HN Simulatornew | past | comments | lists | submit | StableAlkyne's commentslogin

It was always such a treat to find a new website that has information you were interested in.

I still remember how much to sleep I lost when I discovered TVTropes


> Cold war tactics, instead of running a normal business.

There's a bit of a double standard here: When a Chinese company does it, it's a cold war tactic. When an American business does it, it's capitalism. China engages in some shady shit, but this is not a case of that IMO.

Let's not forget that Meta founded the open LLM industry when they released Llama for free. Microsoft, Huggingface, Google, and even Amazon (all American companies) have released their own.

The business strategy here is to disrupt the closed-weight leaders (OpenAI and Anthropic) by delivering an LLM that is good enough for day to day use. This commoditizes AI, driving the cost down and shifting the business model to providing off the in infra.

It's no mistake that the companies releasing these open-weight models are by and large cloud providers.

> their primary objective with distilling western models

This is another double standard: it's simultaneously wrong to train on the output of a closed model without permission, but somehow okay for the closed model to have been trained by reading everything that has ever been written without permission.


Companies and Governments do things to their own interests but those interests are wildly different.

Yes, China as a nation state has different goals than Alibaba, just as OpenAI has different goals than the US.

This is different than taking a basic market play of commodification and painting it as some Commie conspiracy


The difference being that there is a Politburo officer sitting on the Alibaba board. The amount of direct control the Chinese enforce on their businesses to ensure their goals are aligned with the state goes leagues behind what's done in Western Democracies.

Sure, but that still does not change the fact that everyone who isn't OpenAI or Anthropic are trying to commodify AI.

It's completely irrelevant here. It would be just as logical to try and tie them drinking water to some conspiracy to undermine Coke, because some Chinese public ad campaign encouraged people to drink healthy options.


I agree with you that some of the most outlandish attacks (especially the ones directed at open-source/weight) come from US companies that don't appreciate the competition but don't be fooled for a second that just like the US with its chip exports controls, China is not aggressively trying to undermine US interests.

The difference is that the Mona Lisa is art, created by a human. People generally value art because of the human factor - the artist was trying to say something, there is something interesting about the piece, etc. This is part of why generated images get so much pushback.

The same cannot be said for SaaS: your business needs certain productivity tools in order to operate, and the only options tend to be these hyper-expensive SaaS options whose UX has been clunky since 2005. Historically, you could not build these in-house because it's more expensive to hire a full engineering team. Today, the math being done is whether it is worth it to build something 90% as good for 1% the price of the SaaS.


These people are not using AI they're training AI. In a few years it will replace tham and their jobs just like it did programmers

Generally, enterprise contracts with the AI assisted coding companies have protections against training on the inputs.

You or I probably can't fight it if they lie on their terms of service. A Fortune 500 would go for blood (metaphorically speaking) if their IP was violated by one of these companies.

> In a few years it will replace tham

There isn't anybody to replace in this scenario. You're either buying the SaaS, hiring a full engineering team you didn't have before, or slop coding it. But if you're already on a SaaS, someone higher up has already made the decision that you cannot increase headcount to support something internal.

Personally, I'd prefer if they hired more people. But the cold math of your average business is going to choose to ditch the SaaS in favor of the cheap slop job.


problems evolve, hence why a saas is a subscription and not a disc you can buy with a license code

clones are frozen in time and tend to rust and a financial incentive emerges to just purchase something that doesn't rust


Clones are not frozen in time. If you've already slopped out some vibe-coded abomination, you can just modify it again.

The cost differential is so great when it comes to B2B SaaS that you could generate a brand new every month and still come out ahead.

I know there is probably some axiomatic "AI Bad" at play here (I avoid vibe coded solutions as well), but it's myopic to claim that we aren't at a point where the traditional B2B SaaS model is under serious risk.


Costs to maintain many of these items aren't that much different than self-hosting open source. Every week on HN you hear someone saying something has been enshittified. Well then just maintain your own shit with half the spend.

We used to buy software on floppys. It provided value without a sub. Pretending that every single software solution requires a 50mn operation is silly. I suspect HN tends to think of problems framed at massive scale.. but most people just ain't Google.


> I suspect HN tends to think of problems framed at massive scale

I think this is an industry-wide problem. People (especially managers and juniors) tend to read blogs or watch talks by FAANG discussing how they saved millions or achieved 5-nines uptime, and go on to cargo cult it.

There seems to be a perception that everything needs a sub-3 hour recovery (plus on-calls), when realistically, the business will go on just fine if some QC tech can't log their passing test for a couple days (the plant manager makes a call to ship it now and update the record later), or a project manager can't update their portfolio until Monday (they can just message someone if a status update is immediately needed).

Yeah, some things need resiliency. But a key aspect of engineering is balancing the complexity and cost of a certain design, with the risks of that design failing.


> We used to buy software on floppys. It provided value without a sub.

that was a different world with different problems


Yes, in our new world we love subscription treadmills and sexy new frameworks that do mostly the same thing the last one did.

> self-hosting open source

I get update notifications almost every week or so with most of my self hosted software.


Yep, software rusts. And when something like, I dunno, a popular self-hosted BI tool has 1000s of customers with 1000s of more needs then of course you'll have an update every week. They also have to stay ahead of their competitors with features. But when you're supporting 1 customer on 1 platform and have a choice of dependencies the update cadence ain't going to be nearly as bad is it?

A frontier model has no problem keeping a VM and a list of dependencies up to date. When one of those dependencies goes unmaintained it can fork or rewrite it and do the same there.


IMO one cause of this and many other problems in the public education system is that parents have too much influence in the education system.

Nowadays, if they give a kid a failing grade, there's a good chance the parents show up to try and intimidate their way into a pass... Instead of figuring out why their kid is failing. Same goes if their kid gets into trouble - their angel can do no wrong, why are you being mean?

This kind of stuff was happening 20 years ago, I am not surprised it's gotten worse. Parents by and large are unqualified to make decisions about education, but have outsized influence in it. It's probably a huge reason why the US public educational system has been getting smoked by Europe and east Asia, where parents have less insurance.


>Parents by and large are unqualified to make decisions about education

This attitude will drive people away from public schools (and already has). Parents are major stakeholders in their children's education and public schools must be accountable to the public. For the benefit of everyone, there should be a better process for handling the input from parents.

I think we would agree there needs to be a balance. Arguing about changing an F to some passing grade is wrong. But it's also incentivized by the schools and states to meet metrics, so it's unfair to just cast this blame on the parents. But absolutely parents who do this are in the wrong and schools should not enable it.

Some things experts should have broad leeway on are methods. How they structure a reading curriculum or teach long division. Not so much on enforcing certain goals or values.

And the more results that the schools deliver, the more deference from parents they will be able to expect. But the more contention there is between parents and schools, the worse public school support will become.


> This attitude will drive people away from public schools (and already has)

This is a net positive because it will mostly be the people harming the education system through unqualified micromanagement.

> Not so much on enforcing certain goals or values.

Yes, schools shouldn't go beyond the basic virtues society is built on (don't steal, don't bully, etc). But telling them they can't have an addictive device, designed to monetize their attention, is not about virtue. You can't learn how to multiply if you're watching tiktok instead of paying attention.

> the more contention there is between parents and schools, the worse public school support will become.

This is a hard cultural problem. The US seems to have been doing better when parents trusted schools and pushed their kids to succeed. European and East Asian schools similarly trust educators in a way Americans don't. But yes, things would be better if parents trusted schools more.


>This is a net positive because it will mostly be the people harming the education system through unqualified micromanagement.

Well not really, since it pulls support from the public schools. Which are struggling.

>But telling them they can't have an addictive device, designed to monetize their attention, is not about virtue. You can't learn how to multiply if you're watching tiktok instead of paying attention.

Totally agree on that. Phones shouldn't be in schools.

>> the more contention there is between parents and schools, the worse public school support will become.

>This is a hard cultural problem. The US seems to have been doing better when parents trusted schools and pushed their kids to succeed. European and East Asian schools similarly trust educators in a way Americans don't. But yes, things would be better if parents trusted schools more.

How would this be a problem, and not a solution, if unsupportive parents pulled out, as you stated above? It's a two way street. Both parents and schools need to feel the relationship is beneficial for the relationship to sustain itself.


Completely agree. Early education is the time to be protective of social dynamics formation without overeager parents blowing up every PTA meeting and skewing things for everyone. Definitely empathize with educators that have to deal with this balancing act.

> Parents by and large are unqualified to make decisions about education

Especially in the sense of grading and assessment, but I wouldn't generalize this. Both in the sense of do they actually make good decisions on things they can (sometimes) control, such as school (district) selection — barring interesting exceptions like Nikole Hannah-Jones I think they do — and whether they are legally or morally thus qualified. Not that you're implying this, but I bristle a little at the use of the term, since it connotes the removal of parent agency by state action. Caution is advisable there.


They are experts, but they are providing a service. There needs to be a balance. Your doctor is another primary example.

I think one guiding principle is methods vs. values. An expert has more authority on methods, but values need broad alignment between the service provider and the the people they serve. And neither party is entitled to the other - both parties need to feel the relationship is mutually beneficial for the relationship to be sustainable.

So there is some truth in needing to defer to educators, but not absolutely.


> Buggy and bad compared to my code

It probably is, if they didn't have the chops to write it internally. But to them, the trade-off is probably worth the benefit of ditching a component with AGPL.

Personally I just release everything I publish under MIT to make it as easy a as possibly to use.


A major barrier to the improvement of giant FOSS replacements like GIMP and LibreOffice is that the people who develop it and the people who use it have become very used to all of its quirks. It's extremely hard to identify the rough edges in a UI without an external perspective.

Blender is a little different because every 3D modeling program is odd in its own way, and the new UI was close enough that it was easy to learn after the transition (I do still use the legacy keybinds though).

But Blender is very niche hobby software. To "succeed" (if we define success as capturing a significant user share from closed source spyware), Libre and GIMP need to offer workflows that are close to what the average non-coder technical person expects. Maybe that isn't valued by the teams, or maybe they don't have enough resources to make it happen, but the end result is a reputation of jank.

I dislike vibe coded clones because of concerns about the engineering quality. But the reason they're making waves when traditional FOSS hasn't is that the UX is leagues ahead.


> But Blender is very niche hobby software

I don't think this is entirely accurate at this point. It has pretty wide adoption in the educational space, and a number of well-known studios have adopted Blender into their workflows.

The Blender development fund wouldn't be topped-up to the tune of $3 million/year if it were purely funded by hobbyists.


Yes, it is used in a professional capacity.

But when I say "niche hobby software," I'm talking about the relative scale of the user base.

Every working adult who touches a computer in 2026 is using office software. Photo editing is so common that "Photoshop" has become a verb. These two use cases are behemoths compared to 3D modelers.


It's either "niche hobby" or "embraced by professionals".

He seems to be saying that 3D modeling is a niche industry specific application, compared to photo editors and office suites.

Either I'm taking to see the contradiction or this is a false dichotomy.

After 5 years of KDE, I now got Gnome on my office laptop since 3 days. I'm cursing quite a lot on it. All my hotkeys are failing, everything is in the wrong place, and the whole UI feels wrong and broken.

Now partially it is because Gnome is oversimplified, but a big thing is just that I have to unlearn a lot of habits. So I won't switch to KDE for at least a month, as I need to get out of my comfort zone.


I use blender a bit. I still cant figure out the hotkeys. Its the most unpleasant to use software i know of. If the alternatives werent outrageously expensive, I wouldnt touch it

> […] the people who use it have become very used to all of its quirks.

Oh definitely. The way some people seem oblivious to all of Word's faults and weirdness― …Oh! You weren't talking about Microsoft Word!

Software having some quirks is fine. Most people would be fine with LibreOffice if they were taught the rudiments of word processing (that; not using Word, but actual word processing) and never came into a situation where they had to use Word or Google Docs.

The difference between all those tools is that LibreOffice will still do its thing in the future. Just as it has done for a long while, while Word will have forced some new subscription model on you and force-fed you some AI-integration, and Google will have done some similar googley shit.


We have continuously overcorrected in the hundred years since we banned snake oil medicines.

Sure, people should know what they are taking, and they should take things that are safely manufactured and safe to use.

But, there needs to be room for a person to make informed decisions as part of their body autonomy. I personally do not want to start taking peptides because they're so new and the long term risk is unknown, but people who see the initial results and want to be human guinea pigs should be allowed to.

There really needs to be a middle ground between a decade of testing on the off-chance that 1-in-1,000,000 parents gets an upset stomach, and ordering who-knows-what research chem from a sketchy website.


Phase 2 clinical trials establish safety. IMHO anything that passes phase 2 should be prescribable. Phase 3 is the expensive one.

Whatever ideological team jersey we want it to wear, the root cause is that medical insurance has gone from being insurance used in case of emergency, and gone to being a middleman skimming as much as humanly possible while ghoulishly denying healthcare if it is more profitable to let a human die before they can appeal the payout decision.

The only way out of this is to remove the profit motive sparking this avarice: either nationalize the healthcare, or ban healthcare insurance and regulate hospitals so tightly that they aren't allowed to profit as a business, the executives have capped pay, and the business can't be publicly or privately traded.

The laissez-faire approach to healthcare has completely failed in the US. Given it is the wealthiest country in the world, I cannot think of a more damning result for the health insurance model.


Every medical system in the world "ghoulishly den[ies] health care" to patients. If it's not an insurance company, it's the charity providing the care or a government bureaucrat.

And the least accurate label to apply to the US health care system is "laissez-faire approach", since the government is involved in literally every decision.


> The laissez-faire approach to healthcare has completely failed in the US.

Why do you think we have a free market approach to healthcare? Why can I only sign-up once per year unless I have a QLE? Why can I contribute a certain amount to an HSA which is triple tax advantaged with an HDHP but I can't with a PPO? If I have an HDHP+HSA why can I also contribute pre-tax dollars to an FSA that is use-it-or-lose-it for dental or vision only but if I have a PPO I can have a limited purpose FSA that can be used for medical, dental, and vision? Why are there contribution limits? Or even more basic, why is healthcare tied to employment?

The reality is we don't have a free market when it comes to healthcare at all, we have a hampered market where over the years we have politically regulated it into the abomination it is today with some of the worst incentives possible. We can certainly have the debate of whether a free market or nationalized healthcare is best, they both have trade-offs, but it's fallacious to assume we have the former and discount it entirely thus leaving either more regulation or nationalization as the only alternatives. More regulation is almost certainly the worst possible option, we've been on that path since WW2.


What is the market value of your mother's morbidity and mortality? Your spouse's? Your children's? Trying to apply a market to healthcare can become macabre rather quickly.


Assuming a nationalized healthcare model, should the entire productivity of the nation be dedicated towards keeping a single person alive and healthy at the expense of every other conceivable economic activity such as building houses, growing food, manufacturing water treatment facilities, etc? Would it be macabre to deny them any amount of healthcare?

All economic systems allocate scarce resources which have alternative uses and thus rationing must always exists regardless of whether it is a free market, mercantilist, socialist, or as yet unknown economic system. Prices are a rationing mechanism within markets[1] but every other system will have its own set of one or more rationing mechanisms such as by-status, by-lot, by-need, by-survivability, etc. That necessarily means that there aren't enough resources to treat everyone for everything under all circumstances. Do you believe that all economic systems are thus macabre? If so I agree if only because life itself gives us hard choices, never utopic solutions.

[1] In terms of emergency healthcare by-need tends to be used even in market systems.


The answer to all of your questions is that if you are obscenely well off, all of these restrictions go away with enough money. Private insurance exists, or hell - just paying out of pocket. It is a free market, but only for those with enough money to pay.

The average person cannot pay. Health insurance is designed to maximize extraction of wealth from the non-wealthy, while doing just barely enough that they don't immediately get regulated away. The reason for this is greed: sure, greed drives efficiency in capitalism, but if it isn't reined in via regulation we wind up with the abomination we have today.

> More regulation is almost certainly the worst possible option, we've been on that path since WW2

We have de-regulated. In 1950, stock buybacks were illegal, the top corporate tax was 42%, and the top personal income tax was 91%. Today, companies buy back their stocks (transferring wealth to shareholders faster than dividends), top corp tax is 21%, and top income tax is 37%.

Sure, in some cases there is more regulation: factories can't dump waste into the rivers, and employers can't discriminate by race or whatever. But none of these things are the reason why healthcare has become unobtainable to many in the US.


> sure, greed drives efficiency in capitalism, but if it isn't reined in via regulation we wind up with the abomination we have today.

This is exactly where we part ways though. You specifically said "laissez-faire", i.e. "leave it alone" in French; economically this is anti-interventionist, let the markets sort themselves out without interference from government. You are saying regulation is required for the market to function or it leads to what we have today and I'm pointing out that we already have enormous quantities of regulation in healthcare. I pointed out several very easy examples to identify with but by no means is that an exhaustive list of healthcare interventions. The very thing you are advocating for, regulations, has been increasingly attempting to "fix" this market for the past 80 years and our current system exists because of it, not in spite of it. It is not a free market in any sense of the word, and doubling down on more regulation will not improve it. I am more than happy to debate an actual free market or nationalized healthcare model, but further regulating this "free" market is misdiagnosing the system we currently have and assumes the same good intentions that have failed to improve it for four generations will somehow produce better results this time.


>employers can't discriminate by race

Seems to be flaunted in tech.


When you're trying to analyze the failures, perhaps try to not wear any ideological jersey to begin with?

The US does not have a laissez-faire approach to healthcare. The industry has been designed on government interventions. From the financial perspective, the main dynamic is some idea that by being middlemen, health "insurance" companies will work to put pricing pressure on healthcare providers. This was fallacious when it was designed (in 80s or 90s I think), but became patently absurd when the ACA limited insurance company profits to a percentage of revenue.

What other industry do you go to buy a service from a seller, but when you ask about pricing they tell you they don't know anything, and they can't know anything, until they decide to bill you some arbitrary amount months or even years later?

The closest thing is something like auto repair [0]. But mechanics offer fixed fee straightforward services (oil change), straightforward pricing (shop rate at book hours), and written estimates - to be responsive to customers' concerns! Sure, you can drop your car off at a shop and say "fix everything, spare no expense, and call me when it's ready", but that is not how most people operate, by far.

I am definitely not going to tell you that a functioning price dynamic is sufficient for healthcare reform [1] - but it is most certainly necessary!

Single payer supposes to move that pricing dynamic onto the government, where there could be at least a modicum of responsibility. I'm not even against this - our system needs a strong push in some direction. We've currently got the worst of both worlds - centralized planning for provision, coupled with laissez faire extortion for billing (made especially powerful due to the resulting shortages).

But even with single payer, there are going to be cases that exist outside its scope, right? And unscrupulous providers claiming they provided things outside of that scope and demand to be paid directly, rather than settle for their regulated rates [2]. And so regardless what you see as the overall solution, whatever legal bases that create this post-facto-billing-shakedown need to be completely eliminated. Sending fictitious inflated bills at patients needs to be prosecuted as the straightforward attempt to defraud that it is, not trivialized as the standard operating procedure it has become.

[0] who are actually dealing with a much wider gamut of unknown unknowns!

[1] also just preempting the inevitable nonsense response about having to compare hospital prices on your phone before the ambulance arrives. ie tell me you've never significantly interacted with the healthcare system without telling me...

[2] this culture of fraud has become so embedded in the industry, I've had billing agents continue to attempt to defraud me even after confronted with a solid price limit confirmed by both Medicare guidelines and the underwriter (of a medicare advantage plan). This is about as close to single payer as one can get right now, and yet the fraud/extortion dynamic still exists.


It's much closer to nuclear arms-race thinking.

If the US or China developes an AGI that can leverage RSI, the game is over. It is over in the same way that WW3 is over as soon as the missiles are in the air: it doesn't matter who has the fastest or highest yield missiles, because there is a finite amount of time before there is nobody left to care about the advantages of one side. Except at least WW3 is over in about an hour, whereas an AGI could take years or decades to complete whatever thing it has been misaligned to desire.

AI alignment research is still in its alchemy era, and global cyber security is underfunded and bad. As a species, we simply do not have the tools to deal with AGI that can self-improve.


What stops another power from developing its own AGI a year or two later? Or from catching up once there are two competing RSI AGIs? The Soviets were a few years behind on getting nukes but they still got them.

> What stops another power from developing its own AGI a year or two later?

Well, the fact that we're all gone before then if we're unlucky. If we're on a lucky enough timeline where AGI with RSI doesn't immediately end civilization, then maybe we last long enough for the second to be created at least.

But unless we solve the AI assignment problem, it really doesn't matter if we quickly get 2-3 AGIs: an unaligned AI will be the end of us in the long term.

One would hope that self preservation would be a factor, but it seems hubris has deluded many people into thinking they can control a superhuman intelligence.


The first AGI protecting its turf and sabotaging or corrupting the second AI to be its agent.

They are very good research assistants.

I would never trust an AI to make a scientific conclusion, but they're very good at some things:

- Plotting with natural language: "Create a scatter plot of column A versus column B, log scale Y and use stars sized by column C." Every plotting library (except maybe ggplot2 / plotnind) sucks, and the boilerplate never gets less tedius.

- Literature search, e.g. "Have there been any peer reviewed papers where they applied X method to Y domain"

- Complex visualization, e.g. "Make a self contained HTML file with a 3D interactive visualization of this graph. Add ball and spring position optimization and let me move nodes"

- Trying random stuff that would normally be too time consuming to be worth it, eg "Here is a dataset. Try regression methods A, B, and C and report a table." In this case, usually it will screw up and overfit, but it's a good smoke test.

Note how in all of these, it's being used for tedious things that provide some value, but ultimately they are only providing information to make a decision on.

Granted, I would like to have a research assistant I could train, but if your company will not create a position then there is not much you can do.


Yes to all but I have trouble using them effectively for my research side projects or Kaggle competitions. Claude Code/Codex designs and runs experiments quickly, invalidates my initial idea using some obscure criteria than changes direction, and after long hours and days of experiments we end up with poor results. There must be a better agentic workflow for research.

> invalidates my initial idea using some obscure criteria than changes direction

You might be giving it too much leeway.

Opus and Fable are more trustworthy, but even today I find they very quickly go off the rails for research tasks if I don't give them short chunks of work.

They're not ready for prime time automation of research (and before someone mentions the math results, that required absurd amounts of computation time and has immediate feedback in the form of Lean)


Yes, I find DeepSeek can be very precious about its ideas and I have to insist multiple times to do the experiments I ask for despite it thinking them not worth doing. Often my experiments do give good results and I swear the model gets salty about it.

I guess when you ask these things to take on the role of a senior scientist it picks up some natural stubbornness in the training data.


When it comes to difficult optimization/search problems, doing research in a single Claude/Codex thread is definitely a struggle. There are ways to bypass some of the pitfalls by creating a agentic swarm with hard guardrails such as context resets and changing the way the harness works with memory, etc.

So you're not wrong in most of these conclusions, however this is very surface level use case for these tools when it comes to scientific research.

> I would never trust an AI to make a scientific conclusion, but they're very good at some things

> In this case, usually it will screw up and overfit, but it's a good smoke test.

> it's being used for tedious things that provide some value

I see a lot of comments like this about AI use in general, and I suspect there’s a kind of learned helplessness combined with anchoring to earlier experiences. Someone tries a model, encounters a few failures, and forms a fairly permanent view of what these systems can do.

My concern is that people sometimes turn an expectation of failure into a permanent ceiling on how much they’ll even experiment with these tools. Their limited use then keeps confirming their limited expectations.

EDIT: This came off as criticizing the user, but this is entirely to validate my own thesis that for quite a long time nobody has been able to crack the usability challenge when it comes to non-coding knowledge work for these harnesses. It's fascinating because it seems like it's such an easy problem to solve, but in reality it's a complex product problem.


> this is very surface level use case for these tools

Correct. As of Q4 2026, in my experience they are only trustworthy for surface level use cases in the domain of computational chemistry / in-silico high throughout screening, unless you are working on a highly constrained, immediately verifiable and well-studied problem like protein docking or structure-property relations.

> anchoring to earlier experiences.

I have observed the models improving over time. I am basing my post on things at the level of Opus 5.5 / GPT-6.1 Sol / Fable 5.1

> Their limited use then keeps confirming their limited expectations.

Every time a new model comes out or a harness I use gets a new feature, I do try to push it until it breaks. A swarm of GPT Lunas at max context is highly effective at finding information or condensing highly disparate data sources; that wasn't even a thing a year ago in most harnesses.

But even the smartest models like Fable will get sidetracked, or make bad decisions because they do not have enough context of the problem domain. As of Q4 2026.

I'd put their autonomous intelligence in this particular domain at the level of a well-read first year grad student. Phenomenal at knowing things, but unable to question their own assumptions or invalidate their own work unless you hold their hand in the form of a prompt, skill, or direct instruction.


Thanks for the clarification, this is such a spot on response that actually mirrors my experience considerably.

> unless you hold their hand in the form of a prompt, skill, or direct instruction.

The hand-holding and inability to execute automated tasks is a very well known failure mode.

The requirement for me to prod, redirect and continuously steer is the number one pitfall when I did the Muse Open Kernel challenge using a long-horizon search/improvement swarm (or whatever the term is... just don't call it 'RSI').

Using plateau avoidance techniques such as https://arxiv.org/abs/2610.07447 and building harnesses around these is what helped me.


Very interesting article, I'm going to give it a read over lunch. Thanks!

Experimenting with these tools, checking their work and spending that time is not free of cost.

Literature search has been a big improvement for me. Though I have to check all the papers manually (of course, as I want to read them :), it's much more effective for me than searching via Scopus or Web of Science.

Didn’t read your whole post, but the first point is wrong.

“Plotting with natural language: "Create a scatter plot of column A ver….”

No, the tedium of boilerplate code was a feature. Charts were generated by template so styling was consistent, making charts a bit more bland but easier to consume and check. You cannot achieve that same level of consistency or predictability with ai without reinventing a wheel. The models change too fast, and they all want to differentiate themselves for brand reasons, and the emphasis is on fast and easy, not on slow and foolproof, there’s no documentation of the work the way code provides documentation and an easy audit trail. You can restrict the ai by giving specs (use ggplot, use color template foo, create readable code with unit tests), but these are all similar tradeoffs as you might make with human labor too.

What llm do is to simply lower human value, not remove it entirely.

Just as results needed checking and replication before llm, they will continue to need checking and replication by humans. The humans won’t be compensated nearly as much, but humans are humans and so are not.


> Charts were generated by template so styling was consistent, making charts a bit more bland but easier to consume and check

This is why reading the whole post is important :)

You seem to think I'm suggesting the use an image generator for plotting. Don't do that, it will screw that up, image models are not made for that. The use for AI here is to save 5 minutes writing the code to generate the plot, or to edit the style (again via code) of the plot.

If you do plt.scatter(x,y, c="blue", marker="o") that is a function call in Matplotlib. Its API notoriously sucks. Most Python plotting libraries just create convenience functions around this library because although the API (actually there are two and the documentation is not consistent on which to use) is hot garbage, and the documentation is labyrinthine, the plotting engine itself is powerful.

> What llm do is to simply lower human value, not remove it entirely.

The value in a plot is not the exact code used to make it (beyond the code being correct). The value is in the visualization giving the reader an "ah-ha" moment. I would much rather take the 5 minutes I save writing a function call, and use it to spend more time examining the plot itself.


They are great computer interface assistants, but forget scientific conclusion would never trust then for committing in right git branch

> Literature search, e.g. "Have there been any peer reviewed papers where they applied X method to Y domain"

This is something the LLMs are really good for, at least for compiler papers.


And like everything LLM related this is something that people may have had bad experience with in the past but which work well now -- a year or two ago half those papers would be "hallucinated" non-existent papers but this rarely happens these days.

I have a reference checker tool that I have the agent use to make sure nothing is hallucinated. It very often corrects small details like dates or authors on otherwise real references.

Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: