, node-RED and others already exist. There are many tools for automations, and I guess most of them can also do cron-like jobs.
Consumer grade automations built on node-RED? I suppose it depends on the market, but most people aren’t going to want to fiddle with it, I suspect.
A plugin for Chrome might be able to take off though, or some killer mobile app, but it needs to run on a cheap phone and control things without having to keep track of loops and logic and variables and all the fun stuff.
And re: Zapier: yes, that’s the key to Zapier, from my experience: usage in marketing and the “power user” base.
Re: shortcuts: (I live in the Apple ecosystem) Shortcuts + AppleScript is gold on macOS. Shortcuts + iOS is about to be game changing - it already changed the game, it’s just nobody has been playing it, because it’s not “fun”.
After Siri+Gemini+Shortcuts, everyone will be playing it, I suspect, even on Android, it will get built somehow.
And there is also the mindset to avoid boring loops, and prefer event driven solutions for optimal resource-usage. So people also have a kind of blind spot for this functionality.
Same. Sometimes it is just people overeager to play with new toys, but in our case there is a push from the top & outside too: we are in the process of being subsumed into a larger company (completion due on April the 1st, unless the whole thing is an elaborate joke!) and there is apparently a push from the investors there to use "AI" more in order to not "get left behind the competition".
This company already does some pretty cool stuff with statistics for forecasting but now they are pivoting their roadmap to bake in GenAI into their offering over some other features that would be more valuable to their clients.
I wrote this to help people (not just Devs) reason about agent skills
https://alexhans.github.io/posts/series/evals/building-agent...
And this one to address the drift of non determism (but depending on the audience it might not resonate as much)
https://alexhans.github.io/posts/series/evals/error-compound...
Grok has had this feature for some time now. I was wondering why others haven't done it yet.
This feature increases user stickiness. They give 10 concurrent tasks free.
I have had to extract specific news first thing in the morning across multiple sources.
It doesnt allow egress curl, apart from few hardcoded domains.
I have created Cronbox in the cloud which has a better utility than above. Did a "Show HN: Cronbox – Schedule AI Agents" a few days back.
and a pelican riding a bicycle job -
https://cronbox.sh/jobs/pelican-rides-a-bicycle?variant=term...
Is this assuming you give it git commit permission and it just does that? Or it acts through MCP tools you enable?
It's a bit like asking if "an API" was a critical link in some cybersec incident. Yes, it probably was, and?
Prompt injection is "social engineering" but applied to LLMs. It's not a bug, it's fundamentally just a facet of its (LLM/human) general nature. Mitigations can be placed, at the cost of generality/utility of the system.
Fair enough but then that means that MCP is not "a bit like asking if "an API" was a critical link in some cybersec incident"
Because I can secure an API but I can't secure the the "(LLM/human) general nature."
The security risk here is the LLM, not the MCP, and you cannot secure the LLM in such system any more you can secure user - unless you put that LLM there and own it, at which point it becomes a question of whether it should've been there in the first place (and the answer might very well be "yes").
In fact, I re-read the article before submitting this comment just to make sure I wasn’t missing something. What on earth is so polarizing about a prompt being run recurrently? It’s a long-awaited feature that I’ve personally needed.
If you want to win your war, you’ll need better propaganda to recruit people. Start with me. My mind is open. Why should I join?
Please tie your claims concretely to this new feature. I’m interested in how adding this could erode open source software. To me they seem completely independent, and it’s a welcome change.
"A scheduled task runs a prompt on a recurring cadence using Anthropic-managed infrastructure." >> There is no other way to read this as in this context, its just a small feature, but its a land grab to run workflows locked into their cloud not just models, we don't fall for regimes in one go but one tiny piece at a time, like the frog in the water.
Do you not see this at all and this sounds all crazy to you?
It’s also undeniable that Claude is very, very good. I hope that kind of quality comes to open source models. Lots of people have said they’re happy with the experiences they’ve had.
Personally, a middle ground seems like a nice compromise. Use both when it suits you. I don’t view it as a war, but as an inevitable evolution due to the amount of money being poured into the ecosystem.
The thing is, I would be behind you if there was a concrete alternative. Is there? Because one way or another, consumers will want this kind of quality that Claude is providing.
Either way, I didn’t mean to discourage you, only to ground you. Framing things as a war for our freedom is fine, but ultimately the freedom side has to be able to provide the same features as the corporation side. So where are they? “I use X instead of Y” is the best defense against vendor lock in.
Anthropic wants a world where they own your agent where it can't exist outside of the Claude desktop app or Claude Code.
There could exist a world where your agent isn't confined by the whims of a corporation.
Please. I'm sure you're referring to their locking down of subscription keys, which of course they are going to have restrictions on. It's a subsidized subscription model.
You've always been able to create a platform account and use API keys with usage-based billing, and that will never go away. Charging enough to make a profit on inference isn't exactly rent-seeking or whatever language you want to use to villainize a company trying to make enough revenue to survive.
You misspelt ">95% discount relative to API pricing" ;)
I run conferences and I like to have photos of delegates on the page so you can see who else is attending.
I wanted to automate this by having Claude go to the person’s LinkedIn profile and save the image to the website.
But it seems it won’t do that because it’s been instructed not to.
That's not unique to LinkedIn but what is somewhat unique is the strong linkage to real world identities, which raises the cost of Sybil attacks on personal networks with high trust.
It's a game changer.
Edit: my mistake. It's inferior to a Cron job. If my repos happen to be self hosted with Forgejo or codeberg, then it won't even work. If I concede to use GitHub though I don't have to set up any env variables. Schedules lock-in, all over the web.
I feel this is rooted in problems that extend beyond computing. Regular people are not allowed to automate things in their life. Consider that for most people, the only devices designed to allow unattended execution off a timer are a washing machine, some ovens and dishwashers, and an alarm clock (also VCRs in the previous era). Anything else requires manual actuation and staying in a synchronous loop.
Of course a provider can offer convenient shortcuts, but at the cost of getting tied into their ecosystem.
Anthropic is clearly battling an existential threat: what happens when our paying users figure out they can get a better and cheaper model elsewhere.
They solved that with subscriptions. For end-users (and developers using AI for coding), it makes no sense to go for pay-as-you-go API use, as anything interesting will burn more than the monthly subscription worth of $$$ in API costs in few hours to days.
Sure subscription is a sort of tie in, but only if users are fooled into investing in workflows bound to anthropic. That's what the company is hooking them to do with this scheduler, banning open agentic framework and the rest.
The moat, if any, will be the tooling. Token is becoming a commodity, they know it.
Such a service will always be destroyed by the bell-ends who want to run spam or worse activities.
(And on Android, AFAIK there's exactly nothing at all. There's not even common support for any kind of basic automation; only recent exception is Samsung. From third-party apps, there's always been Tasker - very powerful, but the UX almost makes you want to learn to write Android apps instead.)
I think the core problem is not so much that it is not "allowed", but that even the most basic types of automation involves programming. I mean "programming" here in the abstract sense of "methodically breaking up a problem into smaller steps and control flows". Many people are not interested in learning to automate things, or are only interested until they learn that it will involve having to learn new things.
There is no secret conspiracy stopping people from learning to automate things, rather I think it's quite the opposite: many forces in society are trying to push people to automate more and more, but most are simply not interested in learning to do so. See for example the bazillion different "learn to code" programs.
Computing isn't, and has never been, demand-driven. It's all supply-driven. People choose from what's made available by vendors, and nobody bothers listening to user feedback.
https://imgur.com/a/apero-TWHSKmJ
Cron triggers (or specific triggers per connector like new email in Gmail, new linear issue, etc for built in connectors).
Then you can just ask in natural language when (whatever trigger+condition) happens do x,y and z with any configuration of connectors.
It creates an agentic chain to handle the events. Parent orchestrator with limited tools invoking workers who had access to only their specific MCP servers.
Official connectors are just custom MCP servers and you could add your own MCP servers.
I definitely had the most advanced MCP client on the planet at that point, supporting every single feature of the protocol.
I think that's why I wasn't blown away by OpenClaw, I had been doing my own form of it for a while.
I need to release more stuff for people to play around with.
My friends had use cases like "I get too many emails from my kids school I can't stay on top of everything".
So the automation was just asking "when I get an email from my kids school, let me know if there's anything actionable for me in it"
Push events into a running session with channels: https://news.ycombinator.com/item?id=47448524
I use it to:
- perform review of latest changes of code to update my documentation (security policies, user documentation etc.)
- perform review to latest changes of code, triage them, deduplicate and improve code - I review them, close them with comments for over-engoneering / add review for auto-fix
- perform review of open GitHub issue with label, select the one with highest impact, comment with rationale, implement it and make pull request - I wake up and I have a few pull request to fix issues that I can approve /finish in existing Claude Code thread
I want also use it to: - review recent Sentry issues, make GitHub issues for the one with highest priority, make pull request with proposed fix - I can just wake up and see that some crash is ready to be resolved
Limit of 3 scheduled jobs is pretty impactful, but playing with it give me a nice idea on how I can reduce my manual work.
- you can easily edit the prompt
- you can see the prompt execution history
- you don't need any infrastructure to orchestrate it
- it works even when your computer is off
Once you start using it, it turns out to be very convenient.